Delft University of Technology

Proactive Approach to Safe Automated Driving: Control of Evasive Maneuvers in

Hazardous Scenarios

Gharavi, L.

DOI
10.4233/uuid:7d7ed515-f870-4d7c-bfc2-6b72893b11c0
Publication date
2025
Document Version
Final published version
Citation (APA)
Gharavi, L. (2025). Proactive Approach to Safe Automated Driving: Control of Evasive Maneuvers in
Hazardous Scenarios. [Dissertation (TU Delft), Delft University of Technology].
https://doi.org/10.4233/uuid:7d7ed515-f870-4d7c-bfc2-6b72893b11c0

Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.

Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.

This work is downloaded from Delft University of Technology.

For technical reasons the number of authors shown on this cover page is limited to a maximum of 10.
 Proactive Approach to Safe Automated
Driving: Control of Evasive Maneuvers in
Hazardous Scenarios
 Proactive Approach to Safe Automated
Driving: Control of Evasive Maneuvers in
Hazardous Scenarios

Dissertation

for the purpose of obtaining the degree of doctor

at Delft University of Technology
by the authority of the Rector Magnificus Prof.dr.ir. T.H.J.J. van der Hagen
chair of the Board for Doctorates
to be defended publicly on
Monday 3 March 2025 at 10:00 o’clock

by

Leila GHARAVI

Master of Science in Mechanical Engineering,

Amirkabir University of Technology, Iran
born in Qom, Iran
This dissertation has been approved by
the promotors: Prof.dr.ir. B.H.K. De Schutter and Prof.dr.ir. S. Baldi

Composition of the doctoral committee:

Rector Magnificus, chairman
Prof.dr.ir. S. Baldi, Southeast University, China
Prof.dr.ir. B.H.K. De Schutter, Delft University of Technology, Netherlands

Independent members:
Prof.dr. M. Corno, Politecnico di Milano, Italy
Prof.dr. M. Johansson, KTH Royal Institute of Technology, Sweden
Prof.dr. N. van de Wouw Eindhoven University of Technology, Netherlands
Prof.dr. R.R. Negenborn, Delft University of Technology, Netherlands
Dr. B. Shyrokau, Delft University of Technology, Netherlands

Reserve member:
Prof.dr.ir. R. Happee, Delft University of Technology, Netherlands

This dissertation was supported by the Dutch Research Council (NWO-TTW) within the
EVOLVE project (no. 18484) and the EU Horizon 2020 innovation and research program
within the Marie Skłodowska-Curie OWHEEL project (no. 872907).

Keywords: Collision avoidance, Automated driving, Model predictive control, Hybrid

systems, Evasive maneuvers

Printed by: Ridderprint, https://www.ridderprint.nl

Cover: Leila Gharavi
Style: TU Delft House Style, with modifications by Moritz Beller
https://github.com/Inventitech/phd-thesis-template

The author set this thesis in LATEX using the Libertinus and Inconsolata fonts.

ISBN 978-94-6506-978-4
An electronic version of this dissertation is available at https://repository.tudelft.nl and
https://phd.leilagharavi.com.
Let’s keep in mind that in real life the lines are fluid, are not easily drawn, and should not be
rigidly maintained.

— Gabor Maté, The Myth of Normal

 vii

Contents
Summary xi
Samenvatting xiii
Acknowledgments xv
1 Introduction 1
1.1 The Concept of Hazard in Automated Driving . . . . . . . . . . . . . . . 1
1.2 Proactivity in Hazardous Scenarios. . . . . . . . . . . . . . . . . . . . . 3
1.3 Approximation: A Double-Edge Sword . . . . . . . . . . . . . . . . . . . 4
1.4 Contributions and Organization of this Thesis . . . . . . . . . . . . . . . 4

I Hybridization Approach to Computational Efficiency 7

2 Iterative PWA Approximation via Cut-Based Domain Partitioning 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 PWA Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Parametric Definition of Regions . . . . . . . . . . . . . . . . . . 12
2.2.3 Approximation Algorithm . . . . . . . . . . . . . . . . . . . . . 13
2.2.4 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Case Study: Vehicle Dynamics . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Nonlinear Vehicle Model . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Implementation and Results . . . . . . . . . . . . . . . . . . . . 17
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 PWA Approximation of Multi-Dimensional Nonlinear Systems 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Parametric Region Definition. . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Cut-Based PWA Approximation . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.1 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.2 Comparison with Approaches in the Literature. . . . . . . . . . . 31
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4 H4MPC: A Hybridization Toolbox for MPC 35
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.1 Nonlinear Problem Description. . . . . . . . . . . . . . . . . . . 36
4.2.2 Approximation of the Nonlinear Problem . . . . . . . . . . . . . 37
 viii Contents

4.2.3 Approximation Problem Formulation. . . . . . . . . . . . . . . . 38

4.3 Toolbox Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3.1 Grid Generation Module . . . . . . . . . . . . . . . . . . . . . . 39
4.3.2 Model Approximation Module . . . . . . . . . . . . . . . . . . . 41
4.3.3 Constraint Approximation Module . . . . . . . . . . . . . . . . . 41
4.4 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4.1 Nonlinear Vehicle Model . . . . . . . . . . . . . . . . . . . . . . 41
4.4.2 Approximation of the Vehicle Dynamics . . . . . . . . . . . . . . 43
4.4.3 Approximation of the Kamm Circle Constraint . . . . . . . . . . . 43
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5 Sensitivity Analysis for PWA Approximations of NLPs 47
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2.1 Representation of Continuous PWA Functions . . . . . . . . . . . 49
5.2.2 Sensitivity of the Convex Optimization Problem . . . . . . . . . . 50
5.3 Continuous PWA Approximation of NLPs . . . . . . . . . . . . . . . . . 51
5.4 Theoretical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.5 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5.1 Case 1: Finding the Confidence Radii . . . . . . . . . . . . . . . . 56
5.5.2 Case 2: Finding the Approximation Criteria . . . . . . . . . . . . 57
5.5.3 Application for NMPC Optimization . . . . . . . . . . . . . . . . 57
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

II Collision Avoidance in Automated Driving 61

6 Design and Numerical Analysis of Hybridization Benchmark for NMPC 63
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.2.1 Model Approximation . . . . . . . . . . . . . . . . . . . . . . . 67
6.2.2 Constraint Approximation . . . . . . . . . . . . . . . . . . . . . 67
6.2.3 Relation to the State-of-the-Art. . . . . . . . . . . . . . . . . . . 68
6.3 Grid Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.4 Approximation Problem Formulation. . . . . . . . . . . . . . . . . . . . 72
6.4.1 Model Approximation . . . . . . . . . . . . . . . . . . . . . . . 72
6.4.2 Constraint Approximation . . . . . . . . . . . . . . . . . . . . . 73
6.5 Model and Constraint Hybridization for Vehicle Control . . . . . . . . . . 75
6.5.1 Nonlinear System Descriptions. . . . . . . . . . . . . . . . . . . 75
6.5.2 Grid Definition and Coverage . . . . . . . . . . . . . . . . . . . 77
6.5.3 Model Approximation Results . . . . . . . . . . . . . . . . . . . 79
6.5.4 Constraint Approximation and Validation . . . . . . . . . . . . . 81
6.6 Conclusions and Outlook. . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.6.1 Conclusions for Vehicle Control . . . . . . . . . . . . . . . . . . 84
6.6.2 Generalized Hybridization Framework . . . . . . . . . . . . . . . 85
6.6.3 Next Steps and Future Work . . . . . . . . . . . . . . . . . . . . 86
Contents ix

7 Comparative Assessment of Hybridization in Vehicle Control 87

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.2.1 Model and Constraint Hybridization . . . . . . . . . . . . . . . . 90
7.2.2 MPC Optimization Problems . . . . . . . . . . . . . . . . . . . . 91
7.3 Comparison Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3.1 Prediction Model and Physics-Based Constraints . . . . . . . . . . 93
7.3.2 Control Parameters. . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3.3 MPC Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3.4 Reference Trajectory . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3.5 Driving Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.3.6 Solver Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.4.1 Ideal Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.4.2 Friction Offset . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.4.3 Friction Disturbance . . . . . . . . . . . . . . . . . . . . . . . . 101
7.4.4 Analysis of Computational Performance . . . . . . . . . . . . . . 102
7.4.5 Handling Limits . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.5 IPG CarMaker Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.6 Discussions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.6.1 MILP vs. MIQCP . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.6.2 Robustness to Friction Uncertainty . . . . . . . . . . . . . . . . . 105
7.6.3 Performance Close to Handling Limits . . . . . . . . . . . . . . . 106
7.6.4 Overall Computational Performance . . . . . . . . . . . . . . . . 106
7.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8 Efficient Response to Sudden Appearance of Static Obstacles 109
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.2 Vehicle Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.2.1 Kinematic Model . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.2.2 Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . 114
8.3 Proposed Control System. . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.4 Design of the Planning Strategy . . . . . . . . . . . . . . . . . . . . . . 116
8.4.1 MSF Planner . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.4.2 MPC Planner . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.4.3 FBL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.5 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.6 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.6.1 Relaxed cases with 𝑣𝑥 = 5m/s . . . . . . . . . . . . . . . . . . . . 120
8.6.2 Lateral assessment with 𝑣𝑥 = 6m/s . . . . . . . . . . . . . . . . . 121
8.6.3 Extreme cases with 𝑣𝑥 = 7m/s . . . . . . . . . . . . . . . . . . . 121
8.6.4 Comparison with MPC . . . . . . . . . . . . . . . . . . . . . . . 122
8.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
x Contents

9 Proactive Collision Avoidance with Stochastic Obstacle Behavior 125

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
9.1.1 Motion Planning Challenges in Different Scenarios. . . . . . . . . 126
9.1.2 Sources of Uncertainty in Highway Driving . . . . . . . . . . . . 127
9.1.3 Computational Efficiency in Emergencies . . . . . . . . . . . . . 128
9.1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
9.2 Problem Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.3 MMPS Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.4 Problem Reformulation and Solution Approach . . . . . . . . . . . . . . 130
9.4.1 Obstacle Vehicle Model. . . . . . . . . . . . . . . . . . . . . . . 130
9.4.2 Hybrid Ego Vehicle Model . . . . . . . . . . . . . . . . . . . . . 132
9.4.3 Chance Constraints and Collision Risk Function . . . . . . . . . . 134
9.4.4 SMPC Optimization Problem . . . . . . . . . . . . . . . . . . . . 137
9.5 Simulations and Results . . . . . . . . . . . . . . . . . . . . . . . . . . 138
9.5.1 Proactivity Assessment. . . . . . . . . . . . . . . . . . . . . . . 139
9.5.2 Assessment of Attainable Trajectories . . . . . . . . . . . . . . . 143
9.5.3 Comparison with IPG Motion Planner . . . . . . . . . . . . . . . 144
9.5.4 Performance Analysis and Discussion . . . . . . . . . . . . . . . 146
9.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
10 Conclusions 151
10.1 Summary of Research Contributions . . . . . . . . . . . . . . . . . . . . 151
10.2 Future Research Suggestions . . . . . . . . . . . . . . . . . . . . . . . . 152
10.3 Outlook and Recommendations . . . . . . . . . . . . . . . . . . . . . . 153
Bibliography 155
Curriculum Vitæ 171
List of Publications 173
 xi

Summary
Emergency maneuvers on highways present one of the most complex challenges for
automated driving. High speeds pushing the vehicle towards nonlinear regimes, coupled
with the necessity of swift decision making, complicates the collision avoidance problem
to the extent that even expert human drivers may struggle to safely avoid collisions.
Lack of sufficient and reliable data limits applicability of model-free and data-driven con-
trol approaches in hazardous scenarios, opening the door to model-based and optimization-
based control approaches. However, the unknown behavior of other road users, the sensi-
tivity of the handling limits (e.g., tire saturation) to road conditions, and the amplification
of minor steering adjustments on the lateral trajectory due to high speed necessitate the
incorporation of nonlinear models in the design. Such nonlinearities should be balanced
with the increased complexity and the need for swift responses to hazard.
This thesis addresses the critical challenge of controlling evasive emergency maneuvers
for automated driving on highways. As the complexity of the collision avoidance problem
arises from its intrinsic connection to safety definitions and the need for computational
efficiency, we investigate proactive solutions in the sense of computationally rapid and max-
imally safe responses, while avoiding unnecessary conservatism. This thesis is structured
into two main parts.
In Part I, hybridization – i.e. approximation of nonlinear functions using hybrid systems
formalism – is explored as a means to improve the computational efficiency. Chapter 2 in-
troduces a novel PWA approximation technique utilizing parametric cut-based partitioning
of domains, extending the applicability of the state-of-the-art methods to multi-dimensional
systems. Chapter 3 introduces a more generalized PWA approximation method employing a
hinging hyperplane formulation, offering greater flexibility compared to the 2-dimensional
plane-based cutting strategy in domain partitioning. As a result, the generalized cut-based
approach is able to obtain simpler PWA approximations of nonlinearities in comparison
with the parametric cut-based approach for the same desired approximation error. In Chap-
ter 4, we introduce H4MPC, a MATLAB-based open-source toolbox for hybridization of
nonlinear control problems in automated driving, and in Chapter 5, the sensitivity of PWA
approximation of nonlinear optimization problems with polytopic constraints is analyzed.
This analysis can be used in two ways: finding the confidence radius, i.e. a bound on the
distance between the approximated and original minima, for a given approximation error,
as well as obtaining a required bound on the approximation error for a desired confidence
radius.
Part II investigates the problem of proactive collision avoidance in emergency scenar-
ios. Following the results of Part I, we define and provide a hybridization benchmark for
nonlinear Model Predictive Control (MPC) in Chapter 6. Next, we extensively investigate
different hybridization formulations for hybrid approximation of nonlinear models using
the computationally tractable Max-Min-Plus-Scaling (MMPS) systems formalism, as well
as nonlinear inequality constraints such as vehicle stability and time saturation limits. In
xii Summary

Chapter 7, this hybridization framework is assessed in a vehicle control benchmark in vari-

ous driving scenarios such as friction uncertainty, showcasing the efficacy of hybridization
in enhancing the computational speed, as well as improving the search across the decision
space for the globally optimal solution. Chapter 8 addresses the challenge of real-time
collision avoidance on a vehicle with limited computational resources after the sudden
appearance of a static obstacle. More specifically, we propose an integrated approach to
motion planning that combines efficient feed-forward planning signals to mitigate the
feasibility issues of MPC. Finally in Chapter 9, we extend our focus to proactive collision
avoidance in dynamic environments using a hybrid stochastic MPC framework for motion
planning, validated through rigorous simulations with a high-fidelity vehicle model in IPG
CarMaker.
In summary, by developing novel hybridization strategies, comprehensive analytical
and numerical investigations, and introducing assessment benchmarks, this thesis tackles
the challenges of controlling evasive maneuvers in hazardous scenarios. The newly devel-
oped methods throughout this thesis are validated using extensive cases studies and the
effectiveness of hybridization in proactive response to danger is demonstrated through
high-fidelity simulations as well as real-world experiments. Our recommendations for
future work include enhancing robustness, more generalized and extensive sensitivity
analysis, and improving the interpretability of control solutions to mimic human behavior,
aiding in the acceptance of automated driving.
 xiii

Samenvatting
Noodmanoeuvres op snelwegen vormen een van de meest complexe uitdagingen voor
geautomatiseerd rijden. Hoge snelheden, die het voertuig naar niet-lineaire regimes duwen,
gecombineerd met de noodzaak van snelle besluitvorming, bemoeilijken het probleem van
botsingvermijding zodanig dat zelfs ervaren menselijke bestuurders moeite kunnen hebben
om op een veilige manier botsingen te vermijden.
Een gebrek aan voldoende betrouwbare gegevens beperkt de toepasbaarheid van model-
vrije en datagestuurde regelmethoden in gevaarlijke scenarios, waardoor er een duidelijke
nood is aan modelgebaseerde en optimalisatiegebaseerde regelmethoden. Het onbekende
gedrag van andere weggebruikers, de gevoeligheid van de voertuiglimieten (bijvoorbeeld
bandensaturatie) voor wegomstandigheden, en de versterking van kleine stuurcorrecties op
de laterale trajectorie door de hoge snelheid, vereisen echter de opname van niet-lineaire
modellen in het ontwerp. Dergelijke niet-lineariteiten moeten in balans worden gebracht
met de toenemend complexiteit en de behoefte aan snelle reacties op gevaar.
Dit proefschrift behandelt de kritische uitdaging van het controleren van ontwijkende
noodmanoeuvres voor geautomatiseerd rijden op snelwegen. Omdat de complexiteit van
het botsingvermijdingsprobleem voortkomt uit het intrinsieke verband met veiligheids-
definities en de noodzaak van computationele efficiëntie, onderzoeken we proactieve
oplossingen in de zin van computationeel snelle en maximaal veilige reacties, terwijl
onnodige conservativiteit wordt vermeden. Dit proefschrift is gestructureerd in twee
hoofdonderdelen.
In Deel I wordt hybridisatie – dat wil zeggen de benadering van niet-lineaire functies met
behulp van hybride systeemformalismen – verkend als een middel om de computationele
efficiëntie te verbeteren. Hoofdstuk 2 introduceert een nieuwe PWA-benaderingstechniek
die gebruik maakt van parametrische domeinpartitionering, waardoor de toepasbaarheid
van de state-of-the-art methoden op multidimensionale systemen wordt uitgebreid. Hoofd-
stuk 3 introduceert een meer algemene PWA-benaderingsmethode die gebruik maakt
van een hinging-hyperplane formulering, wat meer flexibiliteit biedt vergeleken met de
tweedimensionale snijstrategie bij domeinpartitionering. Als resultaat kan de algemene
snijbenadering eenvoudigere PWA-benaderingen van niet-lineariteiten verkrijgen in ver-
gelijking met de parametrische snijbenadering bij dezelfde gewenste benaderingsfout. In
Hoofdstuk 4 introduceren we H4MPC, een op MATLAB gebaseerde open-source toolbox
voor de hybridisatie van niet-lineaire regelproblemen bij geautomatiseerd rijden, en in
Hoofdstuk 5 wordt de gevoeligheid van PWA-benadering van niet-lineaire optimalisatie-
problemen met polytoop-beperkingen geanalyseerd. Deze analyse kan op twee manieren
worden gebruikt: voor het vinden van de betrouwbaarheidsradius, d.w.z. een grens op de
afstand tussen de benaderde en originele minima, voor een gegeven benaderingsfout, even-
als voor het verkrijgen van een vereiste grens op de benaderingsfout voor een gewenste
betrouwbaarheidsradius.
xiv Samenvatting

Deel II onderzoekt het probleem van proactieve botsingvermijding in noodscenario’s.

Naar aanleiding van de resultaten van Deel I definiëren en bieden we een hybridisatie-
benchmark voor niet-lineaire Model Predictive Control (MPC) in Hoofdstuk 6. Vervol-
gens onderzoeken we uitgebreid verschillende hybridisatieformuleringen voor hybride
benadering van niet-lineaire modellen met behulp van het computationeel beheersbare
Max-Min-Plus-Scaling (MMPS) systeemformalisme, evenals niet-lineaire ongelijkheidsbe-
perkingen zoals voertuigstabiliteit en tijdverzadigingslimieten. In Hoofdstuk 7 wordt dit
hybridisatiekader beoordeeld in een voertuigregelbenchmark in verschillende rijscenario’s,
zoals wrijvingsonzekerheid, waarbij de effectiviteit van hybridisatie in het verbeteren van
de computationele snelheid en het verbeteren van de zoektocht naar de globaal optimale
oplossing wordt aangetoond. Hoofdstuk 8 behandelt de uitdaging van botsingvermijding
in real-time met een voertuig met beperkte computationele middelen na het plotseling
verschijnen van een statisch obstakel. Meer specifiek, we stellen een geïntegreerde aanpak
voor bewegingsplanning die efficiënte feed-forward planningssignalen combineert om de
feasibility problemen van MPC te verminderen. Ten slotte richten we ons in Hoofdstuk 9
op proactieve botsingvermijding in dynamische omgevingen met behulp van een hybride
stochastisch MPC-kader voor bewegingsplanning, gevalideerd door middel van rigoureuze
simulaties met een voertuigmodel van hoge nauwkeurigheid in IPG CarMaker.
Samenvattend, door het ontwikkelen van nieuwe hybridisatiestrategieën, uitgebreide
analytische en numerieke onderzoeken, en het introduceren van beoordelingsbenchmarks,
gaat dit proefschrift de uitdagingen aan van het beheersen van ontwijkende manoeuvres
in gevaarlijke scenario’s. De nieuw ontwikkelde methoden in dit proefschrift worden
gevalideerd met behulp van uitgebreide case study’s en de effectiviteit van hybridisatie in
proactieve reacties op gevaar wordt aangetoond door middel van simulaties van hoge nauw-
keurigheid evenals experimenten in de echte wereld. Onze aanbevelingen voor toekomstig
werk omvatten het verbeteren van robuustheid, meer gegeneraliseerde en uitgebreide
gevoeligheidsanalyse, en het verbeteren van de interpretatie van regeloplossingen om
menselijk gedrag te imiteren, wat kan bijdragen aan de acceptatie van geautomatiseerd
rijden.
 xv

Acknowledgments
“We seek more and more privacy, and we feel more and more alienated when we get it.”

— Philip Slater, The Pursuit of Loneliness

It’s hard to recognize the person I was at the start of my PhD journey. Reflecting on
these four turbulent years feels almost surreal, a testament to how transformative the
experience has been. Yet, none of us grow in isolation, and I want to take a moment to
thank everyone who supported me along the way. Whether you realize it or not, you have
contributed to this dissertation, and for that, I will always be grateful.
I would like to express my gratitude to my supervisors, Bart and Simone, for their
guidance, flexibility, and open-mindedness during my research process. I am also thankful
to the incredible colleagues I worked with closely. Barys, thank you for your support
during EVOLVE user committee meetings and for arranging numerous opportunities that
allowed me to learn and take on new challenges. Laura, Azita and Ton, I’m deeply grateful
for our enriching discussions and brainstorming sessions.
I extend my heartfelt gratitude to my MSc students: Roald, your resilience and dedica-
tion were unforgettable; I’ll never forget the time we discretized a system into two million
pieces! Jelske, your scientific curiosity and critical thinking are inspiring. Pietro, your
attention to detail and perseverance in a new environment were remarkable. Kenrick,
your determination and tireless effort to succeed left a lasting impression on me. During
my dark days, meeting you all gave me the energy I needed to move forward and I’m proud
of the incredible scientists you are today.
I’ll forever cherish the wonderful people I met and collaborated with during my research
visit in the University of Tokyo. Fujimoto-sensei, I deeply appreciate the opportunity
to visit your lab. Binh-Minh-sensei, working with you was a valuable experience, and
spending Nowruz together made it even more special. Thank you Hosomi-san for offering
help in every step of the way and SaTona-san for being the coolest driver ever! To
Mihoko-san, Hedaiko-san, and the sweetest, Michi-chan, you all hold a special place in
my heart and your warmth and kindness will always stay with me.
To my wonderful officemates—Frederik, Max, and David—thank you for being such a
great support system. From your friendly greetings to sharing fun information, I enjoyed
every moment with you. I appreciate how you always checked in on me and offered help
when I needed it. Our office was a safe space, because of you.
A huge thank you to my other colleagues, the amazing crew of X: Nuno, Shamangi,
Olga, Neil, Kjeld, and Henri. Nuno, your energy was contagious, and many dark days
turned brighter with just an hour of spinning with you. Shamangi, your open-mindedness,
courage, and authenticity are truly inspiring. Neil, thank you for occasionally reminding
xvi Acknowledgments

me which leg was left and which was right in the dance class! Olga, I’ll always treasure the
wonderful conversations we shared. My gratitude also goes to Deborah, Ismail, Simona,
Kiarash and everyone else who brought joy, warmth, and smiles to our shared experiences
at X.
Communities thrive when we live by praxis–taking our values and ideals and putting
them into action to create positive change. I am incredibly thankful for the people who
contribute to this process with intention and care, making the world a better place for all
of us. Magda, your unwavering support is invaluable. Giedo, I appreciate all the effort
you put into improving the well-being of employees. Mascha, your thoughtful care for
the PhDs, even when you are overwhelmingly busy, does not go unnoticed and is deeply
appreciated. Bo, your dedication and kindness in your daily work are truly priceless. You
have no idea how many tough days were turned around by simply interacting with you.
Ailishia and Mimi, I hope you know how inspiring your projects are. Kirsten, thank you
for your compassionate approach to your medical profession. Mijntje and Nick, thank
you for your meaningful support and care in your practice. Alba and Camilo, I’m deeply
grateful for the moments we shared. Your mindful, self-aware approach to art inspired
me to grow and deepen my understanding of myself. And Bart, I would like to thank you
again, this time as our department chair. Your management in critical times has made
important and meaningful impact.
To Emilio, thank you for helping me find my first electric guitar and for our fun
moments of non-verbal communication. João, your kindness radiates warmth and calmness,
making every moment special. Thank you Mahendra, for your tasty curries and being
up for any kind of fun activity. Amin, the reels you share are legendary, please keep on
sending more! Vicente, “you’re the best!” I’m grateful for every trip we’ve taken together
and the meaningful discussions we’ve had. Changrui, thank you for making complex
derivations surprisingly fun. Hasti, Mohammad, Fatemeh, Saman, and Nima, thank
you for the wonderful times we’ve shared together. You’ve all contributed something
special to my journey, and I’m grateful for each of you.
To my other friends and colleagues–Shengling, Anil, Lorenzo, Suad, Xiaoyu, Steven,
Athina, Jan, Maolong, Arian, Kiana, Gianpietro, Tian, Dana, Micah, Kanghui, Sasan,
Maria, Lotfi, Alessandro, Leonore, Ayda, Rayyan, Francessco, Sreeshma, Alberto,
Ali, Caio, Maria, and Reza–I owe a heartfelt thanks. Your presence made this difficult
journey not just survivable, but meaningful.
I would also like to thank my parents, Houra, who inspires me to be courageous, and
Ali, who motivates me pursue deeper understanding. I am also grateful to my brother,
Mahdi, whose patience and commitment remind me to stay true to myself. Regardless of
the distance between us, I love you all endlessly.
In the end, a special thank you to Eva for her mindfulness and bravery, and Mohamad
for his inner strength and deep sense of integrity. I am so grateful for you both, and words
cannot express how much I appreciate your presence in my life.

Leila
Delft, January 2025
 1

1
Introduction
As the saying goes, ‘he who has a hammer sees everything as a nail’. If you approach a problem
from a particular theoretical point of view, you will end up asking only certain questions and
answering them in particular ways. You might be lucky, and the problem you are facing might
be a ‘nail’ for which your ‘hammer’ is the most appropriate tool. But, more often than not,
you will need to have an array of tools available to you.

— Ha-Joon Chang, Economics: The User’s Guide

Imagine the following scenario: you are driving on a highway, and the driver in front
of you suddenly starts making erratic movements, possibly due to an unforeseen situation
such as a stroke or even a collision. Another example is the unexpected appearance of an
animal on the road, which is closely related to the well-known moose test in the context
of automated driving. In such hazardous scenarios, an ideal driver would remain calm,
maintain decision-making ability, and be able perform a safe maneuver, balancing between
steering and braking to avoid collision with the obstacle, as well as vehicle instability. In
reality, however, drivers often panic, which leads to loss of focus and degraded performance.
Automated driving systems can mitigate such risks, which is the core of the research in
this thesis. To begin, it is important to define what constitutes a hazard.

1.1 The Concept of Hazard in Automated Driving

Let us first distinguish between highway and urban driving scenarios, as each have their
respective challenges and necessitate different research approaches. Urban driving involves
navigating through complex traffic patterns, interacting with vulnerable participants such
as pedestrians and cyclists, and managing frequent stops and starts. The frequent presence
of traffic signals, crosswalks, and the need for constant vigilance for unexpected movements
by pedestrians or cyclists add layers of difficulty that require comprehensive studies on
human behavior, traffic flow, and safety measures [1–3]. These complexities demand a
high level of situational awareness and quick decision-making from the driver.
In contrast, on highways the environment and the behavior of other road users are less
complex. For instance, there are no junctions nor red lights, and other vehicles have fewer
 2 1 Introduction

behaviors such as decelerating, accelerating, or changing lanes. This reduced complexity

1 might suggest that hazardous scenarios are less dangerous on highways than in urban
areas. However, the opposite is true! In fact, highway driving involves its own intricate
set of challenges, as evidenced by driver training programs where trainees must master
specific urban driving scenarios before progressing to high-speed highway driving. For
instance, in some countries e.g. Iran, new drivers are not permitted to drive on inter-city
highways during their first year after obtaining their license.
On highways, the primary challenges stem from the high speeds, which significantly
amplify the consequences of errors. As a slight disturbance in steering at high speeds
can lead to severe instability and loss of control, the margin for error is drastically
reduced and the reaction time available to drivers is much shorter than in urban
driving scenarios. Moreover, maneuvers options are limited on highways; in urban
environments, drivers can often stop completely to avoid obstacles, e.g. in case pedestrians
or cyclists suddenly enter their path. However, on highways, it is either dangerous or even
impossible to do the same due to the high speeds. For instance, if an animal suddenly
appears on the road, the driver might not be able to stop in time or change lanes safely,
potentially leading to a serious accident.
In such scenarios, human drivers often experience severe anxiety and impaired judg-
ment, which can lead to suboptimal, if not drastic and dangerous, reactions. In contrast, an
automated system does not panic or experience delays in decision-making. Therefore, auto-
mated driving can be a promising solution for human-based errors in hazardous scenarios,
among other benefits e.g. enhanced mobility.
The concept of hazard plays a significant role in understanding the barrier towards high
and full automation levels [4] defined by the Society of Automotive Engineers (SAE) [5].
The SAE classifies automation into six levels shown in Fig. 1.1, ranging from 0 to 5, with
levels 4 and 5 respectively representing high and full autonomy. The term “automated
driving” – in particular, Automated Driving Systems – is used to refer to levels 3 to 5 [5].

Figure 1.1: Vehicle automation levels defined by SAE, inspired by [6]

In level 3 of autonomy, similar to autopilot systems in airplanes, the automated system

is responsible for perception and executing driving tasks [4]. However, the presence of a
skilled driver in the vehicle is still required because the driver must take over the driving
1.2 Proactivity in Hazardous Scenarios 3

task in emergencies [7]. This requirement is eliminated in level 4, meaning that the system
must handle emergencies independently in specific operational design domains e.g. sunny 1
weather and particular routes. Finally, a level 5 automated system functions as a skilled
driver, and is capable of operating in a wide range of scenarios and various operational
design domains without any human intervention. Thus, the primary barrier to reaching
high and full automation levels in automated driving lies in the implementation of a driving
system capable of handling hazards autonomously. This thesis will focus on these critical
scenarios via a proactive approach, i.e. anticipating potential issues rather than simply
reacting to them. We provide a detailed explanation of the concept of proactivity in the
following section.

1.2 Proactivity in Hazardous Scenarios

One of the main challenges in addressing hazards in the design of automated driving
systems is the lack of sufficient and reliable data in such scenarios, which limits
effective employment of data-driven and learning-based control design solutions. Moreover,
a rapidly changing environment and uncertain driving conditions in real-world
scenarios require high levels of adaptivity to change and robustness to disturbances for
candidate solutions. Model Predictive Control (MPC), as evidenced by its increasing
popularity in the field [8–12], is a promising approach for collision avoidance in hazardous
scenarios, by systematically handling constraints within an online control optimization
problem. As the name suggests, in MPC approach, a model of the system is used to predict
its behavior over a finite horizon in the future time steps, which is then incorporated into
the formulation of the control optimization problem.
The effective application of MPC for high levels of automated driving necessitates the
use of complex physics-based models as many simplifying assumptions, such as low
lateral accelerations or linear tire dynamics [13], are no longer valid when performing
evasive maneuvers in high-speed hazardous scenarios. In this sense, the MPC optimization
problem in hazardous scenarios may lead to long computation times, which is a critical
limitation in these situations where swift decision-making is essential.

A proactive control solution respects the entanglement of the aforementioned

challenges by providing computationally fast and optimally safe trajectories
and/or control braking and steering commands in response to an unforeseen
dangerous situation.

For instance, one may think of robust [14] and tube-based [15] MPC design approaches
to tackle the uncertainty challenge. However, a robust approach has a limited ability to
providing a safe solution in case of an unforeseen incident on the road. Moreover, in a
hazardous scenario an overly-cautious and conservative control solution is in general not
optimal and may even lead to the propagation of hazard in time. For example, braking
behind a car with an unconscious driver keeps the ego vehicle in danger for a long period,
while overtaking can help the system get out of the hazardous scenario as soon as possible.
In this sense, a stochastic formulation of the uncertainties allows for a more proactive
response to a hazardous situation on the road. In addition, even the most optimal overtaking
trajectory becomes ineffective if it is computed too slowly, missing the critical window
 4 1 Introduction

of opportunity to avoid a collision. In summary, designing a proactive control strategy

1 for automated driving requires consideration of the stochastic properties of the MPC
optimization problem, alongside the need for computational efficiency. As the latter is the
bottleneck in the field, the following section will focus on this aspect in greater depth.

1.3 Approximation: A Double-Edge Sword

The nonlinearity of the MPC optimization problem is the main source of long computation
times. For instance, while solving a NonLinear Program (NLP) in general yield a local
optimum, state-of-the-art solution methods for a Mixed-Integer Linear Program (MILP)
can guarantee convergence to the global optimum.
Linearization can effectively represent the local behavior of the system, but often
fails to capture critical nonlinear phenomena such as saturation of the tire forces. As a
result, avoiding the nonlinearity by e.g. successive online linearization [16] or grid-based
linearization [17] of the nonlinearity in hazardous scenarios does not allow capturing the
full dynamics of systems such as tires operating in nonlinear regions, where the interaction
between longitudinal and lateral dynamics further complicates the behavior.
With these limitations of linearization-based techniques for our research problem, it is
evident that a sophisticated approximation approach is required for the vehicle control
problem in hazardous scenarios. This approximation must be meticulously defined within
a comprehensive framework to ensure a well-balanced trade-off between computation time
and accuracy, allowing for informed decision-making in the control design process. In this
context, hybridization emerges as a particularly promising approach.
Hybridization involves approximating nonlinearities using hybrid systems modeling
frameworks [18]. While the literature provides examples of this hybridization for compu-
tational efficiency [19–21], developing a comprehensive framework necessitates a more
thorough analysis and synthesis. This development is achieved by analysis of various
approximation methods and parameters while keeping an eye on the dynamic nature of
the problem and its critical aspects including the optimization problem and its number
of decision variables, specific application requirements, numerical stability, and computa-
tional performance analysis, especially in systems with multi-dimensional domains. In line
with our primary research objective to propose a proactive control solution for automated
driving on highways, this thesis includes the examination of hybridization techniques for
automated driving, as well as assessment of their impact on the proactivity of the resulting
MPC formulation.

1.4 Contributions and Organization of this Thesis

This thesis addresses the control of evasive maneuvers in hazardous scenarios. In summary,
the contributions of the thesis are

• presenting novel iterative cut-based PieceWise Affine (PWA) approximation ap-

proaches for a flexible approximation of multi-dimensional nonlinear systems while
ensuring adherence to user-defined error tolerances (Chapters 2 and 3),

• introducing H4MPC, a MATLAB-based hybridization toolbox, offering a user-friendly

1.4 Contributions and Organization of this Thesis 5

interface to formulate and solve optimization problems for approximating nonlinear-

ities for MPC optimization problems (Chapter 4), 1
• analyzing the sensitivity of PWA approximations to obtain quantitative bounds on
the distance between the original and the approximated local minima of a NLP with
polytopic constraints (Chapter 5),

• developing a novel hybridization method for MPC optimization problems using

the Max-Min-Plus-Scaling (MMPS) formalism along with a comparative assessment
benchmark to evaluate the resulting computational performance (Chapters 6 and 7),

• investigating real-time implementation of MPC for emergency collision avoidance

after sudden appearance of a static obstacle with extensive discussion of experimental
insights (Chapter 8), and

• proposing a proactive Stochastic Model Predictive Control (SMPC) approach for

collision avoidance in highway scenarios (Chapter 9).

Figure 1.2: Structure of this thesis

 6 1 Introduction

The structure of this thesis is designed to provide a comprehensive exploration and

1 application of hybridization techniques, as well as applying the developed frameworks to the
control problem at hand by categorizing the chapters in two parts. In Part I, we focus on the
concept of hybridization, providing solid theoretical foundation and a clear understanding
of how hybridization can be effectively implemented. In Part II, we shift our focus to the
practical application of these hybridization frameworks in various aspects of the control
problem in hazardous scenarios to demonstrate how the theoretical principles discussed
in Part I can be translated into practical solutions, addressing real-world challenges in
automated driving.
Fig. 1.2 illustrates the thesis structure. Each chapter proposes solutions to specific
questions within our main research problem. The connection between different chapters is
indicated by dashed arrowhead lines.
 7

I
Hybridization Approach to
Computational Efficiency
 9

2 2

Iterative PWA
Approximation via Cut-Based
Domain Partitioning
“Science is not a theory of reality, but a method of inquiry.”

— Donald D. Hoffman, The Case Against Reality

PieceWise Affine (PWA) approximations are widely used among hybrid modeling frameworks
as a way to increase computational efficiency in nonlinear control and optimization problems.
A variety of approaches to construct PWA approximations have been proposed, most of which
are tailored to specific application areas by using some prior knowledge of the system in
their assumptions and/or steps. In this chapter, a parametric method is proposed to identify
PWA approximations of nonlinear systems, without any prior knowledge of their dynamics or
application requirements. The algorithm defines the regions parametrically using hyperplanes
to cut the domain, and increases the number of regions iteratively until a user-defined error
tolerance criterion is met. General remarks are given on the algorithm’s implementation and
a case study is provided to illustrate its application to vehicle dynamics.

2.1 Introduction
The literature on hybrid systems provides analysis and control synthesis methodology for
systems featuring interacting continuous and discrete dynamics. To do so, a variety of
modeling frameworks have been proposed for hybrid systems, as well as proof of their
equivalence [23], such as PWA, Mixed-Logical-Dynamical (MLD), Max-Min-Plus-Scaling
(MMPS), and linear complementary systems [18]. Several of these frameworks have been
extensively studied, including for control [24] and reachability analysis [25] of MMPS
This chapter has been published in IFAC-PapersOnLine, the proceedings of IFAC World Congress [22].
 10 2 Iterative PWA Approximation via Cut-Based Domain Partitioning

systems, model predictive control design [26] and its explicit solution in some cases [27]
for MLD, as well as for continuous PWA systems [28].
Among all the hybrid modeling frameworks, PWA systems have received extensive
attention due to their simple, yet clear, formulation of the hybrid nature of the system
behavior (i.e. explicit representation of different dynamics and their activation criteria).
2 For example, the performance of discrete-time PWA systems [29], their stability criteria in
presence of uncertainty [30], their periodic solutions [31], and bifurcation phenomena [32]
were analyzed.
The PWA formalism is not only applied in domains where the hybrid nature of the
system is important, but it has also been extensively utilized in a wide range of problems
to increase computational efficiency, such as modeling prostate-specific antigen levels [33],
water motion in sewer networks [34], or cornering behavior in vehicles [35]. In some cases,
PWA approximation of a nonlinear model facilitates reduction of the nonlinear control
optimization problem into a mixed-integer programming one, while still capturing the
complexity of the nonlinear behavior.
There are two main aspects to the problem of finding a PWA approximation: optimal
partitioning of the state space into regions, and finding the optimal affine approximation
in each one. The shape and the number of the regions influence computational complexity,
the accuracy, and potential numerical issues of the final form. A higher number of regions
improves accuracy and reduces the error bound, but leads to computationally more complex
control problems. In addition, the shape and edge of the regions are of importance as the
optimization problem is most likely to encounter numerical problems, if e.g. regions have
redundant edges or gaps exist between them.
In some applications, a proper partitioning strategy is known based on heuristics or
physics-based knowledge of the system [20, 36]. In such cases, finding local affine approxi-
mations is more straightforward and can be achieved using least-squares or other regression
methods. However, a generic PWA-approximation optimization problem is combined, i.e.
both regions and local approximations are decision variables. Some techniques have been
proposed to tackle challenges due to the combined nature of the problem, like partitioning
the domain based on the variations of the nonlinear function [37], learning-based PWA
system identification using recursive adaptive control laws [38] and online observer-based
identifiers [39], or clustering approaches, either based on convex relaxation of sparse
optimization problems [40] or incorporating fuzzy-based outlier rejection and k-means
method [41].
To date, many of the developed techniques, either explicitly or implicitly, limit the
application to low dimensions or a bound on the number of local dynamics/modes [42],
and many require some prior knowledge of the PWA approximation to be found. e.g.
by employing some heuristic clustering steps [43]. Evidently, the effectiveness of the
method depends upon the application area and its requirements; the cited papers have
successfully found computationally efficient PWA models for their respective systems.
However, to the best of our knowledge, no method has been proposed that addresses
generic PWA approximation of a system, without taking specific dimensions, applications,
or assumptions into account.
In this chapter, we propose a novel iterative algorithm to find PWA approximations of
nonlinear systems satisfying a user-defined error tolerance. Our proposed approach solves
2.2 PWA Approximation 11

combined optimization problems in each iteration where parametric hyperplanes are used
to cut the domain into different regions. This results in parametric definition of regions,
which are then directly optimized as a subset of the decision variables. As the algorithm
assumes no prior knowledge of the system, it can be implemented for discrete-time and
continuous-time dynamics, as well as event-driven and time-driven dynamics, in a wide
range of application areas. In any case, the algorithm can still be simplified, curtailed, or 2
easily modified if any information on the system is available. Details of the algorithm and
parametric region definition are described in Section 2.2, accompanied by general remarks
on various steps and considerations. The algorithm is then tested using a nonlinear vehicle
model as a case study in Section 2.3. Finally, concluding remarks and suggestions for future
work are given in Section 2.4.

2.2 PWA Approximation

2.2.1 Problem Formulation
Consider a given nonlinear system with its dynamics expressed in the generic form
𝑠̇ = 𝐹 (𝑠, 𝑢), where 𝑠 ∈ ℝ𝑛 and 𝑢 ∈ ℝ𝑚 respectively represent the state and input vectors
and 𝐹 ∶ ℝ𝑚+𝑛 → ℝ𝑛 is the nonlinear function to be approximated. Without loss of gen-
erality, the augmented state vector 𝑥 = [𝑠 𝑇 𝑢𝑇 ]𝑇 is used to define 𝐹 (𝑥) B 𝐹 (𝑠, 𝑢) since the
approximated function will be selected to be affine in both the state and the input. The
augmented domain is assumed to be bounded and will be defined as dom(𝐹 ) =  ⊂ ℝ𝑚+𝑛 .
The proposed approach approximates the nonlinear function 𝐹 by a PWA function 𝑓
defined as
𝑥 ∈ 𝑝 ⟹ 𝑓 (𝑥) = 𝑓𝑝 (𝑥), 𝑓𝑝 (𝑥) = 𝐴𝑝 𝑥 + 𝐵𝑝 , (2.1)
with 𝑝 ∈ {1, 2, … , 𝑃}, where 𝑃 is the number of regions, each defined by polytope 𝑝 ⊆ ℝ𝑚+𝑛
with

𝑝 ≠ ∅, ∀𝑝 ∈ {1, … , 𝑃}, (2.2a)

int (𝑝 ) ∩ int (𝑞 ) = ∅, ∀𝑝, 𝑞 ∈ {1, … , 𝑃}, 𝑝 ≠ 𝑞, (2.2b)
𝑃
⋃ 𝑝 = , ∀𝑝, 𝑞 ∈ {1, … , 𝑃}, 𝑝 ≠ 𝑞, (2.2c)
𝑝=1

to form a partition of , with int(𝑝 ) denoting the interior of region 𝑝 . By defining the
border hyperplanes 𝐿𝑝,𝑞 ⊂ ℝ𝑚+𝑛−1 as

𝐿𝑝,𝑞 = 𝑝 ∩ 𝑞 , ∀𝑝, 𝑞 ∈ {1, 2, … , 𝑃}, 𝑝 ≠ 𝑞, (2.3)

the set of border hyperplanes forming boundaries of the region 𝑝 are represented by the
set
𝑝 = {𝐿𝑝,𝑞 | 𝑞 ∈ {1, … , 𝑃} ∧ 𝑞 ≠ 𝑝}.
For a fixed 𝑃, both the regions 𝑝 and the corresponding local affine approximations 𝑓𝑝
are obtained by finding the optimal values of the matrices 𝐴𝑝 and 𝐵𝑝 , as well as the set 𝑝
so as to minimize the squared approximation error. This is implemented by solving the
 12 2 Iterative PWA Approximation via Cut-Based Domain Partitioning

optimization problem

‖𝐹 (𝑥) − 𝑓 (𝑥)‖22
min ∫ 𝑑𝑥, s.t. (2.1) − (2.2𝑏), (2.4)
𝐴𝑝 ∈, 𝐵𝑝 ∈, 𝑝 ∈ ‖𝐹 (𝑥)‖22 + 𝜖

2 where , , and  represent the sets containing 𝐴𝑝 , 𝐵𝑝 , and 𝑝 , respectively. The term
‖𝐹 (𝑥)‖22 in the denominator is introduced such that the cost values represent the relative
error and the added scalar 𝜖 > 0 prevents division be very small values where ‖𝐹 (𝑥)‖2 ≈ 0.

2.2.2 Parametric Definition of Regions

Without loss of generality, 𝑚 + 𝑛 is assumed to an even number as 𝑚 + 𝑛 = 2𝑑, and the states
are paired in couples as (𝑥𝑖 , 𝑥𝑗 ) to form 2-dimensional subspaces. The corresponding pairs
(𝑖, 𝑗) are collected in the set Ω and the local domains 𝑖,𝑗 ⊂ ℝ2 , are defined as
𝑇
𝑥 ∈  ⟹ [𝑥𝑖 𝑥𝑗 ] ∈ 𝑖,𝑗 , ∀(𝑖, 𝑗) ∈ Ω.

Remark 2.1. It should be noted that this assumption will not pose any restrictions on the
method since for an odd (𝑚 + 𝑛) value, the cutting procedure can be easily implemented on
the unpaired single dimension as an axis.




 

Figure 2.1: Parametric definition of cutting the domain; two cases are proposed to cover all the cutting angles
within the local domain.

After pairing the states, the regions are defined by cutting  perpendicular to the (𝑥𝑖 , 𝑥𝑗 )
planes as shown in Fig. 2.1. Since the region boundaries are to be optimized, the place of
the cuts needs to be defined parametrically. To do so, two carrier lines are introduced on
opposite sides of 𝑖,𝑗 , on which points 𝛼𝑖,𝑗 and 𝛽𝑖,𝑗 can slide. As an example, Fig. 2.1 shows
three points (in yellow and orange) sliding on the carriers, where the lines connecting the
pairs of (𝛼𝑖,𝑗 , 𝛽𝑖,𝑗 ) are used to cut the domain  perpendicular to 𝑖,𝑗 .
Remark 2.2. The 𝛼𝑖,𝑗 values should be increasing and the same holds for the 𝛽𝑖,𝑗 values, since
otherwise the corresponding cuts collide.
2.2 PWA Approximation 13

Remark 2.3. Given 𝑙𝛼 and 𝑙𝛽 , the location of the points 𝛼𝑖,𝑗 and 𝛽𝑖,𝑗 on carriers parallel to the
diagonal of  can be obtained as

𝑥𝑖𝛼 = 𝑥𝑖min + 𝑋𝑖 sin2 𝜙 + 𝑙𝛼 cos 𝜙,

𝑥𝑗𝛼 = 𝑥𝑗min − 𝑋𝑖 sin 𝜙 cos 𝜙 + 𝑙𝛼 sin 𝜙,
𝑥𝑖𝛽 = 𝑥𝑖min − 𝑋𝑗 sin 𝜙 cos 𝜙 + 𝑙𝛽 cos 𝜙,
(2.5) 2
𝑥𝑗𝛽 = 𝑥𝑗min + 𝑋𝑗 cos2 𝜙 cos 𝜙 + 𝑙𝛽 sin 𝜙,

where the domain parameters 𝑋 , 𝜙, and 𝑥min associated with the 𝑖 and 𝑗 axes are shown in
Fig. 2.2.

Figure 2.2: A schematic view of the connection among various domain parameters and their relation to the
decision variables, i.e. domain cuts.

To cover all possible cutting angles, the two cases in Fig. 2.1 should be investigated
separately with different carriers. Note that requiring two cases stems from the (𝑥𝑖 , 𝑥𝑗 )
plane being 2-dimensional. For a rectangular 𝑖,𝑗 (e.g. due to bound constraints), or a
parallelogram, it is convenient to define the carriers for 𝛼𝑖,𝑗 and 𝛽𝑖,𝑗 points parallel to
one of the diagonals in each case. Nevertheless, this concept can be easily extended for
applications with other 𝑖,𝑗 forms by circumscribing a parallelogram to 𝑖,𝑗 and defining
the carriers parallel to the diagonals for each case. The numbers of cuts perpendicular to
each (𝑥𝑖 , 𝑥𝑗 ) plane, is denoted by 𝑛𝑐𝑖,𝑗 and it is equal to the number of 𝛼𝑖,𝑗 and 𝛽𝑖,𝑗 points
sliding on the carriers.

2.2.3 Approximation Algorithm

As the optimal number of regions is not known a priori, our proposed algorithm tackles
PWA approximation through an iterative loop given in Algorithm 1. The vector 𝑛c contains
the number of cuts 𝑛𝑐𝑖,𝑗 with indices 𝑖 and 𝑗 such that (𝑥𝑖 , 𝑥𝑗 ) ∈ Ω. By getting 𝑛c as input
and solving (2.4) for a fixed 𝑃 as

𝑃 = ∏ (𝑛𝑐𝑖,𝑗 + 1) ,
(𝑖,𝑗)∈Ω
 14 2 Iterative PWA Approximation via Cut-Based Domain Partitioning

the function reg_optimization(⋅) finds the optimal affine approximations and their corre-
sponding regions simultaneously. During each iteration, reg_optimization(⋅) returns both
the minimum objective 𝐽 ∗ and its corresponding optimal decision variables

𝜈∗ = (∗ , ∗ , ∗ ) ,
2
as output. The asterisk indicates the optimal value of the variable, and the border hyper-
planes are defined using the position of the 𝛼𝑖,𝑗 and 𝛽𝑖,𝑗 points as
{ }
 = (𝑙𝛼𝑘 , 𝑙𝛽𝑘 ) | 𝜅 ∈ {1, 2, … , 𝑑} , 𝑘 ∈ {1, 2, … , (𝑛c )𝜅 } .

It should be noted that (2.4) is a nonlinear optimization problem. Therefore, the function
reg_optimization(⋅) can either use a global search solver such as genetic algorithm or
particle swarm, or gradient-based approaches with multiple starting points. In both cases,
the best objective value would be the lowest value among the minima obtained in each
trial.

Algorithm 1 Iterative cut-based PWA approximation

cond ← true
𝑛c ← 01×𝑑
iter ← 0
while cond do
iter ← iter +1
𝑛cuts ← 1𝑑×1 × 𝑛c + 𝐼𝑑×𝑑
for 𝑖 ∈ {1, 2, … , 𝑑} do
𝑛in ← 𝑑-th row of 𝑛cuts
[err(𝑖), sol(𝑖)] ← reg_optimization(𝑛in )
end for
𝑑best ← arg min (err (i))
𝑖
𝑛c ← 𝑑best -th row of 𝑛cuts
if min (err) ⩽ tolerr then
cond ← false
return 𝑛c , sol (𝑑best )
else if ∏𝑑𝑞=1 𝑛c (𝑞) ⩾ 𝑁reg max then
cond ← false
return print(‘Exceeded 𝑁reg ’)
else if iter ⩾ itermax then
cond ← false
return print(‘Exceeded itermax ’)
end if
end while

However, the number of regions may not be sufficient to approximate the nonlinear
function within a particular error bound. In that case, more cuts should be introduced to
partition . To do so, the designed loops runs as follows to investigate different scenarios:
in each iteration, reg_optimization(⋅) is solved for 𝑑 cases, in which only one element in
2.2 PWA Approximation 15

𝑛c is increased by 1, and the 𝑛c with the lowest objective is selected as the best cutting
strategy for the next iteration. The algorithm stops when reaching objective values below
the error tolerance tolerr . To avoid an infinite loop, the procedure can also be stopped by
passing maximum bounds on the number of iterations or the number of regions.
Example 2.1. For  ⊂ ℝ4 , 𝑑 = 2, and Ω = {(1, 2), (3, 4)}, the algorithm starts by setting 2
𝑛c = [0 0], which means no cutting.
In the first iteration, reg_optimization(⋅) is called twice, finding the best approximations
for 𝑛c = [1 0] and 𝑛c = [0 1] which correspond respectively to making only one cut
perpendicular to 1,2 , and only one cut perpendicular to 3,4 .
If 𝑛c = [1 0] gives a lower objective, but fails to satisfy the error tolerance, the next iteration
starts with 𝑛c = [1 0], and two cases 𝑛c = [2 0] and 𝑛c = [1 1] are investigated. In other
words, if one cut on 1,2 is considered a successful cutting strategy, the next step is to improve
the result by adding more cuts to it as a baseline.

2.2.4 General Remarks

The power of Algorithm 1 stems from neither posing limits on system dimensions nor
assuming a required number of regions. Moreover, as the approximation problem (2.4) can
be solved by gridding the domain, our proposed method can also be applied in cases where
the analytical form of the nonlinear model is not available. For instance, training measure-
ment data sets can also be used to find a fitted PWA approximation using Algorithm 1. In
addition, some general notes should be made:

State-based vs. full-state partitioning: PWA approximation of 𝐹 ∶ ℝ𝑚+𝑛 → ℝ𝑛 is done

by running Algorithm 1 independently for each of the 𝑛 states. This leads to the cuts and
subsequently regions that are independently defined and evaluated for each component
of 𝑠.̇ If this is not convenient for certain applications and it is desired to have the same
regions for all the elements of 𝑠,̇ Algorithm 1 can be used in the same fashion or modified
by changing the objective in (2.4) as
‖𝑊 (𝐹 (𝑥) − 𝑓 (𝑥))‖22 + 1
min 𝑑𝑥, s.t. (2.1) − (2.2𝑏),
𝐴𝑝 ∈, 𝐵𝑝 ∈, 𝑝 ∈ ∫ ‖𝐹 (𝑥)‖22


where 𝑊 is a weight matrix.

Selection of a single case: Implementing the proposed approach is completed by run-

ning Algorithm 1 for cases 1 and 2 in parallel and choosing the best result. However,
one of the two cases may always be showing better results from the first iteration. To
avoid unnecessary computation in such instances, the cases can be tested and compared
by running the first iteration of Algorithm 1, identifying the better case (i.e. with a lower
objective), and then implementing Algorithm 1 only for that case.

Pairing the states: Pairing the states as Ω can be done arbitrarily. Prior knowledge
of the system and/or its application may suggest that specific states should be paired.
Nevertheless, the pairing can be also done by testing different combinations of Ω through
one iteration, as was proposed for evaluating cases 1 and 2.
 16 2 Iterative PWA Approximation via Cut-Based Domain Partitioning

Cases with unbounded domain: The proposed algorithm assumes the domains ( and
subsequently 𝑖,𝑗 ) to be bounded. In case of an unbounded domain, a subset of the regions
𝑝 need to be defined unbounded as well. This will not affect the decision variables in (2.4)
as the cutting places are optimized, not the regions’ boundaries. However, the objective in
(2.4) approaches infinity across an unbounded domain. To avoid this, a sufficiently large
2 bounded subset of the unbounded domain  can be used to find the PWA approximation
using our algorithm. The result can then be directly used to approximate the behavior in
the original domain.

Obtaining the cutting hyperplanes: The matrix form of the border hyperplanes ob-
tained from (2.3) can be constructed by extending the definition of the cuts. Using (2.5), a
cut 𝐿 is defined by
𝑥𝑗𝛼 − 𝑥𝑗𝛽 𝑥𝑗𝛼 − 𝑥𝑗𝛽
𝐿 B 𝑥𝑗 = 𝑥𝑖 + 𝑥𝑗𝛼 − 𝑥𝑖𝛼 .
( 𝑥𝑖𝛼 − 𝑥𝑖𝛽 ) ( 𝑥𝑖𝛼 − 𝑥𝑖𝛽 )
As each pair of cuts from different 𝑖,𝑗 are perpendicular, the resulting cutting hyperplanes
in  can be directly combined in a generic matrix form
𝐿𝑝,𝑞 B 𝐻 𝑥 + ℎ = 0.

2.3 Case Study: Vehicle Dynamics

In this section, Algorithm 1 is used to find a PWA approximation of a nonlinear model of
vehicle dynamics, integrating the coupled longitudinal and lateral dynamics in a single-
track configuration, and considering linear tire forces. The model and implementation of
the proposed approach are explained in the following sections.

2.3.1 Nonlinear Vehicle Model

A single-track representation of the vehicle is shown in Fig. 2.3. With the system variables
and parameters respectively defined in Tables 2.1 and 2.2, the nonlinear vehicle model is
described by the following equations:
1
𝑣̇ 𝑥 = [𝐹𝑥f cos 𝛿 − 𝐹𝑦f sin 𝛿 + 𝐹𝑥r ] + 𝑣𝑦 𝑟, (2.6a)
𝑚
1
𝑣̇ 𝑦 = [𝐹𝑥f sin 𝛿 + 𝐹𝑦f cos 𝛿 + 𝐹𝑦r ] − 𝑣𝑥 𝑟, (2.6b)
𝑚
1
𝑟̇ = [𝐹𝑥f sin 𝛿 𝑙f + 𝐹𝑦f cos 𝛿 𝑙f − 𝐹𝑦r 𝑙r ] , (2.6c)
𝐼𝑧𝑧
and the lateral forces are given by the linear tire model
𝐹𝑦f = 𝐶𝛼f 𝛼f , 𝐹𝑦r = 𝐶𝛼r 𝛼r ,
where the slip angles are obtained by
𝑣𝑦 + 𝑙f 𝑟 𝑣 𝑦 − 𝑙r 𝑟
𝛼f = 𝛿 − tan−1 , 𝛼r = tan−1 .
( 𝑣𝑥 ) ( 𝑣𝑥 )
2.3 Case Study: Vehicle Dynamics 17

YGlobal
y
Fxf α δ
Fyf f
vf x
Fyr vx
vy
CoG
ψ 2
r = ψ·
Fxr
XGlobal
αr
vr lf
lr
L

Figure 2.3: Configuration of the single-track vehicle model

Table 2.1: System variables

Variable Definition Unit

𝑣𝑥 Longitudinal velocity m/s
𝑣𝑦 Lateral velocity m/s
𝜓 Yaw angle rad
𝑟 Yaw rate rad/s
𝛿 Steering angle (road) rad
𝐹𝑥f Longitudinal force on the front axis N
𝐹𝑥r Longitudinal force on the rear axis N
𝐹𝑦f Lateral force on the front axis N
𝐹𝑦r Lateral force on the rear axis N
𝐹𝑧f Normal load on the front axis N
𝐹𝑧r Normal load on the rear axis N
𝛼f Front slip angle rad
𝛼r Rear slip angle rad

Table 2.2: System parameters

Parameter Definition Value Unit

𝑚 Vehicle mass 1970 kg
𝐼𝑧𝑧 Inertia moment about z-axis 3498 kg/m2
𝑙f CoG∗ to front axis distance 1.4778 m
𝑙r CoG to rear axis distance 1.4102 m
𝐶𝛼f Front cornering stiffness 126784 N
𝐶𝛼r Front cornering stiffness 213983 N
∗ Center of Gravity

2.3.2 Implementation and Results

Considering system dynamics in (2.6a) to (2.6c), Algorithm 1 is used to find PWA approxima-
tion of 𝑣̇ 𝑥 , 𝑣̇ 𝑦 , and 𝑟̇ independently. MATLAB’s Optimization toolbox is used to implement
the algorithm using lsqnonline for 10 starting points. The system is simulated during an
evasive double lane-change maneuver and the axes corresponding to the augmented state
 18 2 Iterative PWA Approximation via Cut-Based Domain Partitioning

𝑇
vector 𝑥 = [𝑣𝑥 𝑣𝑦 𝑟 𝐹𝑥f 𝐹𝑥r 𝛿 ] are paired as

Ω = {(𝑣𝑥 , 𝑟), (𝑣𝑦 , 𝛿), (𝐹𝑥f , 𝐹𝑥r )},

which results from our physics-based knowledge of the system states, their dimensions,
2 and their order of magnitude. Comparing the first iterations of cases 1 and 2 showed that
case 2 gives lower objectives when cutting  perpendicular to 𝑣𝑥 ,𝑟 and 𝑣𝑦 ,𝛿 , while case
1 is the better one to define cuts on 𝐷𝐹𝑥f ,𝐹𝑥r .

Figure 2.4: Open-loop simulation of an evasive double lane-change maneuver using nonlinear vehicle model and
two PWA approximations: LB and CB approaches

The solution time depends on the number of regions due to an subsequent increase
in the number of decision variables. The algorithm was run for different error tolerances
using the DelftBlue supercomputer [44] with every iteration for the number of regions
between 2 to 10 taking on average 435 minutes.
The approximations obtained for tolerr values in Table 2.3 using our proposed cut-based
algorithm (CB), and the Lebesgue PWA approximation (LB) approach proposed by [37],
have been compared with the nonlinear system for the open-loop system simulation in
Fig. 2.4. In the LB approach, the domain is partitioned perpendicular to each axis and
based on variation of the nonlinear function’s gradient; this results in hypercubic regions.
However, the CB approach cuts the domain perpendicular to 2-dimensional subspaces
which leads to polytopic regions. The same tolerances were selected for both algorithms
2.4 Conclusions 19

for fair comparison, and they converged to the number cuts 𝑛c defined as

𝑛c = [𝑛𝑐(𝑣𝑥 ,𝑟) , 𝑛𝑐(𝑣𝑦 ,𝛿) , 𝑛𝑐(𝐹 ].

𝑥f ,𝐹𝑥r )

The total number of regions 𝑁 is listed as well in Table 2.3. Fig. 2.4 shows that the CB
approach provides a more accurate approximation of the model, and its good performance 2
is better seen in 𝑣̇ 𝑥 which has a higher degree of nonlinearity where CB gives a better
approximation while introducing a smaller number of regions.

Table 2.3: The number of cuts at convergence for case study instances with different error tolerance values

𝐯̇ 𝐱 𝐯̇ 𝐲 𝐫̇
Instance
𝐭𝐨𝐥𝐞𝐫𝐫 0.30 0.10 0.05
𝐧c [1,3,0] [0,0,0] [0,0,0]
𝐏𝐖𝐀 − 𝐋𝐁
𝐍 8 1 1
𝐧c [0,3,0] [1,0,0] [1,0,0]
𝐏𝐖𝐀 − 𝐂𝐁
𝐍 4 2 2

2.4 Conclusions
In this chapter, an iterative algorithm for PWA approximation of nonlinear systems was
proposed assuming no prior knowledge of the application area. By using a cut-based
parametric definition of the regions in the optimization problem, the algorithm aims at
finding an optimal partitioning of the domain into polytopic regions and the corresponding
local affine approximations, simultaneously. This combined optimization problem is solved
in each iteration for several cases of adding new cuts whereas the number of cuts is
increased in each iteration until a user-specified error tolerance is reached. The algorithm
is implemented on a nonlinear vehicle model as a case study where different error tolerances
were selected for each state and the results were compared to another PWA approximation
approach from the literature, where similar to our proposed algorithm, the regions are
included parametrically in the decision variables of the combined optimization problem.
The comparison shows that our approach gives more a accurate approximation of the
nonlinear system, in some cases with fewer number of regions.
In future work, the current algorithm can be improved along two lines. First, the
iteration law can be enhanced for faster convergence to the optimal number of regions
while avoiding introduction of extra and/or redundant cuts. For instance, instead of
increasing the number of cuts in each iteration by one, more cuts can be introduced
based on the difference of the objective functions between the last two iterations. Second,
adjustments or additions to the algorithm structure can be introduced for applications
where discontinuity is problematic, to either avoid discontinuity on the region borders in
the obtained PWA approximation, or to circumvent its undesired consequences (e.g. in
switching analysis or control synthesis) by defining auxiliary affine dynamics or switching
rules along the borders. Moreover, on the application level we aim at investigating the
performance of our proposed approximation method on a wider variety of test cases, i.e.
driving scenarios.
 21

3 3
PWA Approximation of
Multi-Dimensional
Nonlinear Systems
“Behold me - I am a Line, the longest in Lineland, over six inches of Space”.
“Of Length,” I ventured to suggest.
“Fool”, said he, “Space is Length.”

— Edwin Abbott, Flatland: A Romance of Many Dimensions

PieceWise Affine (PWA) approximations for nonlinear functions have been extensively used
for tractable, computationally efficient control of nonlinear systems. However, reaching a
desired approximation accuracy without prior information about the behavior of the nonlinear
systems remains a challenge in the function approximation and control literature. As the
name suggests, PWA approximation aims at approximating a nonlinear function or system by
dividing the domain into multiple subregions where the nonlinear function or dynamics is
approximated locally by an affine function also called local mode. Without prior knowledge of
the form of the nonlinearity, the required number of modes, the locations of the subregions, and
the local approximations need to be optimized simultaneously, which becomes highly complex
for large-scale systems with multi-dimensional nonlinear functions. This chapter introduces a
novel approach for PWA approximation of multi-dimensional nonlinear systems, utilizing a
hinging hyperplane formalism for cut-based partitioning of the domain. The complexity of the
PWA approximation is iteratively increased until reaching the desired accuracy level. Further,
the tractable cut definitions allow for different forms of subregions, as well as the ability to
impose continuity constraints on the PWA approximation. The methodology is explained via
multiple examples and its performance is compared to two existing approaches through case
studies, showcasing its efficacy.
This chapter has been submitted to Automatica.
 22 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

3.1 Introduction
PWA systems are a class of hybrid modeling frameworks where the dynamics is expressed
by multiple subsystems, i.e. local modes, that are affine functions of states and inputs and
are active on a partition of the domain region, i.e. subregions [45]. PWA approximation of
nonlinear functions has long been used in diverse applications, contributing to enhanced
modeling power [46], improved computational efficiency [47], identification of explicit
control laws [48] or serving as a descriptor for neural networks in machine learning [49].
Moreover, complexity of PWA approximations [50] and their verification processes [51]
3 has been examined in the literature.
Approximating a nonlinear function by a PWA form is rather straightforward if the
subregions are known a priori, as the problem boils down to determining local affine
approximations on each subregion. The knowledge of the subregions may arise from the
knowledge of different regimes (refer to the tire model in [52]) or from the knowledge of
different equilibria (refer to the chaos model in [53]). However, in numerous applications,
the difficulty arises when we lack prior information about the location of the subregions
and the quantity of local modes to reach a particular approximation accuracy.
The conceptualization of PWA approximation as an optimization problem becomes
notably intricate when dealing with multi-dimensional nonlinear functions. Even for
known, yet multi-dimensional models such as in resistor networks [54], analytical solution
of the PWA approximation problem may be elusive and the optimization problem should
better be formulated for a set of points sampled from the domain [55]. This idea resembles
PWA approximation approaches learned through experimental data [56]. A question
arise about how to sample the points. A trivial solution is taking as many subregions
as data points [45]; however, this approach easily leads to overfitting. Therefore, for the
optimization problem to become well-posed, the number of local modes is often fixed while
minimizing approximation error [45] or its expectation [57].
Various methods have been used to formulate the PWA approximation problem [45, 58,
59]. A common formulation is a bi-level optimization problem [60] that can be recast into
a mixed-integer program [61], or solved in a recursive manner [62–65]. While recursive
solutions are fast and can be used for online PWA approximation [39], they are often
limited in handling multi-dimensional systems and are most effective when the form of
the subregions is partially known and just needs to be refined [38]. For instance, in [66]
more vertices are iteratively added to the subregions for improved accuracy, but the solver
needs to be properly initialized.
Bi-level optimization arises because PWA approximation essentially has two key aspects:
establishing a partitioning strategy to divide the domain into subregions, and finding the
local affine approximations. A popular partitioning strategy is clustering of the mesh
points [40, 64, 67, 68], which can be sensitive to the mesh quality [65]. Despite the efforts
to reduce the sensitivity to the cluster boundary and outliers [41], the performance of
clustering-based partitioning degrades for multi-dimensional nonlinear functions. Some
formulations use a specific shape for the partition, e.g. using hyper-rectangular subregions
for digital systems [69], using the function gradient [37], which is only applicable for
uni-dimensional domains, or simplical representation [70], which is applicable for low-
dimensional domains [58]. Conversely, the hinging hyperplanes formalism, where the
function is defined as a sum of hinging functions, e.g. min and max, of parameterized
3.2 Problem Formulation 23

hyperplanes, have proven to be an efficient and tractable formulation for partitioning of

multi-dimensional domains [71].
If the approximation problem is solved offline, the main goal is to converge to the
most accurate PWA approximation with minimal complexity. This essentially requires
a partitioning strategy that is flexible enough to divide the domain in various ways and
consequently, allow for finding a PWA approximation of a desired accuracy via defining
the lowest possible number of subregions. Additionally, an appropriate methodology
should be able to yield a continuous PWA approximation, which was handled by using
min and max operators in [72]. However, using max and min operators does not allow
for a discontinuous form. In this sense, using the continuity constraint from [54] is more
3
suitable for more flexibility.
In this chapter, we present a novel approach for PWA approximation of multi-dimensional
nonlinear systems using cut-based partitioning of the domain via a hinging hyperplane
formalism. The complexity level of the PWA form can be iteratively increased if the in-
tended level of complexity is either unknown or not restricted. In our earlier work [73],
we have tackled the PWA approximation of multi-dimensional nonlinear systems with no
prior knowledge. In this chapter, we generalize the cutting definitions and the formulation
of the optimization problem, enhancing the flexibility of the partitioning strategy. These
extensions enable evaluating cuts that were not possible with our prior approach, and allow
for a tractable implementation of continuity constraints in the approximation problem.
In summary, our novel PWA approximation approach improves upon the state-of-the-art
methods by (1) flexible definition of subregions using a generalized hinging hyperplane
formalism to allow for finding PWA approximations with fewer number of subregions
for a given approximation error tolerance, and (2) tractable bi-level formulation of the
optimization problem to facilitate modifications based on applications e.g. imposing the
continuity constraint on the PWA approximation. The efficacy of these extensions will
be shown via illustrative examples. Further, we compare our method to the approach
in [65] where a recursive solution of the PWA approximation problem helps reducing the
complexity and we show that our optimization formulation and cut-based partitioning
allows for convergence to the same accuracy level with fewer number of subregions.
The structure of the chapter is as follows: Section 3.2 covers the main definitions
and the formulation of the approximation problem. The novel cut-based partitioning
strategy is explained in Section 3.3 using examples for clarity, and Section 3.4 presents the
resulting optimization problem and the solution procedure. Case studies and comparisons
are described in Section 3.5, while Section 3.6 concludes this chapter.

3.2 Problem Formulation

Let us consider a nonlinear system with dynamics

𝑠̇ = 𝐹 (𝑠, 𝑢),

where 𝑠 ∈ ℝ𝑛 and 𝑢 ∈ ℝ𝑚 respectively represent the state and input vectors and 𝐹 ∶ ℝ𝑛+𝑚 → ℝ𝑛
is the nonlinear function to be approximated. Without loss of generality, the augmented
state vector 𝑥 = [𝑠 𝑇 𝑢𝑇 ]𝑇 is used to define 𝐹 (𝑥) B 𝐹 (𝑠, 𝑢) since the approximated function
will be selected to be affine in both the state and the input. The augmented domain is
 24 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

assumed to be bounded and is defined as dom(𝐹 ) =  ⊂ ℝ𝑛+𝑚 . For brevity of the expressions,
hereafter we will use 𝑑 = 𝑛 + 𝑚 as the dimension of the domain.

Definition 3.1 (Domain ). The domain  ⊂ ℝ𝑑 is defined by the scalar boundary function
𝑔 ∶ ℝ𝑑 → ℝ as
 B {𝑥 ∈ ℝ𝑑 | 0 ⩽ 𝑔(𝑥) ⩽ 1}.

Remark 3.1. We use the normalized form 0 ⩽ 𝑔 ⩽ 1 instead of 𝑔 ⩾ 0 to avoid numerical

issues in solving the approximation optimization problem.
3
The proposed approach approximates the nonlinear function 𝐹 by a PWA function 𝑓
defined as
𝑥 ∈ 𝑝 ⟹ 𝑓 (𝑥) = 𝑓𝑝 (𝑥), 𝑓𝑝 (𝑥) = 𝐽𝑝 𝑥 + 𝐾𝑝 , (3.1)
with 𝑝 ∈ {1, 2, … , 𝑃}, where the matrices 𝐽𝑝 ∈ ℝ𝑛×𝑑 and the vectors 𝐾𝑝 ∈ ℝ𝑛 describe in total
𝑃 local affine modes, each defined on a polytopic subregion 𝑝 ⊂ 𝑅𝑑 such that the polytopes
form a partition of , i.e. the subregions are nonempty,

𝑝 ≠ ∅, ∀𝑝 ∈ {1, … , 𝑃} (3.2a)

they are non-overlapping,

int (𝑝 ) ∩ int (𝑞 ) = ∅, ∀𝑝, 𝑞 ∈ {1, … , 𝑃}, 𝑝 ≠ 𝑞 (3.2b)

and their union covers the entire domain,

𝑃
⋃ 𝑝 = , (3.2c)
𝑝=1

with int(𝑝 ) denoting the interior of region 𝑝 .

Definition 3.2 (Region set ). The region set  is the ordered set collecting the partition
(i.e. the subregions) as
 B {1 , 2 , … , 𝑃 }.

For a fixed 𝑃, the region set  and the corresponding local affine approximations 𝑓𝑝
are obtained simultaneously via solving the optimization problem

‖𝐹 (𝑥) − 𝑓 (𝑥)‖22
min ∫ 𝑑𝑥, (3.3)
 , ,  ‖𝐹 (𝑥)‖22 + 𝜖

subject to (3.1) − (3.2), (3.4)

to minimize the squared approximation error where  and  represent the ordered sets
containing 𝐽𝑝 and 𝐾𝑝 , respectively. The term ‖𝐹 (𝑥)‖22 in the denominator is introduced such
that the cost values represent the relative error and the added scalar 𝜖 > 0 prevents division
by very small values where ‖𝐹 (𝑥)‖2 ≈ 0.
3.3 Parametric Region Definition 25

3.3 Parametric Region Definition

We partition  into 𝑃 subregions by cutting it using (𝑑 − 1)–dimensional hyperplanes.

Definition 3.3 (Cutting hyperplane 𝐻𝑖 ). The 𝑖-th cutting hyperplane 𝐻𝑖 is an affine subspace
of ℝ𝑑 defined as
𝐻𝑖 B {𝑥 | ℎ𝑇𝑖 𝑥 − 1 = 0},
for 𝑖 ∈ {1, … , 𝑛c } where 𝑛c represents the number of cuts.

Definition 3.4 (Cut arrangement ). The cut arrangement  is the arrangement of 𝑛c 3
cutting hyperplanes defined by the set

 = {𝐻1 , 𝐻2 , … , 𝐻𝑛c }.

Remark 3.2. In principle, the number of subregions generated by cutting ℝ𝑑 via the ar-
rangement  can be calculated using Zaslavsky’s theorem [74] provided that all the possible
0- to (𝑑 − 1)-dimensional intersections of the hyperplanes in  are obtained. As a more
computationally-efficient approach, here we fix the number of cutting hyperplanes and numer-
ically obtain the region set  within  by investigating the existence of the possible subregions
created by  without counting the 0- to (𝑑 − 1)-dimensional intersections. As a result, 𝑃 is not
fixed a priori.

To define each cutting hyperplane, we generate 𝑑 points in ℝ𝑑 and find the hyperplane
passing through them as shown in Fig. 3.1. These points are defined on the surface of a
enclosing hypersphere 𝑆 in  to ensure they are linearly-independent.

Definition 3.5 (Enclosing hypersphere 𝑆). The enclosing hypersphere 𝑆 is the smallest
𝑑–dimensional hypersphere enclosing  defined by

𝑆 B {𝑥 | ‖𝑥‖22 − 𝜌2 = 0},

where the constant 𝜌 is the radius of the enclosing hypersphere.

Remark 3.3. Without loss of generality, one can always define a coordinate shift for the
domain so that the center of the hypersphere is located at the origin. . This allows to simplify
the expressions and mathematical manipulations.

(a) Definition of points on 𝑆 (b) Definition of 𝐻1 using 𝑑 points on 𝑆 (c) Cutting  using 𝐻1

Figure 3.1: A schematic view cutting the domain using a hypersphere.

 26 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

To locate the points on 𝑆, we use spherical coordinates (𝑟 = 𝜌, 𝜙1 , 𝜙2 , … , 𝜙𝑑−1 ) to param-

eterize the locations using spatial angles
𝜙1 , 𝜙2 , … , 𝜙𝑑−2 ∈ [0, 𝜋], 𝜙𝑑−1 ∈ [0, 2𝜋].
Figure 3.1a shows this concept for 𝑑 = 3. The Cartesian coordinates of the 𝑘-th point
𝑥𝑘 = [𝑥1,𝑘 , … , 𝑥𝑑,𝑘 ]𝑇 can then be obtained by [75]
𝑗−1
𝑥𝑗,𝑘 = 𝜌 cos 𝜙𝑗,𝑘 ∏ sin 𝜙𝜈,𝑘 , ∀𝑗 ∈ {1, … , 𝑑 − 1}, (3.5a)
3 𝜈=1

𝑑−1
𝑥𝑑,𝑘 = 𝜌 ∏ sin 𝜙𝜈,𝑘 , (3.5b)
𝜈=1
where 𝑥𝑗,𝑘 is the 𝑗 th component of the 𝑘 th point. Then, to find the hyperplane 𝐻𝑖 passing
by 𝑑 points, we need to solve
𝑋𝑖 ℎ𝑖 = 𝟏𝑑×1 ⟹ ℎ𝑖 = 𝑋𝑖−1 𝟏𝑑×1 , (3.6)
where
⎡𝑥1,1 𝑥1,2 …
𝑥1,𝑑 ⎤
⎢𝑥 𝑥2,2 𝑥2,𝑑 ⎥
…
𝑋𝑖 = ⎢ 2,1 ⎥.
⎢ ⋮ ⎥
⎣𝑥𝑑,1
𝑥𝑑,2 … 𝑥𝑑,𝑑 ⎦
Figure 3.1b illustrates the generation of one hyperplane 𝐻1 for 𝑑 = 3 and how the domain
is cut into two partitions in Fig. 3.1c. For more cuts, we then need to proceed analogously
for all 𝑛c cuts, obtaining  as
 = {𝟏1×𝑑 𝑋𝑖−𝑇 𝑥 = 1}, ∀𝑖 ∈ {1, … , 𝑛c }. (3.7)
Since each hyperplane divides the domain into two half-spaces, we define the map 𝜎
from the domain  to the 𝑛c × 1 Boolean vector 𝜎 as
{
0 if ℎ𝑇𝑖 𝑥 < 1
𝜎𝑖 (𝑥) = , ∀𝑖 ∈ {1, … , 𝑛c }, (3.8)
1 if ℎ𝑇𝑖 ⩾ 1
to indicate which side of the hyperplane 𝐻𝑖 the point 𝑥 lies on. Since the subregions are
also located on one side of each hyperplane, there exist at most 2𝑛c possible partitions that
can be stored in an 𝑛c × 2𝑛c matrix. However, to avoid unnecessary usage of memory, we
suggest generating 𝜎 vectors by investigating the binary vectors corresponding to integer
numbers from 0 to 2𝑛c − 1 without storing all of them in a very large matrix. We use the
prune-and-search paradigm [76] by solving 2𝑛c linear programs to check the feasibility of
each combination
min 1, (3.9)
𝑥
subject to 𝑥 ∈ , (3.10)
ℎ𝑇𝑖 𝑥 <1 if 𝜎𝑖 = 0, 𝑖 ∈ {1, … , 𝑛c }, (3.11)
ℎ𝑇𝑖 𝑥 ⩾ 1 if 𝜎𝑖 = 1, 𝑖 ∈ {1, … , 𝑛c }, (3.12)
3.3 Parametric Region Definition 27

where 𝜎 ∈ ℝ𝑛c is the binary representation of the integer 𝑗 ∈ {0, … , 2𝑛c − 1} in each linear
program. The 𝜎 vectors corresponding to feasible problems are then stored in the feasibility
matrixΣ ∈ ℝ𝑛c ×𝑃 . This procedure is implemented by Algorithm 2.

Algorithm 2 Find feasibility matrix of a cut arrangement as Σ = chambers(, )

Require: , 
𝑛c ← || ⊳ |.| denotes cardinality
𝜎 ← 0𝑛c ×1
Σ ← 0𝑛c ×0 ⊳ 0𝑛c ×0 is an empty matrix 3
for 𝑙 ∈ {0, 1, … , 2𝑛c − 1} do
binary vector
𝜎 ←−−−−−−−−−− 𝑙
if (3.9)-(3.12) feasible for (𝜎, ) then
Σ ← [Σ 𝜎 ]
end if
end for
return Σ

Example 3.1. Consider the 3-dimensional hypercube domain  shown in Fig. 3.2 and two
cut arrangements 𝑎 and 𝑏 respectively shown in Figures 3.2a and 3.2b. We have

binary
𝑛c = 2 ⟹ 𝑙 ∈ {0, 1, 2, 3} −−−−−→ {00, 01, 10, 11}.

Problem (3.9)-(3.12) is not feasible for 𝜎 = 01 for 𝑎 as there is no region within  lying above
𝐻1 and below 𝐻2 . However, (3.9)-(3.12) is feasible for 𝜎 ∈ {00, 10, 11}. Therefore, the feasibility
matrices for 𝑎 and 𝑏 are

0 1 1 0 0 1 1
Σ𝑎 = , Σ𝑏 = .
[0 0 1] [0 1 0 1]

(a) 𝑎 (b) 𝑏

Figure 3.2: A schematic view of two cut arrangements in Example 1.

In the combination geometry literature, a similar concept is used but in a set of tuples called the oriented metroid.
For more details see [77].
 28 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

Next, we aim at finding the neighboring subregions to identify the hyperplane they
share as their boundary. We define the adjacency matrix 𝐴 to store this information.
Definition 3.6 (Adjacency Matrix 𝐴). The adjacency matrix 𝐴 ∈ ℝ𝑃×𝑃 represents the neigh-
boring subregions within the region set  as follows:
{
𝑖 if 𝑝 ∩ 𝑞 = 𝐻𝑖
𝐴𝑝,𝑞 = .
0 otherwise
The adjacency matrix is symmetric by definition.
3 Note that the adjacent subregions share their 𝜎 vector, except for only one element,
which is the element corresponding to their boundary hyperplane. Thus, the adjacency
matrix is constructed by investigating the columns in Σ. Since each column in Σ is a binary
vector, two columns differ only in one element if and only if their subtraction contains
only one ±1 element, i.e.
‖Σ.,𝑝 − Σ.,𝑞 ‖1 = 1.
Each column of Σ represents a subregion and how it relates to the hyperplanes in .
However, defining each region is done only by evaluating its boundaries and not all the
cutting hyperplanes to avoid redundancy. Since each column (or row) of 𝐴 corresponds to
one of the elements in , the subregion 𝑝 can now be formulated as follows:
{ }
|
𝑝 = 𝑥 ∈ , ∀𝑖 > 0, 𝐴𝑝,. = 𝑖 | (−1)Σ𝑖,𝑝 ℎ𝑇𝑖 𝑥 ⩽ (−1)Σ𝑖,𝑝 . (3.13)
|
The subregions are then stored in the region set . Algorithm 3 describes the procedure of
obtaining  from  and Σ using the adjacency matrix 𝐴.
Example 3.2. Consider the 3-dimensional domain shown in Fig. 3.3 as
 B { 𝑥 ∈ ℝ3 | 𝑥𝑗 ∈ [−2, 2] , 𝑗 ∈ {1, 2, 3}},
with the cut arrangement  = {𝐻1 , 𝐻2 , 𝐻3 } shown in Fig. 3.3a where
𝐻1 B { 𝑥 ∈ ℝ3 | − 𝑥1 + 2𝑥2 + 5𝑥3 = 1},
𝐻2 B { 𝑥 ∈ ℝ3 | 0.1𝑥1 − 0.5𝑥2 − 0.2𝑥3 = 1},
𝐻3 B { 𝑥 ∈ ℝ3 | − 𝑥1 + 𝑥2 = 1}.

(a) Cutting the domain (b) Resulting subregions

Figure 3.3: Illustration of the cut arrangement and the resulting subregions in Example 2.
3.4 Cut-Based PWA Approximation 29

Algorithm 3 Find subregions of a cut arrangement as (, 𝐴, 𝑃) = regions(, Σ)

Require: , Σ
𝑛c ← || ⊳ |.| B cardinality
𝑃 ← number of columns in Σ
𝐴 ← 0𝑃×𝑃
←∅
for 𝑝 ∈ {1, … , 𝑃} do
for 𝑞 ∈ {𝑝 + 1, … , 𝑃} do
𝛿 ← Σ.,𝑝 − Σ.,𝑞
if ‖𝛿‖1 = 1 then
3
𝐴𝑝,𝑞 ← index of nonzero component in 𝛿
end if
end for
end for
𝐴 ← 𝐴 + 𝐴𝑇 ⊳ upper-triangular to symmetric
for 𝑝 ∈ {1, … , 𝑃} do
solve (3.13)
𝑝 ←−−−−−−−−− , 𝐴
 ←  ∪ {𝑝 }
end for
return , 𝐴, 𝑃

The feasibility and adjacency matrices for  are

⎡0 3 2 1 0⎤
⎡0 0 0 1 1⎤ ⎢3 0 0 0 1⎥
⎢ ⎥
Σ = ⎢0 0 1 0 0⎥ , 𝐴 = ⎢2 0 0 0 0⎥ ,
⎢ ⎥ ⎢1
⎣0 1 0 0 1⎦ 0 0 0 3⎥
⎢0 1 0 3 0⎥⎦
⎣
which gives the subregions as depicted in Fig. 3.3b expressed by
⎡−1 2 5 ⎤
−1 2 5
1 B { 𝑥 ∈  | ⎢0.1 −0.5 −0.2⎥ 𝑥 ⩽ 1}, 2 B { 𝑥 ∈  | 𝑥 ⩽ 1},
⎢ ⎥ [1 −1 0]
⎣−1 1 0 ⎦
1 −2 −5
3 B { 𝑥 ∈  | [−0.1 0.5 0.2] 𝑥 ⩽ 1}, 4 B { 𝑥 ∈  | 𝑥 ⩽ 1},
[−1 1 0]
1 −2 −5
5 B { 𝑥 ∈  | 𝑥 ⩽ 1},
[1 −1 0]

3.4 Cut-Based PWA Approximation

The PWA approximation of a nonlinear function is obtained by solving (3.3) by optimizing
the partitioning strategy via  and the local affine approximations  and , while satisfy-
ing (3.2). Using the cut-based parametric region definition through Algorithms 2 and 3,
 30 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

constraints (3.2) are automatically satisfied since we cut the domain with hyperplanes
which gives  as a set of non-overlapping partitions that cover the whole .
If the aim is to give a continuous PWA approximation of the system, the following con-
straint should be imposed on the dynamic modes corresponding to neighboring subregions,
which is derived based on [78]:
∃𝑐 ∈ ℝ𝑑 s.t. 𝐽𝑝 − 𝐽𝑞 = 𝑐 ℎ𝑇𝑖 , ∀𝑝, 𝑞 ∈ {1, … , 𝑃}, 𝐻𝑖 = 𝑝 ∩ 𝑞 . (3.14)
Corollary 3.1. The difference of 𝐽𝑝 and 𝐽𝑞 in (3.14) is of rank one [78]. As a result, the
continuity can be imposed on the PWA approximation by considering rank of the Jacobian
3 matrices in neighboring modes.
Even with a fixed number of cutting hyperplanes, the number of subregions (and con-
sequently the dimensions of the decision space) differs for different  arrangements with
the same 𝑛c . Due to its varying-dimensional nature, we formulate the PWA approximation
problem as a bi-level optimization problem. At the lower level, we find the minimum
approximation error for the region set  as
𝑃 ‖𝐹 (𝑥) − 𝐽𝑝 𝑥 − 𝐾𝑝 ‖2
Γ∗ () = min ∑∫ ‖ 2
‖2 𝑑𝑥, (3.15)
 ,
𝑝=1 ‖𝐹 (𝑥)‖ 2 + 1
𝑝

s.t. rank (𝐽𝑝 − 𝐽𝑞 ) = 1, (3.16)

∀𝑝, 𝑞 ∈ {1, … , 𝑃}, 𝐴𝑝,𝑞 ≠ 0,
where (3.16) is the continuity constraint and can be disregarded in case a continuous PWA
form is not required. Problem (3.15)-(3.16) is a nonlinear least-squares optimization problem
that can be solved using e.g. gradient-based methods with multiple starting points.
At the higher level, we solve the following optimization problem:
min Γ∗ () + 𝜆𝑃 (3.17)
𝜙

s.t.  = {𝟏1×𝑑 𝑋𝑖−𝑇 𝑥 = 1}, ∀𝑖 ∈ {1, … , 𝑛c }, (3.18)

Σ = chambers(, ), (3.19)
(, 𝐴, 𝑃) = regions(, Σ), (3.20)
𝜙𝑖,𝑗 ∈ [0, 𝜋], 𝜙𝑖,𝑑−1 ∈ [0, 2𝜋], ∀𝑖 ∈ {1, … , 𝑛c }, ∀𝑗 ∈ {1, … , 𝑑 − 2}, (3.21)
with 𝜙 collecting 𝜙𝑖,𝑗 , being the approximation error obtained in the lower level and the
Γ∗
chambers and regions functions correspond to the Algorithms 2 and 3, respectively.
Here we penalize the number of subregions 𝑃 as well by a regularizing weight 𝜆 > 0. Similar
to the lower-level problem, (3.17)-(3.21) is a nonlinear optimization problem. However,
being not smooth, we propose solving it using global optimization methods such as genetic
algorithm.
Equations (3.15)-(3.21) are solved for a fixed 𝑛c . In our approach, we start by 𝑛c = 1 and
in case the best solution does not reach a user-defined error tolerance, we increase 𝑛c and
solve (3.15)-(3.21) again. Algorithm 4 describes this iterative cut-based PWA optimization
problem.
The term “chambers” is often used in the combinatorial geometry literature for the “subregion” concept. Here, to
distinguish between the functions in Algorithms 2 and 3, we use this term as a label for clarity.
3.5 Results and Discussion 31

Algorithm 4 Iterative cut-based PWA approximation

Require: 𝐹 , 
generate 𝑆
cond ← true
𝑛c ← 0
iter ← 0
while cond do
𝑛c ← 𝑛c + 1
iter ← iter +1 3
(Γ∗ ,  ∗ , ∗ , ∗ ) ← solve (3.15)-(3.21)
if Γ∗ ⩽ tolerr then
cond ← false
return  ∗ , ∗ , ∗
else if iter ⩾ itermax then
cond ← false
return print(‘Exceeded itermax ’)
end if
end while

3.5 Results and Discussion

In this section, we analyze the performance of our proposed cut-based PWA approximation
method in three steps: first, we use a case study as an illustrative example of a nonlinear
function for a user-defined approximation error tolerance. Then, we compare our method
with [73] and [65] to show the flexibility of our approach, leading to lower approximation
errors for a lower complexity of PWA the form.

3.5.1 Case Study

To illustrate our cut-based PWA approximation approach, consider the system
𝑥̇ = sin (𝑥 + 𝑢2 ),
with a 2-dimensional domain
 B { (𝑥, 𝑢) ∈ ℝ2 | 𝑥 ∈ [−2, 2] , 𝑢 ∈ [−2, 2] },
to be able to display the function and the subregions. We approximate the nonlinear
function Using Algorithm 4 by selecting the stopping criterion
tolerr = 5%
and imposing the continuity constraint in (3.16). The algorithm reaches the stopping
criterion with 𝑛c = 8 cuts by partitioning the domain into 𝑃 = 16 subregions. Figure 3.4
shows the resulting PWA approximation, region set, and the cutting hyperplanes.

3.5.2 Comparison with Approaches in the Literature

The comparison analysis is done separately for each approach from the literature, using
case studies form its respective publication.
 32 3 PWA Approximation of Multi-Dimensional Nonlinear Systems

2 𝐻1 𝐻2
𝐻3 𝐻4
0

𝑢
𝐻5 𝐻6
1 −2 𝐻7 𝐻8
−2 0 2
𝑠̇

0 𝑠
2 1 2
−1 0 3 4

𝑢
2 5 6
−2 7 8
3 0
−1 −2
−2 0 2 9 10
𝑠 1 0 𝑠 11 12
−2 2 13 14
𝑢
15 16

(a) Nonlinear and PWA functions (b)  and 

Figure 3.4: Cut-Based PWA approximation of the nonlinear function 𝑠̇ = sin(𝑠 + 𝑢2 ) using 8 cuts. Each local mode
is shown in the same color as its corresponding subregion.

Mesh-based recursive abstractions

Here, we use the same benchmark incorporated in [65]: the Dubins vehicle dynamics
𝐹 ∶ [−2, 2] × [0, 2𝜋] → ℝ given by

𝐹 (𝑥) = 𝑥1 cos(𝑥2 ).

The abstraction error is defined as the distance between the upper and lower PWA approx-
imations of 𝐹 . Therefore, for a fair comparison, we assume the approximation error to be
half of the abstraction error.
Table 3.1 compares the number of required subregions to reach three different user-
defined maximum approximation errors. It is evident that the new cut-based approach can
reach the same accuracy level with significantly fewer subregions. However, it should be
noted that the two methods introduced in [65] converge to their minima faster than our
approach. Therefore, the benefit of the new cut-based method is best realized for offline
computations aimed at reaching a more accurate, yet simpler, PWA approximation forms.

Table 3.1: Number of required subregions using different PWA approximation methods

Maximum error
Approach 10% 5% 2.5%
Cut-based approach 12 24 40
Method I [65] 58 108 210
Method II [65] 61 121 257

Parametric plane-based cutting strategy

In the previous chapter, a cut-based PWA approximation is done by pairing the input-state
vectors to form 2D hyperplanes and the cutting hyperplanes are defined perpendicular to
3.6 Conclusions 33

the selected pairs. As one of the illustrative examples in [73], the longitudinal velocity of a
single-track vehicle model with linear tires is denoted by 𝐹 ∶ ℝ6 → ℝ is given by

𝑢1 cos(𝑢3 ) + 𝑢2 𝑥2 + 𝜂3 𝑥3
𝐹 (𝑥, 𝑢) = + 𝜂2 tan−1 −𝑢 ,
𝜂1 [ ( 𝑥1 ) 3 ]

with 𝜂1 = 1970, 𝜂2 = 64.36, and 𝜂3 = 1.48, incorporating 3 states and 3 inputs. To approximate
𝐹 , 3 non-intersecting cuts are required on (𝑥2 , 𝑢3 ) to partition the domain into 4 subregions
and achieving the – significantly high – approximation error of 30%. Moreover, the resulting
PWA approximation is not continuous. 3
Table 3.2 compares different cut-based PWA approaches in terms of the continuity of
their respective number of regions, their maximum error, and the continuity of their PWA
form. Using our proposed cutting strategy, we are no longer limited to plane-based cutting
of the domain, which significantly improves our flexibility in partitioning the domain,
which results in obtaining a PWA approximation with 3 local modes for a maximum error
of 3%. Moreover, we can now enforce a continuous PWA approximation of the dynamics.
As a result, we are able to reach 3% approximation error using only two cuts: one on 𝑣𝑦 − 𝑟
and one on 𝑣𝑥 − 𝑣𝑦 axis.

Table 3.2: Comparing cut-based PWA approximation approaches

Approach Continuous form Number of regions Maximum error

Current method Yes 4 3%
Current method No 3 3%
Previous method [73] No 4 30%

3.6 Conclusions
In this chapter, we have proposed a novel approach for PWA approximation of nonlinearities
using a hinging-hyperplane formulation to partition the domain into subregions. Our
proposed method does not require prior knowledge of the dynamics, is applicable to
nonlinear systems defined on multi-dimensional domains, and allows for a straightforward
formulation of the continuity constraint for the PWA approximation. To avoid unnecessary
complexity in the final form, the number of cutting hyperplanes is iteratively increased in
case the solutions of the approximation problem are unable to satisfy a user-defined error
tolerance. The flexibility of our proposed approach allows for various polytopic subregion
definitions and adaptability for different approximation requirements. By comparing the
performance of our approach to other state-of-the-art methods from the literature, we have
showcased its potential for practical applications in complex, large-scale systems, paving
the way for future advancements in nonlinear function approximation and control.
 35

4
H4MPC: A Hybridization
4
Toolbox for MPC
Adults are bewitched by their language because they try to apply discrete words to continuous
activities. Before the breakaway of natural language, words were activities. Yet we speak, very
discretely. This is just what, and all, we can do.

— Council of Human Hybrid-Attractors, Incessance: Incesancia

The computational complexity of Nonlinear Model Predictive Control (NMPC) poses a signif-
icant challenge in achieving real-time levels of 4 and 5 of automated driving. This chapter
presents the open-access Hybridization toolbox for Model Predictive Control (MPC) (H4MPC),
targeting computational efficiency of NMPC thanks to several modules to hybridize NMPC
optimization problems commonly encountered in automated driving applications. H4MPC
is designed as a user-friendly solution with a graphical user interface within the MATLAB
environment. The toolbox facilitates intuitive and straightforward customization of the
hybridization process for any given function appearing in the equality or inequality con-
straints within the MPC framework. The initial release, Version 1.0, is freely available from
https:// bit.ly/ H4MPCV1. To provide a clear illustration of the toolbox capabilities, we present
two case studies: one to hybridize a vehicle model and another one to approximate tire satura-
tion constraints.

4.1 Introduction
Nonlinearity of the MPC optimization problem in automated driving is a significant obstacle
towards real-time vehicle control [12]. Approximating the nonlinearities is often done
in many applications [80] to come up with improved computational efficiency in solving
the nonlinear control optimization problem. In this line, hybridization techniques [81]
This chapter has been published in the proceedings of IEEE Conference on Advance Motion Control [79].
 36 4 H4MPC: A Hybridization Toolbox for MPC

approximate a nonlinear function using hybrid systems formalism, with both continuous
and discrete-time dynamics involved in the approximation [18]. For more information on
hybrid systems, the reader is referred to [23, 82].
Hybridization has been extensively employed in automated driving research, e.g.,
in vehicle control [21] by approximating the nonlinear model using a mixed-logical dy-
namics [26], or by approximating nonlinear tire forces using PieceWise Affine (PWA)
dynamics [36, 83, 84]. Efficiency of MPC after hybridization has been recently studied in
[85, 86].
This chapter presents H4MPC [87], a hybridization toolbox in Matlab that provides
a user-friendly interface to formulate and solve optimization problems to approximate
the nonlinearities in NMPC using hybrid systems formalism, in particular PWA modeling
framework. The toolbox exploits the formulation from Max-Min-Plus-Scaling (MMPS)
systems [24] to allow for an intuitive adjustment of the complexity level in the approximated
4 form. Further, H4MPC facilitates approximation of the nonlinear constraints via covering
the resulting non-convex feasible region by a union of convex subregions, namely ellipsoids
or polytopes, where the latter are obtained using MMPS formalism as well.
This chapter describes the H4MPC modules in detail and demonstrates its capabilities
using two case studies: approximating a single-track vehicle model [88], and hybridizing
the non-convex feasible region due to tire saturation limits, known as the Kamm circle
constraint [89] for a Pacejka tire model [90]. The chapter is structured as follows: Section 4.2
covers the preliminary definitions, Section 4.3 presents the architecture of H4MPC and
Section 4.4 illustrates the case studies and analysis of the results. Finally, Section 4.5
summarizes the results of this chapter.

4.2 Preliminaries
4.2.1 Nonlinear Problem Description
Consider a given discrete-time nonlinear system 𝑠(𝑘 + 1) = 𝐹 (𝑠(𝑘), 𝑢(𝑘)) where 𝑠 ∈ ℝ𝑛 and
𝑢 ∈ ℝ𝑚 respectively represent the state and input vectors, and the domain of 𝐹 is denoted
by  ⊆ ℝ𝑛+𝑚 . With the state and input vectors defined over the whole prediction horizon
𝑁p as
𝑇
𝑠̃(𝑘 + 1) = [𝑠̂𝑇 (𝑘 + 1|𝑘) 𝑠̂𝑇 (𝑘 + 2|𝑘) … 𝑠̂𝑇 (𝑘 + 𝑁p |𝑘)] ,
𝑇
𝑢(𝑘)
̃ = [𝑢𝑇 (𝑘) 𝑢𝑇 (𝑘 + 1) … 𝑢𝑇 (𝑘 + 𝑁p − 1)] ,

and 𝑠̂𝑇 (𝑘 + 𝑖|𝑘) with 𝑖 ∈ {1, … , 𝑁p } representing the prediction of the states in step 𝑘 + 𝑖 given
the measured states at step 𝑘, the nonlinear MPC problem at step 𝑘 is formulated in the
general form:

min ‖Θ𝑠 𝑠̃(𝑘)‖𝜌 + ‖Θ𝑢 𝑢(𝑘)‖

̃ 𝜌, (4.1a)
𝑠̃(𝑘),𝑢(𝑘)
̃
s.t. 𝑠̂(𝑘 + 𝑖|𝑘) = 𝐹 (̂𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)), ∀𝑖 ∈ {1, … , 𝑁p }, (4.1b)
𝐺(̂𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)) ⩽ 1, ∀𝑖 ∈ {1, … , 𝑁p }, (4.1c)

where (4.1b) represents the equality constraints due to the prediction model, and (4.1c)
expresses the non-convex feasible region via the normalized nonlinear constraint function 𝐺
4.2 Preliminaries 37

resulting from physics-based constraints such as tire saturation or vehicle stability. Without
loss of generality, we assume 𝐺 to be a scalar function. The objective function (4.1a) is
a sum of the 𝜌−norm of the state and input vectors with 𝜌 ∈ {1, 2, ∞}, induced by weight
matrices Θ𝑠 and Θ𝑢 .

4.2.2 Approximation of the Nonlinear Problem

Hybridization of the NMPC problem is done in two steps: (1) approximating the prediction
model, i.e. 𝐹 , and (2) approximating the nonlinear constraints, i.e. 𝐺. For a more compact
𝑇
representation, we use the augmented state vector 𝑥 = [𝑠 𝑇 𝑢𝑇 ] to define 𝐹 (𝑥) B 𝐹 (𝑠, 𝑢)
and 𝐺(𝑥) B 𝐺(𝑠, 𝑢).

Model Approximation
𝑇
We approximate each component 𝐹𝑤 of 𝐹 = [𝐹1 … 𝐹𝑛 ] by an MMPS function 𝑓𝑤 in the 4
Kripfganz form [91] as
{ } { }
𝑓𝑤 (𝑥) = max 𝜙+𝑤,1 (𝑥), 𝜙+𝑤,2 (𝑥), … , 𝜙+𝑤,𝑃𝑤+ (𝑥) − max 𝜙−𝑤,1 (𝑥), 𝜙−𝑤,2 (𝑥), … , 𝜙−𝑤,𝑃𝑤− (𝑥) , (4.2)
∀𝑤 ∈ {1, … , 𝑛},
𝜂 𝜂
where the vectors 𝜙𝑠 ∶ ℝ𝑚+𝑛 → ℝ𝑃 with 𝜂 ∈ {+, −} are affine functions of 𝑥, also referred
to as dynamic modes. Figure 4.1 shows an illustrative example for a 1-dimensional case
with (𝑃 + , 𝑃 − ) = (3, 4).

𝜙+ 𝑓 𝜙+
3
1
Nonlinear 𝜙+
2

MMPS
𝑥

𝜙−
3
𝜙−
2
𝜙−
4
𝜙−
1

Figure 4.1: MMPS approximation of a nonlinear function using the difference of two max functions

Constraint Approximation
With the nonlinear, non-convex constraints given as 𝐺(𝑥) ⩽ 1, we approximate the feasible
region  B {𝑥 ∈  | 𝐺(𝑥) ⩽ 1} by a union of convex subregions . The shape of the subre-
gions in  can either be polytopic, which we obtain by an MMPS approximation of the
boundary, or ellipsoidal.
Similar to the prediction model, MMPS approximation of the constraints is expressed
by
{ } { }
𝑔MMPS (𝑥) = max 𝛾1+ (𝑥), 𝛾2+ (𝑥), … , 𝛾𝑅++ (𝑥) − max 𝛾1− (𝑥), 𝛾2− (𝑥), … , 𝛾𝑅−− (𝑥) , (4.3)
𝜂
with the vectors 𝛾 𝜂 ∶ ℝ𝑚+𝑛 → ℝ𝑅 and 𝜂 ∈ {+, −} being affine functions of 𝑥.
 38 4 H4MPC: A Hybridization Toolbox for MPC

In the ellipsoidal approach, 𝐺 is approximated by 𝑔ELLP as

{ }
𝑔ELLP (𝑥) = min (𝑥 − 𝑥𝑒,0 )𝑇 𝑄𝑒 (𝑥 − 𝑥𝑒,0 ) − 1 , (4.4)
𝑒∈{1,…,𝑛e }

with 𝑄𝑒 being positive definite matrices and 𝑥𝑒,0 representing the center coordinates of
the (possibly rotated) 𝑛e ellipsoids. Figure 4.2 represents a schematic view of MMPS and
ellipsoidal constraint approximations.
x2

MMPS

Ellipsoidal
4

x1

Nonlinear

Domain

Figure 4.2: MMPS and ellipsoidal constraint approximation.

4.2.3 Approximation Problem Formulation

All the nonlinear functions 𝐻 ∈ {𝐹1 , … , 𝐹𝑛 , 𝐺} are approximated by their respective hybrid
formulations ℎ ∈ {𝑓1 , … , 𝑓𝑛 , 𝑔} for 𝑔 ∈ {𝑔MMPS , 𝑔ELLP } via solving the nonlinear optimization
problem

‖𝐻 (𝑥) − ℎ(𝑥)‖2
min ∫ 𝑑𝑥, (4.5)
 ‖𝐻 (𝑥)‖2 + 𝜖0


where  represents the decision variables used to define ℎ and the positive value 𝜖0 > 0
added to the denominator avoids division by very small values for ‖𝐻 (𝑥)‖2 ≈ 0. For the
nonlinear constraint, (4.5) approximates the boundary of the feasible region, and therefore
we call this approach “boundary-based”.
Another method to formulate the constraint approximation problem is the “region-base”
approach where we formulate the optimization problem as

{ ⧵ } { ⧵ }
min 𝛾c + (1 − 𝛾c ) , (4.6)
 {} { ⧵ }

where the operator  gives the size or “volume” of the region, and 𝛾c ∈ [0, 1] is a tuning pa-
rameter to adjust the relative penalization weight for the misclassification errors regarding
inclusion error  ⧵ , i.e., failing to cover the feasible region, and the violation error  ⧵ 
which corresponds to violating the constraints.
4.3 Toolbox Architecture 39

4.3 Toolbox Architecture

The graphical user interface of the H4MPC toolbox is shown in Fig. 4.3 and consists of
three steps, which correspond to the three modules in H4MPC as shown in Fig. 4.4: grid
generation, model approximation, and constraint approximation. In each module, during
grid generation, model approximation, and constraint approximation, the toolbox saves
the results as separate .mat files in case the user is only interested in the output from
one of the modules. The arrows in Fig. 4.4 illustrate the possible flow of using each module
within H4MPC.

Figure 4.3: Graphical user interface

4.3.1 Grid Generation Module

In the first step, the user is asked to provide information on:
• number of states and inputs,
• function handles for the nonlinear model and constraint functions,
 40 4 H4MPC: A Hybridization Toolbox for MPC

provide nonlinear
Grid
system/constraints Grids.mat
& their properties
Generation

Model Constraint
Approximation Approximation

System.mat Constraints.mat

Figure 4.4: Toolbox architecture

• sampling time for forward Euler discretization if the model function is continuous-
4 time (can be set to 0 if the provided function is discrete-time), and

• input and state bounds.

For the highly-nonlinear model or constraint functions in automated driving, analytical

closed-form solutions for (4.5)–(4.6) often do not exist. Therefore, a grid is generated on
 to solve these optimization problems by minimizing the objective function across the
grid points. For model approximation, the user can select among four methods described
in [85]:

• Domain-based: [points are directly sampled from ]

– Uniform (U): the points are generated by picking 𝑛samp uniformly-spaced

points along each axis in .
– Random (R): a total of 𝑛rand points are randomly selected from .

• Trajectory-based: [𝑛sim open-loop simulations with 𝑛step steps of 𝐹 are run using
random inputs from ]

– Steady-state (S): the initial state of each simulation is selected as the steady-
state solution w.r.t. the initial input, i.e., it is assumed that each simulation
starts from a steady state.
– Randomly-initiated (T): the initial state of each simulation is randomly
selected from .

For constraint approximation, the grid points are sampled from the whole domain . Since
the region close to the boundary of the feasible region where 𝐺(𝑥) = 1 is of more interest,
the constraint approximation grid is generated by combining a uniform grid (U) with a
random grid (R) on the boundary region with width 𝜖b , where |𝐺(𝑥) − 1| ⩽ 𝜖b . The user can
select the number of uniform and boundary grid points, as well as 𝜖b in the user interface.
After clicking on the “Generate Training Grids” button, the parameters are saved in the
params struct and the model and constraint approximation grids, respectively SM and
SC, are generated.
4.4 Case Study 41

4.3.2 Model Approximation Module

In this module, the user is asked to provide (𝑃 + , 𝑃 − ) values for model approximation, for
each state/input separately, i.e. for components of 𝐹 . We solve (4.5) for the user-defined
values of (𝑃 + , 𝑃 − ) by Matlab’s nonlinear least squares optimizer, lsqnonlin, using the
trust-region-reflective algorithm, in view of the number of optimization variables. We solve
the problem with 𝑛multistart random initial guesses, where 𝑛multistart is provided by the user,
and we select the one with the lowest objective value as the optimal solution. By clicking
on the “Approximate Prediction Model” button, the model approximation optimization
problem is solved and the hybridized model is saved in the system struct.

4.3.3 Constraint Approximation Module

The third module approximates the nonlinear constraints using the desired approach and 4
subregion types that can be selected by the user. Similar to the model approximation
module, the boundary-based approach leads to a smooth problem, which H4MPC solves
for the user-defined values of (𝑅+ , 𝑅− ) or 𝑛e . The boundary-based approach offers greater
flexibility in fine-tuning the trade-off between encompassing the non-linear region and
potentially infringing upon it, ultimately resulting in improved coverage of the non-convex
region. However, if the application demands strict adherence to non-linear constraints, it is
advisable to opt for the region-based approach in (4.6), which results in a non-differentiable
objective function. The user can then select the parameter 𝛾c to adjust the weight factor
in (4.6). H4MPC solves the approximation problem using the particle swarm optimizer in
Matlab, which does not require the problem to be differentiable where -based on extensive
numerical experiments- we select the swarm size to be 10 times larger than the number of
decision variables. The user can then click on the “Approximate Constraint” button, to get
the hybridized constraints as the const struct.

4.4 Case Study

In this section, two examples are investigated to showcase model and constraint approx-
imations handled by H4MPC. The first example involves the single-track vehicle model
from [88] as an illustrative model approximation example, and the second considers the
Kamm circle constraint [89] as a function of the lateral and longitudinal slips to represent
a non-convex feasible region. Multiple selections in H4MPC are tested to highlight the
effect of various tuning parameters for hybridization.

4.4.1 Nonlinear Vehicle Model

The lateral dynamics of the single-track vehicle model in Fig. 4.5 is characterized by

1
𝑟̇ = (𝑙f 𝐹𝑦f − 𝑙r 𝐹𝑦r ) , (4.7a)
𝐼𝑧𝑧
𝐹𝑦f + 𝐹𝑦r
𝛽̇ = arctan −𝑟 , (4.7b)
( 𝑚𝑣𝑥 )
 42 4 H4MPC: A Hybridization Toolbox for MPC

Table 4.1: System parameters and variables∗

System parameters
Parameter Definition Value Unit
𝑚 Vehicle mass 1725 kg
𝐼𝑧𝑧 Inertia moment about z-axis 1300 kg/m2
𝑙f CoG∗∗ to front axis distance 1.35 m
𝑙r CoG to rear axis distance 1.15 m
𝜇 Friction coefficient 1 –
𝑣𝑥 Longitudinal velocity 20 m/s
𝐹𝑧f Normal load on the front axis 5000 N
𝐹𝑧r Normal load on the rear axis 5000 N
𝐵𝜅 11.4 –
𝐶𝜅 1.4 –
Pacejka tire coefficients
𝐵𝛼 10.0 –
4 𝐶𝛼 1.6 –
System variables
Variable Definition Bound Unit
𝛽 Sideslip angle [-0.3,0.3] rad
𝑟 Yaw rate [-0.5, 0.5] rad/s
𝛿 Steering angle (road) [-0.3, 0.3] rad
𝐹𝑥f Longitudinal force on the front axis [-5000, 0] N
𝐹𝑥r Longitudinal force on the rear axis [-5000, 5000] N
𝐹𝑦f Lateral force on the front axis [-5000, 5000] N
𝐹𝑦r Lateral force on the rear axis [-5000, 5000] N
𝛼f Front slip angle [-0.4,0.4] rad
𝛼r Rear slip angle [-0.4,0.4] rad
𝜅f Front slip ratio [-1,1] –
𝜅r Rear slip ratio [-1,1] –
𝑇
𝑠 State vector B [𝑟 𝛽 ] – –
𝑢 Input vector B 𝛿 – –
∗ These kinematic parameters are from [88].
∗∗ Center of Gravity

with the tire forces described by the Pacejka tire model [90] as
𝐹𝑥𝑎 = 𝐹𝑧𝑎 𝜇 sin (𝐶𝜅 arctan (𝐵𝜅 𝜅𝑎 )) , (4.8a)
𝐹𝑦𝑎 = 𝐹𝑧𝑎 𝜇 sin (𝐶𝛼 arctan (𝐵𝛼 𝛼𝑎 )) , (4.8b)
the 𝜅𝑎 representing the slip ratio on axle 𝑎 ∈ {f, r}, and the slip angles being
𝑙f 𝑟
𝛼f = arctan 𝛽 + − 𝛿, (4.9a)
( 𝑣𝑥 )
𝑙r 𝑟
𝛼r = arctan 𝛽 − . (4.9b)
( 𝑣𝑥 )
The system parameters are shown in Table 4.1. The tire forces should satisfy the tire
saturation limits, i.e. Kamm circle constraint [89]
2
𝐹𝑥f 2
+ 𝐹𝑦f ⩽ (𝜇𝐹𝑧f )2 , (4.10a)
2
𝐹𝑥r 2
+ 𝐹𝑦r 2
⩽ (𝜇𝐹𝑧r ) . (4.10b)
4.4 Case Study 43

Figure 4.5: Single-track vehicle model.

4.4.2 Approximation of the Vehicle Dynamics 4

We use the vehicle model (4.7) as a case study to investigate model approximation with two
states, represented by the variables 𝑟 and 𝛽, and the input 𝛿. To showcase the impact of the
parameters (𝑃 + , 𝑃 − ) on the level of complexity of the hybrid formulation, we examine two
sets of parameter values: (𝑃 + , 𝑃 − ) ∈ {(2, 2), (4, 3)} for 𝑟̇ and (𝑃 + , 𝑃 − ) ∈ {(2, 2), (5, 5)} for 𝛽.̇
As visualizing a three-dimensional input/state space can be not easy to read, we plot
two specific cases in Fig. 4.6: 𝑟(𝛽,
̇ 𝑟) for 𝛿 = 0 rad and 𝛽(𝛽, ̇ 𝛿) for 𝑟 = −0.3 rad. For more
clarity, we provide two cuts along different paths in each plot with their own distinguished
colors. We compare the original nonlinear function with its MMPS approximations by
using dashed lines for a simpler MMPS approximation and solid lines for a more complex
one. Notably, the more complex MMPS approximation, which includes additional terms,
provides a more accurate representation of the nonlinear function. For a more clear dynamic
comparison, the phase portrait of the more accurate MMPS approximation is compared
with the nonlinear one from [88] in Fig. 4.7 for three values of 𝛿, which shows that the
MMPS approximation can generate sufficiently-accurate trajectories on the phase portrait
as well.

4.4.3 Approximation of the Kamm Circle Constraint

To demonstrate the use of hybrid approximation for non-convex feasible regions, we provide
the following example with a more intuitive interpretation: approximate the feasible region
associated with the Kamm circle constraint defined in (4.10) within the 𝜅-𝛼 plane as its two
input/states. Therefore, we employ region-based approximation techniques while ensuring
that the constraint violation error is maintained at zero. Figure 4.8 showcases the hybrid
approximations using the ellipsoidal and MMPS methodologies for different values of 𝑛e
and (𝑅+ , 𝑅− ). Figure 4.8 illustrates the obtained hybrid approximations.
Additionally, to emphasize the significance of defining boundaries properly, we examine
simpler cases with 𝑛e = 2 for the ellipsoidal approach and (𝑅+ , 𝑅− ) = (2, 2) for the MMPS
method under two scenarios: setting the boundary on the domain in terms of 𝛼 to either 0.1
or 0.4 radians. A comparison between the orange and purple approximations for both the
ellipsoidal and MMPS approaches reveals that, as the optimizer minimizes the error across
the domain, it converges to a more accurate approximation in the vicinity of the origin
when |𝛼| is less than or equal to 0.1. However, when the domain is extended to |𝛼| ⩽ 0.4, the
optimizer converges to an ellipsoidal or polytopic approximation of the feasible region close
 44 4 H4MPC: A Hybridization Toolbox for MPC

Nonlinear (2,2) (4,3)

𝑟̇ (rad/s2 )
0

−5
0
𝑟̇ (rad/s2 )

−5 path
Nonlinear (2,2) (4,3)
0

𝑟̇ (rad/s2 )
0.5
0 0.5
0 −5
−0.5−0.5
𝑟 (rad/s) 𝛽 (rad)
path
Nonlinear (2,2) (5,5)
4 0.4 0.3

𝛽̇ (rad/s)
𝛽̇ (rad/s)

0.2 0
0
−0.3
−0.2
path
−0.4
−0.5 Nonlinear (2,2) (5,5)

0 −0.5
𝛽̇ (rad/s)

0 0
0.5 0.5
𝛿 (rad) 𝛽 (rad)
−0.3

path

Figure 4.6: Comparison of two MMPS approximations of the vehicle model with different (𝑃 + , 𝑃 − ) values. For a
more clear representation, the functions are plotted along four paths as 2-dimensional cuts of the 3-dimensional
function representation.

𝛿 = −0.1 rad 𝛿 = 0 rad 𝛿 = 0.1 rad

0.5 0.5 0.5
Nonlinear
𝛽 (rad)

𝛽 (rad)

𝛽 (rad)

MMPS
0 0 0
Equilibrium

−0.5 −0.5 −0.5

−0.5 0 0.5 −0.5 0 0.5 −0.5 0 0.5
𝑟 (rad/s) 𝑟 (rad/s) 𝑟 (rad/s)

Figure 4.7: Phase portraits of the nonlinear and the MMPS approximation with (𝑃 + , 𝑃 − ) ∈ {(4, 3), (5, 5)}.

to the origin, which cannot adequately capture the complexity of the nonlinear constraint
further from the origin. By increasing the complexity of the approximation, such as using
five ellipsoids or (𝑅+ , 𝑅− ) = (5, 5), the optimizer finds hybrid approximations that provide
better coverage of the feasible region. This improved approximation is represented in pink
in both figures.
4.5 Conclusions 45

0.3 0.3
Nonlinear Nonlinear
|𝛼| ⩽ 0.4
𝛼 (rad)

𝛼 (rad)
0 0
𝑛e = 2 (𝑅+ , 𝑅− )
|𝛼| ⩽ 0.4 = (3, 2)
−0.3 −0.3
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
𝜅 𝜅
0.3 0.3
Nonlinear Nonlinear
𝛼 (rad)

𝛼 (rad)
0 0
𝑛e = 2 (𝑅+ , 𝑅− ) |𝛼| ⩽ 0.1
−0.3
−1
|𝛼| ⩽ 0.1
−0.5 0 0.5 1
−0.3
−1
= (3, 2)
−0.5 0 0.5 1
4
𝜅 𝜅
0.3 𝑛e = 5 0.3 (𝑅+ , 𝑅− )
Nonlinear Nonlinear
|𝛼| ⩽ 0.4 = (5, 5)
𝛼 (rad)

𝛼 (rad)

0 0

|𝛼| ⩽ 0.4
−0.3 −0.3
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
𝜅 𝜅
(a) Ellipsoidal approximation (b) MMPS approximation

Figure 4.8: Illustration of MMPS and ellipsoidal approximation of the Kamm circles (4.10) as a function of 𝛼 and 𝜅.

4.5 Conclusions
H4MPC is a open-source toolbox for hybridization of nonlinear model and constraints in
NMPC to allow a hybrid formulation of the nonlinear optimization problem. The toolbox
includes three modules and provides a user-friendly interface to allow the user to customize
the approximation. In this chapter, a single-track vehicle model and the tire saturation limits
were investigated as two examples to showcase multiple approximation approaches handled
in H4MPC and to highlight the influence of their corresponding parameters. We expect
the toolbox to be useful in a variety of applications such as automated driving or control of
robotic systems. The H4MPC toolbox is freely available from https://bit.ly/H4MPCV1. The
next versions of the toolbox will include controller design using the hybridized form of the
nonlinear model and physics-based constraints to investigate the effect of approximation
complexity on computation time of the NMPC optimization problem.
 47

5
Sensitivity Analysis for PWA
Approximations of NLPs
5
To be a theorist you have to admit the possibility of being wrong – the provisionality of
knowledge – and you know you cannot spin your way out of a theoretical problem.

— Yanis Varoufakis, And the Weak Suffer What They Must?

NonLinear Program (NLP)s are prevalent in optimization-based control of nonlinear systems.

Solving general NLPs is computationally expensive, necessitating the development of fast
hardware or tractable suboptimal approximations. This chapter investigates the sensitivity
of the solutions of NLPs with polytopic constraints when the nonlinear continuous objective
function is approximated by a PieceWise Affine (PWA) counterpart. By leveraging perturbation
analysis using a convex modulus, we derive guaranteed bounds on the distance between the
optimal solution of the original polytopically-constrained NLP and that of its approximated
formulation. Our approach aids in determining criteria for achieving desired solution bounds.
Two case studies on the Eggholder function and control of an inverted pendulum demonstrate
the theoretical results.

5.1 Introduction
NLPs are commonly encountered in optimization-based control of nonlinear systems,
e.g. Nonlinear Model Predictive Control (NMPC) [93]. Solving non-convex NLPs is in-
tractable, posing a great challenge in applying optimization-based controllers in real-time
operations, especially for systems having fast dynamics. Various solutions have been
proposed to address this issue, such as adaptive problem formulations [94], learning-based
methods [95], and sensitivity analysis of NLPs [96].

This chapter has been published in IEEE Control Systems Letters [92].
 48 5 Sensitivity Analysis for PWA Approximations of NLPs

PWA approximations are widely used due to their tractability [65, 70]. To obtain a
continuous PWA approximation, min and max operators can be used to maintain conti-
nuity and to resolve numerical issues in the resulting optimization problem [28, 97, 98].
The approximated problem can be used to obtain a suboptimal solution [99], whose opti-
mality highly depends on the accuracy of the approximation. For example, a warm start
of a non-convex NLP can be obtained by solving the approximated optimization prob-
lem [100, 101]. Optimality guarantees of such approaches can be derived using sensitivity
analysis, establishing an upper bound on the distance between the original solutions and
the approximated ones. As a result, by finding a subset of the decision space around the
approximated solution, one can sample a structured or random warm start to solve the
original non-convex NLP more efficiently.

Quantitative bounds on the distance between the original and the approximated so-
lutions have been studied in the sensitivity analysis of quadratic [102] and convex [103]
optimization problems. Regarding NLPs, there exist several results on their sensitivity to
the parameters in the optimization formulation [104, 105] and the initial solution [106]. In
5 addition, optimality and dissipativity conditions for the perturbed convex NLP problem
have also been established [107, 108]. For a more extensive study, the reader can refer
to [109]. However, obtaining quantitative bounds on the distance between the solutions of
a non-convex NLP and its PWA approximation is still a gap that needs to be filled, and our
work addresses this problem.

In this chapter, we present a method to bound the solutions of a polytopically-constrained

non-convex NLP problem using a continuous PWA approximation of the nonlinear objec-
tive function. We employ the Max-Min-Plus-Scaling (MMPS) formulation of continuous
PWA functions in [28] to construct a piecewise convex approximation formalism. Leverag-
ing some results in [103], we derive guaranteed bounds on the distance between the original
and the approximated optimal solution. Moreover, our approach can not only establish
such bounds but also aid in determining necessary criteria during the approximation stage
to attain a desired solution bound. To demonstrate the theoretical findings, we present a
case study on the Eggholder function [110], a renowned benchmark optimization problem
with multiple local minima.

The rest of the chapter is organized as follows. In Section 5.2, we present preliminaries
regarding the sensitivity analysis of NLPs and the PWA approximation of nonlinear func-
tions. Section 5.3 formulates the problem and Section 5.4 elaborates our proposed approach
to theoretically compute the confidence radius for the local minima of the corresponding
approximated function. In Section 5.5, we then demonstrate the derived confidence radius
through a case study on the Eggholder function and we apply our analysis to an NMPC
optimization example. Section 5.6 concludes this chapter.

Notation: For a positive integer 𝑃, we use 𝑃 to denote the set {1, 2, … , 𝑃}. For a connected
set  ⊆ ℝ𝑛 , the diameter of  is defined as diam() B max𝑥1 ,𝑥2 ∈ ‖𝑥1 − 𝑥2 ‖, where ‖ ⋅ ‖ is the
Euclidean or the 2-norm.
5.2 Preliminaries 49

5.2 Preliminaries
5.2.1 Representation of Continuous PWA Functions
We start with formally defining a continuous PWA function which will be frequently used
throughout this chapter.
Definition 5.1 (Continuous PWA function [111]). A scalar-valued function 𝑓 ∶  ⊆ ℝ𝑛 → ℝ
is said to be a continuous PWA function if and only if the following conditions hold:
1. the domain space  is divided into a finite number of closed polyhedral regions
1 , … , 𝑅 with non-overlapping interiors,
2. for each 𝑟 ∈ 𝑅 , 𝑓 can be expressed as

𝑓 (𝑥) = 𝛼𝑟𝑇 𝑥 + 𝛽𝑟 if x ∈ r ,

with 𝛼𝑟 ∈ ℝ𝑛 and 𝛽𝑟 ∈ ℝ, and

3. 𝑓 is continuous on the boundary between any pair of regions.
5
PWA functions can be expressed in different forms, among which the MMPS form is
powerful for decomposing PWA functions.
Theorem 5.1 (MMPS representation [28]). For a scalar-valued continuous PWA function 𝑓
as in Definition 5.1, there exist non-empty index sets 𝑃 and 𝑄𝑝 such that

𝑓 (𝑥) = min max (𝑎𝑇𝑝,𝑞 𝑥 + 𝑏𝑝,𝑞 ) , (5.1)

𝑝∈𝑃 𝑞∈𝑄𝑝

for real numbers 𝑏𝑝,𝑞 and vectors 𝑎𝑝,𝑞 ∈ ℝ𝑛 .

For convenience, we define the 𝑝-th local convex segment of 𝑓 as

𝑓𝑝 (𝑥) B max (𝑎𝑇𝑝,𝑞 𝑥 + 𝑏𝑝,𝑞 ) , (5.2)

𝑞∈ 𝑄𝑝

where 𝑓𝑝 is convex since it is defined as the maximum of a finite number of affine functions
and its domain is also convex. In addition, we define the region 𝑝,𝑞 in which a certain
affine function is activated and the region 𝑝,. in which a convex PWA function is activated,
that is,

𝑝,𝑞 B {𝑥 ∈  | 𝑓 (𝑥) = 𝑎𝑇𝑝,𝑞 𝑥 + 𝑏𝑝,𝑞 }, (5.3a)

𝑝,. B {𝑥 ∈  | 𝑓 (𝑥) = 𝑓𝑝 (𝑥)}. (5.3b)

Further, we have
𝑄𝑝
𝑝,. = ⋃ 𝑝,𝑞 .
𝑞=1

Lastly, we define 𝜎𝑝 ∶ 𝑝,. → 𝑄𝑝 as the region index function for 𝑓𝑝 as

𝜎𝑝 (𝑥) = 𝑞 ⟺ 𝑥 ∈ 𝑝,𝑞 . (5.4)

 50 5 Sensitivity Analysis for PWA Approximations of NLPs

5.2.2 Sensitivity of the Convex Optimization Problem

The convexity modulus [103], being used to quantify convexity, is useful in the sensitivity
analysis of convex functions. In the following, we define the convexity modulus specifically
for 𝑓𝑝 , the 𝑝-th convex segment of 𝑓 , and its domain 𝑝,. .
Definition 5.2 (Convexity modulus [103]). For a given convex function 𝑓𝑝 , the convexity
modulus ℎ1 ∶ [0, +∞) → [0, +∞) over the domain 𝑝,. is defined as
⎧
⎪
⎪ inf 𝐽 (𝑣, 𝑤) if 𝛾 < diam(𝑝,. )
⎪ 𝑣,𝑤∈𝑝,.
ℎ1 (𝛾) B ⎨ ‖𝑣−𝑤‖=𝛾 , (5.5)
⎪
⎪
⎩ +∞
⎪ if 𝛾 ⩾ diam(𝑝,. )

where 𝑣 and 𝑤 are two points in 𝑝,. satisfying ‖𝑣 − 𝑤‖ = 𝛾, and 𝐽 (𝑣, 𝑤) is given as

𝑓𝑝 (𝑣) + 𝑓𝑝 (𝑤) 𝑣+𝑤

𝐽 (𝑣, 𝑤) = − 𝑓𝑝 ( . (5.6)
2 2 )

5
Theorem 5.2 (Theorem 4.5 in [103]). Suppose that 𝑓𝑝 ∶ 𝑝,. → ℝ is a scalar-valued convex
function and 𝛿𝑝 ∶ 𝑝,. → ℝ is an arbitrary function satisfying

sup |𝛿𝑝 (𝑥)| = Δ𝑝 < ∞. (5.7)

𝑥∈𝑝,.

Let 𝑥𝑝∗ be any global infimizer of 𝑓𝑝 and 𝑥̂𝑝∗ be any global infimizer of 𝑓̂𝑝 = 𝑓𝑝 + 𝛿𝑝 . Then

‖𝑥̂𝑝∗ − 𝑥𝑝∗ ‖ ⩽ ℎ−1

1 (2Δ𝑝 ) , (5.8)

where ℎ1 is the convexity modulus in Definition 5.2.

For a more compact notation, we call the right-hand side of (5.8), the confidence radius,
defined as follows:
Definition 5.3 (Confidence radius). For a given function 𝑓̂𝑝 ∶ 𝑝,. ⊆ ℝ𝑛 → ℝ approximating
the function 𝑓𝑝 ∶ 𝑝,. ⊆ ℝ𝑛 → ℝ with the maximal approximation error Δ𝑝 from (5.7), the
confidence radius is the upper-bound on the distance between 𝑥̂𝑝∗ , the global minimizer of 𝑓̂𝑝 ,
and 𝑥𝑝∗ , the global minimizer of 𝑓𝑝 , and is obtained by

𝜒 = ℎ−1
1 (2Δ𝑝 ),

where ℎ1 is the convexity modulus in Definition 5.2.

Proposition 5.1 (Proposition 2.2, 2.5 in [103]). Given a convex function 𝑓𝑝 on a compact
domain 𝑝,. with convexity modulus ℎ1 defined as (5.5), the following hold:

1. ℎ1 is left-continuous on (0, diam(𝑝,. )), and

2. for 𝛾1 , 𝛾2 ∈ [0, diam(𝑝,. ), if 𝛾1 < 𝛾2 , then ℎ1 (𝛾1 )/𝛾1 ⩽ ℎ1 (𝛾1 )/𝛾2 .
5.3 Continuous PWA Approximation of NLPs 51

5.3 Continuous PWA Approximation of NLPs

Consider an NLP with polytopic constraints

min 𝐹 (𝑥), (5.9)

𝑥∈

where 𝐹 ∶  ⊂ ℝ𝑛 → ℝ is the nonlinear objective function and  ⊆  is the polytopic

feasible region. From now on, we assume that the domain  is compact. We approximate
𝐹 by a continuous PWA function 𝑓 of the MMPS form (5.1) via solving the approximation
problem

| |
| |
min min max (𝑎𝑇𝑝,𝑞 𝑥 + 𝑏𝑝,𝑞 )| 𝑑𝑥,
∫ ||𝐹 (𝑥) − 𝑝∈ (5.10)
,  𝑃 𝑞∈𝑄𝑝 |
 | |

to minimize the absolute approximation error, where the ordered sets  and  respectively
collect 𝑎𝑝,𝑞 and 𝑏𝑝,𝑞 .
5
Example 5.1. Figure 5.1 shows a 1-dimensional example of approximating a nonlinear
objective function 𝐹 by a continuous PWA function 𝑓 using the MMPS form (5.1) as

⎛ ⎞
⎜ ⎟
𝐹 (𝑥) ≈ 𝑓 (𝑥) = min ⎜max (𝑓1,1 , 𝑓1,2 ), max (𝑓2,1 , 𝑓2,2 ), 𝑓3,. ⎟ ,
⎜⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟ ⎟
⎜ 𝑓1,. 𝑓2,.
⎟
⎝ ⎠

with 𝑃 = 3, 𝑄1 = 𝑄2 = 2, and 𝑄3 = 1. The convex segments of 𝑓 are shown by 𝑓𝑝,. which give
the maximum value among 𝑄𝑝 affine functions 𝑓𝑝,𝑞 , 𝑞 ∈ 𝑄𝑝 . The subregions 𝑝,𝑞 are shown
in the same color as their corresponding active affine functions, 𝑓𝑝,𝑞 . In this 1-dimensional
example, diam(𝑝,. ) is the distance between the upper and lower bounds of 𝑝,. on the 𝑥-axis.

𝑦
𝑓3,.
𝑦 = 𝐹 (𝑥) 𝑓2,2
𝑦 = 𝑓 (𝑥)
𝑓1,1 𝑓1,2 𝑓2,1
𝑥

1,1 1,2 2,1 2,2 3,.

1,. 2,.

Figure 5.1: A conceptual example of approximating a nonlinear function 𝐹 with a continuous PWA approximation
𝑓 using the MMPS form in (5.1).
 52 5 Sensitivity Analysis for PWA Approximations of NLPs

5.4 Theoretical Analysis

The main result of this section is the sensitivity bound in Theorem 5.3. Before, we present
some building blocks about continuity and boundedness of the convexity modulus (cf.
Lemma 5.1, 5.2 and 5.3).
Lemma 5.1. For a convex PWA function 𝑓𝑝 expressed by (5.2), 𝜕ℎ1 /𝜕𝛾 is a piecewise constant
function.
Proof. For compactness, let us denote
𝑣+𝑤
𝑖 = 𝜎𝑝 (𝑣), 𝑗 = 𝜎𝑝 (𝑤), 𝑘 = 𝜎𝑝 (
,
2 )
with 𝑖, 𝑗, 𝑘 ∈ 𝑄𝑝 . For 𝑓𝑝 expressed by (5.2), the function 𝐽 in (5.6) can be written as
𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑘 𝑎𝑇𝑝,𝑗 − 𝑎𝑇𝑝,𝑘
𝑏𝑝,𝑖 + 𝑏𝑝,𝑗 − 2𝑏𝑝,𝑘
𝐽 (𝑣, 𝑤) = 𝑣+ 𝑤+ . (5.11)
2 2 2
The necessary Lagrange conditions for optimality at (𝑣∗ , 𝑤∗ ) in (5.5) state that there must
exist 𝜇 ∈ ℝ that satisfies the following simultaneously:
5 |
∇𝐽 (𝑣∗ , 𝑤∗ ) + 𝜇∇ (‖𝑣 − 𝑤‖ − 𝛾) | 𝑣=𝑣∗ = 0, (5.12a)
|𝑤=𝑤∗
‖𝑣∗ − 𝑤∗ ‖ = 𝛾. (5.12b)
By calculating the gradient of (5.11), we have that (5.12) becomes
𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑘 𝑣∗ − 𝑤∗
+𝜇 = 0, (5.13a)
2 ‖𝑣∗ − 𝑤∗ ‖
𝑎𝑝,𝑗 − 𝑎𝑇𝑝,𝑘
𝑇
𝑤∗ − 𝑣∗
+𝜇 = 0, (5.13b)
2 ‖𝑣∗ − 𝑤∗ ‖
‖𝑣∗ − 𝑤∗ ‖ = 𝛾, (5.13c)
which implies the existence of 𝜇 ∈ ℝ satisfying
𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑗 𝑣∗ − 𝑤∗
+ 2𝜇 = 0. (5.14)
2 𝛾
Note that 𝜕ℎ1 /𝜕𝛾 is equal to 𝜕𝐽 /𝜕𝛾 , except where the indices 𝑖, 𝑗, and 𝑘 change. At such
points, ℎ1 is not differentiable with respect to 𝛾, which does not conflict with 𝜕ℎ1 /𝜕𝛾 being
a piecewise constant function. To find the slope of ℎ1 where it exists, the chain rule can be
applied as
𝜕ℎ1 𝜕𝐽 𝜕𝛾 𝜕𝐽 𝜕𝛾 𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑗 ‖𝑣∗ − 𝑤∗ ‖
= ∗/ ∗ + ∗/ ∗ = ,
𝜕𝛾 𝜕𝑣 𝜕𝑣 𝜕𝑤 𝜕𝑤 2 𝑣∗ − 𝑤∗
which, considering (5.14), leads to
𝜕ℎ1
= −2𝜇, (5.15)
𝜕𝛾
which implies that 𝜕ℎ1 /𝜕𝛾 is a function of the 𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑗 values. □
5.4 Theoretical Analysis 53

Lemma 5.2. For a convex PWA function 𝑓𝑝 expressed by (5.2), the convexity modulus ℎ1 in
(5.5) is continuous on [0, diam(𝑝,. )).

Proof. From Proposition 5.1, we know that ℎ1 is left-continuous on [0, +∞). Seeking a
contradiction, let us assume that ℎ1 is not right-continuous in 𝛾0 ∈ [0, diam(𝑝,. )), hence,

lim ℎ1 (𝛾) ≠ ℎ1 (𝛾0 ).

𝛾→𝛾0+

The monotonicity property of ℎ1 in Proposition 5.1 requires

lim ℎ1 (𝛾) > ℎ1 (𝛾0 ).

𝛾→𝛾0+

Therefore, without loss of generality, we assume there exists a gap 𝜖0 > 0 and a point
𝛾0 < 𝛾0+ < diam(𝑝,. ) such that

𝜕ℎ1 |
ℎ1 (𝛾0+ ) = ℎ1 (𝛾0 ) + | +
(𝛾 − 𝛾0 ) + 𝜖0 . (5.16)
𝜕𝛾 |𝛾=𝛾0 0 5
Using (5.5), we define the points 𝑣0 , 𝑤0 , and 𝑤0+ such that

ℎ1 (𝛾0 ) = inf 𝐽 (𝑣, 𝑤) = 𝐽 (𝑣0 , 𝑤0 ),

𝑣,𝑤∈𝑝,.
‖𝑣−𝑤‖=𝛾0

and ‖𝑣0 − 𝑤0+ ‖ = 𝛾0+ . Considering the optimality property in (5.5), we have

ℎ1 (𝛾0+ ) ⩽ 𝐽 (𝑣0 , 𝑤0+ ),

and knowing that 𝐽 is a continuous function by definition, we can deduce

ℎ1 (𝛾0+ ) ⩽ ℎ1 (𝛾0 ) + 𝜌(𝑤0+ − 𝑤0 ),

where 𝜌 ∶ 𝑝,. → ℝ is a function with the following property:

lim 𝜌(𝜈) = 0. (5.17)

𝜈→0+

Substituting (5.16) into (5.17) and taking the limit on both sides when 𝛾0+ approaches 𝛾0
leads to

ℎ1 (𝛾0 ) + 𝜖0 ⩽ ℎ1 (𝛾0 ),

which contradicts the fact that 𝜖0 > 0. Therefore, ℎ1 is right-continuous on [0, diam(𝑝,. )).
□

Proposition 5.2. For a convex PWA function 𝑓𝑝 expressed by (5.2), we have

{ }
ℎ1 (𝛾) = 0, ∀𝛾 ⩽ max diam(𝑝,𝑞 ) .
𝑞∈𝑄𝑝
 54 5 Sensitivity Analysis for PWA Approximations of NLPs

Proof. By seeking a contradiction, let us assume that

{ }
∃𝛾0 ⩽ max diam(𝑝,𝑞 ) , such that ℎ1 (𝛾0 ) > 0,
𝑞∈𝑄𝑝

with the corresponding optimal points 𝑣0∗ and 𝑤0∗ from (5.5) such that

ℎ1 (𝛾0 ) = inf 𝐽 (𝑣, 𝑤) = 𝐽 (𝑣0∗ , 𝑤0∗ ).

𝑣,𝑤∈𝑝,.
‖𝑣−𝑤‖=𝛾0

Let us select two points, 𝑣0 and 𝑤0 , on the largest subregion in 𝑝,. such that

‖𝑣0 − 𝑤0 ‖ = 𝛾0 .

which results in 𝐽 (𝑣0 , 𝑤0 ) = 0. Considering the optimality property in (5.5), we have

𝐽 (𝑣0∗ , 𝑤0∗ ) ⩽ 𝐽 (𝑣0 , 𝑤0 ),

which contradicts the initial assumption that ℎ1 (𝛾0 ) > 0 and ℎ1 (𝛾0 ) = 𝐽 (𝑣0∗ , 𝑤0∗ ).
5 □
Lemma 5.3. For a convex PWA function 𝑓𝑝 expressed by (5.2), the convexity modulus ℎ1 is
bounded by ℎ̂ 1 ⩽ ℎ1 , with
{
̂ℎ1 (𝛾) B 0 if 𝛾 < diam(𝑝,. )
, (5.18)
𝑐1 𝛾 + 𝑐0 if 𝛾 ⩾ diam(𝑝,. )

where
{ }
𝑎𝑇𝑝,𝑖 − 𝑎𝑇𝑝,𝑗
𝑐1 = min , (5.19a)
𝑗∈𝑄𝑝 2
s.t. 𝑖 = arg max diam(𝑝,𝑞 ), (5.19b)
𝑞∈𝑄𝑝

𝑝,𝑖 ∩ 𝑝,𝑗 ≠ ∅, (5.19c)

and 𝑐0 = 𝑐1 diam(𝑝,. ).
Proof. This can be directly deduced from Proposition 5.2, considering the continuity of
ℎ1 from Lemma 5.2, the piecewise-constant property of 𝜕ℎ1 /𝜕𝛾 from Lemma 5.1, and the
increasing property of ℎ1 from Proposition 5.1. □
We are now in the position to state our main result:
Theorem 5.3. Let 𝐹 ∶  → ℝ be a scalar-valued objective function and let 𝑓 be a continuous
PWA function as in Definition 5.1 that approximates 𝐹 with bounded approximation error
𝛿 = 𝑓 − 𝐹 . Let 𝑓𝑝 in (5.2) be the local convex segment of 𝑓 in its MMPS form (5.1) on the set
𝑝,. , and let 𝛿𝑝 ∶ 𝑝,. → ℝ be the corresponding approximation error bounded by

sup |𝛿𝑝 (𝑥)| = Δ𝑝 < ∞.

𝑥∈𝑝,.
5.5 Case Study 55

Let 𝑥𝑝∗ be any global minimizer of 𝑓𝑝 and 𝑥̂𝑝∗ be any global minimizer of 𝐹 on 𝑝,. . Then, the
following condition holds:

2Δ𝑝 { }
‖𝑥̂𝑝∗ − 𝑥𝑝∗ ‖ ⩽ + max diam(𝑝,𝑞 ) , (5.20)
𝑐1 𝑞∈𝑄𝑝

where 𝑐1 is defined in (5.19).

Proof. This can be directly concluded by extending Theorem 5.2 via considering Proposi-
tion 5.2 and Lemma 5.3. □

5.5 Case Study

To showcase the application of Theorem 5.3, we select the 1-dimensional cut of the well-
known Eggholder test function [110] at 𝑥2 = 0 given by
√ √
|𝑥 |
𝐹 (𝑥) = −47 sin
(
| + 47| − 𝑥 sin ( |𝑥 − 47|) .
|2 |) 5
We approximate 𝐹 by a continuous PWA function 𝑓 that we arbitrarily take as

𝑓 (𝑥) = min (𝑓𝑝,. ) .

𝑝∈5

with local convex segments

𝑓1,. = max (𝑓1,𝑞 ) , 1,. = [−512, −385], (5.21a)

𝑞∈3

𝑓2,. = 𝑓2,1 , 2,. = [−385, −330], (5.21b)

𝑓3,. = max (𝑓3,𝑞 ) , 3,. = [−330, −180], (5.21c)
𝑞∈3

𝑓4,. = max (𝑓4,𝑞 ) 4,. = [−180, 180], (5.21d)

𝑞∈3

𝑓5,. = 𝑓5,1 , 5,. = [180, 512]. (5.21e)

Figure 5.2 shows the plots for the nonlinear objective function 𝐹 and its PWA approximation
𝑓 . The subregions 𝑝,. with 𝑝 ∈ 5 are illustrated by different colors.
Theorem 5.3 can be used in two ways:

1. guaranteeing bounds on the distance between the regional minima of 𝐹 and 𝑓 on

each subregion, given a bound on the approximation error, and

2. finding the required criteria for the approximation to allow guaranteeing a desired
bound on the distance between these minima, which we refer to as the confidence
radius.

We discuss each case separately by focusing on the approximation on 3,. .

 56 5 Sensitivity Analysis for PWA Approximations of NLPs

𝑦
1,.
∗
𝑥̂2∗ 𝑥2
2,. 300
−512

512
3,. 𝑥
𝑥∗
−385
−330

−180

180
4,. 𝑥3∗ 𝑥̂4∗ 4
𝑥̂1∗ 𝑥̂3∗ −400 𝑥5∗
5,. 𝑥̂5∗
𝑥1∗
Nonlinear function 𝐹 PWA function 𝑓
Regional minima of 𝐹 Regional minima of 𝑓

Figure 5.2: Plots of the nonlinear objective function 𝐹 and its PWA approximation 𝑓 .

5.5.1 Case 1: Finding the Confidence Radii

Figure 5.3a shows 𝐹 and two PWA approximations on 𝑝,. with two approximation errors.
The first is 𝑓3(1) , which divides 3,. into 3 subregions with maximum approximation error
5 Δ(1) (2)
3 = 19.9. The second approximation is 𝑓3 with 8 affine pieces, the maximum error
(2)
Δ3 = 2.6. The approximations are given by
⎧
⎪ −8.6𝑥 − 2613.1 ⎫
⎪
⎪
⎪ ⎪
⎪
⎪ −6.8𝑥 − 2095.6 ⎪
⎪
⎪ ⎪
⎪
⎧ ⎫ ⎪ −4.6𝑥 − 1477.9 ⎪
⎪ −7.8𝑥 − 2365.7 ⎪ ⎪ ⎪
⎪ ⎪ ⎪ −2.2𝑥 − 829.8 ⎪
𝑓3(1) (𝑥) = max ⎨ −0.9𝑥 − 501.2 ⎬ , 𝑓3(2) (𝑥) = max ⎨ ⎬ .
⎪
⎪ ⎪
⎪ ⎪
⎪ 0.3𝑥 − 191.6 ⎪
⎪
⎩ 6.1𝑥 + 1176.1 ⎭ ⎪
⎪ 2.8𝑥 + 412.5 ⎪
⎪
⎪
⎪ ⎪
⎪
⎪ 5.1𝑥 + 944 ⎪
⎪
⎪ ⎪
⎪
⎩ 6.9𝑥 + 1348.1 ⎭
The inverses of the corresponding convexity moduli are shown in Fig. 5.3b in the same color
as their corresponding 𝑓 in Fig. 5.3a, where 𝜒 is the confidence radius in Definition 5.3.

𝑦 𝜒
−1
200 ℎ(1)
1
𝑓3(2) (1) 70.0 −1
0 𝑥 ℎ(2)
1
−315 (2) −180 44.3
𝐹
−250 𝑓3(1) 2Δ
𝑥3∗ 𝑥3∗ 5.2 39.8
(a) Nonlinear and PWA functions (b) Inverse of the convexity modulus

Figure 5.3: Comparison of two different PWA approximations of the nonlinear function on 3,. .

Using Theorem 5.3, the confidence radius for 𝑓3(1) by is obtained by 𝜒 (1) = 70, which
5.5 Case Study 57

is the same value obtained by finding ℎ1 and its inverse function using (5.5), which is
presented in Fig. 5.3b. The same process can be performed for the second approximation,
𝑓3(2) , which gives 𝜒 (2) = 44.3. Note that Theorem 5.3 is more conservative for larger values
of Δ, compared to directly using the definition of ℎ1 . For instance, if Δ(2)
3 = 12.5, employing
Theorem 5.3 leads to 𝜒 (2) = 65.7, while the computed confidence radius using ℎ1 is 58.4.
The areas within the confidence radii for the PWA approximations are highlighted on the
𝑥-axis in Fig. 5.3a as well.

5.5.2 Case 2: Finding the Approximation Criteria

In this case, we approach the problem from another direction: we select a desired confidence
radius 𝜒 (3) and find the required criteria for the corresponding approximated function, 𝑓3(3) .
Let the desired 𝜒 (3) = 15; then,

2Δ(3)
3
{ }
+ max diam(3,𝑞 ) ⩽ 10,
𝑐1 𝑞∈𝑄𝑝

which means the diameter of the largest subregion 3,. must be smaller than 10. Firstly, 5
given that diam(3,. ) = 150, it can be concluded that the PWA approximation requires
at least 10 partitions. We can then start the approximation by partitioning 𝑝,. into 15
subregions with the same diameter and find the lowest possible error bound Δ(3) 3 for the
approximation, which is obtained as 2.83 with 𝑐1 = 0.0072. For this approximation, 𝜒 (3)
already exceeds diam(3,. ).
To improve upon this example, we add another partition to reduce the largest partition
diameter further and this time we do not aim at partitions of 3,. with the same diameter,
but require
max{diam (3,𝑞 )} ⩽ 10.
𝑞∈16

We find Δ(3)
3 = 2.47 with 𝑐1 = 1.03 and

max{diam (3,𝑞 )} = 9.4.

𝑞∈16

For this values, we obtain 𝜒 (3) = 14.24. In case this value is acceptable, we can use the
corresponding PWA approximation while ensuring that the minimizer of 𝐹 on 3,. lies in
a ball or radius 14.24 around the minimizer of 𝑓 (3) . In case a tighter confidence radius is
desired, the same procedure can be followed by adding more subregions.

5.5.3 Application for NMPC Optimization

To showcase the application of our proposed method in PWA approximation to control
optimization problems, we use an inverted-pendulum NMPC problem as in [93]. Consider-
ing a prediction horizon of 𝑁p = 2 and zero initial rotation velocity 𝜃̇ at time step 𝑘, the
objective function 𝐽NMPC can be formulated as a function of the measured pendulum angle
𝜃𝑘 , and the control inputs over the horizon 𝑢𝑘 and 𝑢𝑘+1 . For instance, for 𝜃𝑘 = 0 we have
√ √ √
𝐽NMPC (0, 𝑢𝑘 , 𝑢𝑘+1 ) = (0.02𝑢𝑘 + 𝜋)2 + 2𝜋 2 + 0.02 𝑢2𝑘 + (𝑢𝑘 + 𝑢𝑘+1 )2 + 0.01 𝑢2𝑘 + 𝑢2𝑘+1 .
 58 5 Sensitivity Analysis for PWA Approximations of NLPs

Moreover, the feasible region is defined as the box constraint |𝑢𝑘+𝑖−1 | ⩽ 20𝑁 , 𝑖 ∈ 2 with
diam(1,. ) = 56.4.
We approximate 𝐽NMPC by two convex MMPS forms 𝑓 (1) and 𝑓 (2) – with 𝑃 (1) = 𝑃 (2) = 1
in (5.1) – with different complexities in terms of the number of affine functions as
{ }
(2)
𝑄 (1) = 4, Δ(1) = 0.19, max diam (1,𝑞 ) = 28.2,
𝑞∈4
{ }
(2)
𝑄 (2) = 24, Δ(2) = 0.01, max diam (1,𝑞 ) = 14.6.
𝑞∈24

The inverse of the convexity modulus and the corresponding confidence regions for both
approximations are shown in Fig. 5.4. While 𝑓 (1) has a low approximation error, its
complexity level does not allow to guarantee a confidence radius lower than the diameter
of the feasible region. However, the more accurate approximation 𝑓 (2) guarantees a smaller
confidence radius. Moreover, a general approximation criterion can be obtained, similar
to the Eggholder NLP example, for an NMPC problem. In this case, it can be observed
from (5.20) and Fig. 5.4a that 𝜒 is lower-bounded by the maximum subregion diameter.
5 Therefore, if a particular confidence radius is desired, the approximation problem (5.10) can
be solved while imposing constraints on the diameter of subregions, e.g. an upper bound
on the maximum subregion diameter.

𝜒
𝑢𝑘+1
56.4
(1)
ℎ−1
1
28.3
ℎ−1
1
(2) 𝑢𝑘
14.6
2Δ
0.02 0.89
(a) Confidence radii (b) Confidence regions

Figure 5.4: Comparison of two PWA approximations of NMPC.

5.6 Conclusions
This chapter has introduced a novel approach for bounding the minimizers of polytopically-
constrained NLPs with nonlinear continuous objective functions using continuous PWA
function approximations. We have leveraged the continuity of the PWA approximations
resulting from employing an MMPS formalism to construct a locally-convex representation
of the PWA approximation, thus facilitating the derivation of guaranteed bounds on the
distance between the original and the approximated optimal solutions of the NLP by
considering the maximal approximation error. Our approach offers a practical tool for
determining criteria to achieve desired solution bounds. The effectiveness of the method
has been demonstrated through a case study on the Eggholder function, highlighting the
practical application of the proposed method and its potential impact in optimization and
control.
5.6 Conclusions 59

For future work, our primary objective is to conduct an in-depth analysis of the conser-
vatism of our approach and reducing it by refining our sensitivity analysis addressing global
minima on the whole domain, as well as extending our method to NLPs with non-convex
constraints. Moreover, we aim at conducting a comprehensive analysis to gain insight into
the impacts of the the improved computational efficiency through PWA approximation in
light of the corresponding solution bounds. Finally, investigating the effects of probabilistic
error bounds would also be an interesting direction to help integrating our approach into
learning-based and data-driven applications.

5
 61

II
Collision Avoidance in
Automated Driving
5
 63

6
Design and Numerical
Analysis of Hybridization
Benchmark for NMPC
Science does not aim at establishing immutable truths and eternal dogmas; its aim is to 6
approach the truth by successive approximations, without claiming that at any stage final
and complete accuracy has been achieved.

— Bertrand Russell, Sceptical Essays

Despite the extensive application of nonlinear Model Predictive Control (MPC) in automated
driving, balancing its computational efficiency with its control performance and constraint
satisfaction remains a challenge in emergency scenarios: in such situations, sub-optimal but
computationally rapid responses are more valuable than optimal responses obtained after long
computations. This chapter introduces a hybridization approach for efficient approximation
of the nonlinear vehicle dynamics and of its non-convex constraints, e.g., arising during
emergency evasive maneuvers. Hybridization, i.e. , the use of hybrid systems modeling, allows
to reformulate the nonlinear MPC problem as a hybrid MPC problem. Max-Min-Plus-Scaling
(MMPS) hybrid modeling is used to approximate the nonlinear vehicle dynamics. Meanwhile,
different formulations for constraint approximation are presented, and various grid-generation
methods are compared to solve these approximation problems. Among these, two novel grid
types are introduced to structurally include the influence of the nonlinear vehicle dynamics on
the grid point distributions in the state domain. Overall, the work presents and compares three
hybrid models and four hybrid constraints for efficient MPC synthesis and offers guidelines for
implementation of the presented hybridization framework in other applications.

This chapter has been submitted to IEEE Transactions on Intelligent Vehicles [85].
 64 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

6.1 Introduction
MPC has become increasingly popular in automated driving research over the past few
decades [112]. This is mainly due to its capability to handle constraints and its ability to
adapt to the system by performing controller synthesis in a rolling-horizon optimization-
based manner. However, high computation loads remain a major obstacle towards real-time
implementation of MPC for high levels of automation. In particular, Level 4 and Level 5
of automation defined by the Society of Automated Engineers [5] must be able to handle
hazardous scenarios without any intervention from the human driver. Clearly, in such
critical situations, sub-optimal but computationally rapid responses are more valuable than
optimal responses obtained after long computations. Thus, improving the computational
efficiency of MPC in critical scenarios remains a crucial research challenge.
Several lines of research have been investigated to deal with this challenge: suggested
approaches to increase computational efficiency include decoupling the lateral and longitu-
dinal vehicle dynamics [113] or using ad-hoc kinematics and dynamics [8]. Partly-related
research lines have looked at how model fidelity affects the control performance during
critical maneuvers in limits of friction [114] or around drift equilibria [115].
Another line of research has been studying computationally more efficient solutions
to the nonlinear optimization problem e.g., via new numerical algorithms [84] or offline
explicit solutions [116]. Nevertheless, Tavernini et al. [117] demonstrated that the offline
explicit MPC approach does not yield significant computational improvements. Adap-
6 tive weights, adaptive prediction horizon [118] or adaptive sampling times [119] have
also been examined, which can sometimes improve computational efficiency although
Wurts et al. [10] argue that varying sampling times can increase the computational burden
due to the resulting change in integration points in the prediction horizon.
Switching-based control designs are another line of research for computational ef-
ficiency of MPC, for instance, by switching among different prediction models [120].
Nevertheless, there is often no systematic way to define a good switching strategy, as the
switching can be defined in different ways such as switching to a higher-fidelity model in
case of uncontrollable error divergence [121], or switching among different drifting/driving
modes [122]. In this sense, a more systematic framework that covers switching-based
design as a special case is hybridization [81]. Hybridization refer to approximating the
control optimization problem using a hybrid systems formulation incorporating both
continuous and discrete dynamics [18]. Hybridization is equivalent to breaking down a
nonlinear possibly complex form into multiple modes with lower complexity, each mode
being valid in a local activation region. By this approach, nonlinearity is traded with the
introduction of discrete dynamics, representing the switching among the different modes
of the system [23].
Hybridization has been used to improve the computational speed in various applica-
tions [33, 35, 41]. In the automated driving literature, different approaches to hybridize
the vehicle dynamics include representing the nonlinear tire forces by a piecewise-affine
function [36, 83, 84], using a grid-based linear-parameter-varying approximation [123],
or using a hybrid equivalent state machine [124]. Nevertheless, to the best of our knowl-
edge, hybridization has not yet been incorporated into emergency evasive maneuvers
and/or highly-nonlinear vehicle dynamics. For example, the hybridization in [21] via a
Mixed-Logical-Dynamical (MLD) formalism [26] is only valid at low-speeds where vehicle
6.1 Introduction 65

nonlinearity can be neglected and the coupling between lateral and longitudinal vehicle
dynamics is weaker.
Indeed, in addition to the nonlinear vehicle model that enters the MPC problem as
equality constraint, another crucial source of nonlinearity in the control optimization
problem is caused by the physics-based inequality constraints such as handling and tire
force limits that are generally non-convex. The hybridization problem in MPC must
necessarily involve both model and constraint approximations, an aspect that is often
neglected in the literature. Despite some similarities, there are clear distinctions in the two
resulting hybridization problems that must be taken into account.
Among different hybrid modeling frameworks, MMPS systems [24] do not require to
explicitly represent the activation regions, which simplifies the approximation by significant
reduction of the number of decision variables. For this reason, the MMPS approach is the
one adopted in this work. As its name suggests, MMPS formulation represents a function
using only (and possibly nested) max, min, adding and scaling operators. Kripfganz [91]
showed that any MMPS function can also be equivalently represented by the difference of
two convex MMPS functions, which can increase computational tractability.
Physics-based non-convex constraints have been dealt with in different ways. For
instance, [125] considers the convex hull of the non-convex polyhedral constraints and
disregards non-optimal solutions using the binary search tree of [126]. In reachability
analysis, [127] computes an inner-approximation of the feasible region using an outer
approximation of the reachable sets. 6
Lossless or successive convexification is a common approach to deal with non-convex
constraints, as often considered in real-time trajectory planning [128–130]. However, the
real-time capability of the convexification method is a crucial and non-trivial aspect, since
the non-convex constraints imposed by the environment are changing at each control time
step.
Convexification problem can be solved offline only when the constraints are known to be
fixed. In some applications such as path planning in cluttered environments, it is important
to find a feasible region for the next control time step, which translates into finding the
largest convex subset of a given cluttered feasible region [131]. Nevertheless, a generic
offline convexification problem can be obtained by approximating a non-convex region by
a union of convex subregions. As defining these subregions manually is unpractical [132],
approaches from computational geometry have been proposed. For instance, it has been
shown that convexification is analogous to the NP-hard problem of Approximate Convex
Decomposition [133] with applications to shape analysis [134] or decision region in pattern
recognition [135]. Indeed, the recent advances in this field have been tailored more and
more toward the specific needs of pattern recognition. For example, more emphasis is
given on shape analysis by concavity matrices [136]: however, in critical automated driving
scenarios, it is rather important to analyze the approximations inaccuracy with respect
to the distance to the non-convex boundary. Existing methods in this sense are mainly
tailored for non-convex polyhedral regions [137], but several physics-based constraints
arising during critical maneuvers are not polyhedral.
In practice, hybridization has rarely been considered for highly complex vehicle models;
e.g., to the best of our knowledge, there are no studies that include hybridization of the
coupled longitudinal and lateral vehicle dynamics. Moreover, controlling evasive maneuvers
 66 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

in critical scenarios requires a systematic analysis of the vehicle model complexity and the
resulting computation trade-off, which has not been conducted as far as we are aware.
In this chapter, we provide a comparison benchmark to analyze and improve the
computational performance of MPC optimization problem for vehicle control in critical
high-velocity scenarios using hybrid formulation of the control optimization problem.
This benchmark is divided in two parts: the first part is dedicated to the hybridization
of the MPC via approximating the constraints, i.e., prediction model and physics-based
constraints, whereas the second part investigates the improvements of the resulting hybrid
MPC controller in comparison with the original nonlinear MPC controller.
The current chapter contributes to the state of the art by:

• presenting of a novel hybrid approximation of the system using an MMPS formula-

tion,

• developing a new generalized formalism for constraint approximation problem

including an approach based on a polytopic definition of the regions by an MMPS
function, and comparing the resulting approximations with two methods from the
literature,

• introducing two trajectory-based grid generation method for model approximation,

• investigating grid-based numerical solutions of the model and constraint approxima-

6 tion with respect to the grid behavior, and

• presenting a novel benchmark for evaluating and comparing the computational

efficiency of various nonlinear MPC controllers.

The chapter is organized as follows: Section 6.2 covers the preliminary definitions of
the model and constraint approximation problems. Section 6.3 describes the grid generation
methods, including the novel trajectory-based approach in non-uniform sampling of the
input/state pairs. Section 6.4 defines the approximation problems. Section 6.5 presents
the hybridization framework for model and constraint approximation using the generated
grids and the validation results of the said approximation problems. Section 6.6 summarizes
the hybridization framework, findings, and outlook for implementation and future work.
The application and analysis of the presented hybridization framework is discussed in
detail in the next chapter.

6.2 Background
Consider a given nonlinear system, either in continuous-time 𝑥̇ = 𝐹 (𝑥, 𝑢) or in discrete-time
𝑥 + = 𝐹 (𝑥, 𝑢) where 𝑥 ∈ ℝ𝑛 and 𝑢 ∈ ℝ𝑚 respectively represent the state and input vectors,
and the domain of 𝐹 is denoted by (𝑥, 𝑢) ∈  ⊆ ℝ𝑚+𝑛 . In many physics-based applications,
the model 𝐹 is valid over a region  ⊆  defined by

 B {(𝑥, 𝑢) ∈  | 0 ⩽ 𝐺(𝑥, 𝑢) ⩽ 1},

which collects a set of physics-based constraints. For instance, most typical vehicle models
in the literature are no longer valid if e.g., the vehicle is rolling over. Here we aim at
6.2 Background 67

approximating both the nonlinear model 𝐹 and the nonlinear, non-convex set . Therefore,
we need to hybridize both 𝐹 and . Both approximation problems can essentially be
expressed as the minimization of the approximation error over their respective domains.
The approximation error, as well as the domain, are different for each problem, as discussed
hereafter.

Remark 6.1. We use the normalized constraint formulation 0 ⩽ 𝐺 ⩽ 1 instead of the generic
form 𝐺 ⩽ 0 to avoid numerical issues in solving the approximation/control optimization
problems.

6.2.1 Model Approximation

The system 𝐹 is approximated by a hybrid formulation 𝑓 via solving the nonlinear opti-
mization problem

‖𝐹 (𝑥, 𝑢) − 𝑓 (𝑥, 𝑢)‖2

min ∫ 𝑑(𝑥, 𝑢), (6.1)
 ‖𝐹 (𝑥, 𝑢)‖2 + 𝜖0


where  represents the decision variables used to define 𝑓 . The positive value 𝜖0 > 0 added
to the denominator is to avoid division by very small values for ‖𝐹 (𝑥, 𝑢)‖2 ≈ 0. Note that
the domain in the model approximation problem is .
6
6.2.2 Constraint Approximation
With the nonlinear, non-convex constraints given as 0 ⩽ 𝐺(𝑥, 𝑢) ⩽ 1, we approximate the
feasible region  by a union of convex subregions .
This approximation problem can be formulated in two ways: region-based and boundary-
based. In the region-based approach, we minimize the misclassification error via solving
the following optimization problem

{ ⧵ } { ⧵ }
min 𝛾c + (1 − 𝛾c ) , (6.2)
𝜈 {} { ⧵ }

where 𝜈 represents the decision variables used to define , the operator  gives the
size or “volume” of the region, and 𝛾c ∈ [0, 1] is a tuning parameter to adjust the relative
penalization weight for the misclassification errors regarding inclusion error  ⧵ , i.e.,
failing to cover the feasible region, and the violation error  ⧵  which corresponds to
violating the constraints.
In the boundary-based approach, we approximate the boundary 𝐺 by a hybrid function 𝑔
and minimize the boundary-approximation error similar to (6.1) via solving the optimization
problem

|𝐺(𝑥, 𝑢) − 𝑔(𝑥, 𝑢)|

min ∫ 𝑑(𝑥, 𝑢). (6.3)
𝜈 |𝐺(𝑥, 𝑢)| + 𝜖0


with 𝜖0 > 0. Note that as 𝐺 is a scalar function, the 2-norm is replaced by the absolute value
here.
 68 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

Remark 6.2. The proposed ideas also apply in case of more inequalities e.g.,

0 ⩽ 𝐺𝑖 (𝑥, 𝑢) ⩽ 1, for 𝑖 ∈ {1, 2, … , 𝑁 },

by simply formulating 𝐺(𝑥, 𝑢) as

𝐺(𝑥, 𝑢) = max {𝐺𝑖 (𝑥, 𝑢)}.

𝑖∈{1,2,…,𝑁 }

Another possibility is to approximate each 𝐺𝑖 independently; however, this may lead to

redundant approximations of boundaries or parts of 𝐺𝑖 that do not belong to the overall
boundary feasible region.

6.2.3 Relation to the State-of-the-Art

The nonlinear non-convex constraints arise from the physics-based limitations of the
system. Therefore,
• the physics-based nature of the constraints results in a connected feasible region,
• the highly-nonlinear (boundary of the) constraints limits the analytical investigation
of “attainability” or optimality,
• the approximation approach is intended to be used within a hybridization benchmark,
6 which means the method should be applicable for systems of higher degree and/or
with high-dimensional feasible regions,
• the constraint violation is evaluated by ensuring that the solution lies within any of
the subregions, which means overlapping subregions are acceptable,
• in light of improving the computational efficiency, it is desired to have a minimal
approximation of the constraints, i.e., approximating the non-convex feasible region
with a union of fewer number of subregions is desired as well as an accurate coverage
of the whole region, which leads to the need for
• a systematic approach to cover the non-convex feasible region by a union of convex
subregions that allows balancing the violation vs. coverage of the approximation
close to the constraint boundaries.
Considering the aforementioned features, the applicability of state-of-the-art methods
based on convex-hull generation [138] is limited for the current case as input-state spaces
for complex vehicle models exceed four dimensions and a systematic division of the feasible
region is not computationally efficient in terms of memory usage and speed for our desired
accuracy. To compare our constraint approximation approach, we consider two state-of-the-
art methods that share the most common elements with the aforementioned considerations
in their respective problems.
The first method is from [133], where a non-convex region is covered by a number
of ellipsoids. There, an optimization problem is solved to minimize the misclassification
error due to the region approximation where the center and radii of the ellipsoids are the
Attainability of a point means that there exists an input such that the point is obtained by the system dynamics.
6.3 Grid Generation 69

decision variables. We refer to this approach as non-parametric elliptical learning, which is

equivalent to region-based approximation of the constraints by a union of ellipsoids. Our
constraint approximation framework can be seen an extension and generalization of this
approach by investigating boundary-based vs. region-based approximations and polytopic
vs. ellipsoidal definition of the subregions.
The second method is from [139], where the gripping limits of the vehicle are approxi-
mated by a convex intersection of second-order cone constraints. There, the constraints are
formulated using the system dynamics and the parameters of the combined formulation
are fitted using experimental data. We refer to this approach as the convex envelope
method, which is equivalent to a boundary-based approximation of the constraints by the
intersection of multiple convex subregion. Since this method approximates the non-convex
feasible region by a convex one, in Section 6.5 we will show its limitation in converging to
an accurate approximation of the constraints in comparison with our proposed framework.
Since analytical closed-form solutions for (6.1)–(6.3) do not exist, we propose solving
them numerically by generating a grid of samples from their regarding domains  and ,
respectively denoted by  ∗ and ∗ . As the grid generation method influences the quality of
the final fit, we provide various grid-generation methods for both approximation problems
in the next section and examine the resulting fits in our results in Section 6.5.

6.3 Grid Generation

We use two main approaches to generate ∗ : domain-based and trajectory-based. In the
6
domain-based approach, both the input and state elements of the grid points are selected
from the input/state domain , regardless of the system’s behavior. While a domain-based
grid can have a good coverage of , it does not take into account the “likelihood” of the
points being visited in a simulation with respect to the system dynamics. The trajectory-
based way of generating ∗ tackles this issue by selecting the input elements of the grid
points 𝑢∗ from , while assigning the state elements to the points from an 𝑛step -step-ahead
simulation of 𝐹 given 𝑢∗ as the input. As a result, the obtained ∗ will have a higher density
in regions of  where the input/state pairs have a higher likelihood of being attainable.
Each of these two approaches can be implemented in two ways, giving rise to a total of
four methods to generate ∗ :

• Domain-based: [points are directly sampled from ]

– Uniform (∗𝑈 , also referred to as U grid type): the points are generated by
picking 𝑛samp uniformly-spaced points along each axis in .
– Random (∗𝑅 , also referred to as R grid type): a total of 𝑛rand points are
randomly selected from .

• Trajectory-based: [𝑛sim open-loop simulations with 𝑛step steps of 𝐹 are run using
random inputs from ]

For instance, another approach to solving the aforementioned approximation problem is the Monte Carlo
integration method.
 70 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

– Steady-state (∗𝑆 , also referred to as S grid type): the initial state of each
simulation is selected as the steady-state solution w.r.t. the initial input, i.e., it
is assumed that each simulation starts from a steady state.

– Randomly-initiated (∗𝑇 , also referred to as T grid type): the initial state

of each simulation is randomly selected from .

Algorithms 5 and 6 respectively explain the domain-based and trajectory-based grid

generation methods. The total number of grid points for each type denoted by  is

 (∗𝑈 ) = (𝑛samp )𝑚+𝑛 ,

 (∗𝑅 ) = 𝑛rand ,
 (∗𝑆 ) =  (∗𝑇 ) = 𝑛sim ⋅ 𝑛step .

6
Algorithm 5 Domain-based grid generation
Require: 𝐹 , , 𝑛samp , 𝑛rand , type ∈ {‘U’, ’R’}
∗type ← {}
if type = ‘U’ then
for 𝑘 ∈ {1, 2, … , 𝑚 + 𝑛} do
𝑘 ← {} ⊳ 𝑘 B sample set
1
for 𝑖 ∈ {0, , … , 1} do
𝑛samp − 1
𝑘 ← 𝑘 ∪ {(𝑘)min + ((𝑘)max − (𝑘)min ) ⋅ 𝑖}
end for
end for
∗U ← 1 × 2 × ⋯ × 𝑚+𝑛 ⊳ Cartesian product
else if type = ‘R’ then
for 𝑘 ∈ {1, 2, … , 𝑛rand } do
random
(𝑥𝑘 , 𝑢𝑘 ) ←−−−−−− 
∗R ← ∗R ∪ {(𝑥𝑘 , 𝑢𝑘 )}
end for
end if
return ∗type
6.3 Grid Generation 71

Algorithm 6 Trajectory-based grid generation

Require: 𝐹 , 𝑥 , 𝑢 , 𝑛sim , 𝑛step , type ∈ {‘S’, ’T’}
∗type ← {}
for 𝑠 ∈ {1, 2, … , 𝑛sim } do
random
𝑢 ←−−−−−− 𝑢 ⊳ 𝑢 B input domain
random
𝑥 ←−−−−−− 𝑥 ⊳ 𝑥 B state domain
if type = ‘S’ then
solve for 𝑥
𝑥1 ←−−−−−−−− 𝐹 (𝑥, 𝑢1 ) = 0 ⊳ steady-state solution
end if
for 𝑘 ∈ {2, 3, … , 𝑛step } do
𝑥𝑘 ← 𝑥𝑘−1 + 𝐹 (𝑥𝑘−1 , 𝑢𝑘−1 )
if (𝑥𝑘 , 𝑢𝑘 ) ∉ 𝑥 × 𝑢 then
break ⊳ stop current simulation
end if
∗type ← ∗type ∪ {(𝑥𝑘 , 𝑢𝑘 )}
end for
end for
return ∗type

6
Domain-based random Feasible region Feasible region
or
Uniform grid point
trajectory-based grid point

Boundary region
Domain-based
Domain uniform grid point
Domain
Boundary grid point

(a) Grid generation for model approximation (b) Grid generation for constraint approximation

Figure 6.1: A schematic view of different implementations of the proposed grid-generation approaches for model
and constraint approximation.

The grid ∗ plays the role of domain in the approximation problem. Therefore, it
should be tailored to the objective of the problem itself. In this sense, Figure 6.1 shows
a schematic view of the implementation of the proposed grid-generation approaches for
both model and constraint approximation problems.
 72 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

For model approximation, the grid should be generated only from , as the points
outside  are infeasible, which translates to zero likelihood of attainability. Therefore,
while Algorithms 5 and 6 are implemented on , only the samples from the feasible region
should be kept. Then, the four resulting grids, 𝑈∗ , 𝑅∗ , 𝑆∗ , and 𝑇∗ can be used to examine
their efficacy.
Contrary to the model approximation problem, the points for constraint approximation
should be distributed in the whole domain  to allow examining the approximation error.
In addition, for constraint approximation, the areas close to the boundary of  are of more
interest than the areas with higher likelihood of attainability. Therefore, while trajectory-
based methods are useful for model approximation, to find the constraints, we are interested
in using a domain-based grid with a higher density in the neighborhood of 𝐺(𝑥, 𝑢) = 0.
This grid can be obtained by combining a uniform grid ∗𝑈 with a random grid ∗𝑅 on the
boundary region  where

 B {(𝑥, 𝑢) ∈  | |𝐺(𝑥, 𝑢)| ⩽ 𝜖𝑏 }.

The resulting generated grid is ∗𝑈 ∪ ∗𝑅 .

Remark 6.3. To ensure that trajectory-based grids are generated by “realistic” inputs, we
impose a bound constraint on the random inputs as

6 |𝑢∗ (𝑘 + 1) − 𝑢∗ (𝑘)| < Δ∗𝑢 .

This can also account for the physical limitations of the actuators and be considered to be part
of the physics-based constraints  and it is best selected based on data from real operation of
the system.

Remark 6.4. Depending on the problem characteristics such as the system dynamics, domain,
and the nature of the input/state signals, some points in the generated grids (except for the U
grid type) can be very close to each other. To avoid these points from having larger importance
than other points during approximation, Algorithms 5 and 6 can further be refined by keeping
only one point from each set of points that are closer to each other than a user-defined distance
threshold.

6.4 Approximation Problem Formulation

6.4.1 Model Approximation
We approximate the nonlinear system 𝐹 by the MMPS function 𝑓 with the Kripfganz
form [91] as { } { }
𝑓 (𝑥, 𝑢) = max + 𝜙+𝑝 (𝑥, 𝑢) − max − 𝜙−𝑞 (𝑥, 𝑢) , (6.4)
𝑝∈{1,2,…,𝑃 } 𝑞∈{1,2,…,𝑃 }

where 𝑃 + and 𝑃 − are user-selected integers, and 𝜙+𝑝 , and 𝜙−𝑞 are affine functions of 𝑥 and 𝑢,
sometimes referred to as dynamic modes, and expressed as

𝜙+𝑝 (𝑥, 𝑢) = 𝐴+𝑝 𝑥 + 𝐵𝑝+ 𝑢 + 𝐻𝑝+ ,

𝜙−𝑞 (𝑥, 𝑢) = 𝐴−𝑞 𝑥 + 𝐵𝑞− 𝑢 + 𝐻𝑞− .

6.4 Approximation Problem Formulation 73

We implement the MMPS approximation in the following fashion: each dimension of the
nonlinear function, i.e., each component of 𝐹 , is approximated independently. Thus, 𝑃 +
and 𝑃 − , as well as the affine functions 𝜙+ and 𝜙− are separately found for each component
of 𝐹 . Therefore, for brevity and without loss of generality, one can assume 𝐹 to be scalar in
the remaining of this section.

For a fixed pair (𝑃 + , 𝑃 − ) that corresponds to the number of affine terms in the first and
second max operators in (6.4), we solve the nonlinear optimization problem (6.1) subject to
(6.4) to find the optimal 𝜙+ and 𝜙− functions where

{ }
 = 𝐴+𝑝 , 𝐴−𝑞 , 𝐵𝑝+ , 𝐵𝑞− , 𝐻𝑝+ , 𝐻𝑞− 𝑝∈{1,2,…,𝑃 + },𝑞∈{1,2,…,𝑃 − } . (6.5)

Remark 6.5. To solve the nonlinear optimization problem in (6.1), we generate a grid  ∗
of feasible samples from  as expressed in Section 6.3, and minimize the objective function
across  ∗ .

Remark 6.6. The Kripfganz form essentially expresses the function using 𝑃 + ⋅ 𝑃 − hyperplanes
6
as there are 𝑃 + and 𝑃 − affine functions in each max operator. Therefore, the hinging hyper-
planes representing the local dynamics are obtained by subtraction of the affine functions 𝜙−
from 𝜙+ which means that the optimal  in (6.1) would not be unique.

Considering Remarks 6.5 and 6.6 and to avoid numerical problems, it is convenient to
add a regularization term to (6.1) by penalizing the 1-norm of the decision vector as

|𝐹 (𝑥, 𝑢) − 𝑓 (𝑥, 𝑢)|

min ∫ 𝑑(𝑥, 𝑢) + 𝛾m ‖‖1 , s.t. (6.4), (6.6)
 |𝐹 (𝑥, 𝑢)| + 𝜖0
∗

where 𝛾m ∈ ℝ+ serves as a weighting coefficient to balance the penalization of the 1-norm

of  with respect to the approximation error.

6.4.2 Constraint Approximation

We approximate the feasible region  by either a union of convex polytopes using the MMPS
formalism, or by a union of ellipsoids. Figure 6.2 depicts both approaches to constraint
approximation.
 74 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

Feasible region

al

n
id ion

io
im S
at
so at

ox P
pr MM
l ip i m
El rox
p
ap
ap

Domain Domain

(a) MMPS constraint approximation (b) Ellipsoidal constraint approximation

Figure 6.2: Illustration of MMPS and ellipsoidal approximation of the nonlinear constraints.

In the MMPS approach, a similar formulation to the MMPS model approximation

problem is used: we approximate 𝐺 by an MMPS function 𝑔MMPS of the Kripfganz form in
(6.4) with
6 𝜙+𝑝 (𝑥, 𝑢) = 𝐶𝑝+ 𝑥 + 𝐷𝑝+ 𝑢 + 𝐼𝑝+ ,
𝜙−𝑞 (𝑥, 𝑢) = 𝐶𝑞− 𝑥 + 𝐷𝑞− 𝑢 + 𝐼𝑞− .
The resulting feasible region MMPS is then expressed as
MMPS B {(𝑥, 𝑢) ∈  | 𝑔MMPS (𝑥, 𝑢) ⩽ 0}, (6.7)
The MMPS approximation of the feasible region is then obtained via solving either the
region-based (6.2) or the boundary-based (6.3) optimization problems subject to
 = MMPS ,
and { }
𝜈 = 𝐶𝑝+ , 𝐶𝑞− , 𝐷𝑝+ , 𝐷𝑞− , 𝐼𝑝+ , 𝐼𝑞− 𝑝∈{1,2,…,𝑃 + },𝑞∈{1,2,…,𝑃 − } , (6.8)
where the matrices 𝐶, 𝐷, and 𝐼 represent the constraint-approximation counterparts of
matrices 𝐴, 𝐵, and 𝐻 in (6.5) and (𝑃 + , 𝑃 − ) stand for the respective number of affine terms.
The second way is to approximate the feasible region by a union of 𝑛e ellipsoids
{ }
| 𝑥 − 𝑥0𝑒 𝑇 𝑥 − 𝑥 0𝑒
𝑒 B (𝑥, 𝑢) ∈  || 𝑄𝑒 ⩽1 , (6.9)
| (𝑢 − 𝑢0𝑒 ) (𝑢 − 𝑢0𝑒 )

with 𝑄𝑒 being a positive definite matrix and (𝑥0 , 𝑢0 ) representing the center coordinates of
the ellipsoid. Note that this notation includes rotated ellipsoids as well. The approximated
region ELLP is
𝑛e
ELLP = ⋃ 𝑒 B {(𝑥, 𝑢) ∈  | 𝑔ELLP (𝑥, 𝑢) ⩽ 0}, (6.10)
𝑒=1
6.5 Model and Constraint Hybridization for Vehicle Control 75

whose boundary can be expressed by

{ }
𝑇
𝑥 − 𝑥0𝑒 𝑥 − 𝑥0𝑒
𝑔ELLP (𝑥, 𝑢) = min 𝑄 −1 . (6.11)
𝑒∈{1,2,…,𝑛e } (𝑢 − 𝑢0𝑒 ) 𝑒 (𝑢 − 𝑢0𝑒 )

The ellipsoidal approximation is found by solving either the region-based (6.2) or the
boundary-based (6.3) optimization problems subject to

 = ELLP ,

and { }
𝜈 = (𝑥0𝑒 , 𝑢0𝑒 ), 𝑄𝑒 𝑒∈{1,2,…,𝑛 } . (6.12)
e

6.5 Model and Constraint Hybridization for Vehi-

cle Control
In this section, the hybridization framework consisting of the model and constraint approx-
imation approaches is implemented on a nonlinear single-track vehicle model with Dugoff
tire forces and varying friction. First, the nonlinear system and physics-based constraints
are described, then the training and validation grids are defined, which are next used for
model and constraint approximation problems within the hybridization framework. The
results are then discussed to evaluate the performance of the different approaches and 6
analyzed for application in other nonlinear problems.

6.5.1 Nonlinear System Descriptions

A single-track representation of the vehicle is shown in Fig. 6.3. With the system variables
and parameters respectively defined in Tables 6.1 and 6.2, the nonlinear vehicle model is
described by the following equations [8]:
1
𝑣̇ 𝑥 = [𝐹𝑥f cos 𝛿 − 𝐹𝑦f sin 𝛿 + 𝐹𝑥r ] + 𝑣𝑦 𝑟, (6.13)
𝑚
1
𝑣̇ 𝑦 =[𝐹𝑥f sin 𝛿 + 𝐹𝑦f cos 𝛿 + 𝐹𝑦r ] − 𝑣𝑥 𝑟, (6.14)
𝑚
1
𝑟̇ = [𝐹𝑥f sin 𝛿 𝑙f + 𝐹𝑦f cos 𝛿 𝑙f − 𝐹𝑦r 𝑙r ] , (6.15)
𝐼𝑧𝑧
and the lateral forces are given by the Dugoff model
𝐶𝛼 𝑎
𝐹𝑦𝑎 = 𝑓𝜆 (𝜆𝑎𝑤 )𝛼𝑎 ,
1 − 𝜅𝑎
with 𝑎 ∈ {f, r} where 𝜇𝑎 is the varying friction coefficient, and 𝜆𝑎𝑤 and 𝑓𝜆 are the weighting
coefficient and function, defined as
√
𝜇𝑎 = 𝜇0 (1 − 𝑒𝑟 𝑣𝑥 𝜅𝑎2 + tan2 𝛼𝑎 ) ,

𝜇𝑎 𝐹𝑧𝑎 (1 − 𝜅𝑎 )
𝜆𝑎𝑤 = √ ,
2 (𝐶𝜅𝑎 𝜅𝑎 )2 + (𝐶𝛼𝑎 tan 𝛼𝑎 )2
 76 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

{
𝜆𝑤 (2 − 𝜆𝑎𝑤 ) 𝜆𝑎𝑤 < 1
𝑓𝜆 (𝜆𝑎𝑤 ) = 𝑎 .
1 𝜆𝑎𝑤 ≥ 1

YGlobal
y
Fxf α δ
Fyf f
vf x
Fyr vx
vy ψ
CoG
r = ψ·
Fxr
XGlobal
αr
vr lf
lr
L

Figure 6.3: Configuration of the single-track vehicle model.

Table 6.1: System variables

6 Variable
𝑣𝑥
Definition
Longitudinal velocity
Unit
m/s
Bounds
[5, 50]
𝑣𝑦 Lateral velocity m/s [-10, 10]
𝜓 Yaw angle rad –
𝑟 Yaw rate rad/s [-0.6, 0.6]
𝛿 Steering angle (road) rad [-0.5, 0.5]
𝐹𝑥f Longitudinal force on the front axis N [-5000, 0]
𝐹𝑥r Longitudinal force on the rear axis N [-5000, 5000]
𝐹𝑦f Lateral force on the front axis N –
𝐹𝑦r Lateral force on the rear axis N –
𝐹𝑧f Normal load on the front axis N –
𝐹𝑧r Normal load on the rear axis N –
𝛼f Front slip angle rad –
𝛼r Rear slip angle rad –
𝜅f Front slip ratio – –
𝜅r Rear slip ratio – –
𝜇f Friction coefficient on the front tire – –
𝜇r Friction coefficient on the rear tire – –
𝑇
𝑥 State vector B [𝑣𝑥 𝑣𝑦 𝑟 ] – –
𝑇
𝑢 Input vector B [𝐹𝑥f 𝐹𝑥r 𝛿 ] – –

Table 6.1 also shows the bounds we impose on state and input vectors for grid generation.
The feasible region is defined by two other physics-based constraints:
1. the working limits of the vehicle (known as the g-g diagram constraint [8]) should be
satisfied to allow derivation of the dynamics equation in (6.13) to (6.15); this entails
2 2 2
(𝑣̇ 𝑥 − 𝑣𝑦 𝑟 ) + (𝑣̇ 𝑦 + 𝑣𝑥 𝑟 ) ⩽ ( min {𝜇𝑎 𝑔}) , (6.16)
𝑎∈{f,r}
6.5 Model and Constraint Hybridization for Vehicle Control 77

Table 6.2: System parameters∗

Parameter Definition Value Unit

𝑚 Vehicle mass 1970 kg
𝐼𝑧𝑧 Inertia moment about z-axis 3498 kg/m2
𝑙f CoG∗∗ to front axis distance 1.4778 m
𝑙r CoG to rear axis distance 1.4102 m
𝐶𝛼f Front cornering stiffness 126784 N
𝐶𝛼r Rear cornering stiffness 213983 N
𝐶𝜅f Front longitudinal stiffness 315000 N
𝐶𝜅r Rear longitudinal stiffness 286700 N
𝜇0 Zero-velocity friction 1.076 –
𝑒𝑟 Friction slope 0.01 –
∗ These values correspond to the IPG CarMaker BMW vehicle model
∗∗ Center of Gravity

2. the tires can provide forces up to their saturation limit, known as the Kamm circle
constraint [8], which means
2
𝐹𝑥𝑎 2
+ 𝐹𝑦𝑎 ⩽ (𝜇𝑎 𝐹𝑧𝑎 )2 , 𝑎 ∈ {f, r}. (6.17)

Therefore, the feasible region  can be expressed as 6

 B {(𝑥, 𝑢) ∈  | (6.16) , (6.17)} .

6.5.2 Grid Definition and Coverage

Table 6.3 shows the grid properties for the model and constraint approximation problems.
For the model, all four U, R, S, and T grid types are used for training and later validated on
a finer U, R, S, T grid type, respectively, plus C grid type, i.e. a grid that combines all of
them. For the constraint approximation, only one combined grid consisting of the union U
and R grids is used for training and the approximations are validated on a finer and more
extended combined grid.
For a visual comparison of the grid-point distribution for different types, we have
plotted the coverage of the model approximation training and validation grids in the
velocity domain (𝑣𝑥 -𝑣𝑦 ) in Fig. 6.4. While the grids have a similar total number of points,
the density of the points among different grid types varies significantly as follows:

1. The domain-based grids cover  with a uniform density compared to the trajectory-
based grids.

2. Compared to its random counterpart, the U grid represents a sparser distribution in

the velocity domain, which stems from the fact that representation of all the possible
combinations of input/state pairs on lower-dimensional sub-spaces of  projects
many points on the exact same location in the viewed plane.

3. Between the trajectory-based grids, the T grid gives a better coverage of . Contrarily,
the S grid favors the regions of  where the states are attainable from a steady-state
 78 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

solution within a bounded number of steps, which explains the high density of points
in low-speed region and the loose coverage of high-speed regions with zero lateral
velocity.

Table 6.3: Properties of the grid used in the approximation problems (training and validation grids)

6
Training Grids for Model Approximation
Type Domain Properties No. Points Feasible
U  𝑛samp = 6 ≈ 7, 000 100%
R  𝑛rand = 7000 ≈ 7, 000 100%
S  𝑛sim = 500, 𝑛step = 1000 ≈ 7, 000 100%
T  𝑛sim = 300, 𝑛step = 1000 ≈ 7, 000 100%
Validation Grids for Model Approximation
Type Domain Properties No. Points Feasible
U  𝑛samp = 7 ≈ 21, 000 100%
R  𝑛rand = 21,000 ≈ 21, 000 100%
S  𝑛sim = 3000, 𝑛step = 1000 ≈ 21, 000 100%
T  𝑛sim = 1200, 𝑛step = 1000 ≈ 21, 000 100%
C  combining all the above ≈ 84, 000 100%
Training Grids for Constraint Approximation
Type Domain Properties No. Points Feasible
U  𝑛samp = 5 ≈ 15, 000 68%
R  𝑛rand = 15,000, 𝜖b = 0.1 ≈ 15, 000 41%
C  combining all the above ≈ 30, 000 55%
Validation Grids for Constraint Approximation
Type Domain Properties No. Points Feasible
U  𝑛samp = 6 ≈ 47, 000 68%
R  𝑛rand = 45,000, 𝜖b = 0.2 ≈ 45, 000 56%
C  combining all the above ≈ 92, 000 62%
6.5 Model and Constraint Hybridization for Vehicle Control 79

Figure 6.4: Location of training and validation grid points in the 𝑣𝑥 − 𝑣𝑦 domain for different grid-generation
approaches in model approximation

The constraint approximation grids in the velocity domain are shown in Fig. 6.5. Besides
generating more grid points in the validation grids, the width 𝜖𝑏 of its boundary region is
selected twice as large as for the training one, which increases the relative density of the
grid points in the high-speed region as visible in Fig. 6.5. Moreover, both grids have 50-60%
of their points in the feasible region, which is a reasonable ratio for a fair comparison.
6

Figure 6.5: Location of the training and validation combined grid points in the 𝑣𝑥 − 𝑣𝑦 domain for constraint
approximation

6.5.3 Model Approximation Results

Using the four model training grids in Table 6.3, we approximate the dynamics of the three
states independently by Kripfganz MMPS functions with (𝑃 + , 𝑃 − ) with 𝑃 + , 𝑃 − ∈ {1, 2, … 8}.
Since the approximated model will eventually be discretized before being incorporated
in the MPC formulation, we already use a discretized form of the dynamics 𝑥̇ in (6.13) to
(6.15) for approximation as
𝑥(𝑘 + 1) = Δ𝑥(𝑘) + 𝑥(𝑘).
Here, Δ𝑥(𝑘) is approximated instead of 𝑥(𝑘 + 1) for two reasons: first, the assumptions
and the approximation procedure remains valid by switching from 𝑥̇ to Δ𝑥, and second,
 80 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

in cases such as 𝑣𝑥 where the state values are of a significantly larger order of magnitude
compared to their rates of change, approximating Δ𝑥 leads to a more numerically-stable
representation of the error.
We solved the optimization problem (6.6) for every fixed pair of (𝑃 + , 𝑃 − ) by Mat-
lab’s nonlinear least squares optimizer, lsqnonlin, using the trust-region-reflective
algorithm. This optimizer further exploits the structure of the nonlinear problem by ap-
proximating the Gauss-Newton direction through minimizing the 2-norm of the function
deviation in the next step. The problem is then solved for 1000 initial random guesses
to provide sufficient accuracy without excessive computational effort, among which we
select the lowest objective value as the optimal solution. The codes for grid generation and
hybrid approximations are available from our published hybridization toolbox [87].
Fig. 6.6 shows the training validation errors of the optimal solutions for Δ𝑣𝑥 , Δ𝑣𝑦 , and
Δ𝑟 on model approximation validation grids in Table 6.3. The lateral dynamics of the
nonlinear model has a higher degree of nonlinearity, which explains the different error
scales in the MMPS approximation. The plots are grouped based on the system and the
type of the training grid to gain a better insight into the behavior of each grid and its effect
on the accuracy of the approximation.
Firstly, it is observed that U and R grids overfit for lower numbers of hyperplanes
compared to their trajectory-based counterparts, which is represented by high oscillations
6 after a certain degree of complexity in the approximation form. The S grid shows the
lowest oscillatory behavior in validation results, which can indicate the inability of this
grid in converging to an accurate fit due to its grid-point distribution with higher density
in regions that are attainable from a steady-state solution of the system dynamics.
For Δ𝑣𝑥 , U and R grids show overfitting behavior for 𝑃 + + 𝑃 − ⩾ 4 modes and T grid
overfits for 𝑃 + + 𝑃 − ⩾ 5. However, the S grid does not show overfitting until 13 modes with
a lower validation error (≈ 0.4%) compared to the other grid types (≈ 0.8%). It is worth
noting that the trajectory-based validation grids start overfitting for a much larger number
of modes compared to the domain-based types.
For Δ𝑣𝑦 , U and R grids again overfit at 4 modes, with 3% and 2% validation errors,
respectively. The S grid overfits at 12 to 14 modes with reaching a validation error that is
slightly above 1%, and the T grid overfits at 11 modes with an error of 2%.
For Δ𝑟, U grid overfits at 4 modes and its validation error remains above 42%. On the
other hand, the R, S, and T grids reach their best fits at 12 to 15 modes, all with an error
of about 9%. The S grid, while having the lowest training error in most cases, has the
highest offset between the validation and the training error. This could be due to the S
grid needing more points to provide a more realistic training error. However, it should be
noted that the steady-state-initiated method’s ability to generate new “distinct” points is
limited; as Table 6.3 shows, to generate a validation grid three-times as large as the training
one, the number of simulations needed to be multiplied by 6, which is not the case for its
randomly-initiated counterpart, T. As the set of points attainable by a random input signal
from a steady-state solution is limited, this difference is understandable. Nevertheless, this
limitation is not restricting the S grid’s ability to fit the model significantly (compared to
e.g., the U grid).
6.5 Model and Constraint Hybridization for Vehicle Control 81

(a) Δ𝑣𝑥 (b) Δ𝑣𝑦

(c) Δ𝑟 (d) Plot legend

Figure 6.6: Cross-validation of the MMPS approximations for different dynamics using four grid types. Since all
the plots share the same legend, it is placed separately.

6.5.4 Constraint Approximation and Validation

For constraint approximation, both training and validation steps are done on the two
constraint approximation C grids defined in Table 6.3. The nonlinear constraints are
approximated by either an intersection of second-order cones, which corresponds to the
implementation of the convex envelope method from [139], or a union of convex subregions,
which gives a non-convex approximation of the feasible region. Based on the formulation
of the approximation problem, i.e., (6.2) or (6.3), the approach is region- or boundary-
based. The shape of the subregions is also either ellipsoidal or polytopic, where the latter
is developed by an MMPS formulation of the nonlinear constraint. This leads to four
methods of constraint approximation as shown in Table 6.5 where the best fits and their
corresponding parameters as well as their approximation errors are presented. It should be
noted that the region-based ellipsoidal approximation is a modified implementation of the
non-parametric ellipsoidal learning method [134].
Similar to model approximation, we solved the boundary-based optimization problems
 82 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

Table 6.4: Best validation fits for different grid types

Grid Δ𝑣𝑥 Δ𝑣𝑦 Δ𝑟

type (𝑃 + , 𝑃 − ) Error∗ (𝑃 + , 𝑃 − ) Error∗ (𝑃 + , 𝑃 − ) Error∗
U (2,3) 0.8% (2,2) 4.3% (2,2) 42.8%
R (2,2) 0.7% (2,2) 3.0% (7,8) 9.2%
S (3,7) 0.3% (6,8) 1.8% (6,6) 9.8%
T (2,3) 0.5% (6,3) 2.6% (7,8) 8.8%

∗ Relative validation error on the C grid

(6.3) for every fixed pair of (𝑃 + , 𝑃 − ) or 𝑛e by Matlab’s nonlinear least squares optimizer,
lsqnonlin for 1000 initial guesses (selected in a similar way as for the model approxima-
tion). However, the region-based approach results in a non-smooth optimization problem
(6.2) which we solved using the particle swarm optimizer in Matlab, which does not
require the problem to be differentiable. The swarm size was selected to be 10 times larger
than the number of decision variables as a sufficiently large number for our experiments,
and the problem was solved 1000 times for each case of (𝑃 + , 𝑃 − ) or 𝑛e and the best solution
was kept as the optimal one. In addition, the convex envelope approach from [139] where
6 the boundary of the nonlinear constraints is approximated by an intersection of 𝑛c second-
order cone constraints is also implemented in the same fashion for different values of 𝑛c .
Figure 6.7 shows the training and validation errors for different constraint approximation
methods.
The convex envelope approach approximates the feasible region by a convex area that
is the intersection of 𝑛c second-order cone constraints. Therefore, for systems where the
concavity measure, i.e., the difference between the feasible region and its convex hull, is
significant compared to its size, this method converges to either high violation or inclusion
misclassification errors, which is visible in the behavior of the training and validation plots
in Fig. 6.7a. Starting from one second-order cone constraint to approximate the feasible
region with, this approach converges to an area covering about 25% of the feasible and 25%
of the infeasible regions. Increasing the number of cone constraints to more than 3 leads
to a significant improvement in the obtained fit. Nevertheless, the best convex envelope
fit is obtained at 𝑛c = 6 with the inclusion and violation errors of 45% and 5% respectively,
both of which are not acceptable as a proper fit. This shows that the method is converging
to more accurate approximations of the largest convex subset of the feasible region, which
is covering about 50% of it.
The difference between the region- and boundary-based approaches is due the fact that
in the region-based approximation (6.2), the inclusion and violation misclassification errors
are penalized, while in the boundary-based approximation (6.3), the error in approximation
of the distance to the boundary is minimized. This difference is more clear in the MMPS
approximation plots where with one binary variable, the boundary is approximated by
an affine function, i.e., a hyperplane. Problem (6.3) then converges to a hyperplane with
the lowest sum of distances from the nonlinear boundary. However, since the violation
error is penalized more than the inclusion error with 𝛾c < 0.5, problem (6.2) converges
6.5 Model and Constraint Hybridization for Vehicle Control 83

to an empty set where the violation error is zero and the inclusion error is 1, giving the
optimal misclassification error of 1 − 𝛾c . In all the cases, it is observed that the region-based
approximation converges to lower violation and higher inclusion errors due to the same
reason.
MMPS approximation of the constraints via the region-based approach shows overfit-
ting behavior after considering 6 binary variables. After 3 binary variables, the fits start
oscillating between a more “inclusive” approximation and a more “violating” one. However,
the best fit is obtained with 7 binary variables. Even by increasing this number, problem
(6.2) keeps converging to the same misclassification error.
Boundary-based MMPS approximation reaches the best fit with 8 binary variables
where again, adding more binary variables and increasing the complexity level of the fit
does not change the inclusion and violation errors significantly and only minor oscillations
between converging to a slightly more inclusive approximation or to a slightly more
violating one are observed.
Ellipsoidal approximation of the feasible region generally converges to fits with lower
accuracy compared to the MMPS approximation. In the region-based approximation, the
training and validation errors stay at the same level with slight oscillations after 𝑛e = 7
with inclusion and violation misclassification errors of respectively 26.7% and 0.6%. In this
sense, for the same number of integer variables, the ellipsoidal region-based approximation
converges to a similar violation error but a 50% higher inclusion error. The boundary-based
ellipsoidal approximation on the other hand shows a different overfitting behavior where 6
increasing the number of ellipsoidal subregions results in convergence to a better coverage
at the expense of a significant increase in violation error. Therefore, the best fit should
be selected before the point where the violation error exceeds a user-defined accepted
threshold. Here we select 𝑛e = 5 since it is the last complexity before the violation error
exceeds 6%. Another observed pattern is the divergence of violation errors in training
and validation, which mirrors the nature of the approximation approach: increasing the
number of ellipsoids translates into generating more ellipsoidal subregions close to the
boundary to minimize the distance-to-boundary sum. However, in the validation phase this
leads to significantly higher violation errors as a result of the approximation overfitting to
the training grid.

Table 6.5: Best constraint approximation fits

Error
Subregions Approach Fit Parameters
Inclusion Violation
Intersection of convex subregions [139]
Cone Boundary 𝑛𝑐 = 6 45.0% 5.0%
Union of convex subregions
MMPS Region (𝑃 + , 𝑃 − ) = (5, 2) 17.5% 0.5%
MMPS Boundary (𝑃 + , 𝑃 − ) = (4, 4) 9.9% 3.5%
Ellipsoidal Region [134] 𝑛e = 7 26.7% 0.6%
Ellipsoidal Boundary 𝑛e = 5 24.0% 6.0%
 84 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

(a) Intersection of convex subregions

6
(b) Union of convex subregions

Figure 6.7: Training and validation plots for different constraint approximation problems. As the axes share the
same legend, it is only presented in the first one.

6.6 Conclusions and Outlook

This chapter has presented a hybridization framework for approximation nonlinear model
and constraints. This framework serves as benchmark for formulating nonlinear MPC opti-
mization problems using a hybrid systems formalism to improve computational efficiency
and to ensure real-time implementation. The conclusions of the research in this chapter
with respect to its contributions, and the hybridization framework are summarized in the
following subsections. The hybrid control comparison benchmark is discussed in detail in
the next chapter.

6.6.1 Conclusions for Vehicle Control

Introduction of the hybridization framework in this chapter is a result of the following steps
where the model and constraint approximation problems were defined by means of several
novel descriptions of the approximation problem. First, for the model approximation, the
Kripfganz MMPS form was used to approximate the nonlinear system to a user-defined error
bound. Second, the nonlinear feasible region resulting from the physics-based constraints
was approximated by a union of ellipsoids and polytopes via region- and boundary-based
formulation of the approximation problem. Third, the model and constraint approximation
6.6 Conclusions and Outlook 85

problems were solved numerically across various grids types sampled from the input/state
domain and their corresponding fit qualities in terms of accuracy and overfitting behavior
were compared. Fourth, among the different grid types, two novel trajectory-based grid
generation methods were introduced to structurally increase the density of the grid points
in regions of the state domain with higher likelihood of the attainability by the system
dynamics. This approach resulted in 15-60% reduction of the approximation error compared
to its domain-based counterpart. Finally, the different grid generation and formulations
of the approximation problems were analyzed to present a hybridization benchmark for
improving the computational performance of the MPC problem for other applications
of nonlinear MPC, as well as tracking control in emergency evasive maneuvers; this
comparative assessment is explained in the next chapter.

6.6.2 Generalized Hybridization Framework

Our proposed hybridization framework can be implemented in other applications of non-
linear MPC to improve computational efficiency by considering the following guidelines:

1. The model approximation problem should be solved by either an R, S, or T grid.

The density of the R-type grid points can vary by sampling using various random
distributions. Additionally, if there is a significant variance in the likelihood of
attainability for different input/state pairs, it is recommended to use the trajectory-
based S or T grids. Depending on the nature of the system dynamics, the S grid
is a proper choice if the attainable subset of the state-domain from steady-state
6
solutions is rich or large enough to ensure coverage of the whole domain by selecting
a sufficiently large number of sampling points over each trajectory. On the other
hand, this will not be an issue for the T grid, at the expense of including input/state
pairs that are only attainable from an unattainable initial state. In general, if such
properties of the system dynamics are not fully known, it is suggested to consider
all three grid types and compare the overfitting behavior as done in this chapter.

2. The Kripfganz MMPS form is a compact and well-formulated way to impose continu-
ity in the hybrid approximation of the nonlinear problem; it provides straightforward
and intuitive control over the accuracy of the approximation with respect to the
number of introduced binary variables that are assigned to each affine local dynamics
appearing in the max operators. The number of affine terms can be increased up until
the point where either the maximum number of binary variables or the maximum
tolerated approximation error are reached. Both of these stopping criteria can be
chosen by the user and based on the application.

3. The nonlinear non-convex feasible region can be approximated by a union of el-

lipsoids or polytopes using region-, as well as boundary-based formulations of the
approximation problem. If the application requires to strictly avoid violating the
nonlinear constraints by the approximated ones, it is recommended to use the region-
based formulation of the approximation problem. However, the boundary-based
formulation leaves more room to balance the trade-off between covering the non-
linear region and violating it, and converges to better coverage of the non-convex
region. This trade-off can also be managed within the region-based formulation by
 86 6 Design and Numerical Analysis of Hybridization Benchmark for NMPC

adjusting the tuning parameter 𝛾c , but its capability in modifying the priority of the
costs of inclusion vs. violation error with respect to the distance from the boundary
is limited.

Using the above guidelines, the hybridization approach can be implemented in different
applications such as motion planning, navigation, or real-time control of systems with
fast dynamics where it is required to balance the computational speed and accuracy of the
MPC problem.

6.6.3 Next Steps and Future Work

In the next part of this chapter, we present the hybrid control comparison benchmark
using this hybridization framework for balancing the computational efficiency of the MPC
optimization problem in vehicle control during emergency evasive maneuvers.
The next steps of the current research can proceed along (but not limited to) the
following lines: investigation of the proposed hybridization framework in applications with
higher dimensions e.g., large-scale control problems, extension of the model approximation
step by incorporating other hybrid modeling frameworks such as piecewise-quadratic
or mixed-logical-dynamical systems as compact models for a good trade-off between
constraint satisfaction, computational complexity, and control performance.

6
 87

7
Comparative Assessment of
Hybridization in Vehicle
Control
We may call it the paradox of the decider: as the circulation of information becomes faster and
more complex, the time available for the elaboration of relevant information becomes shorter.
The more space taken by the available information, the less time there is for understanding
and conscious choice.

— Franco Berardi, After the Future

7

Optimization-based approaches such as nonlinear Model Predictive Control (MPC) are promis-
ing approaches in safety-critical applications with nonlinear dynamics and uncertain envi-
ronments such as automated driving systems. However, the computational complexity of the
nonlinear MPC optimization problem coupled with the need for rapid response in emergency
scenarios is the main bottleneck in realization of automation levels four and five for driving
systems. In this chapter, we construct hybrid formulations of the nonlinear MPC problem for
vehicle control during emergency evasive maneuvers and assess their computational efficiency
in terms of accuracy and solution time. To hybridize the MPC problem, we combine three
hybrid approximations of the prediction model and four approximations of the nonlinear
stability and tire saturation constraints and simulate the closed-loop behavior of the resulting
controllers during five emergency maneuvers for different prediction horizons. Further, we
compare the robustness of the controllers and their accuracy-time trade-off when the friction
of the road is either unknown or has an offset error with respect to the prediction model. This
robustness is investigated for different levels of friction uncertainty and with respect to the
proximity to the vehicle handling limits. Our tests show that the hybridization of the MPC

This chapter has been submitted to IEEE Transactions on Intelligent Vehicles [86].
 88 7 Comparative Assessment of Hybridization in Vehicle Control

problem results in an efficient implementation of MPC for emergency evasive maneuvers,

paving the way for implementation at high levels of automation.

7.1 Introduction
Real-time implementation of nonlinear MPC for high-speed safety-critical evasive ma-
neuvers is an open research problem [12]. Two specific reasons contribute to this: high
computation times for solving a NonLinear Program (NLP) compared to a linear or a
Quadratic Program (QP), and possible convergence to local optima, which is highly sensi-
tive to the initial guess provided to the NLP solver.
Proactive vehicle control in emergency scenarios requires using the full control potential
of the system, which means that some sub-optimal solution techniques for the NLP [11, 140,
141] are not suitable to incorporate [142]. To mitigate the slow convergence of NLP solvers,
an upper bound is often imposed for the computation time as stopping criterion; this bound
can be selected e.g., as a function of the complexity of the problem using prediction horizon,
decision variable, etc. If the solver does not converge to a optimum before hitting this
bound, the solution to the previous step is shifted and used [8]. Nevertheless, if this occurs
repeatedly and the controller does not converge to a solution for consecutive steps, this
may result in a large degree of suboptimality or even infeasibility.
A popular approach for selecting the initial guess is using a warm-start strategy based
on shifting the previous solution to tailor it for the current MPC optimization problem [8,
84, 143], which is suitable provided that the previous step converged to a good solution.
This however is a restrictive condition, for which [144] proposed using a tangential solution
predictor instead of shifting, which is essentially based on using the concept of parametric
sensitivity of the NLP for constructing new initial guesses. Nevertheless, warm start is a
7 suitable strategy only if the solver converged to a “good” solution in the previous step [93].
Other strategies to improve the initial guess include using the reference trajectory [93],
using the inverse static model of the system [100], or selecting the solution to a simpler
approximation of the NLP e.g., a QP [101]. Nevertheless, the mentioned approaches are
not sufficient for real-time control during emergency evasive maneuvers where a more
extensive search in the decision space is required.
During emergency maneuvers, relying on one solution is restrictive: even with the
improvements on the search direction and transformation, the search for the optimum
would be limited within a neighborhood of the solution for the previous time-step. However,
abrupt changes to the reference trajectory e.g., due to sudden appearance of an obstacle
on the road, require a more extended exploration of the search space to increase the
likelihood of finding an acceptable optimum. In this sense, [145] uses a divide-and-conquer
strategy in searching for starting regions based on the current state and then picks the first
solution that satisfies an acceptable bound on the objective. While this method improves
convergence to better optima, it still does not expand the search region in case of abrupt
changes in the reference. In [10], multiple filtered random initial guesses are used to solve
the NLP problem and in [146], the NLP is solved offline and a dataset of “good” initial
guesses to be used in real time is learned, which could be an improvement upon relying on
one solution without wasting additional computational effort on initial guesses with lower
improvement value. However, this approach is only applicable in case there is sufficient
and reliable data to learn such guesses, which is usually not currently available for vehicle
7.1 Introduction 89

control in emergency scenarios.

While multi-start solution of the NLP improves the chances of converging to a suitable
optimum to use the full control potential of the vehicle, it significantly increases the com-
putation time, which is the main obstacle toward real-time implementation of MPC during
emergency scenarios. In this sense, hybridization of the nonlinear control optimization
problem was proposed [81, 147] to balance the computational efficiency via the trade-off
between accuracy and the convergence speed by using a hybrid systems formalism [18] to
express the prediction model and the nonlinear constraints.
Hybrid MPC for vehicle tracking control has attracted attention as a potential solution
to tackle the problem of computational efficiency [121, 124, 148–150]. Nevertheless, to
the best of our knowledge, the capability of the hybridization approach in improving the
computational efficiency of MPC has neither been assessed for highly-nonlinear prediction
models, nor investigated during hazardous scenarios and aggressive evasive maneuvers.
In this sense, such scenarios are particularly important since they require using the full
control potential of the vehicle in its handling limits and the need for fast computation is
critical in collision avoidance.
This work is the squeal of Chapter 6 where we proposed an approach to approximate the
prediction model and nonlinear physics-based constraints using a hybrid system formalism.
The contributions of the current chapter are:

1. improving the computational efficiency of MPC, by exploiting the Max-Min-Plus-

Scaling (MMPS) formulation to obtain a hybrid representation of the MPC optimiza-
tion problem,

2. definition of a comparison benchmark, involving highly-nonlinear vehicle models

and non-convex constraints 7
3. evaluation of several hybridization methods in the context of MPC by assessing
three hybrid approximations of the prediction model and four approximations of the
nonlinear stability and tire saturation constraints, and

4. showcasing the impact and efficacy of hybridization in terms of control performance

and computation speed.

In this chapter, we use the approximated prediction model and constraints to formulate
and to solve the MPC problem as either a Mixed-Integer Linear Program (MILP) or a Mixed-
Integer Quadratically-Constrained Program (MIQCP). We then investigate the trade-off
between the accuracy and the computation speed of the resulting hybrid MPC controllers
against their nonlinear counterparts. The computational performance of the hybrid and
nonlinear controllers are assessed during five aggressive evasive maneuvers, representing
abrupt changes in the reference trajectory due to a hazardous situation such as a sudden
appearance of an obstacle on the road. Further, we investigate the tracking errors in the
presence of uncertainty in the friction coefficient as an offset as well as a disturbance such
as a significant decrease of friction due to the presence of water on a section of the road.
This chapter is organized as follows: the theoretical background such as the formulation
of the nonlinear and hybrid MPC problems is explained in Section 7.2. To make this chapter
self-contained, we recall the hybridization approach in Chapter 6 and its corresponding
 90 7 Comparative Assessment of Hybridization in Vehicle Control

notation, to which the reader is referred. Section 7.3 explains different aspects of the
comparison benchmark and assessment criteria e.g., the choice of driving scenarios and
the prediction horizons. The results of the simulations and the comparative assessment
are discussed in Section 7.4 followed by high-fidelity simulations in IPG CarMaker in
Section 7.5. Finally, Section 7.7 presents the main results and draws an outlook for future
research.

7.2 Background
7.2.1 Model and Constraint Hybridization
Consider a nonlinear discrete-time system

𝑥(𝑘 + 1) = 𝐹 (𝑥(𝑘), 𝑢(𝑘)) ,

where 𝑥 ∈ ℝ𝑛 and 𝑢 ∈ ℝ𝑚 represent the state and input vectors, respectively. We approximate
𝑇
each component 𝐹𝑠 of 𝐹 = [𝐹1 … 𝐹𝑛 ] separately by an MMPS function 𝑓𝑠 with the
Kripfganz form [91] as

𝑓𝑠 (𝑥, 𝑢) = max (𝜙+𝑠 (𝑥, 𝑢)) − max (𝜙−𝑠 (𝑥, 𝑢)) , ∀𝑠 ∈ {1, … , 𝑛}, (7.1)
𝜂 𝜂
where the vectors 𝜙𝑠 ∶ ℝ𝑚+𝑛 → ℝ𝑃 with 𝜂 ∈ {+, −} are affine functions of 𝑥 and 𝑢, also
referred to as dynamic modes, and expressed via matrices
𝜂 ×𝑚 𝜂 ×𝑛 𝜂
𝐴𝜂𝑠 ∈ ℝ𝑃 , 𝐵𝑠𝜂 ∈ ℝ𝑃 , 𝐻𝑠𝜂 ∈ ℝ𝑃 , ∀𝜂 ∈ {+, −}, ∀𝑠 ∈ {1, … , 𝑛},

as
7 𝜙𝜂𝑠 (𝑥, 𝑢) = 𝐴𝜂𝑠 𝑥 + 𝐵𝑠𝜂 𝑢 + 𝐻𝑠𝜂 .
The general form of 𝑓 is then given as

𝑓 (𝑥, 𝑢) = Ψ+ (𝑥, 𝑢) − Ψ− (𝑥, 𝑢),

where Ψ+ and Ψ− are vector-valued functions with

Ψ𝜂𝑠 (𝑥, 𝑢) = max (𝜙𝜂𝑠 (𝑥, 𝑢)) , ∀𝜂 ∈ {+, −}, ∀𝑠 ∈ {1, … , 𝑛}.
𝜂
Note that in this notation, the max operator returns the largest component in the vector 𝜙𝑠 .
Remark 7.1. In Chapter 6, we used a scalar representation of the MMPS formulation to ap-
proximate each component of 𝐹 separately. In this chapter, we use vector-valued representation
to make the formulation of the MPC optimization problem more compact. However, without
loss of generality, we consider 𝐺 to be a scalar function.
For bounded 𝑥 and 𝑢, the physics-based constraints are in general nonlinear and non-
convex and expressed via the normalized boundary function 𝐺 as

 B {(𝑥, 𝑢) ∈ ℝ𝑚+𝑛 | 0 ⩽ 𝐺(𝑥, 𝑢) ⩽ 1},

where  is referred to as the feasible region. It should be noted that we normalize the
constraint function to the interval [0, 1] to avoid numerical issues in the subsequent control
7.2 Background 91

optimization problems. The region  is approximated either by a union of ellipsoids or

by using the MMPS formalism, which corresponds to approximating the  by a union
of convex polytopes. In the MMPS approach, a similar formulation to the MMPS model
approximation problem is used: we approximate 𝐺 by an MMPS function 𝑔MMPS of the
Kripfganz form (7.1). The resulting feasible region MMPS is then expressed as

MMPS B {(𝑥, 𝑢) ∈ ℝ𝑚+𝑛 | 0 ⩽ 𝑔MMPS (𝑥, 𝑢) ⩽ 1}. (7.2)

via the boundary function

𝑔MMPS (𝑥, 𝑢) = max (𝛾 + (𝑥, 𝑢)) − max (𝛾 − (𝑥, 𝑢)) , (7.3)

𝜂
where 𝛾 𝜂 ∶ ℝ𝑚+𝑛 → ℝ𝑅 are affine functions of 𝑥 and 𝑢 as

𝛾 𝜂 (𝑥, 𝑢) = 𝐶 𝜂 𝑥 + 𝐷𝜂 𝑢 + 𝐼 𝜂 ,

and
𝜂 ×𝑚 𝜂 ×𝑛 𝜂
𝐶 𝜂 ∈ ℝ𝑅 , 𝐷𝜂 ∈ ℝ𝑅 , 𝐼 𝜂 ∈ ℝ𝑅 , ∀𝜂 ∈ {+, −}.
The second way is to approximate the feasible region by a union of 𝑛e ellipsoids as

ELLP B {(𝑥, 𝑢) ∈ ℝ𝑚+𝑛 | 0 ⩽ 𝑔ELLP (𝑥, 𝑢) ⩽ 1}, (7.4)

whose boundary can be expressed by

𝑔ELLP (𝑥, 𝑢) = min (𝜔(𝑥, 𝑢)) , (7.5)

where the min operator gives the smallest component in the vector 𝜔, and where

𝑥 − 𝑥0,𝑒
𝑇
𝑥 − 𝑥0,𝑒
7
𝜔𝑒 (𝑥, 𝑢) = 𝑄 − 1, ∀𝑒 ∈ {1, … , 𝑛e }, (7.6)
(𝑢 − 𝑢0,𝑒 ) 𝑒 (𝑢 − 𝑢0,𝑒 )

with 𝑄𝑒 being a positive definite matrix and (𝑥0,𝑒 , 𝑢0,𝑒 ) representing the center coordinates
of the ellipsoid. Note that this representation includes rotated ellipsoids as well.

7.2.2 MPC Optimization Problems

The state and input vectors over the whole prediction horizon 𝑁p are defined as
𝑇
̃ + 1) = [𝑥̂ 𝑇 (𝑘 + 1|𝑘)
𝑥(𝑘 𝑥̂ 𝑇 (𝑘 + 2|𝑘) … 𝑥̂ 𝑇 (𝑘 + 𝑁p |𝑘)] ,
𝑇
𝑢(𝑘)
̃ = [𝑢𝑇 (𝑘) 𝑢𝑇 (𝑘 + 1) … 𝑢𝑇 (𝑘 + 𝑁p − 1)] ,
where 𝑥(𝑘
̂ + 𝑖|𝑘) represents the predicted state of the (𝑘 + 𝑖)-th time step based on the
state measurement at the 𝑘-th time step. For the sake of simplicity, the control horizon is
assumed to be equal to the prediction horizon 𝑁p . In addition and for brevity of expressions,
we introduce the generalized form of the systems dynamics 𝐹 and inequality constraints 𝐺
over the prediction horizon as

̃ + 1) = 𝐹̃ (𝑥(𝑘),
[𝑥(𝑘 + 1) = 𝐹 (𝑥(𝑘), 𝑢(𝑘))] ⟺ [𝑥(𝑘 ̃ 𝑢(𝑘))
̃ ],
[0 ⩽ 𝐺 (𝑥(𝑘), 𝑢(𝑘)) ⩽ 1] ⟺ [0 ⩽ 𝐺̃ (𝑥(𝑘),
̃ 𝑢(𝑘))
̃ ⩽ 1] .
 92 7 Comparative Assessment of Hybridization in Vehicle Control

Note that 𝐹̃ is the generalized counterpart of 𝐹 by extending the notation over the prediction
horizon and not by recursive substitution. For the sake of brevity, 𝑥(𝑘) is not an explicit
argument of 𝐹̃ but note that the dependence of 𝐹̃ on 𝑥(𝑘) is implied within the 𝑥(𝑘) ̃
argument.
Using the 𝓁1 -norm in defining the objective function in tracking 𝑥̃ref , MPC requires
solving the optimization problem

min ‖Θ𝑥 (𝑥(𝑘)

̃ − 𝑥̃ref (𝑘)) ‖1 + ‖Θ𝑢 𝑢(𝑘)‖
̃ 1, (7.7a)
𝑢(𝑘)
̃

s.t. ̃ + 1) = 𝐹̃ (𝑥(𝑘),
𝑥(𝑘 ̃ 𝑢(𝑘))
̃ , (7.7b)
0 ⩽ 𝐺̃ (𝑥(𝑘),
̃ 𝑢(𝑘))
̃ ⩽ 1, (7.7c)

with Θ𝑥 ⩾ 0 and Θ𝑢 ⩾ 0 being normalizing diagonal matrices with non-negative entries for
the state tracking error and input signals, respectively. Note that the 𝓁1 -norm is selected to
allow a mixed-integer linear description of the objective function.
The hybrid MPC problem can then be formulated as:

min Θ𝑥 𝑒̃𝑥 (𝑘) + Θ𝑢 𝑒̃𝑢 (𝑘), (7.8a)

𝑢(𝑘)
̃
s.t. − 𝑒̃𝑥 (𝑘) ⩽ 𝑥(𝑘)
̃ − 𝑥̃ref (𝑘) ⩽ 𝑒̃𝑥 (𝑘), (7.8b)
− 𝑒̃𝑢 (𝑘) ⩽ 𝑢(𝑘)
̃ ⩽ 𝑒̃𝑢 (𝑘), (7.8c)
𝑥(𝑘 +
̃ + 1) = vec (Ψ (𝑘)) − vec (Ψ (𝑘)) , −
(7.8d)
Ψ+𝑖𝑗 (𝑘) = max (𝜙+𝑖 (𝑘 + 𝑗 − 1)) , ∀𝑖 ∈ {1, … , 𝑛}, ∀𝑗 ∈ {1, … , 𝑁p }, (7.8e)
Ψ−𝑖𝑗 (𝑘) = max (𝜙−𝑖 (𝑘 + 𝑗 − 1)) , ∀𝑖 ∈ {1, … , 𝑛}, ∀𝑗 ∈ {1, … , 𝑁p }, (7.8f)
7
where (7.8b)–(7.8c) are introduced to obtain a linear representation of the objective function
by defining

𝑒̃𝑥 (𝑘) = ‖𝑥(𝑘)

̃ − 𝑥̃ref (𝑘)‖1 , 𝑒̃𝑢 (𝑘) = ‖𝑢(𝑘)‖
̃ 1,

and (7.8d)–(7.8e) are the hybridized model approximation to replace (7.7b). The vec(⋅)
operator in (7.8d) converts its matrix argument into a vector by stacking its components
into one column vector. Then, constraint approximation can be hybridized by replacing
(7.7c) by the MMPS constraints (7.9a) for an MILP or the ellipsoidal constraints (7.9b) for
an MIQCP formulation:
𝜂
Λ𝑗 = max (𝛾 𝜂 (𝑘 + 𝑗 − 1)) , ∀𝜂 ∈ {+, −}, ∀𝑗 ∈ {1, … , 𝑁p }, (7.9a)
Ω𝑗 (𝑘) = min (𝜔(𝑘 + 𝑗 − 1)) , ∀𝑗 ∈ {1, … , 𝑁p }. (7.9b)

Remark 7.2. The binary variables of the optimization problem are introduced via activating
the local modes for the hybrid model and constraints (for more details, see Chapter 6). Therefore,
the corresponding MILP problem will have
𝑛
𝑁p 𝑅+ + 𝑅− + ∑(𝑃𝑠+ + 𝑃𝑠− )
( 𝑠=1 )
7.3 Comparison Benchmark 93

binary variables with 𝑅+ and 𝑅− being the constraint-approximation counterparts of 𝑃 + and

𝑃 − , and the MIQCP problem will have

𝑛
𝑁p 𝑛e + ∑(𝑃𝑠+ + 𝑃𝑠− )
( 𝑠=1 )

binary variables.

7.3 Comparison Benchmark

7.3.1 Prediction Model and Physics-Based Constraints
The nonlinear prediction model is a single-track vehicle with a Dugoff tire [151] model
with varying friction as described in Section 6.4, with system variables and parameters
given in Tables 6.1 and 6.2.
In Chapter 6, we hybridized the nonlinear model using different grid–generation
methods via the MMPS formalism (7.1) and obtained three hybrid approximations for the
nonlinear model labeled by their corresponding grid types as R, S, and T models. The
nonlinear physics-based constraints due to the tire force saturation and vehicle stability
were hybridized as well via approximating the feasible region by a union of ellipsoids
and by a union of polytopes (using the MMPS formalism) via boundary-based and region-
based approximations. There, we obtained four approximations labeled by their approach
(R and B) and by the shape of the subregions (MP for MMPS or EL for ellipsoidal) as
RMP, BMP, REL, and BEL. Table 7.1 summarizes the abbreviations used in this chapter for
different hybrid models, constraints, and their corresponding controllers. For more details
of the boundary-based and region-based approximations and their errors, or the number 7
of introduced binary variables by each approach, the reader is referred to Chapter 6.

Table 7.1: Abbreviations for hybrid models and controllers

Hybrid Models
Approximation Grid Type Abbreviation
Domain-based random R
Trajectory-based steady-state initiated S
Trajectory-based randomly initiated T
Hybrid Constraints
Formulation
Ellipsoidal (EL) MMPS (MP)
Region-based (R) REL RMP
Boundary-based (B) BEL BMP
Hybrid MPC Controllers
[Model abbreviation] – [Constraint abbreviation]
Example: R model + BMP constraint → R–BMP controller
 94 7 Comparative Assessment of Hybridization in Vehicle Control

7.3.2 Control Parameters

Given the application, the control sampling time 𝑡sc is often restricted by the capabilities of
the control hardware such as the maximum operation frequency. In addition, the model
sampling time 𝑡sm is either known a priori for a discrete-time system, or obtained for
a continuous-time system with respect to its natural frequency and dynamic behavior.
Therefore, we assume these two parameters to be fixed during all the simulations as known
system parameters 𝑡sm = 0.01s and 𝑡sc = 0.05s.
In the path tracking MPC literature [12], the time span of the prediction often covers
below 1.5s ahead, for control sampling times shorter than 0.1s. Based on our selected
control sampling time 𝑡sc = 0.05s, we therefore test different time spans of the prediction in
the range 0.25s to 1.50s, corresponding to 𝑁p ∈ {5, 10, … , 30}. This is further explained in
Section 7.3.5.

7.3.3 MPC Controllers

In this benchmark, we consider two nonlinear MPC controllers with the nonlinear predic-
tion model and the physics-based constraints. The first one solves the NLP using the warm
start only (i.e. with the shifted solution of the previous time step) labeled as NL–1, and the
second one referred to as NL–5 solves the problem for five different initial guesses and
selects the best solution it has found. The initial guesses for NL–5 are as follows: one warm
start as in NL–1, one random point within the domain, one point at the lower bound, one
at the upper bound, and one in the geometric center. The solution for each control time
step is fed to the system via a receding-horizon strategy: while the control optimization
problem finds the optimal input signal during the next 𝑁p steps, we only use the solution
for the first step.
7 The computational performance of the two NLP controllers is compared against MPC
controllers based on combinations of three hybrid models (R, S, and T) with four constraint
approximations: two polytopic (RMP and BMP) and two ellipsoidal (REL and BEL). In total
we have six MILP and six MIQCP controllers to compare against their NLP counterparts.
The hybrid MPC controllers are labeled by combining the abbreviations for their model
and constraints, separated by a dash (–) symbol, as described in Table 7.1. Further, MILP
and MIQCP controllers can be constructed using our published hybridization toolbox [87].

7.3.4 Reference Trajectory

We compare the computational performance of the nonlinear and hybrid MPC controllers
during five maneuvers of two seconds as reference trajectories. These maneuvers are
selected to represent aggressive evasive maneuvers with different longitudinal velocities
as explained in Table 7.2. Note that for a more intuitive representation of distance to
the handling limits and/or tire saturation, we indicate the normalized distance of the
reference trajectory in the g–g diagram (6.16) and Kamm circles (6.17) in their respective
column where 1 represents the boundary of the circles, i.e. the constraints. The column
𝛽-𝑟 represents the normalized location of the stability envelope often used in the literature
as |𝑟| ⩽ 𝜇𝑔𝑣𝑥 under the assumption of steady-state cornering conditions as a typical yaw
rate constraint [12] and |𝛽| ⩽ 5deg [8]. Note that the steady-state cornering conditions do
not hold in extreme maneuvers. Nevertheless, we provide the trajectory plots on the 𝛽-𝑟
envelope to gain more insight in the cornering behavior of the vehicle during the reference
7.3 Comparison Benchmark 95

10
5 0.5

𝑎𝑦 (m/s2 )

𝑟 (rad/s)
0 0
−5
−0.5
−10
−10 −5 0 5 10 −0.06 0 0.06
𝑎𝑥 (m/s2 ) 𝛽 (rad)
Man. 1 Front Man. 1 Rear
5
Man. 2 Front Man. 2 Rear
𝐹𝑦 (kN)

Man. 3 Front Man. 3 Rear

0
Man. 4 Front Man. 4 Rear
Man. 5 Front Man. 5 Rear
−5

−5 0 5
𝐹𝑥 (kN)

Figure 7.1: Selected maneuvers for the benchmark, represented in terms of the constraints. The green zone in the
g-g diagram represents the safe region and the red one corresponds to the aggressive yet acceptable acceleration
range.

maneuvers. Moreover, the normalized values represent the distance to the boundary with
values between 0 and 1, where 1 indicates the position of the boundary itself. Figure 7.1
shows the schematic view of the five reference maneuvers in terms of these constraints.
The two-seconds simulation time is selected to represent the recovery window for the 7
controller in hazardous scenarios in case of an abrupt change in the reference trajectory.

Table 7.2: Selected maneuvers as reference trajectories for the benchmark. The 𝑣𝑥 column represents the average
longitudinal velocity in km/h.

No. Maneuver 𝐯𝐱 g–g 𝛽-𝐫 Kamm

1 Safe lane change 130 0–0.5 0–0.4 0–0.7
2 Aggressive lane change 128 0–0.6 0–0.9 0–0.9
3 Drift cornering 73 0.2–0.8 0–0.9 0–1.0
4 High-speed cornering 154 0.3–0.7 0.3–1.1 0.4–0.9
5 Low-speed cornering 75 0.2–0.8 0.2–0.8 0–1.0

7.3.5 Driving Scenarios

We compare the computational performance of the nonlinear and hybrid MPC controllers
in the following four driving scenarios:

• Ideal Case: The nonlinear prediction model is selected as the real system. The com-
putational performance of the hybrid controllers is evaluated over 𝑁p ∈ {5, 10, … , 30}.
 96 7 Comparative Assessment of Hybridization in Vehicle Control

• Friction Offset: We use a different tire-road friction in the real system as 𝜇 ∈

{0.70, 0.75, … , 1.00} and compare the computational performance of the controllers
for a selected 𝑁p value over this range of friction offsets.

• Friction Disturbance: We assume the road friction for the second quarter of the
maneuver to be very low, representing a disruption such as a slippery road surface
and we compare the computational efficiencies in the same fashion as the friction
offset case.

• Handling Limits: We investigate the computational performance of the controller

for a fixed friction offset during the lane change maneuver (no. 1 in Table 7.2). We
vary the input steering angle during the maneuver to simulate different levels of
aggressive steering and assess the closed-loop performance in terms of the closeness
to handling limits, i.e., the boundary of the g–g diagram with the acceleration
magnitude between 0.5𝑔 up to the 𝜇𝑔 limit.

Note that 𝛿 is the steering angle on the road and not the steering wheel angle, and since
the scenarios are high speed where compared to an urban scenario, a smaller steering
angle can lead to a higher lateral acceleration. For instance, the steering angle bound is
approximately 10 degrees in [8] and 20 degrees in [152] or the extreme maneuvers in [114].
For drifting, to the best of our knowledge, the maximum bound is 38 degrees in [153] where
the velocity is 54 km/h, while we investigate 75 up to 154 km/h maneuvers. Therefore, we
believe 30 degrees is a suitable upper bound for the steering angle for extreme maneuvers
in highway scenarios.
7
7.3.6 Solver Selection
For a fair comparison in terms of computation time, we select the most efficient known
solvers within the academic community for the MILP/MIQCP and NLP problems.
The MILP/MIQCP and NLP problems are solved by GUROBI [154] and TOMLAB/KNI-
TRO [155] optimizers, respectively, using Matlab as interface and overall computation
environment. To further improve the solution time for the NLP problems, we provided the
objective and constraint functions via MEX files (instead of m-files in Matlab), which in
our experiments reduced the computation time for the NLP problems by around 50% for
all the cases.
The simulations were all run on a PC with a 4-core(s) Intel Xeon 3.60 GHz CPU and 8
GB RAM on Windows 10 64-bit and in a Matlab R2020b environment.

7.4 Simulation Results

Using the benchmark described in Section 7.3, we compare the computational performance
of the hybrid and nonlinear MPC controllers as follows: we first compare these controllers
in the ideal case for different 𝑁p values and then we select the most promising hybrid
controllers and compare their robustness to friction uncertainty and their performance
close to the handling limits in the next subsections.
7.4 Simulation Results 97

7.4.1 Ideal Case

The computational performance of the nonlinear and hybrid MPC controllers is shown in
Figures 7.2 and 7.3 in terms of their average and maximum tracking errors and computation
time per control sampling time.

Aggressive Lane Change Maneuver

First, we start by comparing the hybrid and nonlinear MPC controllers during the aggressive
lane change maneuver in Fig. 7.1.
In this maneuver, all MILP controllers, as well as NL–5, perform equally well with
different 𝑁p values. However, NL–1 shows oscillatory tracking accuracy across the 𝑁p axis.
This happens because extending the prediction horizon, even with the same prediction
model as the ground truth, makes the search for an optimal solution more difficult, which
reduces the chance of reaching an “acceptable” solution when only using the warm-start
strategy, especially when performing the aggressive lane change maneuver as shown in
Fig. 7.1.
Extending the prediction horizon not only enlarges the optimization problem’s search
space but also causes the prediction model error to accumulate, which also contributes
to higher computation times as the reference trajectory may become difficult for the
prediction model to track. Therefore, the suitable prediction horizon for tracking, in terms
of acceptable accuracy for lower computation times, is 10 or 15.
In terms of computation speed, the MILP controllers with the T model show a steady
increase rate similar to the nonlinear ideal models while other hybrid controllers show an
increase in the rate of computation time after a certain 𝑁p value, which stems from the
prediction model accuracy. Longer prediction horizons do not only increase the dimensions
of the search space in the optimization problem, but also lead to accumulation of the 7
prediction model error. This accumulation increases the error as well as computation time
as the reference trajectory can become infeasible to track for the prediction model, leading
to slower convergence. Therefore, the T model yields the best accuracy compared to the R
and S models.
Controllers with the T model in general show the lowest increase of computation time
when compared to the other hybrid controllers. In the MIQCP controllers T–REL and
T–BEL the rate of increase is higher for 𝑁p between 5 and 10, but it converges to the same
rate as the NLP controllers and their MILP counterparts T–RMP and T–BMP. This stems
from the limitations of using the ellipsoidal constraints compared to the polytopic ones:
the ellipsoidal approximations of the feasible region have a lower coverage of the feasible
region (for more details, see Table V in Chapter 6), which limits using the full control
potential for an aggressive maneuver in shorter prediction horizons.
As Fig. 7.2 shows, the MILP controllers exceed the performance of the MIQCP ones in
terms of accuracy, as well as computation speed. Therefore, for the next simulations we
only consider the MILP controllers as prospective suitable hybrid candidates.

Cornering Maneuvers
We compare the MILP and NLP controllers during the three cornering maneuvers as shown
in Fig. 7.3. In maneuvers 3 and 5 where the input forces vary drastically over the maneuver
(see Fig. 7.1), the NL–1 controller shows a poor computational performance and oscillatory
 98 7 Comparative Assessment of Hybridization in Vehicle Control

Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

15 50 103
10 NL–1 NL–5
25
5 R–RMP R–BMP
10 101 S–RMP S–BMP
2
5 T–RMP T–BMP
2 10−1
10 20 30 10 20 30 10 20 30
𝑁p 𝑁p 𝑁p

(a) MILP vs. NLP controllers

Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

15 50 103
10 NL–1 NL–5
25
5 R–REL R–BEL
10 101 S–REL S–BEL
2
5 T–REL T–BEL
2 10−1
10 20 30 10 20 30 10 20 30
𝑁p 𝑁p 𝑁p

(b) MIQCP vs. NLP controllers

Figure 7.2: Computational performance of the nonlinear and hybrid MPC controllers during the aggressive lane
change maneuver (maneuver 2 in Table 7.2) in the ideal scenario.

behavior in the error plots across the 𝑁p axis, which was also observed in the aggressive
lane change maneuver and discussed there.
In all three cornering maneuvers, the controller with the T model yields the best
computational performance with its mean tracking error below 4.5% and maximum error
7 below 10% in all cases.
Just as for the lane change maneuvers, increasing 𝑁p leads to a higher computation time
for all the controllers; however, the rate of increase is the lowest for the nonlinear MPC and
the T–BMP controller. For the T–RMP controller the same behavior is observed for 𝑁p < 20.
Comparing the performance of the hybrid and nonlinear MPC controllers in all the five
maneuvers, the suitable prediction horizon for tracking, in terms of acceptable accuracy
for lower computation times, is 10 or 15. Next, we select 𝑁p = 10 for the comparison of
the robustness of the controllers to friction uncertainty. However, we also have simulated
other 𝑁p values, reaching similar results. Therefore, for a compact presentation, we present
the trends and analyze them for 𝑁p = 10. In addition, since the MILP controllers with the S
prediction model show larger tracking errors and larger computation times, especially for
shorter prediction horizons, we disregard them at this stage and compare the four MILP
controllers with R and T prediction models against their NLP counterparts.
Remark 7.3. We have simulated the scenarios for other 𝑁p values and reached similar results.
Therefore, for a compact presentation, we present the trends and analyze them for 𝑁p = 10.

7.4.2 Friction Offset

The maneuvers 2, 3, and 5 in Table 7.2 are the three most critical ones: here, the vehicle
operates close to the tire saturation and stability limits almost the whole time as shown
in Fig. 7.1. Thus, we used these maneuvers to study the effect of friction offset. The
7.4 Simulation Results
Mean Rel. Err. (%) 99

Max Rel. Err. (%)

Comp. Time (s)

15 50 103
10 NL–1 NL–5
25
5 R–RMP R–BMP
10 101
2 S–RMP S–BMP
5 T–RMP T–BMP
10−1
10 20 30 10 20 30 10 20 30
𝑁p 𝑁p 𝑁p

(a) Drift cornering maneuver (maneuver 3 in Table 7.2)

Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

15 50 103
10 NL–1 NL–5
25
5 R–RMP R–BMP
2 10 101 S–RMP S–BMP
5 T–RMP T–BMP
10−1
10 20 30 10 20 30 10 20 30
𝑁p 𝑁p 𝑁p

(b) High-speed cornering maneuver (maneuver 4 in Table 7.2)

Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

15 50 103
10 NL–1 NL–5
25
5 R–RMP R–BMP
10 101 S–RMP S–BMP
2 5 T–RMP T–BMP
2 10−1
10 20 30 10 20 30 10 20 30
𝑁p 𝑁p 𝑁p

(c) Low-speed cornering maneuver (maneuver 5 in Table 7.2)

Figure 7.3: Computational performance of the nonlinear and MILP MPC controllers during three cornering 7
maneuvers (maneuvers 3, 4, and 5 in Table 7.2) in the ideal scenario.

prediction horizon is selected as 𝑁p = 10 and the simulations are run for different road
friction coefficients in the range 𝜇 ∈ {0.70, 0.75, … , 1.00} to account for uncertain friction in
the prediction model. Figure 7.4 shows the computational performance of the nonlinear
and hybrid MPC controllers during the three reference maneuvers.
While the computation time for the hybrid controllers does not vary by increasing the
friction uncertainty, the nonlinear controllers show an increase in the computation time in
maneuvers 3 and 5 where a significant fraction of the maneuver is performed close to the
tire-saturation and vehicle-handling limits, which are functions of the friction coefficient.
The maximum and mean tracking errors increase for lower friction coefficients for all
the controllers. However, the rate of error increase for the nonlinear controllers is higher.
The difference between the tracking errors of NL–1 and NL–5 once again indicates the
shortcoming of a warm-start strategy during aggressive maneuvers in searching for the
optimal solution in the search space. This however comes at the price of an increase in
computation times, best shown in Fig. 7.4c where solving the NLP for five initial guesses
increases the computation time tenfold.
Remark 7.4. The reason behind the computational increase in NLP is as follows: compared
 100 7 Comparative Assessment of Hybridization in Vehicle Control

to the shifted solution to the previous step, the other initial guesses are generally further away
from a local optimum. As a result, the increase in computation time is more than linear.

In the presence of friction offset, the tracking error of NL–5 converges to that of the
MILP controllers in maneuvers 2, 3, and 5 where a more extensive search over the search
space is required to perform the maneuver from an initial state with an error from the
previous solution. To understand this phenomenon and its two contributing factors, we
look at the NL–5 and T–BMP controllers in more detail.
Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

50 101 NL–1 NL–5
10 25
5 R–RMP R–BMP
10 100 T–RMP T–BMP
2 5
10−1
2
0.7 0.8 0.9 1 0.7 0.8 0.9 1 0.7 0.8 0.9 1
𝜇 𝜇 𝜇

(a) Aggressive lane change maneuver (maneuver 2 in Table 7.2)

Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

50 101 NL–1 NL–5

10 R–RMP R–BMP
25
5 100
10 T–RMP T–BMP
2
5 10−1
0.7 0.8 0.9 1 0.7 0.8 0.9 1 0.7 0.8 0.9 1
𝜇 𝜇 𝜇

(b) Drift cornering maneuver (maneuver 3 in Table 7.2)

7
Mean Rel. Err. (%)

Max Rel. Err. (%)

Comp. Time (s)

50 101 NL–1 NL–5

10 25 R–RMP R–BMP
5
10 100 T–RMP T–BMP
2 5
10−1
0.7 0.8 0.9 1 0.7 0.8 0.9 1 0.7 0.8 0.9 1
𝜇 𝜇 𝜇

(c) Low-speed cornering maneuver (maneuver 5 in Table 7.2)

Figure 7.4: Computational performance of the nonlinear and hybrid MPC controllers during three reference
maneuvers (maneuvers 3, 4, and 5 in Table 7.2) in the friction offset scenario.

Notice that NL–5 and T–BMP have mean tracking errors below 5% in all the maneuvers
of Fig. 7.4 with 𝜇 = 1, which generates the same friction as their prediction models. When
𝜇 on the road is decreased to 0.7, the controllers still seek to find the solution (including
tire forces) close to the boundary of the feasible region of their model, which assumes 𝜇 ≈ 1.
However, these forces cannot be generated by the tire in the real system due to the lower
friction on the road. Therefore, the first contributing factor to the error is the fact that
error accumulation grows exponentially with the number of control time steps and as a
result the controller converges to an infeasible solution for the real system (note that the
real feasible region is shrinking with the friction reduction). Secondly, with larger errors,
7.4 Simulation Results 101

finding a feasible solution to track the reference trajectory from an initial state with an
already large tracking error might not be possible after a certain error bound. This not
only increases the convergence time for the NL–5 controller, but also results in converging
to even worse solutions both in terms of constraint violation and optimality to the point
where we observe the tracking error of NL–5 exceeds the error for T–BMP in Figures 7.4b
and 7.4c as with a similar order of model error, the branch-and-bound approach of the
MILP solver, as opposed to the NLP solver, guarantees convergence to the global optimum
if given enough time, while keeping the computation time as low as for the ideal case.

7.4.3 Friction Disturbance

In this case, we assume a correct knowledge of the road friction during the maneuver,
while exposing the system to a sudden friction reduction to 𝜇 = 0.4 during the second
quarter of the maneuver and we evaluate the tracking error, computation times, and the
recovering ability of the closed-loop system without any estimation or corrections during
the disturbance.
The tracking errors of the controllers are compared in case of sudden reduction of the
friction to 𝜇 = 0.4 in the second quarter of the maneuver to represent a case similar to
pouring water on the road surface. The tracking errors at each time step are shown in
Fig. 7.5 to compare the errors, as well as the ability of the different controllers in recovering
from the friction disturbance. The average computation times per control step for each
controller are presented in a separate plot in Fig. 7.5f.
During the safe lane change maneuver, all controllers recover to a tracking error below
5% after five time steps, while the tracking error is larger during the other maneuvers that
are more aggressive. After the friction disturbance in maneuvers 2 and 5, the tracking
error keeps on increasing as the controllers fail to recover to an acceptable error bound. 7
This could be understood by taking into account that the second quarter of maneuver 2 is
when tracking the reference trajectory requires tire forces that will no longer be feasible
due to the fact that the radius of the Kamm circle for the rear tire has decreased to 40%
of its original value, which means that following the planned trajectory will no longer be
feasible for the prediction model. The same issue arises during maneuver 5 that can be
tracked provided that the rear tire is generating forces close to its saturation limit during
the whole maneuver (Fig. 7.1).
The hybrid controllers show smaller tracking errors after recovery during the low-
speed cornering maneuver while starting with larger tracking errors before the friction
disturbance. This reduction is understandable in light of the fact that the second half of
this maneuver requires the vehicle to operate further from the stability boundaries in 𝛽-𝑟
envelope and the g-g diagram in Fig. 7.1. Meanwhile, the tracking error of NL–1 stays above
5%, which shows the limitations of depending on a warm-start strategy in convergence
to better optima. This limitation is even more clear in Fig. 7.5c for the drift cornering
maneuver: while NL–5 recovers better after the disturbance, it still fails to reach smaller
tracking errors as fast as the hybrid MPC controllers, which suggests that increasing
the number of initial guesses by considering more than five points could improve the
convergence capability of NL–5 to acceptable optima. However, it should be noted that
even for five initial guesses, NL–5 requires 10 times more than the slowest hybrid MPC to
converge to its best solution.
 102 7 Comparative Assessment of Hybridization in Vehicle Control

Rel. Err. (%)

Rel. Err. (%)

100 𝜇 = 0.4 100
25 NL–1 25 𝜇 = 0.4 NL–1
5 NL–5 5 NL–5
1 R–RMP 1 R–RMP
0 0.5 1 1.5 2 R–BMP 0 0.5 1 1.5 2 R–BMP
T–RMP T–RMP
Time (s) Time (s)
T–BMP T–BMP

(a) Safe lane change maneuver (maneuver 1 in Table 7.2) (b) Aggressive lane change maneuver (maneuver 2 in Table 7.2)
Rel. Err. (%)

Rel. Err. (%)

100 100
25 NL–1 25 𝜇 = 0.4 NL–1
5 NL–5 5 NL–5
1 𝜇 = 0.4 R–RMP 1 R–RMP
0 0.5 1 1.5 2 R–BMP 0 0.5 1 1.5 2 R–BMP
T–RMP T–RMP
Time (s) Time (s)
T–BMP T–BMP

(c) Drift cornering maneuver (maneuver 3 in Table 7.2) (d) High-speed cornering maneuver (maneuver 4 in Table 7.2)

Mean Comp. Time (s)

101 NL–1
Rel. Err. (%)

100
25 NL–1
100 NL–5
5 NL–5
R–RMP 10−1 R–RMP
1 𝜇 = 0.4
M1 M2 M3 M4 M5 R–BMP
0 0.5 1 1.5 2 R–BMP
T–RMP Reference Maneuver T–RMP
Time (s)
T–BMP T–BMP

(e) Low-speed cornering maneuver (maneuver 5 in Table 7.2) (f) Computation times

Figure 7.5: Tracking error of the nonlinear and four selected MILP MPC controllers during five reference maneuvers
7 in (a) to (e) in case of friction disturbance. The average computation time for each control time step is shown in
(f).

In the safe lane change maneuver, T–RMP and T–BMP show larger tracking errors
compared to the other controllers, and this is the scenario where the hybrid controllers in
general show the highest error increase of 12% to 30% for 95% computation time reduction.
However, the effectiveness of the hybrid MPC controllers in terms of tracking error is
more observable in more aggressive maneuvers where T–RMP and T–BMP show a better
performance, in some cases even better than that of NL–1 and NL–5, due to the fact that the
shortcomings of convergence to local optima is more clear in hazardous scenarios where
there are sudden changes in the environment that require a more thorough search across
the decision space. Comparing the control performance vs. computation time trade-off
during the four aggressive reference maneuvers shows that choosing the hybrid MPC
controllers R–BMP and R–RMP decreases the computation time to 2% to 5%, while it
increases the maximum error from 5% to 15% in maneuver 5 while decreasing it during
maneuvers 2, 3, and 4 compared to NL–5.

7.4.4 Analysis of Computational Performance

Figure 7.6 plots the range of tracking error and computation times for the NLP and four
MILP MPC controllers during all the five reference maneuvers in both friction offset and
7.4 Simulation Results 103

friction disturbance scenarios as shaded boxes. The average points are shown by square
markings in the shaded area. Comparing the computational performance in Fig. 7.6 shows
the power of the hybrid MPC controllers compared to the nonlinear ones in the presence
of uncertainty. While the NLP controllers have a lower optimum, their maximum tracking
errors reach much higher values while taking more time to converge. In terms of the
average points, not only the best MILP controller in the ideal case (T–BMP) has a lower
maximum error in both friction uncertainty cases compared to the best NLP one (NL–5),
but it also has a higher computational efficiency: in the friction offset case, it trades off an
error increase from 8.7% to 9.7% for reducing the computational time from 10.2s to 0.3s,
and in the disturbance case it reaches a smaller tracking error (from 9.5% to 8.7%) as well
as a lower computation time (from 8.3s to 0.4s).

50 NL–1 50 NL–1
Rel. Err. (%)

20 NL–5 Rel. Err. (%) 20 NL–5

10 R–RMP 10 R–RMP
5 R–BMP
5 R–BMP
2 T–RMP 2 T–RMP
0.1 1 10 T–BMP 0.1 1 10 T–BMP

Comp. Time (s) Comp. Time (s)

(a) Friction offset case (b) Friction disturbance case

Figure 7.6: Relative errors and computation times for the nonlinear and four MILP MPC controllers during the
five reference maneuvers in cases with friction uncertainty.

7.4.5 Handling Limits

7
For a more clear understanding of the computational performance during hazardous
scenarios, we perform a second test in the friction offset case. Here, we fix the road friction
to 𝜇 = 0.85 and compare the computational performance of the controllers during the safe
lane change maneuver over a range of steering actions to get closer to the boundary of
the g–g diagram in Fig. 7.1. Figure 7.7 shows the tracking errors and computation time for
MILP and NLP√ controllers for different levels of aggressiveness in terms of the acceleration
magnitude 𝑎2𝑥 + 𝑎2𝑦 , which is bounded by 𝜇𝑔.
The computational efficiency of hybrid MPC can be easily seen in Fig. 7.7. The closer
the maneuver gets to the boundary of the stability constraint in the g–g diagram, the higher
the tracking errors get for all the controllers. However, the error increase for the NLP
controllers is much higher as they fail to converge to an acceptable optimum with 1 or
even 5 starting points. Therefore, when the acceleration magnitude exceeds 6.2 m/s2 , even
the NL–5 controller reaches higher mean and maximum tracking errors compared to the
T–BMP, R–BMP, and R–RMP controllers. This is happening while NL–5 requires about 20
times more time to converge to its final solution.
Comparing among the MILP controllers, T–BMP can be the best hybrid MPC controller
as it shows the most steady computation time and the lowest increase of tracking error as
the steering action gets more extreme in our simulations.
 104 7 Comparative Assessment of Hybridization in Vehicle Control

Mean Rel. Err. (%)

Max Rel. Err. (%)

25 50
10 20
5
5
2 2

𝜇𝑔
𝜇𝑔
5.33 6.19 7.14 8.21 5.33 6.19 7.14 8.21
√ √
𝑎𝑥 + 𝑎𝑦 (m/s )
2 2 2 𝑎𝑥 + 𝑎𝑦 (m/s )
2 2 2
Comp. Time (s)

10 NL–1 R–RMP
T–RMP NL–5
1 R–BMP T–BMP

𝜇𝑔
0.1
5.33 6.19 7.14 8.21
√
𝑎𝑥 + 𝑎𝑦 (m/s )
2 2 2

Figure 7.7: Computational performance of the nonlinear and hybrid MPC controllers during lane change maneuver
in the friction offset scenario for different levels of aggressive steering in the g–g diagram.

7.5 IPG CarMaker Simulation

38
Reference
𝑣𝑥 (m/s)

36
7 34
R–RMP
R–BMP
32 T–RMP
T–BMP
0 2 4
𝑡 (s)
0.2 10
R–RMP R–RMP
𝑎𝑦 (m/s2 )

0.1 5
R–BMP
𝛿 (rad)

R–BMP
0 T–RMP
0 T–RMP
T–BMP −5 T–BMP
−0.1 −10
0 2 4 −10−5 0 5 10
𝑡 (s) 𝑎𝑥 (m/s2 )
0.5 2 R–RMP
𝑟 (rad/s)

𝑣𝑦 (m/s)

0 R–BMP
0 T–RMP
−2
T–BMP
−0.5 −4
−0.1 0 0.1 −0.15 0 0.15
𝛽 (rad) 𝛿 (rad)

Figure 7.8: Computational performance of the nonlinear and hybrid MPC controllers during lane change maneuver
in the friction offset scenario for different levels of aggressive steering in the g–g diagram.
7.6 Discussions and Outlook 105

To validate the performance of the MILP controllers, we have simulated the most aggressive
double-lane change maneuver in handling-limit scenario using a high-fidelity BMW model
in IPG CarMaker with the RealTime tire model. Figure 7.8 shows the state and input plots,
as well as the tire force and constraint bounds from the simulations. Despite the fact that
the MILP controllers use a much simpler prediction model and friction offset, comparing
the states shows that they are able to provide control inputs that satisfy the constraint
boundaries in the handling limits as shown on the g–g diagram.

7.6 Discussions and Outlook

7.6.1 MILP vs. MIQCP
In general, as MILPs are solved faster than MIQCPs, MILP-based approaches are more
suitable choices in terms of computation speed. The control performance highly depends
on the accuracy of the hybrid approximation. Consequently, the tracking performance
of MIQCP controllers was not as good as MILP controllers due to their less accurate
constraint approximation compared to their MILP counterparts. Nevertheless, for systems
or applications where a lower approximation error with mixed-integer quadratic constraints
is obtained for the nonlinear constraints, MIQCP can be an efficient formulation to hybridize
the nonlinear MPC problem. This can for instance be the case for systems with quadratic
or bi-variate nonlinearities where considering the quadratic terms helps avoid using large
number of local linear or affine modes to obtain the same level of accuracy.

7.6.2 Robustness to Friction Uncertainty

7
When friction uncertainty is present, MILP controllers do not require longer computation
times, whereas NLP controllers take significantly longer to converge. In addition, the
rate of increase in tracking error for the NLP controllers is higher than that of the MILP
controllers in the presence of a friction offset.
Furthermore, MILP controllers are better able to recover from friction disturbances
due to their more comprehensive search for an optimal solution in the decision space and
the fact that they always reach the global optimum if given enough time. This means
that if the error is already very high in the initial state for the current time step, the NLP
controllers may not find a feasible solution at all in the vicinity of their initial guesses,
while the MILP controller will converge to one via branch-and-bound strategy, provided
that they have given sufficient time. As a result, even in cases where both nonlinear and
hybrid controllers recover from high tracking errors during the friction disturbance, the
MILP controller reaches smaller tracking error in fewer control time steps.
The robustness of the MILP controllers compared to the NLP controllers is summarized
in Fig. 7.6. While NLP controllers can reach lower relative errors, their behavior in terms
of accuracy, as well as computation speed, is not as consistent as MILP controllers in the
presence of friction offset or disturbance. However, MILP controllers are considerably (at
least 10 times) quicker to converge to their optimal solution and show significantly less
variations in the relative error when exposed to friction uncertainty.
 106 7 Comparative Assessment of Hybridization in Vehicle Control

7.6.3 Performance Close to Handling Limits

Getting closer to the handling limits leads to larger tracking errors for all the controllers.
However, NLP controllers may deviate from the reference significantly as they may fail to
converge to an acceptable optimum. Meanwhile, the MILP controllers converge to lower
errors in a shorter time, e.g., in the boundary of the acceleration magnitude, the best MILP
controller converges to 30% of the tracking error of the best NLP controller with 95%
reduction of its corresponding computation time. This shows that MILP controllers are
more suitable choices for real-time control in emergency evasive maneuvers.

7.6.4 Overall Computational Performance

The shortcoming of a warm-start strategy in solving an NLP is more clear in emergency
maneuvers, which stems from its limitation in searching for the optimal solution in the
search space during aggressive maneuvers. In addition, even with a multi-start strategy,
the NLP controller may converge to worse solutions if the uncertainty influences the
feasible region and the NLP solutions become infeasible for the real system. However,
the branch-and-bound approach of the MILP solver has a better exploration of possible
solutions than a potentially real-time implementable NLP solver such as gradient-based
solver, while keeping the computation time as low as for the ideal case.
Even in the ideal case where the NLP controllers benefit from employing the same
prediction model as the real system, they show high variations in tracking control and
computation time, which means they are not suitable options for robust control in hazardous
scenarios. In the presence of uncertainty, the NLP controllers suffer from larger tracking
errors as well as an exponential growth in their computation time. However, the MILP
controllers converge to smaller tracking errors, within a much smaller variation bound,
7 and with significantly less computation times.

7.7 Conclusions
This chapter has presented a comparative assessment of nonlinear MPC controllers vs. their
various hybridized counterparts in terms of computational efficiency for vehicle control
during emergency evasive maneuvers. The hybridization of the nonlinear problem was
presented and discussed in Chapter 6, where several guidelines for hybridization are given
in a generalized framework.
The benchmark of this chapter uses three hybridized models and four hybridized
constraint formulations for a nonlinear single track vehicle model considering nonlinear
physics-based constraints for stability and tire-force saturation. Five reference maneuvers
were selected to represent emergency situations where the computational efficiency is
crucial for real-time proactive vehicle control. The hybrid and nonlinear controllers then
were compared in multiple scenarios to compare their control performance and computation
time, and their robustness in the presence of friction uncertainty in the form of an offset or
a disturbance. Further, we studied the tracking behavior of the controllers with respect to
how close the vehicle is operating in handling limits. The conclusions of our comparative
assessment are summarized next with respect to different criteria.
Based on our comparative assessment, we propose combining hybrid MPC and hybrid
predictive estimation techniques (e.g., moving-horizon estimation) as a potential next re-
7.7 Conclusions 107

search step for improving robustness in hazardous driving scenarios. Moreover, as quadratic
forms of nonlinearity are extensively encountered in modeling of physical systems, we
propose investigating piecewise-quadratic-based hybridization of the prediction model and
physics-based constraints for MIQCP formulation of MPC optimization problem. This can
particularly be beneficial for systems with nonlinearities that can better be approximated
using quadratic approximations and can lead to significant improvements in terms of
accuracy and computational efficiency of the hybrid controller. In addition, we suggest
extending the prediction model to include the effects of wheel-speed/torque dynamics
for improved control performance in hazardous scenarios. Finally, we suggest hardware
implementations of the MILP tracking controllers as a topic for future research.

7
 109

8
Efficient Response to Sudden
Appearance of Static
Obstacles
When they invented the car they invented the collision and the darkness of what time leads
the willing body into.

— David Wojnarowicz, Close to the Knives: A Memoir of Disintegration

The sudden appearance of a static obstacle on the road, i.e. the moose test, is a well known
emergency scenario in collision avoidance for automated driving. Model Predictive Control 8
(MPC) has long been employed for planning and control of automated vehicles. However,
real-time implementation of automated collision avoidance in emergency scenarios such as
the moose test is still an open issue due to the high computational demand of MPC to perform
an evasive maneuver in such hazardous scenarios. This chapter brings new insights into
real-time collision avoidance via experimental implementation of MPC for motion planning
after sudden and unexpected appearance of a static obstacle. As state-of-the-art nonlinear
MPC shows limited capability to provide an acceptable solution in real time, we propose a
human-like feedforward planner to help in cases where the MPC optimization problem is either
infeasible or unable to yield a suitable solution due to the poor quality of the initial guess. We
introduce the concept of maximum steering maneuver to design the feedforward planner and
to mimic a human-like reaction after the unforeseen detection of the static obstacle on the road.
Real-life experiments are conducted in a variety of scenarios with different speeds and level of
emergency using an FPEV2-Kanon electric vehicle. Moreover, we demonstrate the effectiveness
of our planning strategy via comparison with a state-of-the-art MPC motion planner.

This chapter has been submitted to IEEE Transactions on Intelligent Transportation Systems.
 110 8 Efficient Response to Sudden Appearance of Static Obstacles

8.1 Introduction
Motion planning in automated driving has been extensively researched during the past
years. Avoiding a collision in hazardous scenarios is particularly challenging on the
operational and stability levels [156], i.e. planning a safe trajectory for the ego vehicle and
tracking it during an emergency scenario.
Among the various testing scenarios [157], one example of an emergency situation is
the sudden appearance of a static obstacle on the road, which can also be considered as the
extreme case of leading vehicle deceleration in [158]. Figure 8.1 shows a schematic view of
influential elements for the planning problem in such scenarios. Given the considerable
relative velocity involved, the time-to-collision criterion [159] in these circumstances
proves insufficient for maintaining a safe distance through braking alone. Hence, executing
a safe evasive maneuver may require performing an extreme steering action near the
boundaries of the vehicle stability constraint, also referred to as driving at handling limits.
In the case of a static obstacle, two critical parameters are the distance to the object –
often reflected in time-to-collision or distance-to-collision threat measures in the litera-
ture [159] – and the width of the obstacle. For instance, shorter distance to the obstacle
with a larger width can both contribute to the criticality of the situation as the width
determines the necessary lateral displacement for collision avoidance. Understanding the
importance of a wider width is closely linked to the current states of the vehicle as the
required braking or steering actions for achieving a specific lateral displacement depend
on factors like the current vehicle speed or sideslip angle.

object width loc

a lly-plan
ned trajectory

object distance globally-planned trajectory

Figure 8.1: Elements of the collision avoidance problem after detecting a static obstacle
8
MPC has become increasingly popular in the field of automated collision avoidance
thanks to its straightforward handling of constraints and its capacity to dynamically adjust
to environmental changes by solving the control optimization problem in a receding horizon
manner [12]. In the current state of the art, trajectory planning and vehicle control are
commonly addressed through one of two architectural frameworks: hierarchical [160–166]
or integrated [9, 167–171].
A hierarchical architecture offers greater flexibility in defining control problems and
enables faster responses for real-time implementation, owing to the differing frequency
and performance requirements at each level. However, the reference trajectory provided
by the planner may not be attainable for the real plant. This is a critical issue in emergency
cases where the feasible area for collision avoidance is limited and the vehicle is operating
at its handling limits. Integrated planning and control circumvents this issue by treating
both problems within a single optimization problem.
While integrated MPC design, in particular when used with Electronic Power Steering
(EPS), allows the handling of two optimization problems in combination, it is essential to
highlight a key distinction between the two architectures: in a hierarchical architecture, ad-
8.1 Introduction 111

dressing the planning and control optimization problems can occur at different frequencies
(e.g. planning at 5-10Hz and control at 50-100Hz). Conversely, an integrated architecture
demands solving the integrated optimization problem at the control frequency, albeit with
a higher computational demand compared to the control optimization problem.
An important source of computational complexity stems from the fact that the inte-
grated MPC optimization problem is nonlinear, which requires employing computationally-
demanding NonLinear Program (NLP) algorithms to find the – locally – optimal solution.
In [169], Gaussian safe envelopes used in the integrated MPC problem are obtained via
Gaussian processes regression; this formulation allows for efficient solution of the inte-
grated MPC problem by Quadratic Programming (QP). Other examples of QP solution of
the MPC optimization problem include [170] where constraining the decision space to the
linear tire force range leads to a quadratic formulation of the problem, and [172] where
the weights of the simplified quadratic problem are adapted online for improved control
performance.
In [9], two models are serially cascaded to handle the two problems simultaneously,
hence facilitating real-time solution of the optimization problem by NLP and a warm-
start strategy. However, this strategy is limited in finding the optimal solution if a static
obstacle suddenly appears on the road. Moreover, physics-based and local convexification
approaches [165, 173] or explicit sub-optimal solution [167] have helped the real-time
realization of integrated planning and control for normal driving. However, achieving
real-time solutions in emergency scenarios remains a primary bottleneck of the current
automated driving research.
Another technique in this area is incorporating parametric methods for trajectory
planning into an integrated control optimization framework, which offers significant ad-
vantages by combining the strengths of both approaches: while the integrated architecture
prevents the generation of unattainable trajectories for the control layer, the parametric
formulation streamlines the optimization process significantly, facilitating fast convergence
to a local optimum, which is essential for real-time planning and control. Parametric tra-
jectories during a lane change maneuver are typically represented by polynomial curves at 8
lower speeds [165, 166, 168, 170] or sigmoid [174] and tangent-hyperbolic functions [172]
that replicate human behavior at higher speeds [175]. For instance, in the case of a sudden
appearance of a static obstacle, the finite-state machine in [174] decides on simultane-
ous braking and steering to perform a lane change. The reference trajectory for such a
maneuver is defined as a parametric sigmoid function and is optimized in real time via
NLP.
Despite the fast-paced progress of the literature, there remains a gap in achieving real-
time implementation of automated collision-avoidance in real-life emergency situations.
This is primarily due to the limited computational capacity of real systems [12] and the low
acceptance of the automated driving systems due to lack of interpretability [176].
In the context of avoiding collision after sudden appearance of a static obstacle, as well
as limited memory and computation time, converging to – even a local – optimum via NLP
is often not possible, which means that in practice the best feasible solution found before
hitting a stopping criteria is used. As a result, the ‘quality’ of such a solution is sensitive to
the provided initial guess. More specifically, the popular warm-start strategy that is often
used in the literature, would be of limited value in such scenarios since the shifted solution
 112 8 Efficient Response to Sudden Appearance of Static Obstacles

of the previous time step after detecting the obstacle is often far from a feasible solution to
avoid the detected obstacle. Moreover, the limited computational resources restrict our
ability to explore the decision space using other starting points. As a result, having a “good
initial guess" is the practical way to improve the quality of the obtained solution by NLP.
In this case, a good initial guess can be defined as an interpretable candidate solution, i.e.
inspired by the normal reaction of a skilled driver to the sudden appearance of an obstacle
ahead.
In this chapter, we tackle the problem of real-time collision avoidance with limited
computational resources after the sudden appearance of a static obstacle. Therefore, we
aim to provide experimental insights into the real-time implementation of MPC-based
collision avoidance by confronting the computational limitation challenge head-on in
real-life scenarios using an electric vehicle, and improving the computational efficiency
of the problem by integration of a physics-based and human-like feedforward planner, to
help convergence to a feasible solution in emergencies. We investigate multiple real-life
emergency scenarios and analyze the effectiveness of our proposed approach in test cases
that render previous solutions infeasible, emphasizing the need for our planning strategy.
Hence, our approach integrates MPC more intelligently and provides effective solutions in
the context of sudden appearance of a static obstacle.
The rest of this chapter is structured as follows: Section 8.2 describes the vehicle model.
followed by an overview of the proposed control system design in Section 8.3. Section 8.4
then covers our proposed planning strategy. Details on system implementation are given in
Section 8.5 and the results of the experimental tests are provided and analyzed in Section 8.6.
Finally, Section 8.7 concludes this chapter.

8.2 Vehicle Model

The single track vehicle model is shown in Fig. 8.2. First, we cover the kinematics of the
8 vehicle and then we apply Newton’s second law to derive the governing equations of
motion. Table 8.1 shows the variables and parameters used in this chapter.

Figure 8.2: Single-track vehicle model

8.2 Vehicle Model 113

Table 8.1: List of variables and parameters

Variable Definition Units

𝑥ego Global longitudinal position of the CoG m
𝑦ego Global lateral position of the CoG m
𝑥obs Global longitudinal position of the obstacle m
𝑦obs Global lateral position of the obstacle m
𝑣𝑋 Global longitudinal velocity of the CoG m/s
𝑣𝑌 Global lateral velocity of the CoG m/s
𝑣𝑥 Local longitudinal velocity of the CoG m/s
𝑣𝑦 Local lateral velocity of the CoG m/s
𝜃 Yaw angle rad
𝛽 Sideslip angle rad
𝛾 Yaw rate rad/s
𝛿 Steering angle rad
𝛼f Front slip angle rad
𝛼r Rear slip angle rad
𝜅f Front slip ratio –
𝜅r Rear slip ratio –
𝐹𝑥f Longitudinal force of the front tire N
𝐹𝑥r Longitudinal force of the rear tire N
𝐹𝑦f Lateral force of the front tire N
𝐹𝑦r Lateral force of the rear tire N
𝐹𝑧f Normal force on the front tire N
𝐹𝑧r Normal force on the rear tire N
Parameter Definition Value
𝑡s Planning sampling time 0.2 s
Vehicle mass 925 kg
𝑚
𝐼𝑧𝑧 Vehicle moment of inertia 617 kgm2 8
𝐽𝜔r Rotational inertia of the rear wheel 1.24 kgm2
𝑟 Wheel radius 0.301 m
𝑙f Distance between the front axle and the CoG 0.99 m
𝑙r Distance between the rear axle and the CoG 0.71 m
𝑇max Maximum torque 200Nm
𝛿max Maximum steering angle 0.3rad
𝛿̇ max Maximum steering angular speed 2618 rad/s
𝐶𝑥 Longitudinal Pacejka tire parameter 1.5
𝐵𝑥 Longitudinal Pacejka tire parameter 8.0
𝐶𝑦 Lateral Pacejka tire parameter 1.4057
𝐵𝑦 Lateral Pacejka tire parameter 7.1138
 114 8 Efficient Response to Sudden Appearance of Static Obstacles

8.2.1 Kinematic Model

The trajectory of the ego vehicle in the global coordinates can be written as

𝑥̇ ego = 𝑣𝑋 , (8.1a)
𝑦̇ ego = 𝑣𝑌 , (8.1b)
𝜃̇ = 𝛾, (8.1c)

and the velocities in the local coordinate system attached to the vehicle body are obtained
by

𝑣𝑥 = 𝑣𝑋 cos (𝜃) − 𝑣𝑌 sin (𝜃), (8.2a)

𝑣𝑦 = 𝑣𝑋 sin (𝜃) + 𝑣𝑌 cos (𝜃). (8.2b)

With the kinematic states and inputs defined by

𝑇
𝑠kin = [𝑥ego 𝑦ego 𝑣𝑋 𝑣𝑌 ] ,
𝑇
𝑢kin = [𝑣̇ 𝑋 𝑣̇ 𝑌 ] ,

the kinematic model of the vehicle can be written as

𝑠̇ kin = 𝐴𝑠kin + 𝐵𝑢kin , (8.3)

where

⎡0 0 1 0⎤ ⎡0 0⎤
⎢0 0 0 1⎥ ⎢0 0⎥
𝐴=⎢ , 𝐵=⎢ .
⎢0 0 0 0⎥⎥ ⎢1 0⎥⎥
⎣0 0 0 0⎦ ⎣0 1⎦

8 8.2.2 Eqations of Motion

The governing equations of motion for the vehicle model in Fig. 8.2 can be written as
follows [8]:

∑ 𝐹𝑥 = 𝐹𝑥f cos (𝛿) − 𝐹𝑦f sin (𝛿) + 𝐹𝑥r , (8.4a)

∑ 𝐹𝑦 = 𝐹𝑥f sin (𝛿) + 𝐹𝑦f cos (𝛿) + 𝐹𝑦r , (8.4b)
∑ 𝑀𝑧 = (𝐹𝑥f sin (𝛿) + 𝐹𝑦f cos (𝛿)) 𝑙f − 𝐹𝑦r 𝑙r . (8.4c)

Considering a small front steering angle, the dynamics of the single-track vehicle model is
obtained as
1
𝑣̇ 𝑥 = 𝐹𝑥f + 𝐹𝑥r ) + 𝑣𝑦 𝛾, (8.5a)
𝑚(
1
𝑣̇ 𝑦 = (𝐹𝑦f + 𝐹𝑦r ) − 𝑣𝑥 𝛾, (8.5b)
𝑚
1
𝛾̇ = (𝐹𝑦f 𝑙f − 𝐹𝑦r 𝑙r ). (8.5c)
𝐼𝑧𝑧
8.3 Proposed Control System 115

Since we expect that the evasive maneuver will generate forces beyond their maximum
peak [177], we consider a nonlinear model for the lateral tire forces on the front and rear
axles using the celebrated Pacejka tire model [90] as

𝐹𝑥f = 𝜇𝐹𝑧f sin (𝐶𝑥 arctan (𝐵𝑥 𝜅f )) , (8.6a)

𝐹𝑥r = 𝜇𝐹𝑧r sin (𝐶𝑥 arctan (𝐵𝑥 𝜅r )) , (8.6b)
𝐹𝑦f = 𝜇𝐹𝑧f sin (𝐶𝑦 arctan (𝐵𝑦 𝛼f )) , (8.6c)
𝐹𝑦r = 𝜇𝐹𝑧r sin (𝐶𝑦 arctan (𝐵𝑦 𝛼r )) , (8.6d)

with the front and rear slip angles respectively described by

𝑙f 𝛾
𝛼f = 𝛿 − 𝛽 − , (8.7a)
𝑣𝑥
𝑙r 𝛾
𝛼r = − 𝛽. (8.7b)
𝑣𝑥

8.3 Proposed Control System

Planning Motion Control System Perception
𝑣̂ 𝑥
initial guess −
+ 𝑇rl 𝑣̂ 𝑦
𝐶𝑣,fb IWMrl
+ 𝛽̂
IMU

[𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ]MPC
State
TDL
MPC 𝑣𝑥∗ observer 𝛾̂
𝐶𝑣,ff + IWMrr
𝑇rr 𝜃̂

Weight function Velocity control

FBL [𝑥ego , 𝑦ego ]
[𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ]
Steering angle control
GPS

𝛿∗ + [𝜏, 𝜈]
𝐶𝛿,fb Obstacle
MSF EPS
[𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ]MSF − 𝛿 detector

8
Figure 8.3: Closed-loop system architecture

Figure 8.3 shows the architecture of the proposed closed-loop system, consisting of the
real system, as well as perception, planner, and controller modules, which will be described
next.

Perception On the perception layer, an observer is incorporated to estimate the sideslip

and yaw angles, while the obstacle detector obtains the position and width of the obstacle
and returns the time to collision 𝜏, and the lateral steering index 𝜈, based on the measure-
ments of the current state of the ego vehicle. The lateral steering index indicates the degree
of extreme steering required to avoid a collision, and is covered in the next section in more
detail.

Planner We propose a planner design strategy that leverages the capabilities of MPC
in finding an optimal maneuver for collision avoidance, as well as an assistive Maximum
Steering Feed-forward (MSF) planner designed to replicate a human-like response to the
 116 8 Efficient Response to Sudden Appearance of Static Obstacles

detection of static obstacles. For the sake of computational efficiency, the kinematic model
of the ego vehicle is used in the planner module. Therefore, both the MPC and the MSF
planners provide their respective references for the longitudinal and lateral accelerations,
𝑣̇ 𝑥∗ and 𝑣̇ 𝑦∗ . A weight function then combines the two references and feeds the resulting
reference signals to a FeedBack Linearization (FBL) controller to obtain the corresponding
reference steering angle and longitudinal velocity.

Controller We use a SPeed Controller (SPC) to obtain the required torque for tracking
the reference longitudinal velocity. The output torque of the speed controller is then
distributed via the Torque Distribution Law (TDL) to the rear left and rear right In-Wheel
Motors (IWMs). In addition, the FeedBacked Electric Power Steering (FB-EPS) system
tracks the steering reference by providing the steering angle signal to the steering motor.

8.4 Design of the Planning Strategy

Given the measured and observed current state of the ego vehicle, along with 𝜏 and 𝜈, the
MPC and MSF planners provide their respective solutions, 𝑣̇ 𝑋∗ and 𝑣̇ 𝑌∗ , distinguished by their
respective subscripts in Fig. 8.3. The same information is fed into the weight function that
acts as a situation-assessment unit by distributing the weights between [𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ]MPC and
[𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ]MSF . For instance, in cases where the solution of the MPC planner is close to the
steering limits or if it fails to provide a solution in time, the weight function assigns a higher
weight to the solution from the MSF planner. The design of each planner is discussed
separately in the following sections.

𝑣𝑥 = 1 m/s
20
𝑣𝑥 = 3 m/s
𝑦max (m)

𝑣𝑥 = 5 m/s

8 10 𝑣𝑥 = 7 m/s

0
0 1 2 3 4 5
𝑡 (s)

Figure 8.4: Maximum steering maneuver for different longitudinal velocities.

8.4.1 MSF Planner

We define the maximum steering maneuver 𝑦max ∶ ℝ2 → ℝ to denote the resulting lateral
position of the vehicle, as a function of time and its longitudinal velocity, after performing
a maximum-possible evasive action without braking. This evasive steering is designed to
mimic one of the human-like responses after detecting an obstacle: steering to the side as
fast as possible.
The function 𝑦max is calculated for time 𝑡 and velocity 𝑣𝑥 by considering (8.1)-(8.7) to
8.4 Design of the Planning Strategy 117

solve
𝑡
𝑦max (𝑡, 𝑣𝑥 ) = ∫ 𝑣𝑌 (𝜙)𝑑𝜙, (8.8a)
0
s.t. 𝑣𝑥 (𝜙) = 𝑣𝑥 , (8.8b)
𝛿(𝜙) = min (𝛿̇ max 𝜙 , 𝛿max ) , (8.8c)
𝑣𝑦 (0) = 0, (8.8d)
𝜃(0) = 0, (8.8e)
𝛾(0) = 0, (8.8f)

where 𝑣𝑌 is obtained by solving (8.8) and 𝛿̇ max and 𝛿max respectively represent the maximum
steering rate and steering angles. The resulting lateral position is shown by solid lines
in Fig. 8.4. Using 𝑦max and the width of the danger zone 𝑤, the lateral steering index 𝜈 is
defined by
𝑤
𝜈= . (8.9)
𝑦max (𝜏, 𝑣𝑥 )
Remark 8.1. The zero initial conditions for the lateral velocity, yaw angle, and yaw rate in
(8.8) represent a specific and extreme case. To define a more realistic model of the maximum
steering maneuver, 𝑦max can be defined as a function of these initial conditions. However, this
will lead to a higher-dimensional domain for the function 𝑦max , hence resulting in higher
computational demand in the next steps. For the sake of computational efficiency, we use the
zero initial conditions to merely account for the most extreme form of the maximum steering
maneuver.

8.4.2 MPC Planner

We define the safety barrier 𝑦safe using the sigmoid barrier function introduced in [174] in
its extreme case as
𝑤
8
𝑦safe (𝑡) = , (8.10)
𝑥obs − 𝑥ego (𝑡) − 4𝑣𝑋 (𝑡)
1 + exp √
( 𝜖 𝑤 𝑣𝑋2 (𝑡) )

where 𝑤 is the width of the danger zone, to be avoided by the ego vehicle’s center of gravity
√
and 𝜖 = 1/ 8.8𝜇𝑔. In the following example, we clarify the behavior of the 𝑦safe function.
Example 8.1. Figure 8.5 shows a schematic view of the sigmoid barrier function (8.10) with
the danger zone shown in red. For a given 𝑣𝑋 , 𝑦safe is defined such that the vehicle’s center of
gravity has traveled 𝑤/2 in the lateral direction 4 seconds before arriving at the longitudinal
position of the danger zone. For instance, assuming a constant longitudinal speed for the
vehicle in the global coordinates, with 𝜏 = 6s shown in blue, 𝑦safe passes through the lateral
mid-point after 2 seconds. If the obstacle is closer, i.e. 𝜏 = 5s as shown in orange, the resulting
𝑦safe has the same curvature and ensures a 4𝑣𝑋 distance before reaching 𝑤/2.
Using a discretized model of the vehicle and considering a prediction horizon of length
𝑁p , we denote the input signal and the resulting predicted states over the prediction window
 118 8 Efficient Response to Sudden Appearance of Static Obstacles

𝑤 𝑌 𝑤 𝑌
2𝑣𝑋 4𝑣𝑋 𝑣𝑋 4𝑣𝑋
𝑋 𝑋

Figure 8.5: Sigmoid barrier function examples

by
𝑇
𝑇 (𝑘 + 1|𝑘)
𝑠̃kin (𝑘) = [𝑠kin … 𝑇 (𝑘 + 𝑁 |𝑘)
𝑠kin p ] , (8.11a)
𝑇
𝑢̃ kin (𝑘) = 𝑇
[𝑢kin (𝑘) … 𝑢𝑇kin (𝑘 + 𝑁p − 1)] , (8.11b)

where 𝑠kin (𝑘 + 𝑖|𝑘) for 𝑖 ∈ {1, … , 𝑁p } represents the predicted state at time step 𝑘 + 𝑖 based on
the state measurement at time step 𝑘, i.e. 𝑠kin (𝑘|𝑘). We define four costs for each predicted
step as

𝐽safe (𝑘) = − log (𝑦safe (𝑘) − 𝑦ego (𝑘)) , (8.12a)

𝐽stable (𝑘) = |𝑣𝑌 (𝑘) /𝑣𝑋 (𝑘) | , (8.12b)
𝐽brake (𝑘) = |𝑣𝑋 (𝑘) − 𝑣des (𝑘)| , (8.12c)
𝐽steer (𝑘) = |𝑣̇ 𝑌 (𝑘)| . (8.12d)

The function 𝐽safe represents the barrier function for safe collision avoidance and the
functions 𝐽stable , 𝐽brake , and 𝐽steer are defined to respectively minimize the sideslip angle for
vehicle stability, the deviation from the desired velocity 𝑣des , and the steering action in the
planning optimization problem. The MPC cost function is then defined as

8 𝑁p
 (𝑘) = ∑ [ 𝜂1 𝐽safe (𝑘 + 𝑖) + 𝜂2 𝐽stable (𝑘 + 𝑖) + 𝜂3 𝐽brake (𝑘 + 𝑖) + 𝜂4 𝐽steer (𝑘 + 𝑖)], (8.13)
𝑖=1

where the weights 0 ⩽ 𝜂𝑗 ⩽ 1 with 𝑗 ∈ {1, … 4} yield a convex combination where ∑4𝑗=1 𝜂𝑗 = 1.
The resulting MPC planning optimization problem is given by

𝑢̃ ∗kin (𝑘) = argmin𝑢̃ kin (𝑘)  (𝑘). (8.14)

Note that the decision variable in (8.14) is the input vector and the state trajectory is defined
within the cost function to get an unconstrained MPC optimization problem and to avoid
infeasibility issues. MPC finds the optimal trajectory to avoid a collision by solving (8.14)
at each time step in a receding-horizon fashion. This is done by solving the problem for the
next 𝑁p time steps, while providing the solution to the first step ahead to the controller.

Example 8.2. Consider a simple case of 𝑁p = 2 with 𝑢kin (𝑘) = 𝑢kin (𝑘 +1), 𝑣𝑋 (𝑘) = 𝑣des (𝑘) = 5m/s,
𝑥ego (𝑘) = 𝑦ego (𝑘) = 𝑦obs (𝑘) = 0m, 𝑥obs = 𝑤 = 1m and 𝑣𝑌 (𝑘) = 0m/s. The solution to (8.14) for
input signals normalized on the bound [−1, 1] with the selections of 𝜂𝑗 = 0.25 is obtained as
8.5 System Implementation 119

follows:
}
𝜂1 = 0.25, 𝜂2 = 0.25 0.00
⟹ 𝑢∗kin (𝑘) = ,
𝜂3 = 0.25, 𝜂4 = 0.25 [0.53]
which indicates the optimal response would be no braking and steering rate equal to 53% of its
maximum value. However, if the weight for 𝐽safe is increased at the expense of reduction of the
weights for 𝐽brake and 𝐽steer , the solution to (8.14) would change to
}
𝜂1 = 0.50, 𝜂2 = 0.25 −0.35
⟹ 𝑢∗kin (𝑘) = ,
𝜂3 = 0.05, 𝜂4 = 0.20 [ 1.00 ]
corresponding to maximum steering rate with 35% of maximum braking.

8.4.3 FBL
To extract the reference steering angle and longitudinal velocity from [𝑣̇ 𝑋∗ , 𝑣̇ 𝑌∗ ], we obtain the
required velocities in the local coordinate using (8.2) and we consider the vehicle dynamics
in a feedback-linearization fashion as
𝛿 ∗ (𝑘) = 𝛽(𝑘) + 𝑙f 𝛾(𝑘) / 𝑣𝑥 (𝑘)
−1
+ 𝐹𝑦f [𝑚(𝑣̇ 𝑦 (𝑘) + 𝛾(𝑘)𝑣𝑥 (𝑘)) − 𝐹𝑦r (𝑙r 𝛾(𝑘) / 𝑣𝑥 (𝑘) − 𝛽(𝑘))]
𝑣𝑥 ∗ (𝑘) = 𝑣̇ 𝑥 (𝑘)𝑡s + 𝑣𝑥 (𝑘). (8.15a)

8.5 System Implementation

To evaluate the proposed planning strategy, we use an experimental vehicle FPEV2-Kanon
driven by rear-IWMs shown in Fig. 8.6. The main parameters of the vehicle and the IWM
are summarized in Table 8.1. The FPEV2-Kanon is equipped with an electric power steering
system consisting of a speed-controlled Maxon brushless motor.
8

Figure 8.6: Experimental vehicle system configuration.

The motion control and planning modules are implemented in AutoBox (PPC 750GX
1 GHz, 32GB SDRAM program memory, 96MB SDRAM data storage) and MicroLabBox
 120 8 Efficient Response to Sudden Appearance of Static Obstacles

Table 8.2: Optimization implementation parameters.

Parameter Value
Maximum iteration 10
Maximum function evaluation 100
Constraint tolerance 0.001
Optimality tolerance 0.001
Step tolerance 0.001

(NXP QorlQ P5020 dual core 2GHz, 1GB DRAM, 128MB flash memory), respectively. The
communication among these modules and the on-board Inertial Measuring Unit (IMU) and
Global Positioning System (GPS) sensors is facilitated via a Controller Area Network (CAN),
and set to 500kbps. The GPS, Autobox, and MicroLabBox measurements are respectively
updated at a frequency of 1 Hz, 10kHz, and 5Hz. The MPC planner is implemented using
fmincon’s Sequential Quadratic Programming (SQP) solver from the Matlab R2017b
Optimization toolbox. The optimization parameters are shown in Table 8.2.
The motion control module consists of two separate controllers for tracking the steering
angle and the longitudinal velocity, as shown in Fig. 8.3. Considering the nominal plant
transfer function
1
𝑃n (𝑠) = ,
2𝑚
(𝐽𝜔r + 𝑟 2 (1 − 𝜆n ))𝑠

with the nominal slip ratio 𝜆n = 0.05, a Proportional Integral (PI) speed controller is designed
by the placing the pole in −1rad/s.

8 8.6 Results and Discussion

In this section, the results are presented and discussed in three categories: the relaxed cases
with 𝑣𝑥 = 5m/s, lateral assessment with 𝑣𝑥 = 6m/s, and the extreme cases where 𝑣𝑥 = 7m/s.

8.6.1 Relaxed cases with 𝑣𝑥 = 5m/s

Figure 8.7 shows the results for the first round of experimental tests. The vehicle accelerates
to reach the longitudinal velocity of 5m/s and then detects a static obstacle with 𝜏 ∈ {3, 4, 5}
seconds. The desired vehicle speed is set to 5m/s as well, and to compensate for the lower 𝜏
values, we select a smaller obstacle for lower 𝜏 values to keep the 𝜈 parameter in a narrow
band. In all three scenarios, the vehicle manages to safely avoid colliding with the static
obstacle with the help of the MPC and feedforward steering commands while keeping the
velocity close to the desired value. In all three cases, the maximum steering command
provided by the combined planner occurs after passing the obstacle. At this point, the
planner decides to steer back to the initial lateral position since there is no longer a risk of
colliding with the obstacle. However, the communication delays in reading the GPS signal
result in a slight delay in responding as well, as can be observed in the cases with 𝜏 = 4s
and 3s.
8.6 Results and Discussion 121

Observation 8.1. While MPC is prone to providing solutions close to the steering bounds in
limited computation time, the combined planning strategy offers a smoother maneuver.

10 7
𝜈 = 0.25 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
6

𝛿 (m/s)
5 𝜏 = 5s
0 5
0 4 𝑣𝑥 ref 𝑣̂ 𝑥
−0.2
−5 3
0 10 20 30 0 10 20 30 0 10 20 30
10 7
𝜈 = 0.26 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
5 𝜏 = 4s
𝛿 (m/s) 6
0 5
0 4
−0.2
−5 3
0 10 20 30 0 10 20 30 0 10 20 30
10 7
𝜈 = 0.32 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
𝛿MPC 6
𝛿 (m/s)

5 𝜏 = 3s
0 𝛿 5
0 4
−0.2
−5 3
0 10 20 30 0 10 20 30 0 10 20 30
𝑥ego (m) 𝑥ego (m) 𝑥ego (m)

Figure 8.7: Experimental assessment of the 𝜏 influence: vehicle trajectory for the desired velocity 𝑣𝑥 = 5m/s.

8.6.2 Lateral assessment with 𝑣𝑥 = 6m/s

Three examples for the next set of experiments are shown in Fig. 8.8. The distance to
the static obstacle at the detection time is set to 𝜏 = 3s for all three cases, while varying
the value of 𝜈 from 0.4 to 0.35. Similar to the relaxed cases, we observe that MPC is still
“overreacting” due to the limited computation time, while the combined planner manages
to avoid any collision without an extreme steering command. The desired velocity is set to
6m/s, which is the same as the longitudinal velocity at the time of obstacle detection in the
first two cases. However, in the third case, the vehicle detects the obstacle at the speed of 8
4.5m/s which means that a much higher steering command is required to achieve the same
lateral displacement. As a result, we observe that the planners opt for an increase in the
longitudinal velocity to allow for a lower and safer steering command.
Observation 8.2. With a higher velocity, we observe that the GPS measurement lag can be
more limiting in a swift return to the initial lateral position after overtaking the obstacle.

8.6.3 Extreme cases with 𝑣𝑥 = 7m/s

Finally, we conduct the experiment for the extreme cases close to the maximum allowed lon-
gitudinal velocity on campus field with setting the desired vehicle speed to 7m/s. Figure 8.9
shows the results for three cases with the same obstacle size, while changing the 𝜏 and 𝜈
values. In all three cases, we observe that the planners need to increase the longitudinal
velocity as the steering limits are not sufficient to guarantee a safe collision-free maneuver.
Observation 8.3. Increasing the longitudinal velocity is a potential solution in scenarios
where the steering limits are not sufficient to guarantee a safe collision-free maneuver.
 122 8 Efficient Response to Sudden Appearance of Static Obstacles

10 8
𝜈 = 0.21 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
7

𝛿 (m/s)
5 𝜏 = 3s 𝑣𝑥 ref
0 6
0 5 𝑣̂ 𝑥
−0.2
−5 4
0 10 20 30 0 10 20 30 0 10 20 30
10 8
𝜈 = 0.26 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
𝛿MPC 7

𝛿 (m/s)
5 𝜏 = 3s
0 𝛿 6
0 5
−0.2
−5 4
0 10 20 30 0 10 20 30 0 10 20 30
10 8
𝜈 = 0.34 0.2
𝑦ego (m)

𝑣𝑥 (m/s)
7
𝛿 (m/s)
5 𝜏 = 3s
0 6
0 5
−0.2
−5 4
0 10 20 30 0 10 20 30 0 10 20 30
𝑥ego (m) 𝑥ego (m) 𝑥ego (m)

Figure 8.8: Experimental assessment of the 𝜈 influence: vehicle trajectory for the desired velocity 𝑣𝑥 = 6m/s.

In the most extreme case with 𝜏 = 1.4s and 𝜈 = 0.95, we observe that the MPC planner
cannot find a feasible solution to avoid colliding with the static obstacle. However, the
feedforward planner offers a solution close to the limits of steering, which helps in avoiding
the highly-probable collision. While returning to the initial lateral position and the desired
velocity after overtaking the obstacle.

10 9
𝜈 = 0.19 0.2
𝑦ego (m)

𝑣𝑥 (m/s)

8
𝛿 (m/s)

5 𝜏 = 2.8s
0 7
0 6
−0.2
−5 5
0 10 20 30 0 10 20 30 0 10 20 30
8 10
𝜈 = 0.40 0.2
9
𝑦ego (m)

𝑣𝑥 (m/s)

8
𝛿 (m/s)

5 𝜏 = 2s 𝑣𝑥 ref
0 7
0 6 𝑣̂ 𝑥
−0.2
−5 5
0 10 20 30 0 10 20 30 0 10 20 30
10 9
0.2
𝑦ego (m)

𝑣𝑥 (m/s)

𝛿MPC 𝛿 8
𝛿 (m/s)

5
𝜈 = 0.95 0 7
0 𝜏 = 1.4s 6
−0.2
−5 5
0 10 20 30 0 10 20 30 0 10 20 30
𝑥ego (m) 𝑥ego (m) 𝑥ego (m)

Figure 8.9: Experimental assessment of the hazard: vehicle trajectory for the desired velocity 𝑣𝑥 = 7m/s.

8.6.4 Comparison with MPC

In the final set of tests, we compare the collision avoidance capabilities of the MPC planner
against our proposed combined strategy. Figure 8.10 shows two scenarios with 𝑣𝑥 = 6m/s
8.7 Conclusions 123

and 𝜏 = 3s with two different obstacle widths. It can be observed that while MPC overreacts
to the presence of the obstacle by providing extreme steering commands due to the limited
computation time, the combined planner can safely avoid colliding with the obstacle while
keeping the ego vehicle within the field limits.

20 20

MPC
𝜈 = 0.26 𝜈 = 0.34
MPC
𝑦ego (m)

𝑦ego (m)
ed
10
Combined
10 Combin

0 0
0 10 20 30 0 10 20 30
𝑥ego (m) 𝑥ego (m)

Figure 8.10: Comparing the MPC and combined planners

8.7 Conclusions
This chapter has offered experimental insights into real-time implementation of MPC
for collision avoidance after the unexpected appearance of a static obstacle. Given the
limitations of state-of-the-art nonlinear MPC in providing feasible solutions in real-time,
we have proposed a human-inspired feedforward planner to support situations where the
MPC optimization problem is either infeasible or converges to a poor local solution due to
a poor initial guess. Our real-world experiments, conducted under various conditions and
speeds using the FPEV2-Kanon electric vehicle, validate the effectiveness of our proposed
planning strategy, also in comparison to a state-of-the-art MPC motion planner.
For future research, we suggest real-time experimental tests considering parametric
uncertainties e.g. due to variations in friction coefficient.

8
 125

9
Proactive Collision
Avoidance with Stochastic
Obstacle Behavior
Then predictions could be scientific, ... only by ceasing to prophesy definitively.

— Albert Camus, The Rebel

Uncertainty in the behavior of other traffic participants is a crucial factor in collision avoidance
for automated driving; here, stochastic metrics could avoid overly conservative decisions. This
chapter introduces a Stochastic Model Predictive Control (SMPC) planner for emergency
collision avoidance in highway scenarios to proactively minimize collision risk while ensuring
safety through chance constraints. To guarantee that the emergency trajectory can be attained,
we incorporate nonlinear tire dynamics in the prediction model of the ego vehicle. Further, 9
we exploit Max-Min-Plus-Scaling (MMPS) approximations of the nonlinearities to avoid
conservatism, enforce proactive collision avoidance, and improve computational efficiency
in terms of performance and speed. Consequently, our contributions include integrating a
dynamic ego vehicle model into the SMPC planner, introducing the MMPS approximation for
real-time implementation in emergency scenarios, and integrating SMPC with hybridized
chance constraints and risk minimization. We evaluate our SMPC formulation in terms of
proactivity and efficiency in various hazardous scenarios. Moreover, we demonstrate the
effectiveness of our proposed approach by comparing it with a state-of-the-art SMPC planner
and we validate that the generated trajectories can be attained using a high-fidelity vehicle
model in IPG CarMaker.

This chapter has been published in IEEE Transactions on Control Systems Technology [98].
 126 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

9.1 Introduction
While robust (worst-case) approaches in Model Predictive Control (MPC) synthesis have
been used in automated driving to ensure safe motion planning in uncertain dynamic
environments [14, 15, 178, 179], they can lead to overly-conservative maneuvers [180]
and eventually fail in reaching the main control objective. For instance, it is recognized
that human drivers do not drive according to worst-case considerations: if they did, an
urban driver may never merge into its desired lane when considering the worst-case
scenario in predicting the behavior of other traffic participants [181], or a highway driver
would activate unnecessary emergency braking when considering the worst-case scenario
in predicting the behavior of a cut-in vehicle. Arguably, the way human drivers avoid
overly-conservative maneuvers is by taking some stochastic metrics into account during
the planning.
As an example, Fig. 9.1 shows a scenario of proactive collision avoidance: the ego
vehicle (pink) is surrounded by other road users (green). If the front vehicle suddenly
brakes, a conservative decision would be to decelerate as well to keep the distance. However,
this decision could lead to collision with the rear vehicle. It would be much safer in this
scenario for the ego vehicle to proactively avoid the collision by moving to the left lane
while keeping a safe distance from all the surrounding road users. In summary, proactive
collision avoidance can be understood by three key features: swift response to disturbance
(i.e. danger), optimality in terms of safety, and avoiding propagation of hazard to future
time steps, which translates into getting out of an emergency situation as fast as possible.

Figure 9.1: Example of proactive collision avoidance in a highway scenario: if its front vehicle suddenly brakes,
the ego vehicle (pink) avoids front and rear-end collision with other road users (green) by safely moving to the
left lane.

9.1.1 Motion Planning Challenges in Different Scenarios

9 SMPC [182] is used in various collision-avoidance applications to generate a reference
trajectory within a dynamic environment e.g. for mobile robots [183–185] or spacing
control in vehicle platoons [186, 187]. In automated driving applications, [159] reviews
different threat metrics for risk assessment during maneuvers from collision probability
to time-to-collision or distance-to-collision between the ego vehicle and other road users.
Since the challenges and requirements of stochastic motion planning in an uncertain
environment depend on the driving scenario, two cases can be distinguished: urban or
highway.
In urban driving, the vehicles drive at lower speeds, which allows using kinematic
models for the ego vehicle [188]. In addition, vehicles can decide among different actions
such as turning to different streets at a junction, stopping to park, or merging into another
lane [1]. Moreover, there is a variety of traffic participants from pedestrians and bicycles
to different drivers with their own driving styles that significantly affect the decision-
9.1 Introduction 127

making outcome [181]. Therefore, the prediction of other participants should be more
comprehensive and intention-aware, and the research in this area has been focusing
on robust estimation of feasible space [189, 190] and tractable MPC formulations in the
presence of uncertainty in the behavior of other traffic participants [2, 3].

Conversely, the planning problem in highway scenarios faces two entangled challenges:
ensuring that the generated trajectory can be attained and solving the planning problem
is computationally efficient. On highways, an emergency maneuver at high speed would
push the vehicle in the nonlinear regime. In this sense, ensuring that the generated refer-
ence trajectory can be attained requires considering the nonlinear tire dynamics within
the ego vehicle model for a more accurate prediction of the available tire forces [13]. A
common solution to avoid unattainable generated trajectories is to design an integrated
planner/tracker incorporating a higher-fidelity prediction model of the ego vehicle, e.g.[9]
proposes serially-cascaded models to allow using different sampling times and prediction
horizons for the planning and tracking sub-problems. However, this technique is appli-
cable to less-aggressive maneuvers only, since both prediction models for the planning
and tracking sub-problems are simple. In this sense, hierarchical control design is still
the most popular choice in the literature for emergency collision avoidance in highway
driving [191, 192] and the kinematic single-track model is often selected as the ego vehicle
prediction model [119, 192, 193]. On the other hand, incorporating nonlinear tire dynamics
significantly increases the computational complexity of the MPC planning problem, which
may prevent a proactive response to danger.

9.1.2 Sources of Uncertainty in Highway Driving

In the highway collision avoidance literature, the stochasticity of the uncertain environment
is expressed via chance constraints in the SMPC planning problem. After observing their
initial position and velocity, the behavior of the obstacles is forecasted over a prediction
horizon by considering a linear, often point mass, model [119]. Stochastic behavior of
obstacles is then modeled by random variables in their prediction model such as their ve-
locity [194, 195] or acceleration [183]. Sometimes, randomness in the lane change decision
is considered as well [193]. In this sense, [196] expresses the trajectory of obstacles using a
Markov jump system description, whereas [194] uses a hybrid obstacle model including 9
stochastic switching decision between continuing along a straight path or following an
arc trajectory. This uncertainty is then propagated over the prediction horizon e.g. by
chaos-based approaches [171] or state updates via Kalman filter [119, 183], leading to
chance constraints in the SMPC problem. The reference trajectory is found by minimizing
a cost function which in the literature has been mainly defined as a convex (often quadratic)
function of the states and inputs [197], such as the velocity-tracking error [119, 196] to en-
force maintaining a constant longitudinal velocity. Unfortunately, in emergency maneuvers
and hazardous scenarios, minimizing the probability of collision is more important than
tracking errors. In this sense, [198] uses a potential field function for collision avoidance,
but the obstacle behavior is not stochastic. However, the objective function of avoiding
collision may have no closed form, such as in [194], due to the stochasticity of the switching
decision. There, the objective function is constructed iteratively via reachable sets.
 128 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

9.1.3 Computational Efficiency in Emergencies

At the same time, tractability is also crucial and must be traded with the accuracy of the
model. For instance, in [19] a hybrid nonlinear prediction model is considered for the ego
vehicle and the exponential growth in computational complexity is compensated by adapt-
ing the prediction horizon accordingly. Further, [199] suggests successive convexification
to improve the initial guess for the Nonlinear Model Predictive Control (NMPC) problem to
reduce the number of iterations and [200] uses a Mixed-Integer Linear Program (MILP) to
find the feasible region and feed it into the nonlinear planning problem to find the optimal
trajectory. To the best of our knowledge, no research has been done incorporating tire
force dynamics for real-time emergency motion planning in highway scenarios, i.e. fast
online solution of the planning optimization problem, while minimizing the probability of
collision which leads to a highly- nonlinear formulation for the SMPC problem.
Hybrid modeling frameworks [18] such as MMPS formalism [24], are effective tools
to reduce the computational complexity of the planning problem while incorporating the
nonlinear behavior regime. In this sense, hybridization refers to the approximation of a
nonlinear function, e.g. the prediction model, using a hybrid systems modeling framework.
In case of a nonlinear control optimization problem, hybridization can lead to an MILP
formulation of the problem that is computationally more efficient to solve, compared
to a NonLinear Program (NLP). Sequential Quadratic Programming (SQP) and real-time
iteration scheme have been used in the literature where the nonlinear dynamics is linearized
at each time step [93]. However, that approach has limited capability to adequately capture
the complexity of the nonlinear behavior along the prediction horizon. The fact that MILPs
can be solved to global optimality in a finite number of iterations [201] makes them a
suitable candidate to formulate the MPC planning problem.

9.1.4 Contributions
In this chapter, we propose an SMPC motion planner for emergency collision avoidance in
highway scenarios. We present a proactive planner design by minimizing the collision risk
as well as improving safety using chance constraints in the SMPC formulation. To avoid
generating unattainable trajectories, we incorporate nonlinear tire dynamics (accounting
for the nonlinear tire behavior close to saturation limits) within the prediction model for
the ego vehicle and we use MMPS approximation to reduce the computational complexity
9 of the planning problem. As a result, the novelties in our work are twofold:

1. introducing the idea of MMPS approximation of the nonlinearities for real-time

implementation, and

2. combining hybridized risk minimization within a stochastic MPC framework for

highway path planning.

Moreover, we provide a comprehensive analysis of how various formulations of the MPC

planner influence the conservatism and efficiency of the algorithm to proactively avoid
a collision in hazardous scenarios and we compare our proposed approach to a method
inspired by the state-of-the-art SMPC planner in [119] during various cases studies. To
verify that the generated trajectories can be attained by our proposed SMPC planner, we
simulate the maneuvers using a high-fidelity vehicle model in IPG CarMaker [202].
9.2 Problem Formulation 129

The chapter is structured as follows: Section 9.2 describes the formulation of the
predictive planning problem. Then Section 9.3 briefly covers the MMPS approximation,
and Section 9.4 explains our approach in reformulating and solving the SMPC problem.
Simulation results and comparisons to the state-of-the-art SMPC planner and the built-in
collision avoidance module in IPG CarMaker are presented in Section 9.5. Finally, we
conclude this chapter in Section 9.6.
The notation is this work is rather standard. The state and input vectors at time step
𝑘 are represented by 𝑠(𝑘) and 𝑢(𝑘), respectively. We use a tilde symbol, e.g. as in 𝑠̃(𝑘), to
denote the trajectory of a signal along the prediction horizon. The probability is expressed
by the Pr symbol.

9.2 Problem Formulation

Given a predicted state trajectory 𝑠̃ at control time step 𝑘 along the next 𝑁p steps as
𝑇
𝑠̃(𝑘) = [𝑠 𝑇 (𝑘 + 1|𝑘) … 𝑠 𝑇 (𝑘 + 𝑁p |𝑘)] , (9.1)

the SMPC planning optimization problem can be formulated by the generic form

min 𝐽 (̃𝑠(𝑘)), (9.2a)

𝑠̃(𝑘),𝑢(𝑘)
̃
s.t. 𝑠(𝑘 + 𝑖|𝑘) = 𝑓 (𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)), ∀𝑖 ∈ {1, … , 𝑁p }, (9.2b)
𝑔 (𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)) ⩽ 0, ∀𝑖 ∈ {1, … , 𝑁p }, (9.2c)
Pr(𝑠(𝑘 + 𝑖|𝑘) ∈ 𝑘 ) ⩾ 1 − 𝜖, ∀𝑖 ∈ {1, … , 𝑁p }, (9.2d)

where 𝐽 represents the cost function, usually formulated as deviations from a desired
velocity or divergence from a globally-planned reference trajectory. Further, the planning
problem is constrained to the prediction model of the ego vehicle 𝑓 (⋅) via (9.2b), general
nonlinear constraints 𝑔(⋅) (9.2c), and the chance constraints (9.2d) where 𝑘 is the safe or
confidence region in step 𝑘 and 𝜖 is the minimum acceptable probability for constraint vio-
lation and is selected to be close to 0. Based on the requirements for highway emergencies,
𝐽 , 𝑓 , 𝑔 and 𝑘 , often need to be selected in such a way that (9.2) would be an NLP, hence
computationally expensive to solve in real time. As explained in Section 9.1, we use MMPS
approximation of the nonlinearities to facilitate obtaining an MILP reformulation of (9.2) 9
and to improve the computational efficiency. This is further discussed in the next section.

9.3 MMPS Approximation

As the name suggests, MMPS systems are modeled using max, min, plus, and scaling
operators and are equivalent to continuous piecewise-affine systems [18]. Any MMPS
function 𝑓MMPS can be described by either a conjunctive or a disjunctive canonical form [28]:

𝑓con (𝜒 ) = min 𝑇
max (𝛾𝑝,𝑞 𝜒 + 𝜈𝑝,𝑞 ) , (9.3a)
𝑝=1,…,𝑃 𝑞=1,…,𝑚𝑝

𝑓dis (𝜒 ) = max min (𝜙𝑇𝑝,𝑞 𝜒 + 𝜔𝑝,𝑞 ) , (9.3b)

𝑞=1,…,𝑄 𝑝=1,…,𝑛𝑞
 130 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

where 𝛾 and 𝜙 are vectors, 𝜈 and 𝜔 are scalars, and 𝑃, 𝑄, 𝑚𝑝 , and 𝑛𝑞 are integers determining
the number of nested min and max operators.
A nonlinear (scalar) function 𝑓 ∶  → ℝ can be approximated by an MMPS form
[𝑓 ]MMPS in compact state domain  via solving the nonlinear optimization problem
‖𝑓 (𝜒 ) − [𝑓 ]MMPS (𝜒 )‖2
min ∫ 𝑑𝜒 , (9.4)
 ‖𝑓 (𝜒 )‖2 + 𝜖0


where [.]MMPS represents the MMPS approximation of the corresponding argument with
either forms in (9.3) and  collects the decision variables for fixed values of 𝑃, 𝑄, 𝑚𝑝 , and
𝑛𝑞 as
⎧ { } { }
⎪
⎪ ( 𝛾𝑝,𝑞 , 𝜈𝑝,𝑞 ) 𝑝=1,…,𝑃 if [ f ]MMPS = fcon
⎪
=⎨ { } { } 𝑞=1,…,𝑚𝑝 . (9.5)
⎪
⎪ 𝜙
( 𝑝,𝑞 , 𝜔 𝑝,𝑞 ) 𝑞=1,…,𝑄 if [ f ]MMPS = fdis
⎪
⎩ 𝑝=1,…,𝑛𝑞

Note that  is a tuple of vector and scalar sets since it is necessary to preserve their order
in the MMPS forms. The positive value 𝜖0 > 0 added to the denominator in (9.4) serves to
avoid division by very small values for ‖𝑓 (𝜒 )‖2 ≈ 0.
In the next steps, we hybridize a suitable nonlinear prediction model for the ego vehicle
by solving (9.4) for the nonlinear terms within the vehicle model and use our information
of the shape and form of each nonlinearities to select their respective approximation forms
in (9.3) and the values of the integer pairs (𝑃, 𝑚𝑝 ) or (𝑄, 𝑛𝑞 ). Problem (9.4) is a smooth NLP
which can be solved by e.g. sequential quadratic programming and multi-start strategy.

9.4 Problem Reformulation and Solution Approach

9.4.1 Obstacle Vehicle Model
Given 𝑁o obstacles on the road, the states of the 𝜂-th obstacle where 𝜂 ∈ {1, … , 𝑁o } at time
step 𝑘 are expressed by the stochastic vector 𝑧 (𝜂) (𝑘) defined as
𝑇
(𝜂)
𝑧 (𝜂) (𝑘) = [𝑥obs (𝜂) (𝜂) (𝜂)
(𝑘) 𝑦obs (𝑘) 𝑥̇ obs (𝑘) 𝑦̇ obs (𝑘)] , (9.6)

9 with the Gaussian distribution

𝑧 (𝜂) (𝑘) ∼  (𝜉 (𝜂) (𝑘), Ξ(𝜂) (𝑘)) , (9.7)

where 𝜉 and Ξ respectively indicate the mean vector and the covariance matrix as
𝑇
𝜉 (𝜂) (𝑘) = [𝜉𝑥(𝜂) (𝑘) (𝜂) (𝜂) (𝜂)
𝜉𝑦 (𝑘) 𝜉𝑥̇ (𝑘) 𝜉𝑦̇ (𝑘)] , (9.8)
(𝜂)
⎡𝜎𝑥 (𝑘) 0 0 0 ⎤
⎢ 0 (𝜂)
𝜎𝑦 (𝑘) 0 0 ⎥⎥
Ξ(𝜂) (𝑘) = ⎢ (𝜂) . (9.9)
⎢ 0 0 𝜎𝑥̇ (𝑘) 0 ⎥
⎢ (𝜂) ⎥
⎣ 0 0 0 𝜎𝑦̇ (𝑘)⎦
9.4 Problem Reformulation and Solution Approach 131

Figure 9.2: Model configuration for the ego vehicle and the obstacles on the road.

Remark 9.1. We use discretized double integrator dynamics to model the obstacle behavior
and update variance and mean using Kalman updates. Note that the actual covariance matrix
does not remain diagonal, but it is customary to consider a reduced or approximated covariance
matrix including the diagonal elements of Ξ associated with the target states [203–206] for
computational efficiency; an approach we use in this chapter as well.
More specifically, we use a point mass model [119] for the obstacles in Fig. 9.2, expressed
by

𝑧 (𝜂) (𝑘 + 1) = 𝐴𝑧 (𝜂) (𝑘) + 𝐵𝑤(𝜂) (𝑘) + 𝜈(𝜂) (𝑘), (9.10a)

(𝜂)
(𝜂)
𝑤 (𝑘) = 𝐾 (𝑧ref (𝑘) − 𝑧 (𝜂) (𝑘)) , (9.10b)

where 𝑤 represents the input signal as

𝑇
(𝜂)
𝑤(𝜂) (𝑘) = [𝑥̈obs (𝑘)
(𝜂)
𝑦̈obs (𝑘)] , (9.11)

the 𝐴 and 𝐵 being state and input matrices resulting from discretized double integrator
dynamics, 𝜈 ∼  (04×1 , Ξ0 ) the process noise, and 𝐾 being the Kalman feedback gain such
that the obstacle tracks its corresponding reference state 𝑧ref . The covariance matrix for
each obstacle is updated at each time step in line with Kalman update by 9
(𝜂)
𝜉 (𝜂) (𝑘 + 1) = (𝐴 − 𝐵𝐾 )𝜉 (𝜂) (𝑘)𝐵𝐾 𝑧ref (𝑘), (9.12a)
(𝜂)
(𝜂) (𝜂)
Ξ (𝑘 + 1) = (𝐴 − 𝐵𝐾 )Ξ (𝑘)(𝐴 − 𝐵𝐾 ) 𝑇
+ Ξ0 . (9.12b)

with Ξ0 being the initial estimate of the covariance matrix of the process noise. Using the
(𝜂)
Gaussian distribution in (9.7), we define 𝑝𝑘 to express the probability density function for
the presence of obstacle 𝜂 ∈ {1, … , 𝑁o } on the road as

(𝜂) 2 (𝜂) 2
𝑥 − 𝜉𝑥 (𝑘) 𝑦 − 𝜉𝑦 (𝑘)
exp − √ (𝜂) − √ (𝜂)
(𝜂)
( ( 2𝜎𝑥 (𝑘) ) ( 2𝜎𝑦 (𝑘) ) )
𝑝𝑘 (𝑥, 𝑦) = (𝜂) (𝜂)
, (9.13)
2𝜋𝜎𝑥 (𝑘)𝜎𝑦 (𝑘)
 132 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

which is used to develop the probability function ℙ for the state vector 𝑠(𝑘) defined in (16)
using a chi-squared distribution (see [119]) and taking into account the unsafe area Ω(𝜂) , as

ℙ(𝜂) (𝑠(𝑘)) = Pr ((𝑥ego (𝑘), 𝑦ego (𝑘) ∈ Ω(𝜂) )) . (9.14)

The unsafe set Ω for each obstacle is defined as an area that the center of gravity of the
ego vehicle must avoid, and it is an ellipse calculated by considering the position and size
of both ego and obstacle vehicles as known parameters [159].

9.4.2 Hybrid Ego Vehicle Model

The ego vehicle prediction model as shown in Fig. 9.2 is described by a dynamic bicycle
model [12] with a small-angle assumption for 𝛿 (reasonable in highway scenarios [15])

𝑥̇ ego = 𝑣 cos(𝜓 + 𝛽), (9.15a)

𝑦̇ ego = 𝑣 sin(𝜓 + 𝛽), (9.15b)
𝜓̇ = 𝑟, (9.15c)
1
𝑣̇ = [𝐹𝑥f − 𝐹𝑦f 𝛿 + 𝐹𝑥r ] + 𝑣𝛽𝑟, (9.15d)
𝑚
̇𝛽 = 1 [𝐹𝑦f + 𝐹𝑦r ] − 𝑟, (9.15e)
𝑚𝑣
1
𝑟̇ = [𝐹𝑥f 𝛿 𝑙f + 𝐹𝑦f 𝑙f − 𝐹𝑦r 𝑙r ] , (9.15f)
𝐼𝑧𝑧
𝛿̇ = 𝑑𝛿 , (9.15g)

with 𝐹𝑥f , 𝐹𝑥r , and 𝑑𝛿 as inputs. All the variables and system parameters are described in
Tables 9.1 and 9.2, and the state vector 𝑠 at time step 𝑘 is expressed by
𝑇
𝑠(𝑘) = [𝑥ego (𝑘) 𝑦ego (𝑘) 𝜓(𝑘) 𝑣(𝑘) 𝛽(𝑘) 𝑟(𝑘) 𝛿(𝑘)] . (9.16)

The tire forces should satisfy the tire saturation limits

2
𝐹𝑥f 2
+ 𝐹𝑦f ⩽ (𝜇𝐹𝑧f )2 , (9.17a)
9 2
𝐹𝑥r 2
+ 𝐹𝑦r ⩽ (𝜇𝐹𝑧r )2 , (9.17b)

also known as Kamm circle constraint [89]. Considering the slip angles
𝑙f 𝑟
𝛼f = 𝛿 − 𝛽 + , (9.18a)
𝑣
𝑙r 𝑟
𝛼r = − 𝛽, (9.18b)
𝑣
we describe the lateral tire forces by MMPS approximations of the Pacejka tire model [90]
shown in Fig. 9.3 as

𝛼
[𝐹𝑦 ]MMPS = 𝐹max min max , −1 , 1 , (9.19)
( ( 𝛼s ) )
9.4 Problem Reformulation and Solution Approach 133

𝐹𝑦
𝐹max

𝛼
𝛼s

Pacejka
−𝐹max MMPS

Figure 9.3: Pacejka tire model and its MMPS approximation

where the nonlinear function representing the tire forces on the front and rear axles is
approximated by a parametric MMPS function where 𝐹max and 𝛼s respectively correspond
to the maximum tire force and the saturation slip angle.
Substituting the front and rear slip angles in (9.19) gives the front and rear lateral tire
forces as
𝛿 𝛽 𝑙f 𝑟
[𝐹𝑦f ]MMPS = 𝐹max min max − + , −1 , 1 , (9.20a)
( ( 𝛼s 𝛼s 𝛼s 𝑣0 ) )
𝑙r 1
[𝐹𝑦r ]MMPS = 𝐹max min max 𝑟 − 𝛽, −1 , 1 . (9.20b)
( ( 𝛼s 𝑣0 𝛼s ) )

Using the MMPS approximation of the other nonlinear terms in the ego vehicle model, we
obtain an MMPS formulation for the ego vehicle model expressed by

𝑥̇ ego = max{𝑣, 𝑣0 [cos(𝜓 + 𝛽)]MMPS }, (9.21a)

𝑦̇ ego = 𝑣0 [sin(𝜓 + 𝛽)]MMPS , (9.21b)
𝜓̇ = 𝑟, (9.21c)
𝐹𝑥f + 𝐹𝑥r [𝛿 𝐹𝑦f ]MMPS
𝑣̇ = − + 𝑣0 [𝛽𝑟]MMPS , (9.21d)
𝑚 𝑚
[𝐹𝑦f + 𝐹𝑦r ]MMPS
𝛽̇ = − 𝑟, (9.21e)
𝑚𝑣0 9
𝑙f 𝛿0 𝐹𝑥f 𝑙f [𝐹𝑦f ]MMPS 𝑙r [𝐹𝑦r ]MMPS
𝑟̇ = + − , (9.21f)
𝐼𝑧𝑧 𝐼𝑧𝑧 𝐼𝑧𝑧
𝛿̇ = 𝑑𝛿 , (9.21g)

Figure 9.4 presents three examples of the nonlinear terms vs. their MMPS approximations.
To find these formulations, we have used information on the form of the nonlinear function
and we have selected the number of max and min operators accordingly. For instance,
in Fig. 9.4a, we use three hyperplanes and two max and min operators based on the
cosinusoidal shape of the nonlinear function.

Remark 9.2. Considering the orders of magnitude of variations of the longitudinal velocity
over the prediction horizon, the velocity 𝑣 in (9.15b), (9.15d) and (9.15e) can be approximated as
 134 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

a fixed parameter over the prediction horizon and can be taken equal to the current measured
velocity. Moreover, in cases where 𝑣 is multiplied by cosine terms with values close to 1, we
take the maximum value between the velocity 𝑣 and the MMPS approximation with 𝑣 = 𝑣0 in
(9.21a) to ensure the inclusion of numerically significant effects resulting from variations in
𝑣 when 𝜓 + 𝛽 ≈ 0 in (9.21a). A similar approach is used for 𝛿 in (9.15f) where its variations
are included in the MMPS tire forces and the current steering angle is used as a parametric
coefficient for the first term.

Remark 9.3. After MMPS approximation of the continuous-time model of the ego vehicle,
(9.21) can be discretized e.g. using forward Euler method and a proper sampling time to be
incorporated in the SMPC formulation in (9.2).

Further, the Kamm circle constraints in (9.17) are approximated using MMPS func-
tion in Figures 9.4c. Note that due to different ranges of 𝐹𝑥f and 𝐹𝑥r , the front and rear
force magnitudes are approximated by the maximum of respectively three and four affine
functions, to appropriately capture the form of the nonlinear function. The maximum tire
forces on the front and rear axles are functions of the online measurements of the friction
coefficient 𝜇, which we assume available via a friction estimator [9, 12], as

𝐹max = min (𝜇𝐹𝑧f , 𝜇𝐹𝑧r ) .

⋅10−2
1
𝑦f
2 + 𝐹2
cos(𝜓 + 𝛽)

5 5
0 0
𝛽𝑟

𝐹𝑥f

−5
√

−1 0
0.5
−2 0 2 −0.10 0.1 −0.1 0 0.1 −0.5 0 −4 −2 0 −5 0 5

𝜓 (rad) 𝛽 (rad) 𝛽 (rad) 𝑟 (rad/s) 𝐹𝑥f (kN) 𝐹𝑦f (kN)

√
(a) cos(𝜓 + 𝛽) (b) 𝛽𝑟 (c) 𝐹𝑥f
2 + 𝐹2
𝑦f

Figure 9.4: Plots of example nonlinear terms in the ego vehicle prediction model and their MMPS approximations

9
9.4.3 Chance Constraints and Collision Risk Function
To hybridize the probability function ℙ in (9.14), we approximate it by the MMPS function
[ℙ]MMPS as illustrated in Fig. 9.5. The MMPS approximation [ℙ]MMPS is a probability
function as well and is used as a chance constraint in the SMPC formulation.
Since the chance constraints must be bounded in such a way that the probability of
constraint violation is very low (to improve safety), the accuracy of the MMPS approx-
imation is more important in regions close to ℙ = 0. Therefore, we obtain [ℙ]MMPS by
approximating the Gaussian probability density function (9.13) on a compact domain 
defined by the road boundaries via solving (9.4) and imposing the constraint

(𝜂)
∫ [𝑝𝑘 ]MMPS (𝑥, 𝑦) = 1,

9.4 Problem Reformulation and Solution Approach 135

Table 9.1: System variables and their bounds in the case study

Variable Definition Unit Bounds

𝑥ego Longitudinal position of the ego vehicle m [0, ∞]
𝑥obs Longitudinal position of the obstacle m [0, ∞]
𝑥̇ obs Longitudinal velocity of the obstacle m/s [5, 50]
𝑦ego Lateral position of the ego vehicle m [-6, 6]
𝑦obs Lateral position of the obstacle m [-6, 6]
𝑦̇ obs Lateral velocity of the obstacle m/s [-5,5]
𝑣 Velocity of the ego vehicle m/s [5, 50]
𝛽 Sideslip angle rad [-0.2,0.2]
𝜓 Yaw angle rad [−𝜋,𝜋]
𝑟 Yaw rate rad/s [-0.5, 0.5]
𝛿 Steering angle (road) rad [-0.2, 0.2]
𝐹𝑥f Longitudinal force on the front axis N [-5000, 0]
𝐹𝑥r Longitudinal force on the rear axis N [-5000, 5000]
𝐹𝑦f Lateral force on the front axis N –
𝐹𝑦r Lateral force on the rear axis N –
𝐹𝑧f Normal load on the front axis N –
𝐹𝑧r Normal load on the rear axis N –
𝛼f Front slip angle rad –
𝛼r Rear slip angle rad –

Table 9.2: System parameters

Parameter Definition Value Unit

Fixed parameters
(IPG CarMaker BMW vehicle model)
𝑚 Vehicle mass 1970 kg
𝐼𝑧𝑧 Inertia moment about z-axis 3498 kg/m2
𝑙f CoG∗ to front axis distance 1.4778 m
𝑙r CoG to rear axis distance 1.4102 m
𝐹𝑧f Normal load on the front axis 7926 N
𝐹𝑧r Normal load on the rear axis 8303 N 9
𝛼s Saturation slip angle 0.09 rad
𝑡s Planner sampling time 0.2 s
𝑁p Prediction horizon 10 –
Varying parameters
(measured online)
𝐹max Maximum tire force – N
𝜇 Friction coefficient – –
𝑠0 Initial EV state vector – –
𝑥ref Globally-planned reference trajectory – –
𝑣0 Initial velocity – m/s
𝛿0 Initial steering angle – rad
∗ Center of Gravity
 136 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

ℙ ℙ
[ℙ]MMPS
̂ MMPS
[ℙ]

𝑥 𝑦

Figure 9.5: Conceptual illustration of the Gaussian probability function ℙ, of its MMPS approximation and of the
MMPS proxy functions. The approximations are valid in the compact domain .

which gives the parametric form for [ℙ]MMPS from (9.14) as

[ℙ]MMPS (𝑠(𝑘)) = max min 𝜙𝑇 𝑠(𝑘) + 𝜔𝑝 ) , 0 , (9.22)

(𝑝=1,…,5 ( 𝑝 )

with 𝜙𝑝 being affine functions of 𝜉𝑥 (𝑘), 𝜉𝑦 (𝑘), 𝜎𝑥 (𝑘) and 𝜎𝑦 (𝑘). Similar to ℙ, the MMPS
approximation [ℙ]MMPS is a probability function that is used in the chance constraints.
However, [ℙ]MMPS under-estimates ℙ in regions close to the peak of ℙ, which is not
desired for deriving the collision risk function. To improve safety, we use the MMPS
function [ℙ]
̂ MMPS in Fig. 9.5 as a proxy of [ℙ]MMPS to obtain the risk of collision for each
point on the road in the presence of other road users. This time, we find 𝜙̂𝑝 by approximating
𝑝𝑘 via (9.4) constrained to
(𝜂)
[𝑝̂ 𝑘 ]MMPS (𝑥, 𝑦) ⩾ 𝑝𝑘 (𝑥, 𝑦), ∀(𝑥, 𝑦) ∈ ,

which gives the proxy function

̂ MMPS (𝑠(𝑘)) = max

[ℙ] min 𝜙̂𝑇 𝑠(𝑘) + 𝜔̂ 𝑝 ) , 0 , (9.23)
(𝑝=1,…,5 ( 𝑝 )
9 serving as an over-estimation of ℙ based on [𝑝̂ 𝑘 ]MMPS . Since [𝑝̂ 𝑘 ]MMPS is not a probability
density function, [ℙ] ̂ MMPS is only used to calculate the risk as the cost and does not
serve as an approximation of the probability in evaluating the chance constraint. This
separation allows to avoid conservatism in [ℙ]MMPS within the constraints while seeking
safer trajectories by minimizing the over-approximation [ℙ] ̂ MMPS .
For each time step, the collision risk depends on the probability of the presence of other
road users in (𝑥, 𝑦). Therefore, the collision risk of 𝑠̃(𝑘) can be defined as
𝑁p
1 ̂ (𝜂) (𝑠(𝑘 + 𝑖|𝑘)).
𝑃(̃𝑠(𝑘)) = ∑ max [ℙ] (9.24)
𝑁p 𝑖=1 𝜂={1,…,𝑁o } MMPS

Remark 9.4. The max operator in (9.24) can be replaced by a sum across the presence
probability of all the 𝑁o road users. However, this sum may result in a more conservative
9.4 Problem Reformulation and Solution Approach 137

estimation of the collision risk (The same argument can be deduced using Boole’s inequality).
For instance, if there are two obstacles with a safe corridor in between where ℙ(1) = ℙ(2) = 𝜌,
the sum would give a risk of approximately 2𝜌 for this area, whereas in a real situation, the
chance of two vehicles getting closer is low; furthermore, the real presence probability for both
obstacles would be even lower than 𝜌 which is an estimate that does not take into account the
effect of the presence of one obstacle on the decisions of other road users.

9.4.4 SMPC Optimization Problem

We incorporate the presence probability of obstacles into the MPC planner in two ways:
(𝜂)
first, we ensure a very low probability for the collision by constraining [ℙ]MMPS to be less
than a small threshold 𝜖 > 0. Secondly, we minimize the collision risk function 𝑃 from (9.24)
in the objective function to not only ensure this safety level, but also to converge to the
safest attainable trajectory and to prevent getting close to high-risk areas in a predictive
manner. This in fact will lead to a more proactive response to danger during a hazardous
scenario, which will be illustrated in an example case later.

The stochastic MPC motion planner is formulated as follows: given a globally-planned

reference velocity profile 𝑣̃ref and the initial states 𝑠0 , we find the optimal trajectory 𝑠̃ref by
solving

min 𝑃(̃𝑠) + 𝜆v ‖̃𝑣 − 𝑣̃ref ‖1 + 𝜆u ‖𝑢‖

̃ 1 + 𝜆𝜏 ‖̃𝜏 ‖1 , (9.25a)
𝑠̃,𝑢̃
s.t. 𝑠(𝑘 + 𝑖|𝑘) = [𝑓ego ]MMPS (𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)), ∀𝑖 ∈ {1, … , 𝑁p }, (9.25b)
[𝑔]MMPS (𝑠(𝑘 + 𝑖 − 1|𝑘), 𝑢(𝑘 + 𝑖 − 1)) ⩽ 0, ∀𝑖 ∈ {1, … , 𝑁p }, (9.25c)
{ }
𝜏(𝑖) = min |𝑦ego (𝑘 + 𝑖|𝑘) − 𝑦cj | , ∀𝑖 ∈ {1, … , 𝑁p }, (9.25d)
𝑗=1,…,𝑁lane
(𝜂)
[ℙ]MMPS (𝑠(𝑘 + 𝑖|𝑘)) ⩽ 𝜖, ∀𝜂 ∈ {1, … , 𝑁o }, ∀𝑖 ∈ {1, … , 𝑁p }, (9.25e)
9

where [𝑓ego ]MMPS represents the discretized form of the MMPS system dynamics in (9.21)
and similarly, [𝑔]MMPS approximates the nonlinear constraints such as the Kamm circle.
The objective is to minimize the cost in (9.25a) which consists of the collision probability,
the deviation from the reference velocity, and the control effort. Moreover, the lane-center
deviation 𝜏 is defined over the prediction horizon as (9.25d) which allows switching to a
“better” lane (among 𝑁lane lanes) if necessary. Here, 𝑦cj values represent the center line
in lanes 1 and 2 for as two available lanes for the vehicle on the road and can be easily
extended to include more lanes. Constraints (9.25b) and (9.25e) respectively account for
the prediction model of the ego vehicle and the chance constraints. The Proactive SMPC
(P–SMPC) problem is solved via Algorithms 7 and 8.
 138 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

Algorithm 7 Probability function development

Require: 𝑍(𝑘), Ξ0 , 𝑁p ⊳ 𝑍 contains states of all the obstacles
for 𝜂 ∈ {1, … , 𝑁o } do
𝑧 (𝜂) (𝑘) ← 𝜂th column in 𝑍(𝑘)
Ξ(𝜂) (𝑘|𝑘) ← Ξ0
for 𝑖 ∈ {1, … , 𝑁p } do ⊳ obstacle prediction
(9.12)
Ξ(𝜂) (𝑘 + 𝑖|𝑘) ←−−−−− Ξ0 , Ξ(𝜂) (𝑘 + 𝑖 − 1|𝑘)
(9.10)
𝜉 (𝜂) (𝑘 + 𝑖|𝑘) ←−−−−− 𝑧 (𝜂) (𝑘 + 𝑖 − 1|𝑘)
(𝜂) (9.13)
𝑝𝑖 ←−−−−− 𝜉 (𝜂) (𝑘 + 𝑖|𝑘), Ξ(𝜂) (𝑘 + 𝑖|𝑘)
develop using (9.14) (𝜂)
ℙ(𝜂) (.) ←−−−−−−−−−−−−−−− 𝑝𝑖
(𝜂) (9.22)
[ℙ]MMPS (.) ←−−−−− ℙ(𝜂) (.)
(9.23)
̂ (𝜂) (.) ←−
[ℙ] −−−− ℙ(𝜂) (.)
MMPS
end for
end for
(𝜂) ̂ (𝜂) (.) ∀𝜂 ∈ {1, … , 𝑁o }
return [ℙ]MMPS (.), [ℙ] MMPS

Algorithm 8 The P–SMPC planner

Require: 𝑠(𝑘), [𝑓ego ]MMPS , 𝑍(𝑘), Ξ0 , 𝑣̃ref , 𝑁p , 𝑦c
𝑁lane ← length of 𝑦c
Algorithm 7
̂ MMPS (.) ←−
[ℙ]MMPS (.), [ℙ] −−−−−−−−− 𝑍(𝑘), Ξ0 , 𝑁p
develop using (9.24)
̂ MMPS (.), 𝑁p
𝑃(.) ←−−−−−−−−−−−−−−− [ℙ]
𝑠̃∗ ← solve (9.25) ⊳ the planning optimization problem
return 𝑠̃∗

Remark 9.5. The chance constraints in the SMPC literature [182] are often expressed by
the generic form in (9.2d). In our planner formulation, we use (9.25e) as a more tractable
9 formulation of chance constraints, which is essentially equivalent to bounding the constraint
violation probability in (9.14) or its MMPS approximation (9.22) by a small value 𝜖. Note that
[ℙ]MMPS over-estimates ℙ for probabilities close to zero as shown in Fig. 9.5, and that in (9.25e)
we make sure the collision probability is smaller than 𝜖 for all the states in  and all the time
steps within the prediction horizon.

9.5 Simulations and Results

In this section, we evaluate the control performance of our proposed P–SMPC planner on
two aspects: proactivity of the planner, and attainable generated trajectories. Here, we
select 𝜖 = 0.001 which is the tightest bound investigated in [119]. The P–SMPC optimization
problem defines 10 continuous and 20 binary decision variables per prediction step to
model the ego vehicle. Further, each detected obstacle adds up to 6 binary variables per
9.5 Simulations and Results 139

prediction step to allow for hybrid representation of the collision probability function
associated with it.
The proactivity assessment is done in four highway scenarios where we investigate the
effect of collision-risk minimization in the objective function (9.25a) in our P–SMPC planner
against the optimization formulation inspired by the state-of-the-art [119] indicated as
Regular SMPC (R–SMPC) planner where the collision-risk is not included in the objective
function and the collision is avoided by only considering the left-hand side of (9.25e).
Note that R–SMPC is not the same planner as in [119] since it incorporates the MMPS
approximation of the nonlinearities, but we only change the objective function while
keeping the same dynamic prediction model for both planners for a fair comparison and a
better analysis of the risk-minimization effects. Further, we simulate the SMPC optimization
problem in its nonlinear form as Nonlinear SMPC (N–SMPC) to compare the computation
time against its MILP counterpart, P–SMPC. However, N–SMPC becomes infeasible in the
complex scenario, which is discussed in more detail later.
To assess if the generated trajectory can be attained, we provide the reference trajecto-
ries provided by the P–SMPC planner to a high-fidelity vehicle model in IPG CarMaker [202]
and compare the position and velocity trajectories of the ego vehicle with their references.
The control frequency for all the simulations is set to 1kHz in accordance with the
real-life applications where the computational capabilities limit the operational frequency
of (digital) controllers [12]. The SMPC problems are all designed with sampling time of
0.2s and 𝑁p = 10. We solve the MILPs using the GUROBI [154] optimizer and the NLPs
using the SQP solver in fmincon in a Matlab R2020b environment. For a fair comparison
between the two solvers, we provide the object codes to speed-up the solution time of the
NLPs, which in our simulations, has resulted in up to 20 times faster convergence compared
to providing the objective and the constraints as Matlab functions. The simulations were
run on a PC with a 8-core(s) Intel Xeon 3.60 GHz CPU and 8 GB RAM on Windows 10
64-bit the codes are available from [207].

9.5.1 Proactivity Assessment

In real life, some of the most dangerous situations on a highway are sudden appearance of
a static object or extreme deceleration of a front vehicle. Therefore, we define different
conceptual scenarios with slow-moving vehicles in all of them to present scenarios where
the obstacle is so slow (or even static) that slowing down to keep the distance for collision 9
avoidance is either impossible for the ego vehicle or extremely dangerous. As a result, we
can test the ability of the planner in finding a safe, yet aggressive, evasive maneuver to
avoid the collision. For this, the initial longitudinal velocity of the ego vehicle is considered
to be 22m/s (≈ 80km/h), while the dynamic obstacles are assumed to have initial velocities
between 8 to 11m/s (≈ 30 to 40km/h). Nevertheless, we select the scenarios in a way to
represent challenging, yet possible cases where e.g. other drivers do not aim to collide with
each other, but may behave carelessly.
We use four conceptual scenarios to assess the solutions of the P–SMPC planner:

1. Single obstacle: A slow-moving obstacle is in front of the ego vehicle on the same
lane. We expect the ego vehicle to avoid collision with this obstacle by performing
an evasive maneuver, instead of slowing down to keep a safe distance.
 140 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

2. Dynamic corridor: In addition to an obstacle in the lane as the single-obstacle

scenario, there is another slow-moving vehicle on the other lane to present a situa-
tion where the ego vehicle needs to pass through a corridor between two dynamic
obstacles with stochastic behavior. Here we expect the ego vehicle to pass that
corridor along an optimal trajectory.
3. Static/dynamic corridor: This scenario is similar to the dynamic corridor, except
here we have a static object on the road instead of another slow-moving obstacle.
4. Complex scenario: Here we assess the planner in a situation where there are four
slow-moving vehicles (two on each lane) and one static object present on the road.
There exists a safe corridor between the dynamic and static obstacles, in which we
expect the planner to find an optimal trajectory.
Moreover, each scenario is investigated twice: first as realization (i) where the obstacles
behave ideally as the P–SMPC planner calculates 𝑧ref , i.e. they keep their longitudinal
velocity and lateral position, and secondly as realization (ii) where some/all of them either
change their speed or their lane. Note that in realization (ii), the obstacle’s intention to
change lanes is not known a priori to the ego vehicle, as a result, the SMPC planners keep
the assumption that the obstacle behaves as realization (i).
In total, we have conducted 400 Monte-Carlo simulations by perturbing the initial
speed and the longitudinal distance between the ego vehicle and the obstacles with uniform
sampling within a ±5% range as an acceptable bound from the literature [185]. In Fig. 9.6 ,
four examples are selected as most clear cases to showcase the efficacy of our approach
in a more clear way. The statistical information regarding the Monte-Carlo simulations
can be found in Fig. 9.11a. The ego vehicle is shown in red, while the obstacles are labeled
by the letter “O” and a number to distinguish among them. The solid lines represent the
case where the obstacles move according to the obstacle prediction model and keep their
longitudinal velocity and lateral position. The dashed lines correspond to a case where the
obstacles behave differently than the obstacle prediction model in P–SMPC, e.g. some of
the obstacles on the road are accelerating/decelerating or intending to change their lanes.
The solid red line shows the generated reference trajectory in cases with realization (i),
while the dashed one shows the solution in realization (ii). R–SMPC and N–SMPC results
are shown respectively in gray and blue in a similar fashion.
9
Single obstacle
In the first scenario (Fig. 9.6a), the P–SMPC and R–SMPC planners avoid collision when the
obstacle behaves as predicted by an evasive maneuver. However, the P–SMPC planner keeps
a larger distance with a higher speed compared to the R–SMPC planner that converges to
a trajectory that only satisfies the chance constraints (left-hand side of (9.25e)) and favors
a solution that is closer to the middle of the lane. Note that the higher average velocity is
visible by comparing the length of the red and gray trajectories. If the obstacle intends to
change lanes which is not known to the ego vehicle a priori, the P–SMPC and R–SMPC
planners both keep on assuming that the obstacle will keep its lateral position in each
control step, but after the initial control steps, the planners converge to trajectories on the
same lane as the obstacle merges into the next one. The difference between the planners
is that the P–SMPC planner converges to slightly higher speed (since the red dashed line
9.5 Simulations and Results 141

1–i 1 – ii
𝑦

𝑦
O1 O1
𝑥 𝑥
(a) Scenario 1: single obstacle

2–i 2 – ii
O2 O2
𝑦

𝑦
O1 O1
𝑥 𝑥
(b) Scenario 2: dynamic corridor

3–i 3 – ii
O5 O5
𝑦

O1 O1
𝑥 𝑥
(c) Scenario 3: static/dynamic corridor

4–i O2 4 – ii O2
O3 O3
O5 O5
𝑦

O1 O4 O1 O4
𝑥 𝑥
(d) Scenario 4: complex scenario

P–SMPC case (i) P–SMPC case (ii)

R–SMPC case (i) R–SMPC case (ii)
N–SMPC case (i) N–SMPC case (ii) 9
Obstacle 1 case (i) Obstacle 1 case (ii)
Obstacle 2 case (i) Obstacle 2 case (ii)
Obstacle 3 case (i) Obstacle 3 case (ii)
Obstacle 4 case (i) Obstacle 4 case (ii)
(e) Legend

Figure 9.6: Simulation results for proactivity assessment of the planners. The ego vehicle is shown by a red
rectangle and the fading represents the trajectory evolution over time. Note that obstacle 5 (O5) is static.
 142 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

extends more to the right) to keep more distance from the obstacle. In this sense, P–SMPC
is more proactive as its manages to get out of the hazardous situation while ensuring a
higher safety level.

Dynamic corridor
Figure 9.6b shows the simulation results during the dynamic corridor scenario where both
obstacles are moving. If the obstacles behave as predicted by the ego vehicle and intend to
keep driving on the same lane, the P–SMPC and R–SMPC planners avoid the collision by
overtaking O1 and returning to the center of the right lane. Here, the P–SMPC planner
keeps more distance with O1 since it succeeds in finding a trajectory that has a lower
collision risk than the left-hand side of (9.25e). However, if O2 actually intends to move
to the right lane, after a few control steps when the ego vehicle observes the updated
lateral position of O2, P–SMPC keeps more distance from the center of the right lane and
eventually merges into the left lane as it detects this area to be the safest option. It should
be noted that this is possible due to allowing switching between lanes in (9.25d). Otherwise,
the planners would keep aiming for staying on the right lane which means driving on the
center line between the two lanes until the right lane is risk-free. The R–SMPC planner,
however, is not able to use this potential since it keeps a closer trajectory to the obstacles
and does not search for other trajectories with lower collision-risk, as long as (9.25e) is
satisfied. As a result, P–SMPC is more proactive in the sense of avoiding the propagation
of hazard to the next time steps.

Static/dynamic corridor
In the dynamic/static corridor scenario, both the P–SMPC and R–SMPC planners avoid
colliding with the obstacles by overtaking O1 as shown in Fig. 9.6c, where the P–SMPC
planner keeps a larger distance with the “more uncertain” obstacle (O1). However, if O1
intends to increase its longitudinal velocity, the R–SMPC planner still converges to the
same trajectory since it still satisfies the (9.25e), whereas P–SMPC changes lanes to the safer
track and avoids the collision by overtaking the static obstacle O5 from the left. Similar
to the dynamic corridor, this may lead to hazard propagation to the next steps, a problem
which P–SMPC mitigates by proactive collision avoidance via finding a solution with a
lower collision risk for future time steps.
9
Complex scenario
Figure 9.6d shows the simulations for the complex scenario. If obstacles behave as predicted
by the ego vehicle, the P–SMPC and R–SMPC planners manage to find a solution within
the attainable corridor to avoid collision with the road users. In the final control steps, the
left lane is empty and safer, therefore the P–SMPC planner decides to merge to the left
lane, whereas R–SMPC keeps the same lane. However, if O1 steers to the right and O4
intends to merge into the left lane, the P–SMPC planner decides to stay in the same lane as
the right lane is the safer one and suggests a similar trajectory as planned by the R–SMPC
planner. Figure 9.7 shows the force plots during the complex scenario as an example to
show the capability of the SMPC to operate close to the tire saturation limits. Note that the
velocity of the ego vehicle during the maneuvers is not always constant and is discussed in
more detail in the next section, accompanied by corresponding plots.
9.5 Simulations and Results 143

Note that the N–SMPC planner reaches infeasibility before the end of simulations in
the last three cases, which leads to incomplete trajectories. This phenomenon is a result of
using a warm start strategy (or solution using limited and insufficient number of initial
guesses) which in turn leads to accumulation of errors after a few time steps as follows: in
the complex scenario, the ego vehicle detects the obstacles 2 seconds before reaching their
current position, e.g. O4 is detected after the ego initiates steering to avoid colliding with
O1. Using the shifted solution of the previous time step in such cases leads to a poor result:
as the previous solution was to go back to the initial lateral position after overtaking O1,
by detecting O4, the planner converges to a solution that suggest going back to the initial
lateral position after overtaking O4. Conversely, R–SMPC and P–SMPC planners are able
to find a better solution thanks to their search for a global optimum, which is to brake and
steer to the center of the lane to keep more distance from O4. In the next time steps, O5
is detected, and R–SMPC and P–SMPC manage to find a trajectory to steer to the center
of the lane faster now that an obstacle is in the way. However, the poor solution in the
previous time steps from the N–SMPC planner has resulted in higher longitudinal velocity.
Therefore, the time to collision with O5 is shorter and it is infeasible to find a trajectory to
avoid colliding with O5 with the current velocity.

Front
Rear

3,000
𝐹𝑦 (N)

−3,000

−3,000 0 3,000
𝐹𝑥 (N)

Figure 9.7: Force plot of the complex maneuver with the Kamm circle shown by dashed line. 9

9.5.2 Assessment of Attainable Trajectories

To assess if the trajectories generated by the planner can be attained, we check whether
they can be tracked by a high-fidelity vehicle model. In the proactivity test, eight reference
trajectories were generated in total by the P–SMPC planner. To avoid repetition, we select
four of these trajectories as distinct maneuvers and we simulate the high-fidelity BMW
model in IPG CarMaker [202] to track them. It should be noted that the other trajectories
produced similar results. The selected maneuvers are:

a) Constant-speed overtake: scenario (2-i), the solid red line in Fig. 9.6b,
 144 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

b) Decelerating overtake: scenario (3-i), the solid red line in Fig. 9.6c,

c) Double overtake: scenario (3-ii), the dashed red line in Fig. 9.6c, and

d) Lane change: scenario (2-ii), the dashed red line in Fig. 9.6b.

In each simulation, we give the velocity vector in the four maneuvers to the longitudinal
controller in IPG as the reference velocity profile, and provide the steering angles to the
lateral controller for lateral motion. Figure 9.8 shows comparisons of the 𝑥ego , 𝑦ego , and
𝑣 trajectories obtained by the P–SMPC planner and the resulting trajectory of the IPG
vehicle.

Remark 9.6. We start each IPG simulation from 𝑥ego = 0𝑚 and run a steady, constant
velocity maneuver for 200m to allow for the IPG model to stabilize before tracking the reference
maneuver. As a result, the attainability tests start at 𝑥ego = 200m.

Figure 9.8 shows that the reference trajectories provided by P–SMPC planner are
attainable for the high-fidelity IPG model to track, with slight mismatch along the 𝑋 axis,
which is reasonable considering the larger complexity of the higher-fidelity model in IPG
CarMaker, as compared to the prediction model in the P–SMPC planner.

9.5.3 Comparison with IPG Motion Planner

As the final step, we showcase the proactivity and efficiency of the P–SMPC planner by
comparing its behavior against the built-in collision avoidance module in IPG CarMaker
simulation environment. The test scenario is similar to the complex scenario in Fig. 9.6d
where one static and four slow-moving obstacles are present on the road. This time, we
decrease the obstacle velocities even further down to 2-7 m/s. Moreover, we simulate a
sudden braking by the last obstacle on the road until it stops in a dangerous way.
For a fair comparison, we set a “normal” but “risk-taking” driver behavior in IPG by
selecting a standard driver and the maximum overtaking rate, which means that the driver
always favors evading the obstacles rather than braking. This case shows how an overly-
conservative planning strategy can lead to higher risk and propagating the hazard to other
road users.
In the proactive collision avoidance case, we first ran the simulation in Matlab and
9 used the same TestRun in CarMaker. We arrange the maneuver in IPG such that the
IPG driver merely tracks the speed profile and the steering wheel angle generated by the
P–SMPC planner in Matlab. Notably, we intentionally excluded considerations of other
traffic participants in this scenario to prevent any interference with the operation of the
IPG motion planner. The video of the comparison simulation is accessible online from
https://youtu.be/UacmQDjQ2vI.
Figure 9.9a compares the velocity profiles for the overly-conservative IPG motion
planner and the P–SMPC planner. While the P–SMPC planner manages to keep the velocity
close to the cruising speed, the IPG planner dangerously brakes in multiple occasions.
This issue becomes more critical when the IPG planner decides for a full stop behind the
last obstacle on the right lane as shown in Fig. 9.10a: on the other hand, the P–SMPC
planner manages to safely guide the ego vehicle outside of the risky zone between two
slow-moving vehicles by taking a proactive strategy to overtake the stopping vehicle as
9.5 Simulations and Results 145

5
Planner IPG
22
𝑦ego (m)

𝑣 (m/s)
0 21
20
−5
220 240 260 280 300 320 340 1 3 5
𝑥ego (m) 𝑡 (s)

(a) Constant-speed overtake

5
Planner IPG
22
𝑦ego (m)

𝑣 (m/s)
0 21
20
−5
220 240 260 280 300 320 340 1 3 5
𝑥ego (m) 𝑡 (s)

(b) Decelerating overtake

5
22
𝑦ego (m)

𝑣 (m/s)

0 21
Planner IPG 20
−5
220 240 260 280 300 320 340 1 3 5
𝑥ego (m) 𝑡 (s)

(c) Double overtake

5
Planner IPG
22
9
𝑦ego (m)

𝑣 (m/s)

0 21
20
−5
220 240 260 280 300 320 340 1 3 5
𝑥ego (m) 𝑡 (s)

(d) Lane change

Figure 9.8: Simulation results for attainability assessment of the P–SMPC planner.
 146 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

well as by keeping a safe distance from the other slow-moving obstacle on the left lane in
Fig. 9.10b.

25 P–SMPC IPG

20
𝑣 (m/s)

15

10

0
400 600 800 1,000
𝑥ego (m)

(a) Velocity profiles of the ego vehicle along the road

⋅10−3
𝑃
4 max{[ℙ]MMPS }
Probability/Risk

0
0 10 20 30 40
𝑡 (s)

(b) Risk and maximum [ℙ]MMPS values for P–SMPC

Figure 9.9: Plots of comparative test between overly-conservative and proactive collision avoidance.

9
9.5.4 Performance Analysis and Discussion
In the previous sections, we showed the proactivity of our proposed P–SMPC motion
planner by comparing its performance against the state-of-the-art SMPC formulation (R–
SMPC) and the built-in motion planner in a high-fidelity modeling and simulation platform.
To gain a more clear view of the planning performance of P–SMPC, we have collected the
data from all the aforementioned simulations and plotted the time evolution of chance
constraints and the risk function values and the density histogram for computation time in
Fig. 9.11. Since the simulations have various lengths in terms of time, we have scaled their
data to a risky zone and a safe zone in Figures 9.11a and 9.11b to allow for a meaningful
comparison. The risky zone represents the section of the simulations where the ego vehicle
observes sudden appearance of the obstacles and ends when it does not detect any obstacles
9.5 Simulations and Results 147

(a) Overly-conservative collision avoidance: the ego vehicle (b) Proactive collision avoidance by P–SMPC: the ego vehicle
slows down to keep distance until a full stop behind the obstacle. manages to get out of the risky zone before its front vehicle
stops.

Figure 9.10: Snapshots of overly-conservative (a) and proactive (b) collision avoidance planning strategies.

ahead on the road.

Figure 9.11a shows the statistical information of [ℙ]MMPS values in the Monte-Carlo
simulation results for the R–SMPC and P–SMPC planners. The maximum values for both
planners is 0.001 (0.1%) as shown in gray. Both planners show a reduction of the maximum
[ℙ]MMPS value by getting out of the risky zone. However, the mean for [ℙ]MMPS values
for P–SMPC are significantly lower than the mean values for R–SMPC, which shows the
effectiveness of minimizing a risk function based on over-approximation of the ℙ within
the SMPC formulation. The peak in the mean value for P–SMPC corresponds to the riskiest
time steps during the simulation, which occur where the vehicle is closest to the obstacle,
e.g. during an overtaking maneuver. Further, the InterQuartile Range (IQR) distance for
the planners is shown by the width of a shaded area around the mean values, using their
corresponding colors.
The risk function values for P–SMPC planner are plotted in Fig. 9.11b. Since the risk
function is an over-approximation of ℙ, its value are higher than [ℙ]MMPS . Nevertheless,
the P–SMPC planner manages to keep the risk function below 0.0045 (0.45%) at all times in
Fig. 9.9b due to its predictive proactive collision avoidance. In addition, while convergence 9
to a global optimum cannot be guaranteed for an NLP, an MILP solver can reach its global
optimum when it is given sufficient time. As a result, the MILP formulation of the (originally
nonlinear) SMPC planning optimization problem improves the computational efficiency by
a speed-up in computations and a better coverage of the decision space.
Lastly, the density histogram for computation time per planning step is shown in
Fig. 9.11c. Compared to the planner sampling time of 0.2s, the MILP solver could find the
global optimum 96% of the times within 0.15s (75% of the time step) on the PC used for
the test and only 4% of the times required more than 0.2s to find the global optimum. This
shows the computational efficiency of the P–SMPC planning formulation, which can be
further improved by imposing a time limit for the solver (and trading the global optimality)
or running the simulations on a faster machine. Note that this level of computational
efficiency is achieved for the assumed model and approximation accuracy adopted in this
 148 9 Proactive Collision Avoidance with Stochastic Obstacle Behavior

⋅10−3
1 Maximum value
R–SMPC IQR
P–SMPC IQR

[ℙ]MMPS
R–SMPC Mean
0.5 P–SMPC Mean

0
Risky zone Safe zone
Simulation time/step
(a) Evolution of [ℙ]MMPS values

⋅10−3

4
Risk function

0
Risky zone Safe zone
Simulation time/step
(b) Evolution of maximum risk function values (P–SMPC)

P–SMPC
Frequency density

10
N–SMPC
Sampling time

0
0.1 0.2 0.5 1 2 3.5
Computation time per step (s)
9 (c) Density histogram for computation times. The N–SMPC computation times for steps
that the NLP was infeasible are not considered and the data only account for the duration
of sampling times where the planner converged to a solution.

Figure 9.11: Performance analysis of the P–SMPC planner in terms of safety and computation time. The data
in these plots represent the density histograms of their corresponding variables considering all the performed
simulations in this study.

chapter. For a more comprehensive study of control performance vs. computational speed
trade-off in hybridization of NMPC using MMPS formalism, the reader is referred to our
previous study [85, 86].
9.6 Conclusions 149

9.6 Conclusions
This chapter has presented a novel SMPC motion planner for emergency collision avoidance
during hazardous highway scenarios. The proposed planner proactively avoids collision by
static and dynamic obstacles on a highway by avoiding conservatism and swift response
to sudden appearance of road users with uncertain behavior, thus improving the safety of
the ego vehicle.
The novelties of our proposed approach can be summarized as follows: first, the proac-
tive SMPC planner uses a tractable formulation of chance constraints for safe collision
avoidance, while minimizing a risk function formulated as an over-estimation of the prob-
abilities while facilitating the incorporation of a dynamic model for the ego vehicle as
well as exploiting the tire-force potential close to the vehicle handling limits. Secondly,
hybrid approximations of the nonlinearities in the system dynamics by the MMPS for-
malism are used to allow for an MILP formulation of the SMPC problem and facilitate
real-time implementation and convergence to the global optimum. Safety, proactivity, and
computational efficiency of our proposed planned were shown via various simulations of
emergency scenarios and compared against the state-of-the-art SMPC formulation and a
high-fidelity vehicle modeling and simulation environment.
For future work, we aim at improving the model for dynamic obstacles on the road
and extending the uncertainty regarding the intention of the other road users. While the
model employed in this chapter for the obstacles helped obtain an efficient computational
accuracy-speed trade-off, more comprehensive models of obstacle behavior are influential
for implementation of levels 4 and 5 of automated driving. Further, we aim at integrated
planning and control design for emergency scenarios for improved accuracy and computa-
tional efficiency, in addition to investigating an efficient control structure to integrate our
proposed SMPC planner with hybrid vehicle control and a friction estimator to account for
the uncertainties of the environment as well. Moreover, in-depth calibration of probability
bounds, investigation of suboptimality bounds, feasibility analysis of the SMPC problem
for different probability formulations, and proof of recursive feasibility will be important
topics for our future research, as well as designing a back-up mode in cases where the
feasibility of the planning optimization problem cannot be guaranteed.

9
 151

10
Conclusions
Sometimes the answer to a problem is neither complex nor simple, just unexpected.

— Rouzbeh Moein, The Cold Coffee of Mr. Writer

This chapter serves as the conclusion to the thesis, starting with a summary of the
research contributions. We then provide recommendations for future research, along with
a discussion on the automated driving outlook, keeping an eye on its societal relevance
and technological progress.

10.1 Summary of Research Contributions

In this thesis, we have addressed the problem of controlling evasive maneuvers in hazardous
scenarios for automated driving. In particular, we approached the problem focusing on a
proactive response to hazard, taking into account the need for swift and optimal response
to unforeseen emergencies on highways during a driving task. In summary, our research
contributions can be categorized as follows:

• two cut-based PieceWise Affine (PWA) approximation approaches have been

proposed to allow for a flexible approximation of multi-dimensional nonlinear sys-
tems in Chapters 2 and 3,
• an open-source toolbox named H4MPC has been introduced in Chapter 4 to offer
a user-friendly interface for hybridizing Model Predictive Control (MPC) problems, 10
• a sensitivity analysis of PWA approximations was conducted in Chapter 5 for a
polytopically-constrained NonLinear Program (NLP) to obtain quantitative bounds
on the distance between the original and the approximated local minima,
• a hybridization benchmark for MPC optimization problems using the Max-
Min-Plus-Scaling (MMPS) formalism has been presented in Chapter 6, followed by a
comparative assessment of the computational performance of the hybridized MPC
optimization problems in Chapter 7,
 152 10 Conclusions

• experimental insights into real-time implementation have been discussed for

emergency collision avoidance after the sudden appearance of a static obstacle in
Chapter 8,

• a proactive Stochastic Model Predictive Control (SMPC) approach for colli-

sion avoidance in highway scenarios has been presented to minimize the collision
risk and enhances safety using chance constraints in Chapter 9.

10.2 Future Research Suggestions

For the continued progress of automated driving research, it is essential that future research
explores certain aspects in greater detail. For instance, the following points should be
considered:

Realistic scenarios acknowledge a wide range of uncertainties. Therefore, it is

essential to introduce further levels of uncertainty, such as tire-road friction coefficients,
pneumatic pressure of the tires, and load distribution in the ego vehicle. Additionally,
unforeseen events, such as abrupt weather changes or road blockage caused by a sudden
snow accumulation, should be considered in future research steps.

Hazard extends beyond configuration of the vehicles. Refining, as well as expanding

the definition of hazardous cases is essential for advancing automated driving research.
Although various threat assessment measures have been proposed in the literature, the
successful implementation of automated vehicles demands a broader understanding of
what constitutes a hazardous scenario. For example, in a Level 3 automated vehicle, a
danger alarm triggered by detecting a pedestrian may cause the human driver to panic and
lose control; resulting in a situation with a higher collision risk than the mere presence of
the pedestrian. Therefore, hazards must account for not only metrics such as time/distance
to collision but also other influential parameters, such as the state of the driver.

Human-replacing technology calls for a human-inspired design. In this context,

studying human behavior, such as the decision-making process of average, as well as highly-
skilled drivers like Formula 1 participants, allows for incorporating crucial insights, from
modeling to the application of more realistic probability distributions, into the automated
driving planning and control systems.

10 It is time for nomenclature convergence. The development of more comprehensive

and improved approximation techniques is becoming increasingly important, especially
as the complexity of the control problems rises with the targeted automation level. The
new approximation methods should focus on delivering faster results while minimizing
computational complexity. In light of this, closer collaboration with the learning community
is essential as the fields of function approximation and learning often deal with similar
types of optimization problems. However, the use of different terminologies (e.g. error
function vs. loss function, parameter estimation vs training, and adaptive approximation vs.
online learning) in each field creates a significant barrier. This mismatch in nomenclature
10.3 Outlook and Recommendations 153

not only makes communication difficult but also causes researchers to miss out on insights
from the other field. To create true synergy, these gaps should be bridged by aligning the
vocabularies among the fields of optimization, approximation, and learning.

Technology matters. Technological advancements have substantial impact on the

progress of the field. To name a few, enhancing sensing technologies, hardware pro-
cessing speeds and memory capacity, and reducing overall energy consumption are all
significant aspects of advancing automated driving. Additionally, rethinking mechanical
design could provide further advancements in this area as well. For instance, similar to
how adaptive suspension transformed chassis control in ground vehicles or load-sensing
bearings developed by SKF significantly improved slip-control, introducing concepts such
as moving loads in vehicle dynamics could enhance safety during evasive maneuvers.

10.3 Outlook and Recommendations

Moving beyond the technical aspects, the successful implementation of automated driving
necessitates addressing the following higher-level considerations, as they touch on the
broader view of the future of automated driving and its integration into everyday life.

Global problems require global solutions. The implementation of automated driving

presents a global challenge: a Level 5 automated vehicle must perform well in diverse
environments—from the highways of San Diego and Santiago to the congested streets of
Dar es Salaam, Kathmandu, and Amsterdam, and even the rural areas surrounding Brisbane.
Achieving this will require the collaboration among different global sectors, as each region
comes with its own set of regulations, levels of public trust, and environmental challenges
such as weather patterns and infrastructure variations.

Let us not forget who automated cars are made for. It is crucial to investigate the
foundations of public trust and ensure that it is addressed within a global, diversity-informed
framework. This means that the diverse backgrounds and needs of future users must be
considered at every stage, from system design to production. For instance, understanding
how individuals from different cultural, social, and demographic backgrounds perceive and
trust autonomous vehicles is essential.

Humans do not trust black boxes. In other words, people are less likely to trust
something they do not understand. To ensure the acceptance of automated driving systems,
it is crucial to provide the public with clear and transparent information about how these
technologies work. In this sense, research findings must be communicated effectively,
not only via open science, but also through public talks and workshops to address public
concerns and align with their expectations.
 155

Bibliography
[1] S. Liu, K. Zheng, L. Zhao, and P. Fan. A driving intention prediction method based
on hidden Markov model for autonomous driving. Computer Communications,
157:143–149, 2020.

[2] I. Batkovic, M. Ali, P. Falcone, and M. Zanon. Safe trajectory tracking in uncertain
environments. IEEE Transactions on Automatic Control, 68(7):4204–4217, 2023.

[3] X. Zhang, A. Liniger, and F. Borrelli. Optimization-based collision avoidance. IEEE

Transactions on Control Systems Technology, 29(3):972–983, 2021.

[4] Introduction: From highly automated to autonomous driving. In G. Dimitrakopoulos,

A. Tsakanikas, and E. Panagiotopoulos, editors, Autonomous Vehicles, pages xiii–xxiii.
Elsevier, 2021.

[5] Society of Automotive Engineers (SAE International). Taxonomy and definitions

for terms related to on-road motor vehicle automated driving systems, Sep, 3,
2017. https://web.archive.org/web/20170903105244/https:/www.sae.org/misc/pdfs/
automated_driving.pdf.

[6] I. Dickson. How the supply chain is pioneering autonomous driving, 2013. [Online;
accessed October 2, 2024].

[7] U. Ju and S. Kim. Willingness to take responsibility: Self-sacrifice versus sacrificing

others in takeover decisions during autonomous driving. Heliyon, 10(9):e29616, 2024.

[8] N. Chowdhri, L. Ferranti, F. S. Iribarren, and B. Shyrokau. Integrated nonlinear model

predictive control for automated driving. Control Engineering Practice, 106:104654,
2021.

[9] V. A. Laurense and J. C. Gerdes. Long-horizon vehicle motion planning and control
through serially cascaded model complexity. IEEE Transactions on Control Systems
Technology, 30(1):166–179, 2022.

[10] J. Wurts, J. L. Stein, and T. Ersal. Design for real-time nonlinear model predictive
control with application to collision imminent steering. IEEE Transactions on Control
Systems Technology, 30(6):2450–2465, 2022.

[11] H. Liu, L. Zhang, P. Wang, and H. Chen. A real-time NMPC strategy for electric
vehicle stability improvement combining torque vectoring with rear-wheel steering.
IEEE Transactions on Transportation Electrification, 8(3):3825–3835, 2022.
156 Bibliography

[12] P. Stano, U. Montanaro, D. Tavernini, M. Tufo, G. Fiengo, L. Novella, and A. Sorniotti.

Model predictive path tracking control for automated road vehicles: A review. Annual
Reviews in Control, 55:194–236, 2023.

[13] G. Bellegarda and Q. Nguyen. Dynamic vehicle drifting with nonlinear MPC and a
fused kinematic-dynamic bicycle model. IEEE Control Systems Letters, 6:1958–1963,
2022.

[14] Manan S. Gandhi, Bogdan Vlahov, Jason Gibson, Grady Williams, and Evangelos A.
Theodorou. Robust model predictive path integral control: analysis and performance
guarantees. IEEE Robotics and Automation Letters, 6(2):1423–1430, 2021.

[15] S. Dixit, U. Montanaro, M. Dianati, D. Oxtoby, T. Mizutani, A. Mouzakitis, and

S. Fallah. Trajectory planning for autonomous high-speed overtaking in structured
environments using robust MPC. IEEE Transactions on Intelligent Transportation
Systems, 21(6):2310–2323, 2020.

[16] Q. Shi, Ji. Zhao, A. E. Kamel, and I. Lopez-Juarez. MPC based vehicular trajectory
planning in structured environment. IEEE Access, 9:21998–22013, 2021.

[17] H. Atoui, O. Sename, V. Milanés, and J. J. Martinez. LPV-based autonomous vehicle

lateral controllers: A comparative analysis. Technical Report hal-03197895, 2021.
https://hal.archives-ouvertes.fr/hal-03197895.

[18] J. Lunze and F. Lamnabhi-Lagarrigue. Handbook of Hybrid Systems Control: Theory,

Tools, Applications. Cambridge University Press, 2009.

[19] M. Pčolka, E. Žáčeková, S. Čelikovský, and M. Šebek. Toward a smart car: Hybrid
nonlinear predictive controller with adaptive horizon. IEEE Transactions on Control
Systems Technology, 26(6):1970–1981, 2018.

[20] Y. Zheng and B. Shyrokau. A real-time nonlinear MPC for extreme lateral stabilization
of passenger vehicles. In IEEE International Conference on Mechatronics (ICM 2019),
pages 519–524, 2019.

[21] X. Sun, Y. Cai, S. Wang, X. Xu, and L. Chen. Optimal control of intelligent vehicle
longitudinal dynamics via hybrid model predictive control. Robotics and Autonomous
Systems, 112:190–200, 2019.

[22] L. Gharavi, B. De Schutter, and S. Baldi. Parametric piecewise-affine approximation

of nonlinear systems: A cut-based approach. IFAC-PapersOnLine, 56(2):6666–6671,
2023.

[23] W. P. M. H. Heemels, B. De Schutter, and A. Bemporad. Equivalence of hybrid

dynamical models. Automatica, 37(7):1085–1091, 2001.

[24] B. De Schutter, T. van den Boom, J. Xu, and S. S. Farahani. Analysis and control of
max-plus linear discrete-event systems: An introduction. Discrete Event Dynamic
Systems: Theory and Applications, 30:25–54, 2020.
Bibliography 157

[25] R. M. F.ra Cândido, L. Hardouin, M. Lhommeau, and R. S. Mendes. Conditional

reachability of uncertain max plus linear systems. Automatica, 94:426–435, 2018.

[26] A. Bemporad and M. Morari. Control of systems integrating logic, dynamics, and
constraints. Automatica, 35(3):407–427, 1999.

[27] Richard Oberdieck and Efstratios N. Pistikopoulos. Explicit hybrid model-predictive

control: The exact solution. Automatica, 58:152–159, 2015.

[28] B. De Schutter and T. J.J. van den Boom. MPC for continuous piecewise-affine
systems. Systems and Control Letters, 52(3-4):179–192, 2004.

[29] G. Ferrari-Trecate, F. A. Cuzzola, D. Mignone, and M. Morari. Analysis of discrete-

time piecewise affine and hybrid systems. Automatica, 38(12):2139–2146, 2002.

[30] M. Hovd and S. Olaru. Parameter-dependent PWQ Lyapunov function stability

criteria for uncertain piecewise linear systems. Modeling, Identification and Control,
39:15–21, 2018.

[31] V. Sessa, L. Iannelli, F. Vasca, and V. Acary. A complementarity approach for the
computation of periodic oscillations in piecewise linear systems. Nonlinear Dynamics,
85:1255–1273, 2016.

[32] D. Ito, T. Ueta, T. Kousaka, and K. Aihara. Bifurcation analysis of the Nagumo-Sato
model and its coupled systems. International Journal of Bifurcation and Chaos, 26,
2016.

[33] T. Suzuki and K. Aihara. Nonlinear system identification for prostate cancer and
optimality of intermittent androgen suppression therapy. Mathematical Biosciences,
245:40–48, 2013.

[34] B. Joseph-Duran, M. N. Jung, C. Ocampo-Martinez, S. Sager, and G. Cembrano.

Minimization of sewage network overflow. Water Resources Management, 28:41–63,
2014.

[35] X. Sun, We. Hu, Y. Cai, P.K. Wong, and L. Chen. Identification of a piecewise affine
model for the tire cornering characteristics based on experimental data. Nonlinear
Dynamics, 101:857–874, 2020.

[36] D. Jagga, M. Lv, and S. Baldi. Hybrid adaptive chassis control for vehicle lateral
stability in the presence of uncertainty. In Mediterranean Conference on Control and
Automation (MED 2018), pages 529–534, 2018.

[37] S. Azuma, J. Imura, and T. Sugie. Lebesgue piecewise affine approximation of

nonlinear systems. Nonlinear Analysis: Hybrid Systems, 4(1):92–102, 2010.

[38] S. Kersting and M. Buss. Recursive estimation in piecewise affine systems using
parameter identifiers and concurrent learning. International Journal of Control,
92(6):1264–1281, 2019.
158 Bibliography

[39] Y. Du, F. Liu, J. Qiu, and M. Buss. Online identification of piecewise affine systems
using integral concurrent learning. IEEE Transactions on Circuits and Systems I:
Regular Papers, 68(10):4324–4336, 2021.
[40] L. Bako. Subspace clustering through parametric representation and sparse opti-
mization. IEEE Signal Processing Letters, 21(3):356–360, 2014.
[41] E. Khanmirza, M. Nazarahari, and A. Mousavi. Identification of piecewise affine
systems based on fuzzy PCA-guided robust clustering technique. Eurasip Journal on
Advances in Signal Processing, 2016:1–15, 2016.
[42] L. Q. Thuan, T. van den Boom, and S. Baldi. Online identification of continuous
bimodal and trimodal piecewise affine systems. In European Control Conference (ECC
2016), pages 1075–1070, 2016.
[43] A. Hartmann, J. M. Lemos, R. S. Costa, J. Xavier, and S. Vinga. Identification of
switched ARX models via convex optimization and expectation maximization. Jour-
nal of Process Control, 28:9–16, 2015.
[44] Delft High Performance Computing Centre (DHPC). DelftBlue Supercomputer (Phase
1), 2022. ARK: ark:/44463/DelftBluePhase1.
[45] F. Lauer and G. Bloch. Hybrid System Identification: Theory and Algorithms for
Learning Switching Models, volume 478 of Lecture Notes in Control and Information
Sciences. Springer International Publishing, 2019.
[46] V. V. Gorokhovik, O. I. Zorko, and G. Birkhoff. Piecewise affine functions and
polyhedral sets. Optimization, 31(3):209–221, 1994.
[47] F. Bayat, T. A. Johansen, and A. A. Jalali. Flexible piecewise function evaluation
methods based on truncated binary search trees and lattice representation in explicit
MPC. IEEE Transactions on Control Systems Technology, 20(3):632–640, 2012.
[48] A. Gersnoviez, M. Brox, and I. Baturone. High-speed and low-cost mplementation of
explicit model predictive controllers. IEEE Transactions on Control Systems Technology,
27(2):647–662, 2019.
[49] D. Bertsimas and J. Dunn. Machine Learning Under a Modern Optimization Lens.
Dynamic Ideas LLC, 2019.
[50] F. Lauer. On the complexity of piecewise affine system identification. Automatica,
62:148–153, 2015.
[51] P. Brox, J. Castro-Ramirez, M. C. Martinez-Rodriguez, E. Tena, C. J. Jimenez, I. Batur-
one, and A. J. Acosta. A programmable and configurable ASIC to generate piecewise-
affine functions defined over general partitions. IEEE Transactions on Circuits and
Systems I: Regular Papers, 60(12):3182–3194, 2013.
[52] X. Sun, P. Wu, Y. Cai, S. Wang, and L. Chen. Piecewise affine modeling and hybrid
optimal control of intelligent vehicle longitudinal dynamics for velocity regulation.
Mechanical Systems and Signal Processing, 162, 2022.
Bibliography 159

[53] G. F. V. Amaral, C. Letellier, and L. A. Aguirre. Piecewise affine models of chaotic

attractors: The Rössler and Lorenz systems. Chaos: An Interdisciplinary Journal of
Nonlinear Science, 16(1), 2006.

[54] T. Ohtsuki, T. Fujisawa, and S. Kumagai. Existence theorems and a solution algorithm
for piecewise-linear resistor networks. SIAM Journal on Mathematical Analysis,
8(1):69–99, 1977.

[55] C. Wen and X. Ma. A max-piecewise-linear neural network for function approxima-
tion. Neurocomputing, 71:843–852, 2008.

[56] A. Bemporad. A piecewise linear regression and classification algorithm with appli-
cation to learning and model predictive control of hybrid systems. IEEE Transactions
on Automatic Control, 68(6):3194–3209, 2023.

[57] X. Tang and Y. Dong. Expectation maximization based sparse identification of cyber-
physical system. International Journal of Robust and Nonlinear Control, 31(6):2044–
2060, 2021.

[58] S. Paoletti, A. L. Juloski, G. Ferrari-Trecate, and R. Vidal. Identification of hybrid

systems: A tutorial. European Journal of Control, 13:242–260, 2007.

[59] A. Moradvandi, R.E.F. Lindeboom, E. Abraham, and B. De Schutter. Models and

methods for hybrid system identification: a systematic survey. IFAC-PapersOnLine,
56(2):95–107, 2023.

[60] S. Paoletti, I. Savelli, A. Garulli, and A. Vicino. A bilevel programming framework

for piecewise affine system identification. In IEEE Conference on Decision and Control
(CDC), pages 7376–7381. IEEE, 2019.

[61] J. F. Bard. Practical bilevel optimization: algorithms and applications. Springer, 2011.

[62] P. Mattsson, D. Zachariah, and P. Stoica. Recursive identification method for piece-
wise ARX models: A sparse estimation approach. IEEE Transactions on Signal
Processing, 64(19):5082–5093, 2016.

[63] P. Mattsson, D. Zachariah, and P. Stoica. Recursive nonlinear-system identification

using latent variables. Automatica, 93:343–351, 2018.

[64] V. Breschi, D. Piga, and A. Bemporad. Piecewise affine regression via recursive
multiple least squares and multicategory discrimination. Automatica, 73:155–162,
2016.

[65] Z. Jin, Q. Shen, and S. Z. Yong. Mesh-based piecewise affine abstraction with polytopic
partitions for nonlinear systems. IEEE Control Systems Letters, 5(5):1543–1548, 2021.

[66] A. Oliveri, A. Canepa, L. Queirolo, and M. Storace. An algorithm for automatic

domain partitioning of piecewise-affine model predictive control laws. In IEEE
International Conference on Electronics, Circuits, and Systems (ICECS), pages 217–220,
2013.
160 Bibliography

[67] E. Amaldi, S. Coniglio, and L. Taccari. Discrete optimization methods to fit piecewise
affine models to data points. Computers & Operations Research, 75:214–230, 2016.
[68] Y. Du, F. Liu, J. Qiu, and M. Buss. A semi-supervised learning approach for iden-
tification of piecewise affine systems. IEEE Transactions on Circuits and Systems I:
Regular Papers, 67(10):3521–3532, 2020.
[69] F. Comaschi, B. A. G. Genuit, A. Oliveri, W. P. M. H. Heemels, and M. Storace.
FPGA implementations of piecewise affine functions based on multi-resolution
hyperrectangular partitions. IEEE Transactions on Circuits and Systems I: Regular
Papers, 59(12):2920–2933, 2012.
[70] A. Bemporad, A. Oliveri, T. Poggi, and M. Storace. Ultra-fast stabilizing model
predictive control via canonical piecewise affine approximations. IEEE Transactions
on Automatic Control, 56(12):2883–2897, 2011.
[71] J. Roll, A. Bemporad, and L. Ljung. Identification of piecewise affine systems via
mixed-integer programming. Automatica, 40(1):37–50, 2004.
[72] J. Xu, T. J.J. van den Boom, B. De Schutter, and S. Wang. Irredundant lattice repre-
sentations of continuous piecewise affine functions. Automatica, 70:109–120, 2016.
[73] L. Gharavi, B. De Schutter, and S. Baldi. Parametric piecewise-affine approximation
of nonlinear systems: A cut-based approach. IFAC-PapersOnLine, 56(2):6666–6671,
2023.
[74] T. Zaslavsky. Facing Up to Arrangements: Face-Count Formulas for Partitions of Space
by Hyperplanes, volume 154. American Mathematical Society, 1975.
[75] L. E. Blumenson. A derivation of n-dimensional spherical coordinates. The American
Mathematical Monthly, 67(1):63–66, 1960.
[76] H. Edelsbrunner. Algorithms in Combinatorial Geometry, volume 10. Springer Science
& Business Media, 1987.
[77] P. Orlik and H. Terao. Arrangements of hyperplanes. Springer Science & Business
Media, 2013.
[78] T. Fujisawa and E. S. Kuh. Piecewise-linear theory of nonlinear networks. SIAM
Journal on Applied Mathematics, 22(2):307–328, 1972.
[79] L. Gharavi, B. De Schutter, and S. Baldi. H4MPC: A hybridization toolbox for model
predictive control in automated driving. In IEEE International Conference on Advanced
Motion Control (AMC), pages 1–6, 2024.
[80] N. Groot, B. De Schutter, and H. Hellendoorn. Integrated model predictive traffic and
emission control using a piecewise-affine approach. IEEE Transactions on Intelligent
Transportation Systems, 14(2):587–598, 2013.
[81] E. Asarin, T. Dang, and A. Girard. Hybridization methods for the analysis of nonlinear
systems. Acta Informatica, 43(7):451, 2007.
Bibliography 161

[82] F.D. Torrisi and A. Bemporad. HYSDEL-a tool for generating computational hybrid
models for analysis and synthesis problems. IEEE Transactions on Control Systems
Technology, 12(2):235–249, 2004.

[83] S. Di Cairano, H. E. Tseng, D. Bernardini, and A. Bemporad. Vehicle yaw stability

control by coordinated active front steering and differential braking in the tire sideslip
angles domain. IEEE Transactions on Control Systems Technology, 21(4):1236–1248,
2013.

[84] N. Guo, X. Zhang, Y. Zou, B. Lenzo, and T. Zhang. A computationally efficient

path-following control strategy of autonomous electric vehicles with yaw motion
stabilization. IEEE Transactions on Transportation Electrification, 6(2):728–739, 2020.

[85] L. Gharavi, B. De Schutter, and S. Baldi. Efficient MPC for emergency evasive maneu-
vers, part I: Hybridization of the nonlinear problem. arXiv preprint arXiv:2310.00715,
2023.

[86] L. Gharavi, B. De Schutter, and S. Baldi. Efficient MPC for emergency evasive
maneuvers, part II: Comparative assessment for hybrid control. arXiv preprint
arXiv:2310.00716, 2023.

[87] L. Gharavi. Hybridization Toolbox for Model Predictive Control. 4TU.ResearchData,

2023.

[88] C. G. Bobier-Tiu, C. E. Beal, J. C. Kegelman, R. Y. Hindiyeh, and J. C. Gerdes. Vehicle

control synthesis using phase portraits of planar dynamics. Vehicle System Dynamics,
57(9):1318–1337, 2019.

[89] R. Rajamani. Vehicle Dynamics and Control. Springer Science & Business Media,
2011.

[90] H. Pacejka. Tire and Vehicle Dynamics. Elsevier, 2005.

[91] A. Kripfganz and R. Schulze. Piecewise affine functions as a difference of two convex
functions. Optimization, 18(1):23–29, 1987.

[92] L. Gharavi, B. De Schutter, and S. Baldi. Sensitivity analysis for piecewise-affine

approximations of nonlinear programs with polytopic constraints. IEEE Control
Systems Letters, (XX), 2024.

[93] S. Gros, M. Zanon, R. Quirynen, A. Bemporad, and M. Diehl. From linear to nonlinear
MPC: bridging the gap via the real-time iteration. International Journal of Control,
93(1):62–80, 2020.

[94] D. W. Griffith, L. T. Biegler, and S. C. Patwardhan. Robustly stable adaptive horizon

nonlinear model predictive control. Journal of Process Control, 70:109–122, 2018.

[95] L. Dong, J. Yan, X. Yuan, H. He, and C. Sun. Functional nonlinear model predictive
control based on adaptive dynamic programming. IEEE Transactions on Cybernetics,
49(12):4206–4218, 2019.
162 Bibliography

[96] V. M Zavala and L. T. Biegler. The advanced-step NMPC controller: Optimality,

stability and robustness. Automatica, 45(1):86–93, 2009.

[97] J. Xu, X. Huang, X. Mu, and S. Wang. Model predictive control based on adaptive
hinging hyperplanes model. Journal of Process Control, 22(10):1821–1831, 2012.

[98] L. Gharavi, A. Dabiri, J. Verkuijlen, B. De Schutter, and S. Baldi. Proactive emergency

collision avoidance for automated driving in highway scenarios. arXiv preprint,
2310.17381, 2023.

[99] Y. Chen, M. Bruschetta, D. Cuccato, and A. Beghi. An adaptive partial sensitivity

updating scheme for fast nonlinear model predictive control. IEEE Transactions on
Automatic Control, 64(7):2712–2726, 2019.

[100] M. Ławryńczuk, P. M. Marusak, P. Chaber, and D. Seredyński. Initialisation of

optimisation solvers for nonlinear model predictive control: Classical vs. hybrid
methods. Energies, 15(7):2483, 2022.

[101] T. Ghandriz, B. Jacobson, P. Nilsson, and L. Laine. Trajectory-following and off-

tracking minimisation of long combination vehicles: A comparison between non-
linear and linear model predictive control. Vehicle System Dynamics, pages 1–34,
2023.

[102] J. W. Daniel. Stability of the solution of definite quadratic programs. Mathematical

Programming, 5(1):41–53, 1973.

[103] H. X. Phu. Minimizing convex functions with bounded perturbations. SIAM Journal
on Optimization, 20(5):2709–2729, 2010.

[104] C. Büskens and H. Maurer. Sensitivity analysis and real-time optimization of paramet-
ric nonlinear programming problems. In M. Grötschel, S. O. Krumke, and J. Rambau,
editors, Online Optimization of Large Scale Systems, pages 3–16. Springer Berlin
Heidelberg, 2001.

[105] I. Subotic, A. Hauswirth, and F. Dorfler. Quantitative sensitivity bounds for nonlinear
programming and time-varying optimization. IEEE Transactions on Automatic Control,
67(6):2829–2842, 2022.

[106] S. Shin and V. M. Zavala. Diffusing-horizon model predictive control. IEEE Transac-
tions on Automatic Control, 68(1):188–201, 2023.

[107] H. Pirnay, R. López-Negrete, and L. T. Biegler. Optimal sensitivity based on IPOPT.

Mathematical Programming Computation, 4(4):307–331, 2012.

[108] M. A. Muller, D. Angeli, and F. Allgower. On necessity and robustness of dissipativity

in economic model predictive control. IEEE Transactions on Automatic Control,
60(6):1671–1676, 2015.

[109] Y. Puranik and N. V. Sahinidis. Domain reduction techniques for global NLP and
MINLP optimization. Constraints, 22(3):338–376, 2017.
Bibliography 163

[110] D. Whitley, S. Rana, J. Dzubera, and K. E. Mathias. Evaluating evolutionary algorithms.

Artificial Intelligence, 85(1):245–276, 1996.

[111] L.O. Chua and A.-C. Deng. Canonical piecewise-linear representation. IEEE Transac-
tions on Circuits and Systems, 35(1):101–111, 1988.

[112] P. Stano, U. Montanaro, D. Tavernini, M. Tufo, G. Fiengo, L. Novella, and A. Sorniotti.

Model predictive path tracking control for automated road vehicles: A review. Annual
Reviews in Control, 2022.

[113] Y. Huang and Y. Chen. Vehicle lateral stability control based on shiftable stability re-
gions and dynamic margins. IEEE Transactions on Vehicular Technology, 69(12):14727–
14738, 2020.

[114] J. K. Subosits and J. C. Gerdes. Impacts of model fidelity on trajectory optimization

for autonomous vehicles in extreme maneuvers. IEEE Transactions on Intelligent
Vehicles, 6(3):546–558, 2021.

[115] V. Z. Patterson, F. E. Lewis, and J. C. Gerdes. Optimal decision making for automated
vehicles using homotopy generation and nonlinear model predictive control. In IEEE
Intelligent Vehicles Symposium, pages 1045–1050, 2021.

[116] M. Metzler, D. Tavernini, P. Gruber, and A. Sorniotti. On prediction model fidelity in

explicit nonlinear model predictive vehicle stability control. IEEE Transactions on
Control Systems Technology, 29(5):1964–1980, 2021.

[117] D. Tavernini, M. Metzler, P. Gruber, and A. Sorniotti. Explicit nonlinear model

predictive control for electric vehicle traction control. IEEE Transactions on Control
Systems Technology, 27(4):1438–1451, 2019.

[118] K. Oh and J. Seo. Development of an adaptive and weighted model predictive control
algorithm for autonomous driving with disturbance estimation and grey prediction.
IEEE Access, 10:35251–35264, 2022.

[119] T. Brudigam, M. Olbrich, D. Wollherr, and M. Leibold. Stochastic model predictive

control with a safety guarantee for autonomous driving. IEEE Transactions on
Intelligent Vehicles, 8(1):22–36, 2023.

[120] M. Rokonuzzaman, N. Mohajer, and S. Nahavandi. Effective adoption of vehicle

models for autonomous vehicle path tracking: a switched MPC approach. Vehicle
System Dynamics, pages 1–24, 2022.

[121] K. Zhang, J. Sprinkle, and R. G. Sanfelice. Computationally aware control of au-

tonomous vehicles: a hybrid model predictive control approach. Autonomous Robots,
39:503–517, 2015.

[122] T. Zhao, E. Yurtsever, R. Chladny, and G. Rizzoni. Collision avoidance with transi-
tional drift control. In IEEE International Intelligent Transportation Systems Conference
(ITSC), pages 907–914, 2021.
164 Bibliography

[123] M. Corno, G. Panzani, F. Roselli, M. Giorelli, D. Azzolini, and S. M. Savaresi. An LPV

approach to autonomous vehicle path tracking in the presence of steering actuation
nonlinearities. IEEE Transactions on Control Systems Technology, 29(4):1766–1774,
2020.

[124] M. Amir and T. Givargis. Hybrid state machine model for fast model predictive
control: Application to path tracking. In IEEE/ACM International Conference on
Computer-Aided Design (ICCAD), pages 185–192, 2017.

[125] E. Pérez, C. Ariño, F. X. Blasco, and M. A. Martínez. Explicit predictive control with
non-convex polyhedral constraints. Automatica, 48(2):419–424, 2012.

[126] P. Tøndel, T. Johansen, and A. Bemporad. An algorithm for multi-parametric

quadratic programming and explicit MPC solutions. Automatica, 39(3):489–497,
2003.

[127] N. Kochdumper and M. Althoff. Computing non-convex inner-approximations of

reachable sets for nonlinear continuous systems. In IEEE Conference on Decision and
Control (CDC), pages 2130–2137, 2020.

[128] X. Miao, Y. Song, Z. Zhang, and S. Gong. Successive convexification for ascent
trajectory replanning of a multistage launch vehicle experiencing nonfatal dynamic
faults. IEEE Transactions on Aerospace and Electronic Systems, 58(3):2039–2052, jun
2022.

[129] P. Scheffe, T. Mario Henneken, M. Kloock, and B. Alrifaee. Sequential convex

programming methods for real-time optimal trajectory planning in autonomous
vehicle racing. IEEE Transactions on Intelligent Vehicles, 2021.

[130] B. Açıkmeşe, J. M. Carson, and L. Blackmore. Lossless convexification of nonconvex

control bound and pointing constraints of the soft landing optimal control problem.
IEEE Transactions on Control Systems Technology, 21(6):2104–2113, 2013.

[131] R. Deits and R. Tedrake. Computing large convex regions of obstacle-free space
through semidefinite programming. In Algorithmic Foundations of Robotics XI: Se-
lected Contributions of the Eleventh International Workshop on the Algorithmic Foun-
dations of Robotics, pages 109–124. 2015.

[132] K. Okamoto and P. Tsiotras. Optimal stochastic vehicle path planning using covari-
ance steering. IEEE Robotics and Automation Letters, 4(3):2276–2281, 2019.

[133] L. Yao. Nonparametric learning of decision regions via the genetic algorithm. IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 26(2):313–321,
1996.

[134] J. Xu. Morphological decomposition of 2-D binary shapes into conditionally maximal
convex polygons. Pattern Recognition, 29(7):1075–1104, 1996.
Bibliography 165

[135] L. Yao and K. S. Weng. Learning decision regions based on adaptive ellipsoids.
International Journal of Uncertainty, Fuzziness and Knowlege-Based Systems, 22(1):41–
73, 2014.

[136] X. Wei, M. Liu, Z. Ling, and H. Su. Approximate convex decomposition for 3D meshes
with collision-aware concavity and tree search. ACM Transactions on Graphics,
41(4):1–18, 2022.

[137] R. Bulbul and A. U. Frank. AHD: The alternate hierarchical decomposition of

nonconvex polytopes (generalization of a convex polytope based spatial data model).
In International Conference on Geoinformatics, pages 1–6, 2009.

[138] Q. Zhang, I. E. Grossmann, A. Sundaramoorthy, and J. M. Pinto. Data-driven construc-

tion of convex region surrogate models. Optimization and Engineering, 17:289–332,
2016.

[139] P. Duhr, A. Sandeep, A. Cerofolini, and C. H. Onder. Convex performance envelope for
minimum lap time energy management of race cars. IEEE Transactions on Vehicular
Technology, 71(8):8280–8295, 2022.

[140] T. Fu, H. Zhou, and Z. Liu. NMPC-based path tracking control strategy for au-
tonomous vehicles with stable limit handling. IEEE Transactions on Vehicular Tech-
nology, 71(12):12499–12510, 2022.

[141] N. Zhang, S. Yang, G. Wu, H. Ding, Z. Zhang, and K. Guo. Fast distributed model
predictive control method for active suspension systems. Sensors, 23(6):3357, 2023.

[142] G. Zhu, H. Jie, and W. Hong. Nonlinear model predictive path tracking vontrol for
autonomous vehicles based on orthogonal collocation method. International Journal
of Control, Automation and Systems, 21:257–270, 2023.

[143] M. M. Ghazaei Ardakani, B. Olofsson, A. Robertsson, and R. Johansson. Model pre-

dictive control for real-time point-to-point trajectory generation. IEEE Transactions
on Automation Science and Engineering, 16(2):972–983, 2019.

[144] M. Diehl, H. J. Ferreau, and N. Haverbeke. Efficient numerical methods for nonlinear
MPC and moving horizon estimation. Nonlinear Model Predictive Control: Towards
New challenging Applications, pages 391–417, 2009.

[145] J. Liu, P. Jayakumar, J. L. Stein, and T. Ersal. Combined speed and steering control
in high-speed autonomous ground vehicles for obstacle avoidance using model
predictive control. IEEE Transactions on Vehicular Technology, 66(10):8746–8763,
2017.

[146] Y. Vaupel, N. C. Hamacher, A. Caspari, A. Mhamdi, I. G. Kevrekidis, and A. Mitsos.

Accelerating nonlinear model predictive control through machine learning. Journal
of Process Control, 92:261–270, 2020.
166 Bibliography

[147] B. C. van Huijgevoort, S. Weiland, and S. Haesaert. Temporal logic control of

nonlinear stochastic systems using a piecewise-affine abstraction. IEEE Control
Systems Letters, 7:1039–1044, 2023.
[148] T. Besselmann and M. Morari. Hybrid parameter-varying model predictive control
for autonomous vehicle steering. European Journal of Control, 14(5):418–431, 2008.
[149] M. Rokonuzzaman, N. Mohajer, S. Nahavandi, and S. Mohamed. Review and perfor-
mance evaluation of path tracking controllers of autonomous vehicles. IET Intelligent
Transport Systems, 15(5):646–670, 2021.
[150] H. Zhao, H. Yang, Z. Wang, and H. Li. Nonlinear switched model predictive control
with multiple Lyapunov functions for trajectory tracking and obstacle avoidance
of nonholonomic systems. International Journal of Robust and Nonlinear Control,
33(11):6171–6187, 2023.
[151] H. Dugoff, P. S. Fancher, and L. Segel. An analysis of tire traction properties and
their influence on vehicle dynamic performance. SAE Transactions, 79:1219–1243,
1970.
[152] M. Brown, J. Funke, S. Erlien, and J. C. Gerdes. Safe driving envelopes for path
tracking in autonomous vehicles. Control Engineering Practice, 61:307–316, 2017.
[153] J. Y. Goh, T. Goel, and J. C. Gerdes. Toward automated vehicle control beyond
the stability limits: drifting along a general path. Journal of Dynamic Systems,
Measurement and Control, 142(2):021004, 2020.
[154] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023.
[155] K. Holmström, A. O. Göran, and M. M. Edvall. User’s guide for TOMLAB 7. Technical
report, Tomlab Optimization Inc, 2010.
[156] F. Garrido and P. Resende. Review of decision-making and planning approaches in
automated driving. IEEE Access, 10:100348–100366, 2022.
[157] M. Antkiewicz, M. Kahn, M. Ala, K. Czarnecki, P. Wells, A. Acharya, and S. Beiker.
Modes of automated driving system scenario testing: Experience report and recom-
mendations. SAE International Journal of Advances and Current Practices in Mobility,
2(4):2248–2266, 2020.
[158] E. De Gelder, H. Elrofai, A. K. Saberi, J. Paardekooper, O. Op Den Camp, and
B. De Schutter. Risk quantification for automated driving systems in real-world
driving scenarios. IEEE Access, 9:168953–168970, 2021.
[159] J. Dahl, C. De Campos, G. R.and Olsson, and J. Fredriksson. Collision avoidance: A
literature review on threat-assessment techniques. IEEE Transactions on Intelligent
Vehicles, 4(1):101–113, 2019.
[160] J. Fan, Y. Zhan, and J. Liang. A hierarchical control strategy for reliable lane changes
considering optimal path and lane-changing time point. IET Intelligent Transport
Systems, page itr2.12460, 2023.
Bibliography 167

[161] X. T. Yang, K. Huang, Z. Zhang, Z. A. Zhang, and F. Lin. Eco-driving system

for connected automated vehicles: Multi-objective trajectory optimization. IEEE
Transactions on Intelligent Transportation Systems, 22(12):7837–7849, 2021.

[162] R. Chai, A. Tsourdos, S. Chai, Y. Xia, A. Savvaris, and C. L. P. Chen. Multiphase over-
taking maneuver planning for autonomous ground vehicles via a desensitized trajec-
tory optimization approach. IEEE Transactions on Industrial Informatics, 19(1):74–87,
January 2023.

[163] T. Qie, W. Wang, C. Yang, Y. Li, Y. Zhang, W. Liu, and C. Xiang. An improved model
predictive control-based trajectory planning method for automated driving vehicles
under uncertainty environments. IEEE Transactions on Intelligent Transportation
Systems, 24(4):3999–4015, 2023.

[164] Y. Wang, X. Cao, and Y. Hu. A trajectory planning method of automatic lane change
based on dynamic safety domain. Automotive Innovation, 6(3):466–480, 2023.

[165] F. Wang, T. Shen, M. Zhao, Y. Ren, Y. Lu, B. Feng, and G. Yin. Lane-change trajectory
planning and control based on stability region for distributed drive electric vehicle.
IEEE Transactions on Vehicular Technology, 73(1):504–521, 2024.

[166] H. Dong, Q. Wang, W. Zhuang, G. Yin, K. Gao, Z. Li, and Z. Song. Flexible eco-
cruising strategy for connected and automated vehicles with efficient driving lane
planning and speed optimization. IEEE Transactions on Transportation Electrification,
10(1):1530–1540, 2024.

[167] X. Shang and A. Eskandarian. Emergency collision avoidance and mitigation using
model predictive control and artificial potential function. IEEE Transactions on
Intelligent Vehicles, 8(5):3458–3472, 2023.

[168] H. Guo, C. Shen, H. Zhang, H. Chen, and R. Jia. Simultaneous trajectory planning and
tracking using an MPC method for cyber-physical systems: A case study of obstacle
avoidance for an intelligent vehicle. IEEE Transactions on Industrial Informatics,
14(9):4273–4283, 2018.

[169] K. Li, Z. Yin, Y. Ba, Y. Yang, Y. Kuang, and E. Sun. An integrated obstacle avoidance
controller based on scene-adaptive safety envelopes. Machines, 11(2):303, 2023.

[170] S. Li and Q. Zhao. Research on the emergency obstacle avoidance strategy of

intelligent vehicles based on a safety distance model. IEEE Access, 11:7124–7134,
2023.

[171] Y. K. Nakka and S. Chung. Trajectory optimization of chance-constrained nonlinear

stochastic systems for motion planning under uncertainty. IEEE Transactions on
Robotics, 39(1):203–222, 2023.

[172] Z. Sun, R. Wang, X. Meng, Y. Yang, Z. Wei, and Q. Ye. A novel path tracking system
for autonomous vehicle based on model predictive control. Journal of Mechanical
Science and Technology, 38(1):365–378, 2024.
168 Bibliography

[173] H. Sun, S. Zhang, L. Dai, and S. V. Raković. Locally convexified rigid tube MPC. IET
Control Theory & Applications, 17(4):446–462, 2023.
[174] M. Ammour, R. Orjuela, and M. Basset. A MPC combined decision making and
trajectory planning for autonomous vehicle collision avoidance. IEEE Transactions
on Intelligent Transportation Systems, 23(12):24805–24817, 2022.
[175] S. Yang, H. Zheng, J. Wang, and A. Kamel. A personalized human-like lane-changing
trajectory planning method for automated driving system. IEEE Transactions on
Vehicular Technology, 70(7):6399–6414, 2021.
[176] H. Lu, Y. Liu, M. Zhu, C. Lu, H. Yang, and Y. Wang. Enhancing interpretability of
autonomous driving via human-like cognitive maps: A case study on lane change.
IEEE Transactions on Intelligent Vehicles, pages 1–11, 2024.
[177] M. Galvani, F. Biral, B. M. Nguyen, and H. Fujimoto. Four wheel optimal autonomous
steering for improving safety in emergency collision avoidance manoeuvres. In IEEE
International Workshop on Advanced Motion Control (AMC), pages 362–367, 2014.
[178] J. Ji, A. Khajepour, W. W. Melek, and Y. Huang. Path planning and tracking for
vehicle collision avoidance based on model predictive control with multiconstraints.
IEEE Transactions on Vehicular Technology, 66(2):952–964, 2017.
[179] I. Batkovic, U. Rosolia, M. Zanon, and P. Falcone. A robust scenario MPC approach
for uncertain multi-modal obstacles. IEEE Control Systems Letters, 5(3):947–952, 2021.
[180] Y. Chen, U. Rosolia, W. Ubellacker, N. Csomay-Shanklin, and A. D. Ames. Interactive
multi-modal motion planning with branch model predictive control. IEEE Robotics
and Automation Letters, 7(2):5365–5372, 2022.
[181] K. Liu, N. Li, H. E. Tseng, I. Kolmanovsky, and A. Girard. Interaction-aware trajectory
prediction and planning for autonomous vehicles in forced merge scenarios. IEEE
Transactions on Intelligent Transportation Systems, 24(1):474–488, 2023.
[182] M. Cannon, B. Kouvaritakis, and X. Wu. Probabilistic constrained MPC for multi-
plicative and additive stochastic uncertainty. IEEE Transactions on Automatic Control,
54(7):1626–1632, 2009.
[183] H. Zhu and J. Alonso-Mora. Chance-constrained collision avoidance for MAVs in
dynamic environments. IEEE Robotics and Automation Letters, 4(2):776–783, 2019.
[184] S. X. Wei, A. Dixit, S. Tomar, and J. W. Burdick. Moving obstacle avoidance: A
data-driven risk-aware approach. IEEE Control Systems Letters, 7:289–294, 2023.
[185] R. Chai, A. Tsourdos, A. Savvaris, S. Wang, Y. Xia, and S. Chai. Fast generation of
chance-constrained flight trajectory for unmanned vehicles. IEEE Transactions on
Aerospace and Electronic Systems, 57:1028–1045, 2021.
[186] D. Moser, R. Schmied, H. Waschl, and L. del Re. Flexible spacing adaptive cruise
control using stochastic model predictive control. IEEE Transactions on Control
Systems Technology, 26(1):114–127, 2018.
Bibliography 169

[187] H. Kazemi, H. N. Mahjoub, A. Tahmasbi-Sarvestani, and Y. P. Fallah. A learning-based

stochastic MPC design for cooperative adaptive cruise control to handle interfering
vehicles. IEEE Transactions on Intelligent Vehicles, 3(3):266–275, 2018.

[188] B. Paden, M. Čáp, S. Z. Yong, D. Yershov, and E. Frazzoli. A survey of motion

planning and control techniques for self-driving urban vehicles. IEEE Transactions
on Intelligent Vehicles, 1(1):33–55, 2016.

[189] M. Muñoz Sánchez, D. Pogosov, E. Silvas, D. C. Mocanu, J. Elfring, and M. J. G. van

de Molengraft. Situation-aware drivable space estimation for automated driving.
IEEE Transactions on Intelligent Transportation Systems, 23(7):9615–9629, 2022.

[190] M. Bujarbaruah, X. Zhang, M. Tanaskovic, and F. Borrelli. Adaptive stochastic MPC

under time-varying uncertainty. IEEE Transactions on Automatic Control, 66(6):2840–
2845, 2021.

[191] V. Fors, P. Anistratov, B. Olofsson, and L. Nielsen. Predictive force-centric emergency

collision avoidance. Journal of Dynamic Systems, Measurement and Control, 143(8):1–
12, 2021.

[192] R. Hajiloo, M. Abroshan, A. Khajepour, A. Kasaiezadeh, and S. K. Chen. Integrated

steering and differential braking for emergency collision avoidance in autonomous
vehicles. IEEE Transactions on Intelligent Transportation Systems, 22(5):3167–3178,
2021.

[193] T. Benciolini, D. Wollherr, and M. Leibold. Non-conservative trajectory planning for

automated vehicles by estimating intentions of dynamic obstacles. IEEE Transactions
on Intelligent Vehicles, 8(3):2463–2481, 2023.

[194] N. Malone, H. Chiang, K. Lesser, M. Oishi, and L. Tapia. Hybrid dynamic moving
obstacle avoidance using a stochastic reachable set-based potential field. IEEE
Transactions on Robotics, 33(5):1124–1138, 2017.

[195] L. Petrović, I. Marković, and I. Petrović. Mixtures of Gaussian processes for robot
motion planning using stochastic trajectory optimization. IEEE Transactions on
Systems, Man, and Cybernetics: Systems, 52(12):7378–7390, 2022.

[196] M. Schuurmans, A. Katriniok, C. Meissen, H. E. Tseng, and P. Patrinos. Safe, learning-

based MPC for highway driving under lane-change uncertainty: A distributionally
robust approach. Artificial Intelligence, 320:103920, 2023.

[197] S. Paternain and A. Ribeiro. Stochastic artificial potentials for online safe navigation.
IEEE Transactions on Automatic Control, 65(5):1985–2000, 2020.

[198] J. Wang, Y. Yan, K. Zhang, Y. Chen, M. Cao, and G. Yin. Path planning on large
curvature roads using driver-vehicle-road system based on the kinematic vehicle
model. IEEE Transactions on Vehicular Technology, 71(1):311–325, 2022.
170 Bibliography

[199] L. Schäfer, S. Manzinger, and M. Althoff. Computation of solution spaces for

optimization-based trajectory planning. IEEE Transactions on Intelligent Vehicles,
8(1):216–231, 2023.
[200] F. Eiras, M. Hawasly, S. V. Albrecht, and S. Ramamoorthy. A two-stage optimization-
based motion planner for safe urban driving. IEEE Transactions on Robotics, 38(2):822–
834, 2022.
[201] A. Caregnato-Neto, M. R. O. A. Maximo, and R. J. M. Afonso. Real-time motion plan-
ning and decision-making for a group of differential drive robots under connectivity
constraints using robust MPC and mixed-integer programming. Advanced Robotics,
37(5):356–379, 2023.

[202] CarMaker. IPG Automotive.

[203] Y. Chen, C. Hu, and J. Wang. Motion planning with velocity prediction and composite
nonlinear feedback tracking control for lane-change strategy of autonomous vehicles.
IEEE Transactions on Intelligent Vehicles, 5(1):63–74, 2020.
[204] C. Kuo, A. Schaarschmidt, Y. Cui, T. Asfour, and T. Matsubara. Uncertainty-aware
contact-safe model-based reinforcement learning. IEEE Robotics and Automation
Letters, 6(2):3918–3925, 2021.
[205] M. Tiger, D. Bergström, A. Norrstig, and F. Heintz. Enhancing lattice-based motion
planning with introspective learning and reasoning. IEEE Robotics and Automation
Letters, 6(3):4385–4392, 2021.

[206] T. Brüdigam, M. Olbrich, D. Wollherr, and M. Leibold. Stochastic model predictive

control with a safety guarantee for automated driving: Extended version. arXiv,
(2009.09381), 2022.
[207] L. Gharavi. Proactive Motion Planning Codes for Emergency Collision Avoidance in
Highway Scenarios. 4TU.ResearchData, 2023.
 171

Curriculum Vitæ
Leila Gharavi
[ lejlA ÈæræVi ]

Born on May 5, 1995 in Iran, Leila obtained her B.Sc. and M.Sc.
degrees in Mechanical Engineering from Amirkabir Univer-
sity of Technology, with research experience in automatic
manufacturing and production, vibration analysis, control of
nonlinear dynamics, and soft rehabilitation robotics.
She started her Ph.D. research in February 2021 at Delft
Center for Systems and Control, Delft University of Tech-
nology, The Netherlands. Her doctoral research focuses
on nonlinear and hybrid systems, optimization, and model-
predictive control, with applications to adaptive and proactive
control of automated vehicles in hazardous scenarios. More-
over, Leila was a visiting research fellow at the e-Mobility
and Control lab in The University of Tokyo in Japan from
February to April 2024.
Leila thrives on exploring complexity in every aspect of her life. She delights in solving
intricate jigsaw puzzles, finding joy in uncovering hidden patterns, and channels her
creativity into arts and crafts projects that challenge her problem-solving skills. Her love
for sociopolitical literature reflects this same curiosity, as she seeks to understand the
intricacies of human life.
Openly vocal about diversity, Leila advocates for awareness and inclusion across all
dimensions of human identity, including neurological, gender, orientation, racial, and
cultural aspects. Through her work and her voice, she strives to foster understanding and
create a more inclusive world.
 173

List of Publications
Journals
9. J. Verkuijlen, L. Gharavi, A. Dabiri, B. De Schutter, and S. Baldi, Advancing Safety in Highway
Automated Driving: a Stochastic Model Predictive Approach to Emergency Motion Planning,
Under review.
8. L. Gharavi, S. Baldi, Y. Hosomi, T. Sato, B.M. Nguyen, and H. Fujimoto, Dodging the Moose:
Experimental Insights in Real-Life Automated Collision Avoidance, Under review.
7. L. Gharavi, B. De Schutter and S. Baldi, Iterative Cut-Based PWA Approximation of Multi-
Dimensional Nonlinear Systems, Under review.
6. L. Gharavi, A. Dabiri, J. Verkuijlen, B. De Schutter and S. Baldi, Proactive emergency collision
avoidance for automated driving in highway scenarios, IEEE Transactions on Control Systems
Technology, early access.
5. L. Gharavi, B. De Schutter and S. Baldi, Efficient MPC for Emergency Evasive Maneuvers,
Part II: Comparative Assessment for Hybrid Control, Under review.
4. L. Gharavi, B. De Schutter and S. Baldi, Efficient MPC for Emergency Evasive Maneuvers,
Part I: Hybridization of the Nonlinear Problem, Under review.
3. L. Gharavi, C. Liu, B. De Schutter and S. Baldi, Sensitivity analysis for piecewise-affine
approximations of nonlinear programs with polytopic constraints, IEEE Control Systems Letters,
8, pp. 1271-1276, 2024.
2. L. Gharavi, M. Zareinejad, and A. Ohadi, Continuum analysis of a soft bending actuator
dynamics, Mechatronics, 83, p. 102739, 2022.
1. L. Gharavi, M. Zareinejad, and A. Ohadi, Dynamic Finite-Element analysis of a soft bending
actuator, Mechatronics, 81, p. 102690, 2022.

Conferences
3. L. Gharavi, C. Liu, B. De Schutter and S. Baldi, Sensitivity analysis for piecewise-affine ap-
proximations of nonlinear programs with polytopic constraints, IEEE International Conference
Decision and Control (CDC), 2024.
2. L. Gharavi, B. De Schutter and S. Baldi, H4MPC: a hybridization toolbox for model predictive
control in automated driving, IEEE International Conference on Advance Motion Control (AMC),
pp. 1–6, 2024.
1. L. Gharavi, B. De Schutter and S. Baldi, Parametric piecewise-affine approximation of nonlin-
ear systems: a cut-based approach, IFAC-PapersOnLine, 56 (2), pp. 6666–6671, 2023.
Included in this thesis.