This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

Model-Free Safety Critical Model Predictive

Control for Mobile Robot in Dynamic Environments
Minh-Nhat Nguyen, Mien Van, Stephen McIlvanna, Yuzhu Sun, Jack Close, Kabirat Olayemi, Yan Jin, Member,
IEEE

Abstract—Within the context of Nonlinear Model Predictive dynamic environments with changing references and obsta-
Control (NMPC) design for autonomous mobile robots, which cles complicates the switch between distinct controller types.
face challenges such as parametric uncertainty and measurement Moreover, integrating constraints to maintain the robot within
inaccuracies, focusing on dynamic modelling and parameter
identification becomes crucial. This paper presents a novel a specific area or to avoid obstacles during operation presents
safety-critical control framework for a mobile robot system that substantial difficulties.
utilises NMPC with a prediction model derived entirely from To overcome these challenges, NMPC is utilised for its ability
noisy measurement data. The Sparse Identification of Nonlinear to handle soft and hard constraints and provide simultaneous
Dynamics (SINDY) is employed to predict the system’s state tracking and regulation. NMPC solves an Optimal Control
under actuation effects. Meanwhile, the Control Barrier Function
(CBF) is integrated into the NMPC as a safety-critical constraint, Problem (OCP) in real-time, determining control actions based
ensuring obstacle avoidance even when the robot’s planned path on the system’s predicted state and adhering to constraints.
is significantly distant from these obstacles. The closed-loop The advent of recently developed open-source optimisation
system demonstrates Input-to-State Stability (ISS) with respect to toolkits [14] - [16], coupled with the enhanced processing
the prediction error of the learned model. The proposed frame- capabilities of modern computers, has facilitated the rapid
work undergoes exhaustive analysis in three stages, training,
prediction, and control, across varying noise levels in the state implementation of NMPC for real-time applications. Various
data. Additionally, validation in Matlab and Gazebo illustrates NMPC schemes have been proposed to address both control
that the NMPC-SINDY-CBF approach enables smooth, accurate, problems using a singular NMPC approach, as detailed in [17]
collision-free movement, even with measurement noise and short - [19]. These studies typically assume detailed knowledge of
prediction times. Our findings, supported by tests conducted with the robot’s dynamic model. However, obtaining an accurate
the Husky A200 robot, confirm the approach’s applicability in
real-time scenarios. dynamic model is challenging, especially in systems affected
by parametric uncertainties, measurement errors, and noise.
Index Terms—Nonlinear model predictive control, Nonholo- The success of NMPC implementation heavily relies on
nomic mobile robot, SINDY, Control barrier function, Safety-
critical control. accurately and efficiently identifying a dynamic model that
reflects the system’s behaviour under control influences. Con-
sequently, data-driven or learning-based control strategies have
I. I NTRODUCTION gained increasing attention, offering a substantial opportunity
Recent advancements in robotics have expanded Wheeled to effectively use vast amounts of available data. Integrating
Mobile Robots’ (WMRs) applications in fields like logistics these methodologies with NMPC can enhance control system
and search and rescue [1]. WMR control is divided into performance while reducing modelling and engineering efforts
point stabilisation and trajectory tracking, with various feed- needed in controller design, as discussed in [20]. Our analysis
back controllers developed for localisation, including smooth assumes the robot’s kinematic model is unknown; however,
time-varying feedback [2], [3], discontinuous time-invariant measurement data is available. An offline prediction model is
feedback [4], [5], and hybrid feedback [6], [7]. Addition- developed using identification techniques to approximate the
ally, extensive research on WMR tracking control involves system’s kinematics at a specific operating point.
dynamic feedback linearisation [8], [9], adaptive reinforcement Numerous methodologies exist for developing learning-based
learning [10], and sliding mode techniques [11], [12]. A models, including state-space models via the eigensystem
novel discontinuous controller was developed and presented in realisation algorithm [21], Volterra series [22], autoregressive
[13]. However, these methods necessitate continuously excited models [23], and Neural Network (NN) models [24], [25].
reference trajectories, making them unsuitable for stabilisation Among these, NN models are selected for a comparative
problems. The need for WMRs to function autonomously in benchmark in our study due to their proven ability to learn
and model the complex nonlinear relationships inherent in
This work was partly supported by the EPSRC DTP programme,
Natural Environment Research Council, United Kingdom [grant number robotic systems, which align closely with the challenges faced
NE/V008080/1] and by the Royal Society [grant number IEC/NSFC/211236 in NMPC applications. Given the critical influence of system
and grant number RGS/R1/221356]. (Corresponding author: Mien Van.) noise, the SINDY model in [26] uses sparsity-optimisation
Minh-Nhat Nguyen, Mien Van, Stephen McIlvanna, Yuzhu Sun, Jack Close,
Kabirat Olayemi are with the School of Electronics, Electrical Engineering techniques to develop parsimonious models from limited data,
and Computer Science, Queen’s University Belfast, Belfast, United Kingdom proving notably resistant to measurement noise. The applica-
(email: [email protected]; [email protected]; [email protected]; tion of the NMPC-SINDY framework in various systems is
[email protected]; [email protected]; [email protected]). Yan Jin
is with the School of Mechanical and Aerospace Engineering, Queen’s reviewed in [27]. Still, this study lacks further investigation
University Belfast, Belfast, United Kingdom (email: [email protected]). into tracking control scenarios and the stability proof of the

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

closed-loop system. covers the control approach, encompassing the development

For applications where safety is paramount, various obstacle of NMPC, safety constraints, and proof of the stability of
avoidance algorithms are detailed in [28]. Among these, the the closed-loop system. Section 5 presents the numerical
Bug-Type (BT) algorithm ensures a robot maintains a safe simulations and experimental implementations, including a
distance from obstacles via a Euclidean distance constraint, comparative analysis and discussion. Finally, Section 6 con-
as outlined in [28]. This algorithm, integrated into NMPC cludes the article.
as a safety constraint, activates only when the robot’s pro-
jected path intersects with obstacles, potentially delaying
II. P ROBLEM FORMULATION
avoidance actions. Extending the prediction horizon could
address this, but as [27] shows, learning-based model accuracy The nonholonomic mobile robot state is defined by a state
declines with increased noise in training data. Therefore, using vector ξ = [x, y, θ]T ∈ R3 , which contains the spatial
CBF properties, as proposed in [29], appears advantageous position ξxy = [x, y]T (m,m) and the orientation angle θ
for learning-based NMPC with noisy measurements. This is (rad) of the robot. The robot control inputs are described by
achieved by selecting the CBF hyperparameter, ensuring that a vector u = [v, w]T ∈ R2 where v (m/s) and w (rad/s)
the robot’s reachable set is consistently constrained upon are linear velocity and angular velocity, respectively. In the
encountering obstacles. presence of parametric uncertainties, the kinematic model of
Motivated by the preceding discussion, this article presents a a nonholonomic mobile robot with angle measurement errors
learning-based NMPC framework integrated with CBF for the can be described as:
safety-critical control of a WMR in dynamic environments.  ∗
δ1 cos(θ(t) + ϕ∗ ) 0 


Initially, an appropriate control sequence is applied to the ˙ = f (ξ(t), u(t)) =  δ ∗ sin(θ(t) + ϕ∗ ) 0  v(t) (1)
ξ(t) 1
WMR for data collection. This is followed by the development w(t)
0 δ2∗
of a library of candidate functions in a generalised form to
capture the WMR’s nonlinear response to control inputs. A where δ1∗ and δ2∗ are unknown bounded positive parameters
subsequent comparative analysis between the SINDY and NN determined by the radius of the rear wheels, and ϕ∗ is a
models focuses on their training and predictive performance small bias orientation [30]. To derive the control problems,
with clean and noisy data. The learning-based models are then a reference robot is assumed to have the same constraints as
incorporated into the NMPC to facilitate various control tasks. (1), and its kinematic model is written as:
Multiple challenging scenarios with obstacles are devised to  ∗
δ1 cos(θr (t) + ϕ∗ ) 0 

evaluate the safety performance of the NMPC-SINDY-CBF vr (t)

approach relative to existing methods. Averaged performance ˙ ∗ ∗
ξr (t) = f (ξr (t), ur (t)) = δ1 sin(θr (t) + ϕ ) 0
 
wr (t)
indices are used to underline the advantages and limitations of 0 δ2∗
all proposed frameworks. The contributions of this manuscript (2)
are outlined as follows: The control challenge aims to minimise the difference between
a robot’s observed state and control inputs and those of a
• A learning-based model for WMR that contains para- reference robot. To align the observed robot with the reference
metric uncertainties and measurement errors is developed model, we use a discretised kinematic model with a time-step
using the SINDY framework. The model focuses on kine- τ > 0. Denoting ξ(k) = ξ(tk ), the integration over the fixed
matics discovery purely from noisy data. A comparative interval is numerically approximated with a piecewise constant
analysis highlights the noise robustness of the identified control during each sampling interval and the Runge-Kutta4
model. method.
• A safety-critical control framework is proposed for the
first time by integrating NMPC-SINDY with CBF. This ξ(k + 1) = f (ξ(k), u(k))
Z tk +τ
framework guarantees the effective operation of the robot (3)
= ξ(k) + f (ξ(ϱ), u(ϱ))dϱ
in dynamic environments, even with noisy feedback sig- tk
nals.
• The efficiency of the NMPC-SINDY-CBF approach is where k ∈ N0 is the current sampling instant and f (·) : R3 ×
validated through various complex scenarios involving R2 → R3 is the discrete nonlinear dynamical mapping. In
static and dynamic obstacles positioned randomly, con- the point stabilisation control problem, the feedback control is
ducted in Matlab. These tests demonstrate the proposed designed such that the solution of (3) starting from the initial
framework’s enhanced safety performance under a short condition ξ0 := ξ(0) ∈ X stays close to a desired set point,
prediction horizon. ξr ∈ X, and converges, i.e.
• The practical applicability of the controller is highlighted
by rigorous testing in simulated environments using the lim ∥ξe (k)∥ = lim ∥ξ(k) − ξr ∥ = 0 (4)
k→∞ k→∞
Husky model within Gazebo, as well as through real-time
experiments with the Husky A200 robot. For the trajectory control problem, the feedback control task is
to steer the solution of (3) to track the time-varying reference
The paper is organized as follows: Section 2 formulates the such that
control problems for the WMR’s kinematic model. Section
3 introduces the learning-based methodologies. Section 4 lim ∥ξe (k)∥ = lim ∥ξ(k) − ξr (k)∥ = 0 (5)
k→∞ k→∞

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

III. L EARNING -BASED T ECHNIQUES FOR N ONLINEAR and sparse. The following candidate library functions provide
M ODEL WITH ACTUATION a good approximation model for our study:
This study addresses the challenge of control design for Θ(S, C) =[1 S C sin(S) · · · cos(C) · · ·
an unknown kinematic model that suffers from parametric (7)
· · · (S ⊗ sin(S)) · · · (C ⊗ cos(C))]
uncertainties, deviations in the measured angles, and mea-
surement noise. Inspired by the SINDY method [26], a data- where S ⊗ sin(S) and C ⊗ cos(C) define all product
driven model discovery framework is formulated utilising the combinations of the components in S and C, given by:
available input-output data of the system. The SINDY model  
is deployed to replace the mapping function f (·) in (3), x sin x(t1 ) · · · y sin x(t1 ) · · · θ sin θ(t1 )
facilitating the estimation of the future state of the robot based  .. .. .. 
 . ··· . ··· . 
on the given control inputs and its current state. In addition, a  
S ⊗ sin(S) =  x sin x(ts ) · · · y sin x(ts ) · · · θ sin θ(ts ) 

,
Nonlinear Autoregressive Exogenous Model (NARX), which  .. .. .. 
can represent any continuous function under mild conditions  . ··· . ··· . 
[31], is also designed and tested to evaluate the performance x sin x(tm ) · · · y sin x(tm ) · · · θ sin θ(tm )
and capabilities of the SINDY. 
v cos v(t1 ) · · · w cos v(t1 ) w cos w(t1 )

 .. .. .. .. 

 . . . . 

C ⊗ cos(C) =  v cos v(ts ) · · · w cos v(ts ) w cos w(ts ) 

.
 .. .. .. .. 
 . . . . 
v cos v(tm ) · · · w cos v(tm ) w cos w(tm )
(8)
The sparse model, which uses the linear and nonlinear terms
in the chosen SINDY library to replace the kinematic model of
the mobile robot f (ξ(k), u(k)), can be presented as follows:
ds
X
fˆi (ξ(k), u(k)) = λid Γd (ξ(k), u(k)), i = 1, 2, 3 (9)
d=1

where ds is the number of candidate functions in the library.

Fig. 1. Schematic of the SINDY algorithm for the mobile robot system. λid is the weighted coefficient of i-th state variable, i.e.,
λi = [λi1 , λi2 , ..., λids ]T and Γd (ξ(k), u(k)) are the candidate
functions of the d-th column in Θ(ξ, u). Throughout this paper,
A. Learning-Based Model Using SINDY (ˆ·) denotes variables or functions used to make predictions.
Using the standard procedure of SINDY with control [27], In the next step, we employ sparse regression to identify a
a sparse nonlinear dynamical system is identified from ob- sparse matrix Ψ that achieves an optimal approximation with
servational data. The sparse regression process discovers the the fewest nonlinearities within the library Θ(S, C).
active terms within a library Θ(ξ, u), comprising both linear 2
1
and nonlinear model terms, essential for approximating the λi = arg min Ṡi − Θ(S, C)λ̂i + ε λ̂i (10)
function f in (3). The SINDY with control algorithm for 2 2 1
λ̂i
mobile robot system is illustrated in Fig. 1. The initial step
involves recording m snapshot of ξ and u and constructing where Ṡi represents the i-th column of Ṡ and λi is the i-
them into two data matrices S and C: th row of Ψ. The theoretical results for convergences of the

x(t1 ) y(t1 ) θ(t1 )
 
v(t1 ) w(t1 )
 SINDY framework have been provided in [32]. The discretised
 .. .. ..   .. ..  learning-based model of the nonholonomic mobile robot using
 . . .   . .  SINDY can be written in the following form:
   
S= x(t
 s ) y(t s ) θ(t )
s 
 , C = v(t
 s
 ) w(t s )  (6)

 . . .  . . ˆ + 1) = f̂S (ξ(k), u(k))
ξ(k
 .. .. ..   .. .. 
 
Z tk +ts
(11)
x(tm ) y(tm ) θ(tm ) v(tm ) w(tm ) = ξ(k) + ΨΘ(ξ(ϱ), u(ϱ))T dϱ
tk
where ts is the sampling time. The time derivatives of the state,
denoted as Ṡ = [ξ˙i (t1 ) · · · ξ˙i (ti ) · · · ξ˙i (tm )]T , are numerically
approximated from S using Total Variation Regularized Dif- B. Learning-Based Model Using NARX
ferentiation. Subsequently, the library Θ(S, C) can effectively A multi-layer perceptron network is used to approximate
encompass various functions to characterise the data. The the function f (·) in (3) within a NARX model. The resulting
selection of the appropriate library is crucial. It is advisable structure is known as a NARX network [33]. For this study,
to start with a basic selection, such as low-order polynomials a two-hidden layer network is designed with 7 neurons in
and simple trigonometrics, and gradually increase the library’s the first layer, 5 neurons in the second layer, and hyperbolic
complexity and order until the resulting models are accurate tangent activation functions. The network is trained using the

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

data (6) in series-parallel mode. The output prediction is given and minimising a cost function and constraints on the predic-
by the following equation: tion horizon grids N . The OCP is presented in parametric form
as follows:
ˆ + 1) = f̂N (ξ(k), ..., ξ(k − nξ ), u(k), ..., u(k − nu )) (12)
ξ(k  

 JN (ξ(k), u(k), p(k)) = V (ξ(k + N ), p(k))

 
N −1
where nξ and nu denote the number of previous outputs OCP : min X
ξ,u  + ℓ(ξ(k + j), u(k + j), p(k))
and inputs, respectively. After training, the NARX network is  

 
j=0
converted to the parallel mode for making predictions, which
(16a)
can be written as:
Subject to
ˆ + 1) = f̂N (ξ(k),
ξ(k ˆ ˆ − nξ ), u(k), ..., u(k − nu )) (13)
..., ξ(k ξ(k) − ξfb = 0, (16b)
ξ(k + j + 1) − F(ξ(k + j), u(k + j), e(k + j)) = 0, (16c)
If the data used for training is not corrupted by noise, the
network is trained using the Levenberg-Marquardt algorithm. ξ(k + j) ∈ X , u(k + j) ∈ U, (16d)
However, if the data is noisy, Bayesian regularisation is ξ(k + N ) ∈ Xf ⊂ X. (16e)
employed instead.
where Xf is a terminal constraint domain. The multiple
shooting approach [34] is used to transform the OCP into a
IV. N ONLINEAR MODEL PREDICTIVE CONTROL WITH Nonlinear Program (NLP) with ξ N +1 = [ξ(k)T , ..., ξ(k+N )T ]
CONTROL BARRIER FUNCTION
and uN = [u(k)T , ..., u(k+N −1)T ] are sequences of the state
vector and control input vector, respectively. p is a problem
A. Nonlinear Model Predictive Control Strategy parameter vector that contains the reference state vector ξr ,
the reference control input vector ur , the feedback state vector
NMPC operates by resolving an open-loop optimisation ξfb , and other necessary information for the control strategy.
problem at each measurement interval. In every iteration, the Constructing a vector that contains all decision variables as
process aims to minimise a predefined cost function. This Ξ = [ξ N +1 , uN ]T , the NLP is introduced as below:
section explains how either the SINDY or NARX model, as (
identified in Section II, is integrated as a state estimator within H (Ξ, p) = 0,
the control design. Moreover, to ensure the robot’s safety in min A (Ξ, p) s.t. (17)
Ξ G (Ξ, p) ⩽ 0.
dynamic environments, a safety constraint via the CBF or BT
algorithm is incorporated into the OCP. The control scheme is The primary objective of the NLP is to minimise the objective
depicted in Fig. 2. function A(Ξ, p) = JN (ξ(k), u(k), p(k)) while satisfying
certain constraints. These constraints can be divided into two
categories: the equality constraint vector, denoted as H(Ξ, p),
ensures that the system dynamics in (16b) and (16c) are met,
and the inequality constraint vector, represented by G(Ξ, p),
imposes restrictions of (16d) and safety constraints on the de-
cision variables to maintain the system within safe boundaries.

B. Safety-Critical Control Algorithms

Inspired by the works done in [29] and [35], this section
Fig. 2. Diagram of the proposed framework. derives a safety constraint for the optimal control input that
guarantees the robot’s spatial position, ξxy , always lies within
In practice, data acquisition processes are generally affected by a defined safe set, Z ∈ R2 . That is, Z is forward invariant,
noise, which also reduces the accuracy of identified models. i.e. if ξxy (0) = ξxy0 ∈ Z then ξxy = ξxy (t) ∈ Z , ∀t. The set
To examine the robustness of these models, we reformulate Z is defined as:
the state-space representation, including the noise corruption.
Z = {ξxy ∈ X ⊂ R2 |h(ξxy ) ⩾ 0}, (18a)
ξ(k + 1) = F(ξ(k), u(k), e(k)) 2
∂Z = {ξxy ∈ X ⊂ R |h(ξxy ) = 0}, (18b)
(14) 2
= f̂ (ξ(k), u(k)) + e(k) Int(Z ){ξxy ∈ X ⊂ R |h(ξxy ) > 0}. (18c)

where e(k) denotes the difference between the true model f (·) Let Z denote the superlevel set of a continuously differen-
in (3) and the identified model f̂ (·) in (11) or (13) tiable function h : R2 → R. The function h qualifies as a CBF
if there exists an K∞ -function γ for the control system (14)
e(k) = f (ξ(k), u(k)) + n(k) − f̂ (ξ(k), u(k)) (15) to satisfy:
∃ u s.t. ḣ(ξxy , u) ⩾ −γh(ξxy ), γ ∈ K∞ (19)
here n(k) is the state measurement noise, which is assumed
to lie in a compact set, n(k) ∈ χ ⊂ R. where K∞ -function is defined in Definition 1. Extending
The conventional NMPC scheme is formulated by constructing this inequality constraint to the discrete-time domain and

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

employing γ as a scalar, the set comprising all control values Algorithm 1 IPOPT - NMPC
at a given point ξxy (k) ∈ X: Given Ξ 0
Initialize (j, Ξ j ) ← (0, Ξ 0 )
κcbf (ξxy (k)) = {u ∈ U : ∆h(ξxy (k), u) + γh(ξxy (k)) ⩾ 0,
while ControllerIsRunning() do
0 ⩽ γ ⩽ 1} ξfb ← StateFeedback()
(20)
ξr ← Commands()
Remark 1. By Theorem 2 in [29], if h is a CBF in X and
∂h φob ← Sensors
∂ξxy (ξxy ) ̸= 0 for all ξxy ∈ ∂Z , then any control signal p ← [ξfbT
, ξrT , φTob ]T
u ∈ κcbf (ξxy (k)) for the system (14) renders the set Z safe. Ξ j+1
← IPOPT solver(p, Ξ j )
Additionally, the set Z is asymptotically stable in X. ∗
uN ← ExtractInputSequence(Ξ j+1 )
The rigid body obstacles are conceptualised as the union of u∗ (k) ← ExtractFirstInput(u∗N )
circles with centroids (xob,i , yob,i ) and a fixed radius rob,i . f (ξ, u) + n ← ApplyInput(u∗ (k))
Similarly, the robot safety zone is defined by a circle centred Ξ j ← Shift(ξ ∗N , u∗N )
on the robot with a radius rrb . Safety sets for obstacles are j ← j+1
denoted as follows: end while
Zi = {ξxy (k) ∈ X ⊂ R2 : hi (ξxy (k), p(k)) ⩾ 0} (21)
C. Stability Analysis
where Consider that system (14) is controlled by the NMPC
hi (ξxy (k), p(k)) =(x(k) − xob,i (k))2 + (y(k) − yob,i (k))2 control law κ(p(k)) = u∗ (k), obtained from NLP (17), then
a closed loop system can be expressed as:
− (rrb + rob,i )2
(22) ξ(k + 1) = F(ξ(k), κ(p(k)), e(k)) (26)
Assuming that the robot can detect the position and size
of obstacles, φob = [xob , yob , rob ]T , through its sensors at The ISS property [36] is recalled in this section to present
each sampling time, this information is then stored inside the its benefits for the analysis of the robustness of a controlled
parameter vector p. To ensure safety while regulating the target system. Some required definitions and assumptions are also
state, the condition in (20) is added to the NMPC in (16) as introduced.
a safety constraint. Definition 1. Let R⩾0 denotes the non-negative real. A func-
tion α : R⩾0 → R⩾0 is a class K-function if it is continuous,
hi (ξxy (k + 1), p(k)) − (1 − γ)hi (ξxy (k), p(k)) ⩾ 0, strictly increasing, and α(0) = 0. A function α : R⩾0 → R⩾0
(23) is of class K∞ if it is a K-function and α(s) → +∞ as
k = 0, ..., N − 1.
s → +∞. A function β : R⩾0 × Z⩾0 → R⩾0 is a class KL-
In [28], various obstacle avoidance techniques were examined, function if, for each fixed t ⩾ 0, β(·, t) is of class K, for each
among which the BT algorithm is particularly compatible with fixed s ⩾ 0, β(s, ·) is decreasing and β(s, t) → 0 as t → ∞.
NMPC. This method mandates the robot to navigate tangen- Definition 2. A system (26) is ISS if there exists a class KL-
tially along the obstacle’s perimeter, subsequently reverting function β and a class K-function α such that for all initial
to its initial route post-obstacle circumnavigation. The BT state ξ(0) and modeling uncertainties e(k),
algorithm is formulated as follows:
∥ξe (k)∥ ⩽ β(∥ξe (0)∥, k) + supi∈[0,k] α(|e(i)|) (27)
q
li (ξxy (k), p(k)) = (x(k) − xob,i (k))2 + (y(k) − yob,i (k))2 The stability of the nominal case is first derived, in which the
prediction error and signal noise are assumed to be zero so
− (rrb + rob,i ) ⩾ 0
(24) that the predicted state is the same as the actual value. Then,
The obstacle avoidance performance will be compared be- robust stability is considered when the prediction error and
tween the CBF constraint in (23) and the BT constraint in signal noise are non-zero but bounded. For the nominal case,
the following equation: the following stability proof shows that the closed-loop system
(26) is asymptotically stable to the desired state within a finite
li (ξ(k)xy , p(k)) ⩾ 0, k = 0, ..., N. (25) prediction horizon.
Definition 3. Xf is a forward invariant set for the system
The proposed NMPC-SINDY strategy with safety constraints (26), i.e. for all ξ(k) ∈ Xf there exists κ(p(k)) such that
is solved using an Interior Point OPTimizer (IPOPT). Based on F(ξ(k), κ(p(k)), 0) ∈ Xf . V is a Control Lyapunov Function
the current feedback state and environmental data, the IPOPT (CLF) in Xf such that for all ξ(k) ∈ Xf there exist two
solver derives an optimal trajectory and control sequence, ξ ∗N K∞ -function α1 and α2 satisfying α1 (∥ξ(k) − ξr (k)∥) ⩽
and u∗N . The first element of u∗N , labelled u∗ (k), is then V (ξ(k), p(k)) ⩽ α2 (∥ξ(k) − ξr (k)∥) and
applied to the system. The residual part of the optimised V (F(ξ(k),κ(p(k)), 0), p(k)) − V (ξ(k), p(k))
sequence serves as an initial guess for the decision variable (28)
vector in the subsequent iteration. Thereafter, the solver is re- + ℓ(ξ(k), κ(p(k), p(k))) ⩽ 0
engaged to deduce new control input values. An outline of this The stage cost has two quadratic terms that penalise the
control methodology is delineated in Algorithm 1. tracking error of inputs and outputs, ℓ : R3 × R2 → R⩾0 .

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

∗
The terminal cost is a quadratic function of the state variables, Proof. Considering the optimal value function JN (ξ(k)) as a
V : R3 → R⩾0 . Lyapunov function candidate, recall the cost function in (16a)
N −1
ℓ(ξ(k), κ(p(k)), p(k)) = ξe (k)T Qx ξe (k) + ue (k)T Qu ue (k), X
JN (ξ(k), u(k), p(k)) = ℓ(ξ(k + j), u(k + j), p(k))
V (ξ(k), p(k)) = ξe (k)T Qf ξe (k). j=0
(29)
where Qx , Qu , and Qf are positive definite symmetric + V (ξ(k + N ), p(k))
N −2
matrices. X
Assumption 1. There exists a control input κ(p(k)) ∈ U for = ℓ(ξ(k + j), u(k + j), p(k))
j=0
all ξ, ξr ∈ X that satisfy ∥ξ(k + 1) − ξr ∥2 < ∥ξ(k) − ξr ∥2 .
Remark 2. This assumption refers to the controllability + V (ξ(k + N ), p(k))
properties of the wheeled mobile robot, where the robot + ℓ(ξ(k + N − 1), u(k + N − 1), p(k))
exhibits drift-free behaviour. The accessibility rank condition (33)
is fulfilled, ensuring that all points in the state space are Taking into account the cost function for (N − 1) prediction
reachable. Therefore, there exists a trajectory from the current horizon as follows:
configuration ξ(k) to a new configuration ξ(k + 1), aligned N
X −2
with the desired state ξr . The steady decrease of the cost JN −1 (ξ(k), u(k), p(k)) = ℓ(ξ(k + j), u(k + j), p(k))
function A(Ξ, p) and the step-by-step convergence of the j=0
system’s state variables towards the targeted configuration in + V (ξ(k + N − 1), p(k))
control Algorithm 1 will be substantiated in Theorem 2. (34)
Theorem 1. Let Assumption 1 hold, then the terminal cost Substituting (34) into (33) yields:
function V is a CLF on D ∈ R3 , where D is a small
neighborhood of the reference state. JN (ξ(k), u(k), p(k)) − JN −1 (ξ(k), u(k), p(k)) =
Proof. Considering the diagonal matrices, V (ξ(k + N ), p(k))
Qx = diag(qx , qx , qx ), Qu = diag(qu , qu ), and − V (ξ(k + N − 1), p(k))
Qf = diag(qf , qf , qf ), where qx , qu , qf are strictly positive
+ ℓ(ξ(k + N − 1), u(k + N − 1), p(k))
constants, the stage cost and terminal cost are rewritten as: (35)
ℓ(ξ(k), κ(p(k)), p(k)) = qx ∥ξ(k) − ξr (k)∥2 From Theorem 1, it is inferred that
+ qu ∥u(k) − ur (k)∥2 , (30) JN (ξ(k), u(k), p(k)) < JN −1 (ξ(k), u(k), p(k)) (36)
V (ξ(k), p(k)) = qf ∥ξ(k) − ξr (k)∥2 .
This inequality carries over to the corresponding optimal value
Substituting (30) into (28) yields: functions:
∗
qf ∥ξ(k + 1) − ξr (k)∥2 − qf ∥ξ(k) − ξr (k)∥2 JN (ξ(k)) = inf JN (ξ(k), u(k), p(k))
u(·)∈U
(31)
+ qx ∥ξ(k) − ξr (k)∥2 + qu ∥u(k) − ur (k)∥2 ⩽ 0 ⩽ inf JN −1 (ξ(k), u(k), p(k)) (37)
u(·)∈U
Under the Assumption 1, it can be proven that: = JN∗
−1 (ξ(k))
2 2
qf Λ + qx ∥ξ(k) − ξr (k)∥ + qu ∥u(k) − ur (k)∥ Bellman’s principle of optimality [38] is used to obtain the
2 2
(32)
⩽ qf Λ + qx ∥ξe (k)∥ + qu ∥umax − ur (k)∥ ⩽ 0 relationship of optimal value functions at different horizons
N and points in space. This approach yields
where Λ = ∥ξ(k + 1) − ξr (k)∥2 − ∥ξ(k) − ξr (k)∥2 . ∗
2
∥umax −ur (k)∥2 JN (ξ ∗ (k)) =JN
∗ ∗
−1 (ξ (k + 1))
By choosing qf ⩾ qx ∥ξe (k)∥ +qu|Λ| the inequality (38)
(28) is satisfied and the terminal cost V is a CLF. + ℓ(ξ ∗ (k), κ(p(k)), p(k))
Remark 3. Within a small proximity of the reference state, From Assumption 2, the lower bound of the optimal value
as Λ → 0 (ξ → ξr ), the inequality in (32) is not satisfied. function is defined as:
Hence, selecting a higher value for qf ensures that this
∗
vicinity remains close to ξr . Parameters qx , qu , qf can be fine- JN (ξ(k)) ⩾ ℓ(ξ(k), κ(p(k)), p(k)) ⩾ α3 (∥ξ(k) − ξr (k)∥)
tuned using a trial and error approach to achieve satisfactory (39)
performance. Expanding (37) consecutively from N to 1 leads to:
Assumption 2. There exists a K∞ -function α3 such that ∗ ∗ ∗
the state cost satisfies ℓ(ξ(k), κ(p(k)), p(k)) ⩾ α3 (∥ξ(k) −
JN (ξ(k)) ⩽ JN −1 (ξ(k)) ⩽ · · · ⩽ J1 (ξ(k))
(40)
ξr (k)∥) for all ξ ∈ X and u ∈ U . This is a common ⩽ V (ξ(k + 1), p(k)) ⩽ α2 (∥ξ(k + 1) − ξr (k)∥)
assumption in the design of NMPC [36], [37]. Based on Proposition 2.34 in [39], there exists a control input
Theorem 2. Assume that Qx , Qu , and Qf are selected based u(k) ∈ U that the following is satisfied:
on Theorem 1 and Assumptions 1 and 2 are satisfied, then
∗ ∗
the closed-loop system (26) is asymptotically stable in a JN (ξ(k + 1))−JN (ξ(k)) ⩽ −ℓ(ξ(k), κ(p(k)), p(k))
neighbourhood of the desired state D : ∥ξ − ξr ∥ < ζ. (41)
⩽ −α3 (∥ξ(k) − ξr (k)∥)

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

∗
By Definition 3, JN (ξ(k)) is a control Lyapunov function, V. R ESULTS AND DISCUSSION
and the closed-loop system (26) is asymptotically stable to This section considered the system identification and control
the desired state. of a continuously mobile robot system that operated in a
Remark 4. It should be noted that the terminal cost function dynamic environment containing random static and moving
V (·) might not qualify as a CLF as the robot’s state nears the obstacles. The robot’s kinematics in (1) were employed for
vicinity D : ∥ξ − ξr ∥ < ζ. Consequently, the aforementioned simulation and generating training data but were otherwise
property is not applicable in this domain. However, the initial assumed to be unknown. Parametric uncertainties and an-
values of the control sequence in Ξ can be set to zero for gle measurement error were defined as follows: δ1∗ = 0.9,
all ξ(k) ∈ D. The resulting OCP of the next step becomes δ2∗ = 0.95, and ϕ∗ = 0.2. Control input constraints were
JN (ξ(k + 1), u(k), p(k)) ⩽ (N − 1)qx ζ + qf ζ, which is [−0.7 − π/4]T ⩽ [v w]T ⩽ [0.7 π/4]T .
∗
an upper bound for JN (ξ(k + 1)). Therefore, selecting a
sufficiently large qf ensures the system’s trajectory is confined
within the region D. A. Prediction and Control Performance of Mobile Robot Sys-
After demonstrating nominal asymptotic stability, the next tem without Measurement Noise
step is establishing that the robot controlled by the proposed
NMPC is ISS w.r.t. model mismatches based on the following
assumption.
Assumption 3. There exists a constant µ < ∞, such that the
prediction error (15) is bounded ∀k |e(k)| ⩽ µ.
Remark 5. As shown in (15), the prediction error contains
the approximation mismatch and the measurement noise.
Measurement noise is assumed to be bounded. Due to the
approximation ability of the NARX and SINDY models, the
analysis in [40] for the NN and [41] for the SINDY model
has shown that the prediction error is bounded. Moreover, the
NMPC-NARX and NMPC-SINDY frameworks will operate
within the region where they were initially trained. The
bounded prediction error on the training set can guarantee that
Assumption 3 holds.
Assumption 4. The nominal model F(ξ(k), κ(p(k)), 0) is
uniformly continuous in ξ(k) for all ξ(k) ∈ Xf .
Remark 6. As shown in (14), Assumption 4 is guaranteed
by the uniformly continuous output of the prediction model Fig. 3. Prediction and control performance without measurement noise: (a)
f̂ (ξ(k), κ(p(k))). For the SINDY model, this condition is NARX model and (b) SINDY model. The control inputs are shifted to -2.
achieved through the use of continuously differentiable func-
tions in the model library. Meanwhile, in the NARX model, A raw dataset was obtained from an open-loop simulation of
the uniform continuity is established by the hyperbolic tangent ˙ = f (ξ(t), u(t)) excited by the Schroederphased Harmonic
ξ(t)
activation functions at the neurons in the NARX network Sequence (SPHS) signals with the number of harmonics,
layers since their derivatives are uniformly bounded. Nh = 3, period, P = 4, linear velocity scale factor, Av = 0.4,
Theorem 3. Let κ(p(k)) be the predictive controller derived angular velocity scale factor, Aw = 0.3, and sampling time
from the NLP (17) and let Assumptions 1-4 hold. Then ts = 0.01 for 40 s. For clarity in visualisation, only the
Ω ⊆ Xf is a robust invariant set for a sufficiently small bound first half of the training dataset was plotted from 0-20 s in
on the uncertainty. The system (26) fulfills the ISS property Fig. 3. Multiple NARX models were trained, but the optimal
w.r.t. the model mismatch e(k) in (15). estimation was achieved with an input delay of nξ = 1 and an
Proof. The fulfilment of Assumptions 1-2 ensures the asymp- output delay of nu = 1. Meanwhile, the sparsity-promoting
totic stability of the nominal system, as delineated in Theorems parameter of the SINDY model determines the balance be-
1-2. Should Assumptions 3 and 4 be met, then both f̂S and tween model complexity and accuracy. ε was evaluated over a
f̂N exhibit uniform continuity, and consequently, the model range of values, [0.01, 0.1, 1, 10], and ε = 1 was chosen for its
F(·, ·, ·) is also uniformly continuous. Given that the stage superior performance with validation data. Different control
cost function ℓ(·, ·, ·) and the terminal cost function V (·, ·) are inputs, u(t) = [0.8 sin(2t)0.3 , 0.9 sin(3t)0.3 ]T , were used in
continuous and considering Xf as a compact set, these func- the state estimation to assess the generalizability of the trained
tions are also uniformly continuous. Referencing Proposition models. The prediction performances were displayed between
1, case C1, in [36] it is inferred that the closed-loop system 20-30 s in Fig. 3, where the SINDY model outperformed
exhibits ISS w.r.t. the estimation error e(k). Furthermore, the NARX in both forecast precision and training time (0.09 s
proposed NMPC is also recursively feasible under the model vs 11.5 s).
mismatch, as proven by Theorem 1 in [42]. To ensure a fair comparison, the NLP in (17) of both methods
used the same optimisation routine via CasADi, as detailed
in [43]. The setup was on MATLAB Simulink with an Intel®

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

Fig. 4. Crossvalidated prediction performance for increasing measurement noise: (a) Accuracy threshold, and (b) State predictions percentiles.

CoreTM i7-1185G7 and 16GB RAM. The test controlled the shaded areas for fifty models. Larger shaded areas of NARX
robot from ξ0 = [−0.08, 2.7, 3]T (m,m,rad) to ξr = [2, 1, 4]T models indicated poor performance with noisy data and were
(m,m,rad) with τ = 0.1 s and a prediction horizon of N = 30. highly dependent on initial network weights. Conversely, the
NMPC parameters were chosen as Qx = diag(7, 13, 0.1), SINDY method exhibited remarkable robustness to noise and
Qu = diag(0.1, 0.01), and Qf = 1000 ∗ diag(1, 1, 1). The significantly greater accuracy in state estimation. Notably,
control performance of NMPC-NARX and NMPC-SINDY even at a noise level of η = 0.03, the accuracy threshold
were shown in the last 10 s of Fig. 3 (a) and (b), respectively. remained sufficiently extensive to enable successful NMPC
Despite the increase in computation time from 15 ms for stabilisation.
NMPC-NARX to 22 ms for NMPC-SINDY, this increase was
justifiable due to the resulting smaller steady-state errors and
smoother control inputs. The trade-off led to enhanced overall 1) Point Stabilization with Static Obstacles Avoidance:
system performance. The most accurate model from fifty iterations with noise
level η = 0.01 was used for NMPC evaluation, minimising
NLP (17) with safety constraints CBF (23) or BT (25). The
B. Prediction and Control Performance of Mobile Robot Sys- robot aimed to move from ξ0 = [1, 1, 0]T (m,m,rad) to
tem with Measurement Noise ξr = [3.7, 6, π/2]T (m,m,rad), negotiating three static obsta-
Noise-affected data acquisition was simulated in fifty cles located at (2m,3.5m), (3.5m,2.6m), and (3.1m,4.6m) with
open-loop runs of ξ(t) ˙ = f (ξ(t), u(t)) + n(t) using a radius of 0.5 m, 0.5 m, and 0.4 m, respectively. The robot’s
the same SPHS signal, where n ∈ N (0, σ 2 ) with a safe radius was 0.27 m, and its sensor range was assumed
standard deviation σ. The noise magnitudes were defined to be 1.5 m from its centre. If an obstacle partially entered
as η = σ/max(std(x), std(y)) ∈ (0.005, 0.07), with std the detection zone, its position and size were identified and
denoted standard deviation. Our investigations revealed that recorded in the parameter vector p; otherwise, it was ignored.
both methods required substantially more data to match the The measurement signal was corrupted by the noise magnitude
performance attainable without noise. Hence, the simulations η = 0.01 and updated every 0.05 s. To reduce the prediction
were extended to 200 s, yielding 20000 training points. The error, the prediction horizon was shortened to N = 10. The
training procedures mirrored those of the previous section. model time-step τ = 0.1 s, leading to a prediction horizon
To handle the measurement noise, the prediction horizon in time Tp = 1 s. The CBF parameter was chosen based on the
our proposed controller was chosen to balance the length of observed robot performance, and a smaller value of γ enabled
the prediction horizon time, Tp = N τ , and the accuracy of the earlier obstacle avoidance. Therefore, γ = 0.2 was chosen,
identified models. Fig. 4 (a) presents the accuracy thresholds which effectively generated smooth trajectories and improved
at various noise levels. These thresholds were defined as the the robot’s ability to avoid obstacles. For point stabilisation,
period during which the total prediction error for all states a small weight on θ enhanced the robot’s manoeuvrability
remains below 15%. This threshold indicated a model that among obstacles, while low control input weights minimised
maintains a prediction error of approximately 5% for each state noise effects as it reached the target. NMPC parameters were
variable. Various identified models were verified with NMPC. selected as Qx = diag(3, 5, 0.1), Qu = diag(0.5, 0.01), and
A model that provided good control performance was one with Qf = 1000 ∗ diag(1, 1, 1).
P3 ξi −ξ̂i
i=1 | ξi | · 100% ⩽ 15%. Accuracy thresholds correlated The cross-validated control efficacy of NMPC-SINDY and
well with the visible divergence of the true and predicted NMPC-NARX is shown in Fig. 5. As depicted in Fig. 5 (a),
state, as shown in Fig. 4 (b), which illustrated predictive NMPC-NARX failed at point stabilisation and safety due to
performances for noise levels η = [0.005, 0.01, 0.03] using noisy signals. Conversely, with a prolonged prediction horizon
control inputs u(t) = [0.8 sin(7t)0.3 , 0.9 sin(5t)0.3 ]T . The N ⩾ 20, both NMPC-SINDY-CBF and NMPC-SINDY-BT
trajectories of the state were normalized by [1, 1, 0.5] and successfully navigated the robot to the target point without
offset by [0, −2.6, −14.3]T . The median performance was compromising safety. Notably, CBF surpassed BT in scenarios
shown by a thick line, with 25-75 % data ranges shown as with a shorter prediction horizon, as CBF restricts the robot’s

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

Fig. 5. Crossvalidated control performance for point stabilisation with obstacles and measurement noise: (a) Trajectories under three controllers, (b) Control
signals, (c) Tracking errors, (d) Safety distances and constraints.

Fig. 6. Trajectory tracking with measurement noise in a dynamic environment.

movements, prompting early action at a substantial distance

from obstacles. This property helped the NMPC generate a
smooth avoidance trajectory even with a shorter prediction
Fig. 7. State variable trajectories
period. In contrast, the BT algorithm was only effective at
close range, risking the robot getting stuck between obstacles. constraints and Euclidean distances of the robot w.r.t. the
As the performance of the NARX model was poor, only obstacle at (2m,3.5m). Safety constraints were deactivated
the control signals and state errors of NMPC-SINDY are until the obstacle was within the detection range.
shown in Fig. 5 (b) and (c), respectively. As presented in 2) Trajectory Tracking with Dynamic Obstacles Avoid-
Fig. 5 (b), the control signals stay within their limits. The ance: Due to the NARX model’s poor performance un-
measurement noise and prediction error induced the state der noise, the focus was shifted to comparing CBF and
error of NMPC-SINDY-CBF to converge to a neighbourhood BT in trajectory tracking. The reference trajectory is cir-
of the origin asymptotically. Fig. 5 (d) presents the safety cular, defined by [xr (t) = 1.6 + 2.3 sin(0.02t), yr (t) =

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

10

ξe for all controllers, showing asymptotic convergence to

the neighbourhood of the origin. However, BT experienced
prolonged overshoots due to the robot getting stuck at the
boundary of the obstacles. Control signals remained within
their limitations in Fig. 8 (b) and (c).
In Table 1, the comparison includes optimisation time, norm-
P∥ξP
2 error r − ξ∥2 for point stabilisation, and integral absolute
error k i |ξei (k)| for trajectory tracking, based on averages
from 10 simulations for each control approach under identical
setups. In all scenarios tested with measurement noise, the
SINDY model outperformed the NARX model. It was noted
that both identified methods required more time to derive
optimal solutions. Although the identified model contained
prediction error, NMPC-SINDY could obtain reasonable per-
formance compared to NMPC using the true kinematic model
of the robot. Even considering the case with high noise
magnitude, if the prediction horizon was chosen suitably to
access the most accurate prediction period of the SINDY
model, NMPC could generate a promising response. Addi-
tionally, the properties of CBF enabled the robot to operate
safely with a short prediction horizon. Thus, the NMPC-
SINDY-CBF framework showed itself to be an effective and
efficient nonlinear control in a dynamic environment where
the dynamic model was partially or entirely unknown.
Fig. 8. Trajectory tracking performance: (a) Tracking errors, (b) Control
Remark 7. It’s worth noting that NMPC-True-CBF outper-
signals. formed NMPC-SINDY-CBF, as shown in Table 1. NMPC-
True-CBF was designed with full knowledge of the kinematic
4.2 + 2.3 cos(0.02t)]T , with the robot starting from ξ0 = model of the mobile robot, while NMPC-SINDY-CBF was
[5, 3, −π/2]T (m,m,rad). The test included two static obstacles designed based on a model-free concept. Additionally, NMPC-
at (3m,6m) and (0.5m,2.3m), with the radius 0.4 m and SINDY-CBF outperformed NMPC-NARX-CBF because the
0.3 m, and two moving obstacles with initial pose, linear SINDY library could incorporate expert knowledge on which
velocity, and radius (1, 13, −π/2, 0.11, 0.25) (m,m,rad,m/s,m) category of nonlinearities (i.e., polynomials and trigonometric
and (2.5, 1.5, 0.75π, 0.047, 0.25) (m,m,rad,m/s,m). Due to the functions) to include, providing better prediction capacity than
complexity of the test, the sensor range was increased to NARX, which treated the system as a black box.
2 m and the prediction horizon to N = 20. Measurement
noise, update rate, and CBF parameter γ were the same as
C. Experimental Validation of Controller Performance
in the previous test. For the persistent excitation trajectory,
the weights for the state and control inputs were increased to
enhance tracking performance. This adjustment was quantified
as Qx = diag(6, 3, 0.3), Qu = diag(1, 0.1), and Qf =
1000 ∗ diag(1, 1, 1).
Fig. 6 depicts the safe trajectories in the xy plane regulated by
CBF and BT, with moving obstacles marked by a pink dashed
line. Both controllers exhibited good tracking performance
when the robot did not encounter obstacles, which could
be seen from 4-10 s and 33-40 s in Fig. 7. As shown
in the zoomed-in area of Fig. 6 and Fig. 7, CBF initiated
earlier deviation around static obstacles, leading to smoother
navigation. Meanwhile, the BT algorithm always approached
the boundary of the obstacle, leaving the robot stuck there
for a period of time. Numerous footprints left by the robot Fig. 9. Crossvalidated control performance on Husky A200
(See Video 1 at https://youtu.be/e UodhaXbi8)
were observed in that area. Nonetheless, it was noticed that
BT with a large prediction horizon of N > 30 could drive the In this section, the experimental validations of the proposed
robot to avoid obstacles smoothly. In the left zoomed-in of Fig. NMPC-SINDY-CBF and NMPC-SINDY-BT controllers were
6, CBF demonstrated effective handling of moving obstacles conducted utilising the Husky A200 robot, as shown in Fig. 9.
by generating an optimal predicted trajectory, allowing the The selected control parameters were Qx = diag(15, 15, 0.1),
robot to safely navigate around well before close approach. Qu = diag(0.001, 0.001), Qf = 104 ∗ diag(5, 5, 1), τ = 0.2,
Fig. 8 (a) illustrates the evolution of the error state vector N = 10, γ = 0.1. Additionally, validations were also

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

11

TABLE I
P ERFORMANCE INDEXES OF ALL INVESTIGATED METHODS WITH MEASUREMENT NOISE AND OBSTACLES

Control problem Performance indexes NMPC NMPC NMPC NMPC NMPC NMPC
True-BT NARX-BT SINDY-BT True-CBF NARX-CBF SINDY-CBF

Averaged optimization time

Point stabilization 15 17 18 12 25 14
(millisecond)
(N = 10)
∥ξr − ξ∥2 1.5 3.14 2.12 0.004 6.92 0.032

Averaged optimization time

Trajectory tracking 13 13 16 12 22 17
(millisecond)
(N = 20) P P
k i |ξei (k)| 928.6 10980 1188 615.8 11720 685

performance across all control scenarios. Additionally, the

SINDY model has the advantage of requiring less training
data and faster model identification, making it more advan-
tageous for autonomous systems, enabling them to adapt to
variations over time. The analysis also indicated that the
accuracy threshold decreases with increasing noise levels.
Fortunately, an investigation into the obstacle avoidance ability
of CBF and BT illustrated that a smaller N for NMPC-CBF
with a smaller γ could achieve comparable performance to
NMPC-BT with a larger N . The results demonstrate that the
proposed NMPC-SINDY-CBF effectively provides the robot
with safe and accurate operation in dynamic environments with
short prediction time. Moreover, using a shorter prediction
horizon has reduced the computational burden of NMPC,
enabling real-time application optimization. The experimental
Fig. 10. Crossvalidated control performance on Husky model in Gazebo validations have confirmed the effectiveness of the proposed
performed on the Husky model within the Gazebo envi- controllers in real-world settings.
ronment, depicted in Fig. 10, where the control parameters In future work, we will investigate the extension of SINDY
were set as Qx = diag(3, 3, 0.1), Qu = diag(0.01, 0.01), to the capability of modifying the identified model based on
Qf = 104 ∗ diag(1, 1, 1), τ = 0.15, N = 10, γ = 0.12. online feedback data to generate rapid model recovery from
A video demonstrating the experiment in detail is available abrupt disturbances.
online (Video 1). The experimental results aligned with the
outcomes obtained from the Matlab simulations, showing that R EFERENCES
both controllers exhibited good tracking performance. The [1] Spyros G. Tzafestas, Introduction to Mobile Robot Control. London:
CBF controller generated a better obstacle negotiation strategy, Elsevier, 2014.
while the BT controller led the robot to approach closer to [2] Y. P. Tian and S. Li, “Exponential stabilisation of nonholonomic dynamic
systems by smooth time-varying control,” Automatica, vol. 38, no. 7, pp.
obstacles, necessitating frequent stops and direction reversals 1139–1146, Jul. 2002.
to avoid collisions. This behaviour was evidenced by the [3] A. Lorı́a, E. Panteley, and K. Melhem, “UGAS of skew-symmetric time-
negative values in the linear velocity of the BT controller. varying systems: Application to stabilization of chained form systems,”
European Journal of Control, vol. 8, no. 1, pp. 33–43, 2002.
Moreover, Fig. 10 (c) and (d) illustrated that the BT required [4] B. Wang, S. Nersesov, and H. Ashrafiuon, “Formation regulation and
higher energy consumption to follow the reference path than tracking control for nonholonomic mobile robot networks using polar
the CBF, as indicated by the frequent reversals of velocities. coordinates,” IEEE Control Systems Letters, vol. 6, pp. 1909-1914, 2022.
[5] X. Peng, Z. Sun, M. Chen, and Z. Geng, “Arbitrary configuration
stabilization control for nonholonomic vehicle with input saturation: A
VI. C ONCLUSIONS C-nonholonomic trajectory approach,” IEEE Transactions on Industrial
Electronics, vol. 69, no. 2, pp. 1663-1672, 2022.
This paper presents a safety-critical control strategy for mo- [6] T. Bak, J. D. Bendtsen, and A. P. Ravn, Hybrid Control Design for a
bile robots facing parametric uncertainties and measurement Wheeled Mobile Robot. Aalborg: Department of Control Engineering,
Aalborg University, 2003.
errors, utilizing NMPC with either SINDY or NARX pre- [7] A. I. Cahyadi, P. Nugroho, and Y. Yamamoto, “Hybrid design of passive
diction models. Although both approaches effectively predict mobile robot teleoperation system,” ECTI Trans. Electr. Eng. Electron.
and control nonlinear systems under ideal conditions, analysis Commun., vol. 8, no. 2, pp. 232–238, 2010.
[8] P. Zhang, T. Liu, and Z. -P. Jiang, “Tracking control of unicycle
of fifty trials at various noise levels has shown that SINDY mobile robots with event-triggered and self-triggered feedback,” IEEE
models outperform NARX models in terms of robustness Transactions on Automatic Control, vol. 68, no. 4, pp. 2261-2276, 2023.
to noise and initial parameters. Consequently, NMPC with [9] C. Tiriolo, G. Franzè, and W. Lucia, “A receding horizon trajectory track-
ing strategy for input-constrained differential-drive robots via feedback
the NARX model struggled to control the robot under noisy linearization,” IEEE Transactions on Control Systems Technology, vol.
feedback, whereas the SINDY model maintained excellent 31, no. 3, pp. 1460-1467, 2023.

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Transactions on Intelligent Vehicles. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIV.2024.3389111

12

[10] L. Ding, S. Li, H. Gao, C. Chen, and Z. Deng, “Adaptive Partial [32] L. Zhang and H. Schaeffer, “On the convergence of the SINDy Algo-
Reinforcement Learning Neural Network-Based Tracking Control for rithm,” Multiscale Modelling and Simulation, vol. 17, pp. 948–972, 2019.
Wheeled Mobile Robotic Systems,” IEEE Transactions on Systems, Man, [33] K. S. Narendra and K. Parthasarathy, “Identification and control of
and Cybernetics: Systems, vol. 50, no. 7, pp. 2512-2523, 2020. dynamical systems using neural networks,” IEEE Transactions on Neural
[11] Z. Xu, T. Yan, S. X. Yang, and S. A. Gadsden, “Distributed leader Networks and Learning Systems, pp. 4–27, 1990.
follower formation control of mobile robots based on bioinspired neural [34] H. G. Bock and K. J. Plitt, “A multiple shooting algorithm for direct
dynamics and adaptive sliding innovation filter,” IEEE Transactions on solution of optimal control problems,” Proceedings 9th IFAC World
Industrial Informatics, vol. 20, no. 2, pp. 1180–1189, 2024. Congress Budapest, Pergamon Press, pp. 243-247, 1984.
[12] H. Li, L. Bi, and J. Yi, “Sliding-mode nonlinear predictive control of [35] J. Zeng, Z. Li, and K. Sreenath, “Enhancing feasibility and safety
brain-controlled mobile robots,” IEEE Transactions on Cybernetics, vol. of nonlinear model predictive control with discrete-time control barrier
52, no. 6, pp. 5419-5431, 2022. functions,” 60th IEEE Conference on Decision and Control (CDC), IEEE
[13] J. Hu, P. Bhowmick, and A. Lanzon, “Group coordinated control of Press, pp. 6137–6144, 2021.
networked mobile robots with applications to Object Transportation,” [36] L. Magni, D. Martino. Raimondo, and F. Allgower, Nonlinear Model
IEEE Transactions on Vehicular Technology, vol. 70, no. 8, pp. 8269- Predictive Control: Towards New Challenging Applications. Berlin, Hei-
8274, 2021. delberg: Springer Berlin Heidelberg, 2009.
[14] A. Wächter and L. T. Biegler, “On the implementation of an interior- [37] J. B. Rawlings and M. J. Risbeck, “Model predictive control with
point filter line-search algorithm for large-scale nonlinear programming,” discrete actuators: Theory and application,” Automatica, vol. 78, pp. 258-
Mathematical Programming, vol. 106, no. 1, pp. 25–57, 2006. 265, 2017.
[15] R. Verschueren et al., “Towards a modular software package for embed- [38] L. Grune and J. Pannek, Nonlinear Model Predictive Control. London:
ded optimization,” IFAC World Congress, vol. 51, no. 20, pp. 374 – 380, Springer, 2011.
2018. [39] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory and
[16] P. Sopasakis, E. Fresk, and P. Patrinos, “OpEn: Code generation for Design. Madison: Nob Hill Publishing, 2009.
embedded nonconvex optimization,” IFAC-PapersOnLine, vol. 53, no. 2, [40] A. R. Barron, “Approximation and estimation bounds for artificial neural
pp. 6548–6554, 2020. networks,” Machine Learning, vol. 14, pp. 115–133, 1994.
[17] P. N. Dao, H. Q. Nguyen, T. L. Nguyen, and X. S. Mai, “Finite Horizon [41] R. Benjamin and M. P. Laiu, “Convergence of weak-SINDy Surrogate
robust nonlinear model predictive control for wheeled mobile robots,” Models,” ArXiv abs/2209.15573, 2022.
Mathematical Problems in Engineering, vol. 2021, pp. 1–8, 2021. [42] X. Fang and W. H. Chen, “Model predictive control with preview:
[18] H. Li, L. Bi, X. Li and H. Gan, “Robust Predictive Con- recursive feasibility and stability,” IEEE Control Systems Letters, vol.
trol for EEG-Based Brain–Robot Teleoperation,” IEEE Transactions 6, pp. 2647-2652, 2022.
on Intelligent Transportation Systems, pp. 1–11, Jan. 2024, doi: [43] J. A. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl,
doi.org/10.1109/tits.2024.3359216. “Casadi: A software framework for nonlinear optimization and optimal
[19] Y. Zhang, X. Zhao, B. Tao, and H. Ding, “Point stabilization of nonholo- control,” Mathematical Programming Computation, vol. 11, no. 1, pp.
nomic mobile robot by Bézier Smooth Subline constraint nonlinear model 1–36, 2018.
predictive control,” IEEE/ASME Transactions on Mechatronics, vol. 26,
no. 2, pp. 990-1001, 2021. VII. B IOGRAPHY S ECTION
[20] L. Hewing, K. P. Wabersich, M. Menner, and M. N. Zeilinger. “Learning-
based model predictive control: Toward safe learning in control,” Annual
Review of Control, Robotics, and Autonomous Systems, vol. 3, no. 1, pp. Minh-Nhat Nguyen is currently working toward the Ph.D. degree in robotics
269–296, 2020. and intelligent control with the School of Electronics, Electrical Engineering
[21] F. Igea and M. Chatzis and A. Cicirello, “On the Combination and Computer Science, Queen’s University of Belfast, Belfast, U.K.
of Random Matrix Theory with Measurements On a Single Struc-
ture,” ASCE. ASME J. Risk Uncertainty Part B, Dec. 2022, doi: Mien Van is a Senior Lecturer (Associate Professor) in Robotics and
doi.org/10.1115/1.4054172. Control Engineering with the School of Electronics, Electrical Engineering
[22] V. Yury, S. Solodusha, E. Markova, E. Antipina, and V. Boeva, “Identifi- and Computer Science, Queen’s University Belfast, United Kingdom. He has
cation of Quadratic Volterra Polynomials in the ‘Input–Output’ Models of authored and co-authored for over 65 journals and conference papers and some
Nonlinear Systems,” Mathematics, vol. 10, no. 11, pp. 1836–1836, 2022. are published in prestigious journals such as IEEE Transactions on Fuzzy
[23] K. K. Ahn and H. P. H. Anh, “Inverse Double NARX Fuzzy Modeling Systems, IEEE/ASME Transactions on Mechatronics, IEEE Transactions on
for System Identification,” IEEE/ASME Transactions on Mechatronics, Industrial Informatics, IEEE Transactions on Cybernetics, etc. His current
vol. 15, no. 1, pp. 136-148, 2010. research interests include manufacturing robotics, robot control, robot-vision
[24] T. Wang, H. Gao, and J. Qiu, “A combined adaptive neural network and systems, fault diagnosis and fault tolerant, machine learning, and sensing and
nonlinear model predictive control for Multirate Networked Industrial perception.
Process Control,” IEEE Transactions on Neural Networks and Learning
Systems, vol. 27, pp. 416–425, 2016.
Stephen Mcllvanna is currently working toward the Ph.D. degree in robotics
[25] E. Aggelogiannaki and H. Sarimveis, “Nonlinear model predictive
and intelligent control with the School of Electronics, Electrical Engineering
control for distributed parameter systems using data driven artificial neural
and Computer Science, Queen’s University of Belfast, Belfast, U.K.
network models,” Computers and Chemical Engineering, vol. 32, no. 6,
pp. 1225–1237, 2008.
[26] S. L. Brunton, J. L. Proctor, and J. N. Kutz, “Discovering governing Yuzhu Sun is currently working toward the Ph.D. degree in robotics and
equations from data by sparse identification of nonlinear dynamical intelligent control with the School of Electronics, Electrical Engineering and
systems,” Proceedings of the National Academy of Sciences, vol. 113, Computer Science, Queen’s University of Belfast, Belfast, U.K.
no. 15, pp. 3932–3937, 2016.
[27] E. Kaiser, J. N. Kutz, and S. L. Brunton, “Sparse identification of Jack Close is currently working toward the Ph.D. degree in robotics and
nonlinear dynamics for model predictive control in the low-data limit,” intelligent control with the School of Electronics, Electrical Engineering and
Proceedings of the Royal Society A, vol. 474, no. 2219, p. 20180335, Computer Science, Queen’s University of Belfast, Belfast, U.K.
2018.
[28] Vayeda Anshav Bhavesh, “Comparison of various obstacle avoidance Kabirat Olayemi is currently working toward the Ph.D. degree in robotics
algorithms,” International Journal of Engineering Research and Technol- and intelligent control with the School of Electronics, Electrical Engineering
ogy, vol. V4, no. 12, 2015. and Computer Science, Queen’s University of Belfast, Belfast, U.K.
[29] A. D. Ames et al., “Control barrier functions: Theory and applications,”
18th European Control Conference (ECC), pp. 3420-3431, 2019.
[30] J. G. Carvalho Filho, E. A. Carvalho, L. Molina, and E. O. Freire, “The Yan Jin (Member, IEEE) Dr Yan Jin is a Professor in the School of Mechan-
impact of parametric uncertainties on mobile robots velocities and pose ical and Aerospace Engineering at Queen’s University Belfast. His research
estimation,” IEEE Access, vol. 7, pp. 69070-69086, 2019. interest is in robotics, smart manufacturing, and production management. He
has published over 150 peer-reviewed technical papers. He is a Fellow of
[31] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward
Institution of Mechanical Engineers UK (IMechE). He is a member of IEEE.
networks are universal approximators,” Neural networks 2, pp. 359–366,
1989.

Authorized licensed use limited to: CHONNAM NATIONAL UNIVERSITY. Downloaded on April 24,2024 at 01:46:10 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.