IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO.

12, DECEMBER 2023 8075

Safe Consensus Tracking With Guaranteed Full State and Input

Constraints: A Control Barrier Function-Based Approach
Junjie Fu , Senior Member, IEEE, Guanghui Wen , Senior Member, IEEE,
and Xinghuo Yu , Fellow, IEEE

Abstract—In this article, we consider the safe consensus effect should be taken into consideration in order to guarantee the
tracking problem for uncertain second-order nonlinear multiagent stability of the closed-loop system. In [5] and [6], global nonlinear
systems subject to position, velocity, and input constraints. The
agents are required to cooperatively track a desired leader’s trajec- consensus and formation tracking controllers have been proposed
tory while always satisfying their local state and input constraints. for agents with simple dynamics. Low-gain-based method has been
Therefore, conflicting objectives may exist for an agent when the employed in [7] and [8] to achieve global or semiglobal consensus of
desired trajectory violates its local constraints. We propose to linear multiagent systems. Model predictive control (MPC) method
solve this problem using a control barrier function (CBF)-based
approach. The cooperative tracking objective is encoded by a novel
has been generalized to handle consensus and formation control of
control Lyapunov function-based condition while the state and nonlinear multiagent systems in[9] and [10].
input constraints are handled by CBF-based constraints. For the Apart from the input saturation, the state of the agents may also
relative degree two position constraint, two classes of CBF-based have constraints due to performance or safety concerns. For example,
conditions are proposed based on high-order CBFs and a modified for multiple fixed-wing aircraft, the velocities of the agents should
CBF design, respectively. It is proven that, with the modified CBF,
there always exist feasible control inputs that satisfy all the CBF- have lower bound in order to keep flying and for multiple autonomous
based constraints. Then, unified quadratic programming-based vehicles, the velocities usually have upper bound due to collision
controllers are formulated and the performances are analyzed. avoidance requirements. For multiple manipulators in the working
Simulation examples are provided to verify the obtained results. space, both the positions and velocities of the agents may be con-
Index Terms—Control barrier function, input saturation, multia- strained due to the safety operation protocols. In these situations,
gent system, safe consensus tracking, state constraint. all the state and input constraints should be simultaneously taken
into consideration when designing the consensus tracking controllers.
Several methodologies have been proposed to handle these constraints
I. INTRODUCTION in existing literature. The first class of methods aims to directly propose
Due to the rapid development of sensing, communication, and com- nonlinear distributed controllers such that both the state and input
putation technologies, multiagent system has seen increasing use in var- constraints can be satisfied. In [11] and [12], nonlinear distributed
ious fields, such as environmental sensing, cooperative transportation, controllers have been proposed, which achieve consensus tracking
target tracking, and so on [1]. Among the various cooperative control of second-order integrator systems with both the velocity and input
tasks of the multiagent systems, consensus tracking plays a fundamental constraints. In [13], a control barrier function (CBF)-based method
role as many other tasks, such as formation, containment, and encir- was employed to achieve collision avoidance between the agents and
clement, can be solved based on the consensus tracking controllers [2]. the obstacles. However, these results were limited to integrator systems
Many important results have been obtained for consensus tracking of without uncertainties, and the position constraints were not considered
multiagent systems handling different agent dynamics, communication simultaneously. The second class of methods generalizes the barrier
graphs, and convergence speed requirement [3], [4]. Lyaupnov function (BLF) method to multiagent systems, which could
It should be noted that most of the existing results have focused on solve the full state and input-constrained consensus problems [14],
the consensus tracking of multiagent systems without considering the [15]. Moreover, to relax the feasibility conditions, nonlinear state
physical constraints of the agents in practical applications. Since all transformation-based methods have been given in [16] and [17] to
physical plants can only have limited actuation power, input saturation achieve constrained consensus tracking of multiagent systems. Another
class of methods for state and input constrained cooperative control
of multiagent systems is the MPC-based controllers [18], [19], which
Manuscript received 24 December 2022; accepted 4 June 2023. can handle nonlinear dynamics and constraints better. Note that even
Date of publication 7 June 2023; date of current version 5 December
though most of the existing MPC-based results have not considered the
2023. This work was supported in part by the National Natural Science
Foundation of China under Grant 62173085, Grant U22B2046, Grant position constraints, they can be incorporated into the local optimization
62073079, and Grant 62088101, and in part by the Australian Research problem formulation. However, the MPC-based methods generally
Council under Grant DE180101268 and Grant DP200101199. Recom- require strict initial feasibility conditions to maintain the feasibility of
mended by Associate Editor Giuseppe Notarstefano. (Corresponding the optimization problems and an accurate model of the agent dynamics
author: Junjie Fu.)
Junjie Fu and Guanghui Wen are with the Jiangsu Key Laboratory is usually needed to guarantee the performance. Moreover, the computa-
of Networked Collective Intelligence, School of Mathematics, Southeast tion costs may hinder real-time implementation on practical multiagent
University, Nanjing 211189, China (e-mail: [email protected]; wen- systems.
[email protected]). Although there have been some initial progress in addressing the
Xinghuo Yu is with the School of Engineering, RMIT University, Mel-
state and input constrained consensus tracking of multiagent systems as
bourne, VIC 3001, Australia (e-mail: [email protected]).
Color versions of one or more figures in this article are available at abovementioned, a common limitation of these results is that the desired
https://doi.org/10.1109/TAC.2023.3283697. trajectory has to lie in the constrained region of all the follower agents.
Digital Object Identifier 10.1109/TAC.2023.3283697 This condition is central for the construction of Lyapunov functions

0018-9286 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
8076 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023

in the BLF-based and nonlinear state transformation-based methods, −α(h(x)) for all x ∈ D [20]. Consider the set Kcbf (x) =
and for the stability analysis in the MPC-based methods. However, {u ∈ U : Lf h(x) + Lg h(x)u + α(h(x))  0}.
since the leader’s trajectory may be given by an external agent or Lemma 2 ([22]): For the nonlinear control system (1), if h is a CBF
human operator, it may easily be the case that the desired trajectory on D and ∂h∂x
(x) = 0 for all x ∈ ∂C, then any locally bounded and
violates the constraints of the follower agents occasionally. In these measurable controller u(x) ∈ Kcbf (x) renders the set C safe.
cases, a more desired behavior of the multiagent system should be safe Based on the CLF and CBF, the following QP-based controller can
consensus tracking, which means the agents should always satisfy their be formulated to unify stability and safety:
local constraints while achieving conventional consensus tracking only 1 T
u(x) = arg min 2
u H(x)u + pδ 2
when the desired trajectory is feasible for all the agents in the network. (u,δ)∈Rm+1
However, to the best of authors’ knowledge, an effective control strategy s.t. Lf V (x) + Lg V (x)u + γ(V (x))  δ
that achieves this objective is still lacking. Lf h(x) + Lg h(x)u + α(h(x))  0
Inspired by the control Lyapunov function (CLF)-CBF-based control
framework proposed in [20] for single systems, we propose a novel where H(x) is any positive definite matrix and δ is a relaxation
CBF-based safe consensus tracking strategy for second-order uncer- variable penalized by p > 0, which relaxes the condition on stability
tain nonlinear multiagent systems. The consensus tracking objective to guarantee safety.
is encoded by a novel CLF-based condition relying on only local
relative state information between neighboring agents. The state and B. Problem Formulation
input constraints of the agents are handled by several CBFs, which Consider a second-order nonlinear multiagent system
provide the constraints on the local control inputs at each time instant.
Finally, quadratic programming (QP)-based controllers are proposed to ẋi1 = xi2
achieve the safe consensus tracking control objective. The main contri- ẋi2 = fi (xi ) + gi (xi )ui , i = 1, . . . , N (2)
butions of this work lie in the following aspects. First, we formulate a
2
safe consensus tracking problem for second-order uncertain nonlinear where xi = [xi1 , xi2 ] ∈ R is the state of agent i, ui ∈ R is the
T

multiagent systems where the leader’s trajectory is not required to lie control input, fi (xi ) : R2 → R is the nonlinear dynamics, and gi (xi ) :
in the constrained region of all the followers. Second, a CBF-based R2 → R is the control gain. The safe consensus tracking problem
approach is proposed, which solves the safe consensus tracking problem of (2) is considered, and the leader’s trajectory is given by a dif-
successfully. A new type of CBF is then proposed, which guarantees ferentiable function x01 (t) with |ẋ01 (t)| ≤ vL , where vL ≥ 0. For
the feasibility of local QP problems. Third, the proposed method does agent i, suppose that the position and velocity constraints |xi1 | ≤ pi
not need extensive computation as in the MPC-based methods and, and |xi2 | ≤ vi are required for performance or safety concerns. The
therefore, is more suitable for practical applications. input constraint |ui | ≤ qi is required due to the actuator saturation
The rest of this article is organized as follows. In Section II, some constraint. For the agent dynamics, suppose that the nonlinear dynamics
preliminaries and the problem formulation are given. In Section III, the fi is unknown and only an approximate fˆi is known, which satisfies
CBF-based safe consensus tracking control strategy is presented. Sim- |fi (xi ) − fˆi (xi )| ≤ αi and |fˆi (xi )| ≤ βi with two positive constants
ulation examples are given in Section IV. Finally, Section V concludes αi , βi > 0 on the region {xi : |xi1 | ≤ pi , |xi2 | ≤ vi }. The control gain
this article. gi (xi ) > 0 is assumed to be known and has the lower and upper bounds
of g i and ḡi , respectively.
Remark 1: The assumption of known gain function gi (xi ) is not
II. PRELIMINARIES AND PROBLEM FORMULATION
restrictive since if only an approximate ĝi is available, which satisfies
A. Preliminaries |ĝi − gi | ≤ ᾱi , then by redefining fi as fi + (gi − ĝi )ui , we can trans-
form the agent dynamics into the form that we consider. Furthermore,
In this section, some preliminaries on the CLF-CBF-based control
the obtained results can be easily generalized to the cases xi ∈ R2n
method for single systems is presented. Consider a nonlinear affine
where n > 1.
control system
With the input transformation ui = g1i wi , we have ẋi2 = wi + fi .
ẋ = f (x) + g(x)u (1) Moreover, when |wi | ≤ qi g i , we have |ui | ≤ qi .
Let the interaction relation among the N agents be described by
where f and g are locally Lipschitz, and x ∈ D ⊂ Rn and u ∈ U ⊂ Rm an undirected and connected graph G. The adjacency matrix A :=
are the set of admissible inputs. V is a CLF for (1) if it is positive def- [aij ] ∈ RN ×N satisfies aij > 0 if agents i and j can obtain information
inite and satisfies inf u∈U [Lf V (x) + Lg V (x)u]  −γ(V (x)), where from each other and aij = 0, otherwise. The Laplacian matrix L :=

γ is a class K function. Consider the set Kclf (x) = {u ∈ U : Lf V (x) [lij ] ∈ RN ×N is defined by lii = N j=1 aij and lij = −aij for i = j.
+ Lg V (x)u  −γ(V (x))} for every point x ∈ D. The following Moreover, ai0 > 0 if agent i can directly obtain the information of the
lemma can be derived from [21]. leader and ai0 = 0, otherwise. Let A0 = diag{a10 , . . . , aN 0 }. Then,
Lemma 1: For the nonlinear control system (1), if there exists a the matrix H = L + A0 is positive definite under the condition that at
CLF V : D → R≥0 , then any locally bounded and measurable feed- least one agent can directly obtain the information of the leader [7].
back controller u(x) ∈ Kclf (x) asymptotically stabilizes the system to In the conventional consensus tracking problem, the control objective
x∗ = 0. is usually to design the control input ui for each agent such that the
Safety of (1) can be framed in the context of enforcing in- states of all the agents gradually approach the common trajectory of
variance of a set. Consider a set C defined as the superlevel set the leader, i.e., xi1 (t) − x01 (t) → 0 as t → ∞. When the state and
of a continuously differentiable function h : D ⊂ Rn → R with input constraints of the agents are presented, this consensus tracking
C = {x ∈ D ⊂ Rn : h(x) ≥ 0)}. The system (1) is said to be requirement may be in conflict with the local constraints of the agents.
safe with respect to the set C if the set C is forward invari- In these cases, the safe consensus tracking objective is preferred, which
ant. The function h is a CBF if there exists an extended class means when the leader’s trajectory is not in conflict with the constraints
K∞ function α such that it holds supu∈U [Lf h(x) + Lg h(x)u]  of the agents, consensus tracking can be achieved. When conflict exists,

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023 8077

the constraints of the agents are prioritized and compromised consensus The desired velocity (3) has the nice feature that it is globally
tracking performance is tolerated. As a result, a multiobjective control bounded by k1 + k2 . Therefore, it is possible to achieve this velocity
problem is considered in this article, and we solve this problem using while satisfying the velocity constraint by tuning the controller gains.
a CBF-based approach. Furthermore, the desired velocity only depends on the relative position
information from neighbors. This fact can be utilized to obtain a local
III. CBF-BASED SAFE CONSENSUS TRACKING CONTROL CLF-based consensus tracking condition for each agent. Specifically,
we define the following Lyapunov function for the consensus tracking
A. CLF-Based Consensus Tracking Condition objective:
In order to transform the distributed consensus tracking problem 
N
into a form suitable for the CLF-based method, we first consider the Vf = [xi2 + k1 tanh(ξi ) + k2 tanh(ξi /ε)]2 .
following desired dynamics for the multiagent system: i=1

ẋi1 = −k1 tanh (ξi ) − k2 tanh (ξi /ε) (3) Note that when Vf converges to zero exponentially, we have xi2 =
xdi2 + ei2 , where xdi2 = −k1 tanh(ξi ) − k2 tanh(ξi /ε) and ei2 is a

where ξi = N j=0 aij (xi1 − xj1 ), k1 > 0, k2 ≥ vL , and ε > 0 is a
term, which converges to zero exponentially. Following similar analysis
small positive parameter. as in the proof of Theorem 1, one can show that practical consensus
Remark 2: The desired dynamics (3) is motivated by the distributed tracking for the agents can be achieved. From Section II-A, the condition
consensus tracking controller for first-order multiagent systems. It will
V̇f + λVf ≤ 0 (5)
be used to construct the CLFs in the later development. The first
hyperbolic tangent function can be replaced by any globally bounded is imposed to achieve exponential convergence of the Lyapunov func-
strictly increasing odd function while the second term is a continuous tion Vf where λ > 0. Note that one can obtain a decoupled sufficient
approximation of the sign function with the small positive ε > 0. condition for (5) with
We show that under the desired dynamics (3), practical consensus
tracking of the multiagent system (2) can be achieved. V̇if + λVif ≤ 0 (6)
Theorem 1: Suppose the interaction graph G is undirected and for i = 1, . . . , N , where Vif = [xi2 + k1 tanh(ξi ) + k2
connected and at least one follower can directly obtain the information tanh(ξi /ε)]2 . Based on (3) and (6), we have the CLF-based consensus
of the leader. Along the desired dynamics (3), the consensus tracking er- tracking condition 2(xi2 − xdi2 )[wi + fi − ẋdi2 ] + λ(xi2 − xdi2 )2 ≤ 0.
rors x̃i1 = xi1 − x01 , i = 1, . . . , N are uniformly ultimately bounded Noting that fi is unknown, the following robust CLF-based condition
by a positive number, which approaches zero as ε approaches zero. on wi is employed:
Proof: Let x̃ = [x̃11 , . . . , x̃N 1 ]T and ξ = [ξ1 , . . . , ξN ]T . We have  
that ξ = H x̃ and ξ 22 = x̃T H 2 x̃  λ2 x̃ 22 , where λ = λmin (H) > 2(xi2 − xdi2 )wi + 2αi xi2 − xdi2 
− −
0. From (3), we have + 2(xi2 − xdi2 )(fˆi − ẋdi2 ) + λ(xi2 − xdi2 )2 ≤ 0 (7)
x̃˙ = −k1 tanh (ξ) − k2 tanh (ξ/ε) − 1N ẋ01 (4) which is distributed since only local velocity and relative position, and
velocity information with respect to neighbors are needed.
where 1N ∈ R is the vector of all ones. Consider the Lyapunov
N
Remark 3: The CLF-based condition (7) can be generalized to han-
function V = 12 x̃T H x̃. It follows that
dle other types of coordination tasks, such as distributed containment
V̇ = −k1 ξ T tanh(ξ) − k2 ξ T tanh(ξ/ε) − ξ T 1N ẋ01 by employing different local desired dynamics (3). Moreover, general
directed communication graphs may also be handled by showing the

N convergence of the desired dynamics with graph related Lyapunov
 −k1 ξ T tanh(ξ) − [k2 ξi tanh(ξi /ε) − vL |ξi |]. functions.
i=1

From [23, Lemma 1], we have that for any η ∈ R and ε > 0, it holds B. CBF-Based Constraints Handling
0 ≤ |η| − η tanh(η/ε) ≤ κε, where κ = 0.2785. Then, it follows:
1) Velocity Constraint: The velocity constraint is to be handled
V̇  −k1 ξ T tanh(ξ) + k2 N κε by the CBF candidate hsi = vi2 − x2i2 . Note that hsi has relative degree
one with respect to the agent dynamics (2). Therefore, the CBF-based
where we have used the condition k2 ≥ vL . Let c > 0 be a constant, condition ḣsi + ρ1 hsi ≥ 0 can be used to obtain the local input con-
which satisfies c tanh(c) > k2kN1κε . It is easy to see that c can be straint, which leads to −2xi2 (fi + wi ) + ρ1 (vi2 − x2i2 )  0. Since fi
taken arbitrarily small as ε approaches zero. Under the condition is unknown, the following robust condition on wi is considered:
V  2λλ̄2 N c2 , noting that V  12 λ̄ x̃ 22 where λ̄ = λmax (H) > 0, we  
−
−2xi2 wi − 2xi2 fˆi − 2αi |xi2 | + ρ1 vi2 − x2i2  0. (8)
2
 λ2 x̃ 2
 N c2 , which leads to ξ
ξ
have ξ 2 2 ∞  √ 2  c. Then, Note that under the condition αi + βi < qi g i , one can always choose
− N

it follows that V̇  −k1 c tanh(c) + k2 N κε < 0. the control input wi = −qi g i sgn(xi2 ) such that (8) is satisfied. That
Therefore, the set V  2λλ̄2 N c2 is globally attractive and positive is, the CBF-based condition (8) is compatible with the input constraint
−
|wi | ≤ qi g i . (9)
invariant. Noting that 12 λ x̃ 22  V , it follows that x̃ 2 is uniformly ul-
− 2) Position Constraint: For the position constraint, one can
timately bounded by N λ̄/λ3 c, which leads to that x̃i1 , i = 1, . . . , N
−
 similarly consider the CBF candidate hri = p2i − x2i1 . Note, however,
are uniformly ultimately bounded by N λ̄/λ3 c. It is easy to see that that hri has relative degree two with respect to the agent dynamics (2).
−
the consensus tracking error bound approaches zero as ε approaches Therefore, high-order CBF should be utilized to obtain the constraint
zero.  on the control input. Inspired by the exponential high-order CBF given

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
8078 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023

in [24], the CBF-based condition ḧri + (ρ2 + ρ3 )ḣri + ρ2 ρ3 hri  0 [−qi g i , qi g i ] such that h+ −
ri (t) > 0, hri (t) > 0, and hsi (t) > 0 hold
is considered, where ρ2 , ρ3 > 0. It follows that −2xi1 (wi + fi ) − for ∀t ≥ 0.
2x2i2 − 2xi1 xi2 (ρ2 + ρ3 ) + ρ2 ρ3 (p2i − x2i1 )  0. Noting that fi is un- Proof: We prove the conclusion by showing that there always exists
known, the following robust condition on wi is derived: control input wi such that the CBF-based conditions ḣ+ +
ri + ρ4 hri 
− −
0, ḣri + ρ4 hri  0, ḣsi + ρ1 hsi ≥ 0 and the input constraint wi ∈
− 2xi1 wi − 2xi1 fˆi − 2αi |xi1 | − 2x2i2 [−qi g i , qi g i ] are satisfied simultaneously. When xi2 > 0, from (8),
 
− 2xi1 xi2 (ρ2 + ρ3 ) + ρ2 ρ3 p2i − x2i1  0. (10) (9), and (11), it can be seen that wi = −qi g i sgn(xi2 ) is a feasible
input, which satisfies all the constraints simultaneously. Similarly, when
Under (10) and the initial conditions ḣri (0) + ρ2 hri (0) > 0 and xi2 < 0, wi = −qi g i sgn(xi2 ) satisfies all the constraints (8), (9), and
hri (0) > 0, the position constraint can be enforced. (12) simultaneously. Finally, when xi2 = 0, there only exists the input
Note that for the CBF-based condition (10), one cannot guarantee constraint (9), which means all wi ∈ [−qi g i , qi g i ] are feasible. There-
that it is compatible with both the velocity constraint induced condition fore, under the conditions h+ −
ri (0) > 0, hri (0) > 0, and hsi (0) > 0,
(8) and the input constraint (9). When there are conflicts, there exists no there exists wi ∈ [−qi g i , qi g i ] such that h+ −
ri (t) > 0, hri (t) > 0 and
control input, which satisfies all the constraints simultaneously and the hsi (t) > 0 hold for ∀t ≥ 0. 
safety of the agents cannot be guaranteed. Therefore, we consider an
alternative position constraint CBF design, which guarantees feasibility
in the presence of all the position, velocity, and input constraints. For C. CBF-Based Safe Consensus Tracking
agent i, with the initial position xi1 ∈ [−pi , pi ] and the initial velocity With the CLF-based consensus tracking condition and the CBF-
xi2 ∈ [−vi , vi ], when xi2 > 0, under the condition αi + βi < qi g i , based constraints handling conditions, a unified QP-based controller
one can guarantee a deceleration of at least γi = qi g i − αi − βi with can be formulated to achieve safe consensus tracking of the multiagent
wi = −qi g i sgn(xi2 ) until the agent stops. Then, under the condition system. Based on the two different CBF designs for the position
x2
pi − xi1 − 2γi2i  0, the position constraint will be satisfied. When constraints, two different control strategies are proposed.
xi2 < 0, with the similar control strategy, under the condition xi1 + First, we use the exponential CBF to handle the position constraint,
x2 which leads to the following QP problem for each agent:
pi − 2γi2i  0, the position constraint can be guaranteed. Therefore, the
following two CBF candidates are considered:
 QP1 : min δi2 + wi2
x2 (wi ,δi )
hri (xi1 , xi2 ) = pi − xi1 − 2γi , xi2  0
i2
+
 
pi − xi1 , xi2 < 0 s.t. 2(xi2 − xdi2 )wi + 2αi xi2 − xdi2 
and + 2(xi2 − xdi2 )(fˆi − ẋdi2 ) + λ(xi2 − xdi2 )2  δi (13)

x2
h− pi + xi1 − i2
, xi2 0 − 2xi1 wi − 2xi1 fˆi − 2αi |xi1 | − 2x2i2
ri (xi1 , xi2 ) =
2γi
pi + xi1 , xi2 > 0.  
− 2xi1 xi2 (ρ2 + ρ3 ) + ρ2 ρ3 p2i − x2i1  0 (14)
Both h+ −
ri and hri are continuously differentiable functions with respect
 
− 2xi2 wi − 2xi2 fˆi − 2αi |xi2 | + ρ1 vi2 − x2i2  0 (15)
to xi1 and xi2 . Furthermore, when both h+ −
ri > 0 and hri > 0 are
maintained, we have −pi < xi1 < pi , i.e., the position constraint is |wi |  g i qi (16)
satisfied. Moreover, both h+ −
ri and hri have relative degree one with
respect to the control input wi . where (13) is the relaxed CLF-based condition with the slack variable
The CBF-based conditions corresponding to h+ − +
ri and hri are ḣri + δi , (14) is the exponential CBF-based position constraint condition, (15)
+ − −
ρ4 hri  0 and ḣri + ρ4 hri  0, respectively. We can show that at is the CBF-based velocity constraint condition, and (16) is the control
any time instant, at most, one of the conditions is active. For exam- input constraint. The performance index is to minimize the norm of
ple, when xi2 > 0, the CBF-based condition ḣ+ +
ri + ρ4 hri  0 leads the control input wi and the violation of the CLF-based condition. The
x2
to − xγi2 (wi + fi ) − xi2 + ρ4 (pi − xi1 − 2γi2i )  0. Since fi is un- slack variable δi is introduced to put priority on guaranteeing safety of
i
known, the following sufficient condition is considered: the agents. The following conclusion is obtained under the condition
that there exists no conflict between the CBF-based conditions.
xi2 xi2 ˆ |xi2 | x2
− wi − f i − αi − xi2 + ρ4 pi −xi1 − i2  0. Theorem 3: Suppose that hsi (0) > 0, ḣri (0) + ρ2 hri (0) > 0, and
γi γi γi 2γi hri (0) > 0. Under the condition that the QP 1 is feasible at each time
(11)
instant, safe consensus tracking of (2) with guaranteed full state and
For h− , the CBF-based condition is automatically satisfied. Similarly, 1
ri input constraints is achieved with the control input ui = gi (x wi∗ ,
when xi2 < 0, only the CBF-based condition ∗
i)
where wi is the optimal solution of QP 1.
xi2 xi2 ˆ |xi2 | x2 Proof: When the QP 1 is feasible at each time instant, the CBF-
− wi − fi −αi + xi2 + ρ4 pi +xi1 − i2 0
γi γi γi 2γi based conditions (14) and (15) are satisfied along the trajectories of the
(12) agents. Therefore, under the initial conditions hsi (0) > 0, ḣri (0) +
is active while ḣ+ri + ρ4 h +
ri  0 is automatically satisfied. Moreover, ρ2 hri (0) > 0, and hri (0) > 0, it follows that hri (t) > 0 and hsi (t) >
when xi2 = 0, both the CBF-based conditions are automatically satis- 0 for all t ≥ 0, which means safe consensus tracking with the full state
fied. and input constraints is achieved. 
A crucial feature of the proposed CBFs h+ −
pi and hpi for the position However, even with the relaxed CLF-based condition, the QP 1 is
constraint is that feasible control input, which satisfies all the position, not guaranteed to be feasible at all time since there may exist conflict
velocity, and input constraints induced conditions, is guaranteed. between the position-based condition (14) and the velocity- and input-
Theorem 2: Under the conditions h+ −
ri (0) > 0, hri (0) > 0, and based conditions (15) and (16). To ensure the safety of the agents and
hsi (0) > 0, there always exists feasible control input wi ∈ the feasibility of the local QP problem, we propose the second control

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023 8079

strategy based on the CBFs h+ −

pi and hpi

QP2 : min δi2 + wi2

(wi ,δi )
 
s.t. 2(xi2 − xdi2 )wi + 2αi xi2 − xdi2 

+ 2(xi2 − xdi2 )(fˆi − ẋdi2 ) + λ(xi2 − xdi2 )2  δi (17)

xi2 xi2 ˆ |xi2 |
− wi − f i − αi − xi2
γi γi γi
x2i2
+ ρ4 pi − xi1 −  0, if xi2 > 0 (18)
2γi
xi2 xi2 ˆ |xi2 |
− wi − f i − αi + xi2
γi γi γi
x2i2
+ ρ4 pi + xi1 −  0, if xi2 < 0 (19)
2γi Fig. 1. States and inputs of the agents with the CLF-based consensus
  tracking controller.
− 2xi2 wi − 2xi2 fˆi − 2αi |xi2 | + ρ1 vi2 − x2i2  0 (20)
|wi |  g i qi . (21)

Compared with QP 1, the position constraint condition is replaced by

(18) and (19). From Theorem 2, we know that the QP 2 is always
feasible, which leads to the following result.
Theorem 4: Under the conditions that h+ −
ri (0) > 0, hri (0) > 0, and
hsi (0) > 0, safe consensus tracking of (2) with guaranteed full state
1
and input constraints is achieved with the control input ui = gi (x i)
wi∗ ,
∗
where wi is the optimal solution of QP 2. Moreover, for a static leader,
which satisfies |x01 | < pi , i = 1, . . . , N , asymptotic consensus track-
ing is achieved while the full state and input constraints are guaranteed.
Proof: Note that with the control input ui , all the CBF-based condi-
tions are satisfied. Therefore, under the initial conditions h+ ri (0) > 0,
+
h− −
ri (0) > 0, and hsi (0) > 0, we have that hri (t) > 0, hri (t) > 0, and
hsi (t) > 0 for all t ≥ 0. It follows that safe consensus tracking of (2)
with guaranteed full state and input constraints is achieved.
When x01 is a constant, we have ẋ01 = 0. Then, for the desired
dynamics (3), we can set k2 = 0 and the error dynamics (4) reduces Fig. 2. States and inputs of the agents with the QP 1-based controller.
to ξ˙ = −k1 H tanh(ξ). It is easy to show that ξ → 0 as t → ∞. With
the QP 2, the states of the agents will move toward the trajectory of feasibility of the optimization problems can be guaranteed under both
ξ. Therefore, xi2 approaches zero as t → ∞. Note that the CBF-based the local state and input constraints.
conditions are automatically satisfied as xi2 → 0. Therefore, the opti-
mal control input wi∗ will lead to Vif → 0 and asymptotic consensus
IV. SIMULATION
tracking is, therefore, achieved while the full state and input constraints
are guaranteed. In this section, we consider a multirobot system with the second-
Remark 4: Due to the presence of the unknown leader’s information order dynamics ṙi = si and mi ṡi = Fi , where ri = [ri1 , ri2 ]T ∈ R2
and the uncertain dynamics, robust relaxation terms are present in is the position of robot i, si = [si1 , si2 ]T ∈ R2 is the velocity, mi ∈ R
the CLF-CBF conditions of QP 1 and QP 2 such that the optimal is the mass, and Fi = [Fi1 , Fi2 ]T ∈ R2 is the control input. Suppose
solution wi∗ may not be Lipschitz continuous. When the leader’s ve- that only a nominal mass m̂i is known for each robot, then the dynamics
locity and acceleration are available and the local dynamics fi (xi ) can be transformed into ṙi = si , ṡi = m̂1i Fi + m̂mii−mm̂i
i
Fi .
of the agents are known, the desired dynamics (3) can be taken as Since the two dimensions of the position ri are decoupled, we
ẋi1 = −k1 tanh(ξi ) + ẋi0 , and the linear conditions in QP 1 and QP only illustrate the controller design for the first dimension. De-
2 can be rewritten with fˆi (xi ) = fi (xi ) and αi = 0. Then, following note xi1 = ri1 , xi2 = si1 , gi = 1/m̂i , fi = m̂mii−m m̂i
i
Fi1 , and ui =
from the Mangasarian–Fromovitz regularity conditions [20], [25], the Fi1 , then we have the agent dynamics (2). Suppose that there are
optimal control input is Lipschitz continuous in this case. five followers. The trajectory of the leader is given by x01 (t) =
Remark 5: The proposed CBF-based distributed control strategy 5 sin(t/4). The interaction relation among the agents is determined by
may also be extended to handle interagent constraints, such as the the matrix H = [4, −1, 0, −1, −1; −1, 4, −1, −1, 0; 0, −1, 2, −1, 0;
connectivity maintenance and collision avoidance constraints. By prop- −1, −1, −1, 4, −1; −1, 0, 0, −1, 2]. In the following, we will consider
erly constructing the safety sets for connectivity and collision avoid- the effects of different position constraints. The velocity and input
ance [13], similar CBF conditions can be derived and used to formulate constraints of the agents are given by [v1 , . . . , v5 ] = [3, 3, 3, 3, 3] and
the local QP problems to obtain the final control input for each agent. [q1 , . . . , q5 ] = [20, 30, 30, 25, 35], respectively. The true and nomi-
The main difficulty would be to define suitable safety sets such that the nal masses of the robots are given by [m1 , . . . , m5 ] = [5, 6, 7, 5, 8]

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
8080 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023

Fig. 5. Practical consensus tracking under the position constraints

Fig. 3. States and inputs of the agents with the QP 2-based controller.
[p1 , . . . , p5 ] = [5, 5, 5, 5, 5].

Fig. 6. Safe consensus tracking under the different position constraints

Fig. 4. States and inputs of the agents with the QP 2-based controller [p1 , . . . , p5 ] = [2, 3, 2, 4, 2].
under known leader’s information and local dynamics.

The simulation results are shown in Fig. 2. It can be seen that due
and [m̂1 , . . . , m̂5 ] = [4, 6.5, 6, 5.5, 9], respectively. By taking fˆi = to the conflict between the position, velocity, and input constraints,
0, we have [α1 , . . . , α5 ] = [1, 0.39, 0.72, 0.46, 0.49] and βi = 0, safe consensus tracking cannot be achieved with the QP 1. Then, the
i = 1, . . . , 5. For the desired dynamics (3), the controller parameters position constraint is handled by the proposed CBFs h+ −
pi and hpi , and
are chosen as k1 = 2, k2 = 1.5, and ε = 0.01. For the CLF-based QP 2 is used to produce the control input. The simulation results are
conditions (13) and (17), λ = 0.1. For the CBF-based conditions (14), shown in Fig. 3. It can be seen that safe consensus tracking is achieved
(15), and (18)–(20), ρ1 = ρ2 = ρ3 = ρ4 = 1. while the position, velocity, and input constraints are guaranteed. The
First, we consider the position constraint [p1 , . . . , p5 ] = agents cooperatively track the trajectory of the leader only when the
[3, 3, 3, 3, 3]. Note that under these position constraints, the robots leader’s trajectory lies in the position constraints of all the robots. On
cannot track the leader’s trajectory outside the region [−3, 3]. To the simulation machine with Intel Core i7-6700 [email protected] GHz and
show the effectiveness of the CLF-based consensus tracking method 8 GB RAM, the maximum solving time of the local QP problems
and the necessity of the CBF-based constraints, we first consider for each agent is under 0.05 s, which shows the practical application
the QP-based control input where only the CLF-based condition is potential of the proposed control strategy.
imposed, i.e., we solve the QP 1 with only the constraints (13) and (16) Note that in Fig. 3, there exists chattering phenomenon due to the
or the QP 2 with only the constraints (17) and (21). The simulation unknown leader’s information and the uncertain dynamics. When the
results are shown in Fig. 1. It can be seen that practical consensus leader’s information is available and the dynamics of the agents are
tracking is successfully achieved with the CLF-based controller. known, the simulation results are shown in Fig. 4, which shows that
However, the position constraints of the agents are violated during safe consensus tracking with continuous control inputs is achieved in
the tracking process. Then, we consider the case with the proposed this case. Next, we consider the position constraints [p1 , . . . , p5 ] =
CBF-based controller. For the position constraint, the high-order [5, 5, 5, 5, 5] and [p1 , . . . , p5 ] = [2, 3, 2, 4, 2], respectively. In the for-
CBF is first employed, i.e., QP 1 is used to obtain the control input. mer case, the desired trajectory lies in the position constraints of the

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 12, DECEMBER 2023 8081

robots, whereas in the latter case, not all the agents could track the [10] J. Zhan and X. Li, “Flocking of multi-agent systems via model predictive
desired trajectory without violating the position constraint. With the QP control based on position-only measurements,” IEEE Trans. Ind. Informat.,
2-based controllers, the position tracking results are shown in Figs. 5 vol. 9, no. 1, pp. 377–385, Feb. 2013.
[11] P. Lin, W. Ren, C. Yang, and W. Gui, “Distributed consensus of second-
and 6, respectively. It verifies that in both cases, safe consensus tracking order multiagent systems with nonconvex velocity and control input
is achieved while all the position, velocity, and input constraints are constraints,” IEEE Trans. Autom. Control, vol. 63, no. 4, pp. 1171–1176,
satisfied. Apr. 2018.
[12] J. Fu, G. Wen, W. Yu, T. Huang, and X. Yu, “Consensus of second-order
V. CONCLUSION multi-agent systems with both velocity and input constraints,” IEEE Trans.
Ind. Electron., vol. 66, no. 10, pp. 7946–7955, Oct. 2019.
In this work, we have considered the safe consensus tracking problem [13] J. Fu, Y. Lv, G. Wen, X. Yu, and T. Huang, “Velocity and input constrained
for second-order uncertain nonlinear multiagent systems subject to both coordination of second-order multi-agent systems with relative output
state and input constraints. The leader’s trajectory is not required to information,” IEEE Trans. Netw. Syst. Eng., vol. 7, no. 3, pp. 1925–1938,
Jul./Sep. 2020.
satisfy all the followers’ state and input constraints. By constructing [14] Y. Zhang, H. Liang, H. Ma, Q. Zhou, and Z. Yu, “Distributed adaptive
the CLF-based consensus tracking condition and the CBF-based state consensus tracking control for nonlinear multi-agent systems with state
and input constraints conditions, two QP-based controllers have been constraints,” Appl. Math. Comput., vol. 326, pp. 16–32, 2018.
proposed to achieve the safe consensus tracking objective. It was proven [15] F. Yuan, Y. Ma, Y. J. Liu, J. Lan, and T. Xu, “Adaptive distributed tracking
control for non-affine multi-agent systems with state constraints and dead-
that with one of the QP-based controller, consensus tracing with guar-
zone input,” J. Franklin Inst., vol. 359, no. 1, pp. 352–370, 2022.
anteed state and input constraints can be ensured. Simulation examples [16] T. Zhang, M. Xia, and Y. Yi, “Adaptive neural dynamic surface control of
verify the effectiveness of the proposed control strategy. Future work strict-feedback nonlinear systems with full state constraints and unmod-
includes considering collision avoidance with respect to obstacles in eled dynamics,” Automatica, vol. 81, pp. 232–239, 2017.
the environment and handling more general cooperative tasks. [17] Y. Liu, H. Zhang, J. Sun, and Y. Wang, “Adaptive fuzzy containment
control for multiagent systems with state constraints using unified trans-
formation functions,” IEEE Trans. Fuzzy Syst., vol. 30, no. 1, pp. 162–174,
REFERENCES Jan. 2022.
[18] F. Zhang, G. Tan, C. Yu, N. Ding, C. Song, and M. Liu, “Fair transmission
[1] J. Cortes and M. Egerstedt, “Coordinated control of multi-robot systems: rate adjustment in cooperative vehicle safety systems based on multi-agent
A survey,” SICE J. Control, Meas., Syst. Integration, vol. 10, no. 6, model predictive control,” IEEE Trans. Veh. Technol., vol. 66, no. 7,
pp. 495–503, 2017. pp. 6115–6129, Jul. 2017.
[2] Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress [19] H. Li and Y. Shi, “Robust distributed model predictive control of con-
in the study of distributed multi-agent coordination,” IEEE Trans. Ind. strained continuous-time nonlinear systems: A robustness constraint ap-
Informat., vol. 9, no. 1, pp. 427–438, Feb. 2013. proach,” IEEE Trans. Autom. Control, vol. 59, no. 6, pp. 1673–1678,
[3] J. Qin, Q. Ma, Y. Shi, and L. Wang, “Recent advances in consensus of Jun. 2014.
multi-agent systems: A brief survey,” IEEE Trans. Ind. Electron., vol. 64, [20] A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function
no. 6, pp. 4972–4983, Jun. 2017. based quadratic programs for safety critical systems,” IEEE Trans. Autom.
[4] K. K. Oh, M. C. Park, and H. S. Ahn, “A survey of multi-agent formation Control, vol. 62, no. 8, pp. 3861–3876, Aug. 2017.
control,” Automatica, vol. 53, pp. 424–440, 2015. [21] R. Grandia, A. J. Taylor, A. Singletary, M. Hutter, and A. D. Ames, “Non-
[5] Y. Li, J. Xiang, and W. Wei, “Consensus problems for linear time-invariant linear model predictive control of robotic systems with control Lyapunov
multi-agent systems with saturation constraints,” IET Control Theory functions,” 2020, arXiv:2006.01229.
Appl., vol. 5, no. 6, pp. 823–829, May 2011. [22] P. Glotfelter, I. Buckley, and M. Egerstedt, “Hybrid nonsmooth barrier
[6] W. Ren, “On consensus algorithms for double-integrator dynamics,” IEEE functions with applications to provably safe and composable collision
Trans. Autom. Control, vol. 53, no. 6, pp. 1503–1509, Jul. 2008. avoidance for robotic systems,” IEEE Robot. Autom. Lett., vol. 4, no. 2,
[7] Z. Zhao and Z. Lin, “Global leader-following consensus of a group of pp. 1303–1310, Apr. 2019.
general linear systems using bounded controls,” Automatica, vol. 68, [23] M. M. Polycarpou, “Stable adaptive neural control scheme for nonlin-
pp. 294–304, 2016. ear systems,” IEEE Trans. Autom. Control, vol. 41, no. 3, pp. 447–451,
[8] H. Su, M. Z. Q. Chen, J. Lam, and Z. Lin, “Semi-global leader-following Mar. 1996.
consensus of linear multi-agent systems with input saturation via low [24] Q. Nguyen and K. Sreenath, “Exponential control barrier functions for
gain feedback,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 60, no. 7, enforcing high relative-degree safety-critical constraints,” in Proc. Amer.
pp. 1881–1889, Jul. 2013. Control Conf., 2016, pp. 322–328.
[9] G. Ferrari-Trecate, L. Galbusera, M. P. E. Marciandi, and R. Scattolini, [25] B. Morris, M. J. Powell, and A. D. Ames, “Sufficient conditions for the
“Model predictive control schemes for consensus in multi-agent systems Lipschitz continuity of QP-based multi-objective control of humanoid
with single- and double-integrator dynamics,” IEEE Trans. Autom. Con- robots,” in Proc. IEEE 52nd Conf. Decis. Control, 2013, pp. 2920–2926.
trol, vol. 54, no. 11, pp. 2560–2572, Nov. 2009.

Authorized licensed use limited to: Ecole Centrale de Nantes. Downloaded on April 30,2024 at 09:35:48 UTC from IEEE Xplore. Restrictions apply.