Available online at www.sciencedirect.

com

Journal of the Franklin Institute 354 (2017) 2068–2085

www.elsevier.com/locate/jfranklin

Distributed MPC for formation of multi-agent systems

with collision avoidance and obstacle avoidance
Li Dai, Qun Cao, Yuanqing Xian, Yulong Gao
School of Automation, Beijing Institute of Technology, Beijing 100081, PR China
Received 9 February 2016; received in revised form 26 November 2016; accepted 25 December 2016
Available online 31 December 2016

Abstract

The paper is concerned with the problem of distributed model predictive control (DMPC) for formation
of multiple linear second-order agents with collision avoidance and obstacle avoidance. All the agents are
permitted to implement optimization simultaneously at each time step. The assumed input trajectory and
state trajectory are introduced to obtain a computationally tractable optimization problem in a distributed
manner. As a result, a compatibility constraint is required to ensure the consistency between each agent's
real operation and its plan and to establish the agreement among agents. The terminal ingredients are
tailored by making use of the specific form of the system model and the control objective. The terminal set
is ensured to be positively invariant with the designed terminal controller. The collision avoidance
constraint and the obstacle avoidance constraint are satisfied for any state in the terminal set. The weighted
matrix of the terminal cost is determined by solving a Lyapunov equation. Moreover, recursive feasibility of
the resulting optimization problem is guaranteed and closed-loop stability of the whole system is ensured.
Finally, a numerical example is given to illustrate the effectiveness of the proposed algorithm.
& 2017 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.

1. Introduction

The problem of formation control of multi-agent systems is an important and well-studied

problem [1–5]. In this problem, the system is always subject to physical constraints (e.g.
saturation of input) and task constraints (e.g. collision avoidance) [6]. These constraints have
n
Corresponding author.
E-mail addresses: [email protected] (L. Dai), [email protected] (Q. Cao),
[email protected] (Y. Xia), [email protected] (Y. Gao).

http://dx.doi.org/10.1016/j.jfranklin.2016.12.021
0016-0032/& 2017 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2069

been addressed by many strategies, in which model predictive control (MPC) stands out due to
its capability of handling constraints and good control performance [7,8]. For multi-agent
systems with constraints, distributed MPC (DMPC) has attracted much attention in terms of the
structural flexibility, less computation cost and lower communication burden despite its
suboptimal performance compared with centralized MPC [9,10]. However, some challenges in
DMPC still remain to be solved for formation of multi-agent systems with collision avoidance
and obstacle avoidance. For example, how to reformulate the collision avoidance constraints for
each agent involving only the local information and further construct a computationally tractable
MPC optimization problem in a distributed manner; how to design a terminal controller, a
terminal set as well as a terminal cost for each agent; how to establish recursive feasibility of the
MPC optimization problem and closed-loop stability of the whole system.
Up to now, some works on DMPC are available for the formation problem of multi-agent
systems [11–16]. The result in [11] develops a new form of DMPC algorithm for multi-vehicle
formation stabilization. By adopting a synchronous update strategy and augmenting each
distributed optimal control problem with a compatibility constraint, the convergence to a
neighborhood of the desired equilibrium point is guaranteed for each vehicle. However, this
work does not consider the issue of coupled constraint handling, such as the collision avoidance
constraints. Later contributions [12,13] take into account both formation control problem and
collision avoidance. In [12], all the agents solve their corresponding problems independently and
synchronously. Although the complexity of the problem is reduced by division, the satisfaction
of coupled constraint cannot be guaranteed. In [13], a DMPC approach is presented to achieve
organic air vehicle (OAV) formation flight and avoid collisions. Each vehicle uses its neighbor
information to plan conflict-free trajectories that maintain coordination and realize team
objectives, yet the analysis of recursive feasibility and stability in closed-loop operation is not
provided. A DMPC approach is proposed in [14] to achieve the cooperation task for unmanned
aerial vehicles (UAVs) with coupled constraints. By solving neighbors’ problems iteratively
within one sampling step for each UAV, it is shown to ensure recursive feasibility and achieve
system-wide cooperation with guaranteed convergence. This iterative approach usually requires
more time for computation and higher capacity for communication. Following the synchronous
update strategy adopted in [11], the authors of [15,16] investigate DMPC algorithms with less
computation cost and communication for multi-agent systems. In [15], the collision avoidance
constraint as well as the communication distance constraint are addressed whereas recursive
feasibility and closed-loop stability cannot be ensured. The improvement of [15] appears in [16]
with guaranteed feasibility and stability, in which a reference trajectory is utilized to design the
terminal ingredients for each agent.
Motivated by the above discussion, we aim to develop a DMPC algorithm for formation of
multiple linear second-order agent systems subject to both collision avoidance constraints and
obstacle avoidance constraints. By applying the synchronous update strategy as in [11,16], the
assumed trajectories are introduced to formulate a tractable MPC optimization problem in the
distributed setting. Different from [16], the control objective of this paper is to regulate all the
agents to their respective desired positions with the desired formation while avoiding collisions
and obstacles, thereby resulting in no reference trajectory available for all the agents. The main
contributions of this paper are summarized as follows.

(a) Compared with [11], a relatively non-conservative compatibility constraint is designed to

cater for the stability analysis and the satisfaction of collision avoidance constraints.
2070 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

(b) With the tailored terminal controller and terminal cost, the terminal set is simplified to be a
positively invariant set. Particularly, collision avoidance and obstacle avoidance are
guaranteed in this set. Note that the difference in the design of the terminal ingredients
between our paper and [16] derives from the difference in control objectives.
(c) By implementing the proposed DMPC algorithm, both recursive feasibility of the
optimization problem and closed-loop stability of the whole system are guaranteed.

The remainder of the paper is organized as follows. Section 2 presents the problem
formulation. Section 3 develops a DMPC algorithm including the design of compatibility
constraints and the construction of terminal ingredients, and then presents the main results of this
work. One numerical example is given to demonstrate the effectiveness of the algorithm in
Section 4, while the conclusions are drawn in Section 5.
Notation. For a vector x and a positive definite matrix W, ‖x‖ and ‖x‖W are the 2-norm and the
weighted 2-norm of x, respectively. For a matrix A A Rn
n , A40 and Ao0 mean that the matrix
A is positive definite or negative definite, respectively. For a set A, jAj denotes the number of
elements in A. 0 is the zero vector or zero matrix and I is the identity matrix with appropriate
dimensions.

2. Problem formulation

Consider a system consisting of Na discrete-time linear second-order agents

zi ðk þ 1Þ ¼ Azi ðkÞ þ Bui ðkÞ; 8 iA f1; 2; …; N a g; ð1Þ

T  T
where zi ðkÞ ¼ xi ðkÞ vTi ðkÞ A R2n and ui ðkÞ A Rn are the ith agent's state and input, respectively;
h i
xi ðkÞA Rn and vi ðkÞ A Rn denote, respectively, the position and the velocity of agent i; A ¼ I0n II nn
h i
and B ¼ 0:5I I n . Let Na 9 f1; 2; …; N a g be the index set of all the agents. The state and input of
n

agent i are subject to the constraints

z i A Zi ; ui A Ui ;
where Zi is a compact set and Ui is a closed set.
The control objective is to regulate all the agents to their own desired positions with a certain
formation while avoiding collisions and obstacles, which can be formulated by, for agent i,
 
lim xdi  xi ðkÞ ¼ 0; ð2Þ
k-1

lim ðxj ðkÞ xi ðkÞÞ ¼ dij ; 8 jA N i ; ð3Þ

k-1
 
xj ðkÞ xi ðkÞ Z 2R; 8 jA Na ⧹fig; k Z 0; ð4Þ
 
 o 
xp  xi ðkÞZ 2R; 8 p A No ; k Z 0; ð5Þ

where xdi denotes the desired position of agent i; dij denotes the desired relative position between
agent i and agent j; N i 9 f jA Na ∣dij is an a priori information for agent ig denotes the set of
neighbors of agent i; R is the safety radius of each agent; xop is the position of obstacle p;
No 9 f1; 2; …; Mg and M is the number of obstacles. To ensure the consistency of this control
objective, we will proceed our consideration under the following assumptions.
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2071

Assumption 1. The desired relative positions between agents are consistent with the desired
position of each agent and the collision avoidance constraint, i.e.
dij ¼ xdj  xdi ; ‖d ij ‖ Z 2R; iA Na ; jA N i : ð6Þ
In addition, the desired position of each agent is consistent with the obstacle avoidance
constraint, i.e.
 
 d o 
xi  xp  Z 2R; i ANa ; p A No : ð7Þ

Assume that the set Zi contains ½xdi 0T and the set Ui contains the origin. According to the control
objective, given the weight parameters αi 40, βi 40, and ρi 40, the ideal cost function of agent i
is defined as
  NX
1    
J i k; zi ; z  i ; zdi ; ui ¼ Li k þ ljk; zi ; z  i ; zdi ; ui þ Lif zi ðk þ NjkÞ; zdi ð8Þ
l¼0

with the stage cost

 
Li k þ ljk; zi ; z  i ; zdi ; ui
 2 X  2
  
¼ αi zdi  zi ðk þ ljkÞ þ βi zj ðk þ ljkÞ zi ðk þ ljkÞ d zij  þ ρi ‖ui ðk þ ljkÞ‖2
Xj A N i
¼ αi ‖Δzi ðk þ ljkÞ‖2 þ βi ‖zij ðk þ ljkÞ‖2 þ ρi ‖ui ðk þ ljkÞ‖2 ð9Þ
jANi

  h T iT h iT
and the terminal cost Lif zi ðk þ NjkÞ; zdi , where zdi ¼ xdi 0 A R2n , dzij ¼ dTij 0 A R2n ,
Δzi ðk þ ljkÞ ¼ zdi  zi ðk þ ljkÞ, and zij ðk þ ljkÞ ¼ zj ðk þ ljkÞ zi ðk þ ljkÞ dzij . zi ¼
h iT
zTj1 zTj2 ⋯ zTjjN j , jk A N i , k ¼ 1; 2; …; jN i j, denotes a vector consisting of the states of
i
neighbors of agent i and zi ðk þ ljkÞ denotes the value of z at time step k þ l predicted at time
step k.
Below the MPC optimization problem of agent i is formulated.
Problem 1. At time step k, given zi(k) and zdi, the following MPC optimization problem is
solved.
 
min J i k; zi ; z  i ; zdi ; ui
ui ðkþljkÞ;
l ¼ 0;1;…;N  1

subject to for l ¼ 0; 1; …; N  1
zi ðkjkÞ ¼ zi ðkÞ; ð10Þ

zi ðk þ l þ 1jkÞ ¼ Azi ðk þ ljkÞ þ Bui ðk þ ljkÞ; ð11Þ

zi ðk þ ljkÞA Zi ; ð12Þ

ui ðk þ ljkÞA Ui ; ð13Þ
 
xj ðk þ ljkÞ  xi ðk þ ljkÞ Z 2R; jA Na ⧹fig; ð14Þ
2072 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

Table 1
Variables of agent i.

ui ðk þ ljkÞ Optimal control input

zi ðk þ ljkÞ Optimal state
u~ i ðk þ ljkÞ Feasible control input
z~ i ðk þ ljkÞ Feasible state
u^ i ðk þ ljkÞ Assumed control input
z^ i ðk þ ljkÞ Assumed state
x^ i ðk þ ljkÞ Assumed position

 
 o 
xp  xi ðk þ ljkÞ Z 2R; p A No ; ð15Þ

zi ðk þ NjkÞA Zif ; ð16Þ

where N is the horizon length, Eq. (10) is the initial state constraint, Eq. (11) is the dynamics
equation, Eqs. (12)–(13) are the state and input constraints, Eqs. (14)–(15) are the collision
avoidance and obstacle avoidance constraints, and Eq. (16) is the terminal constraint which
requires that the state of agent i should lie in Zif at the end of the predictive horizon.

Since both the cost function in Eq. (8) and the collision avoidance constraint of agent i in Eq.
(14) involve the states of its neighbors over the whole prediction horizon, Problem 1 cannot be
solved directly online in a distributed way. To handle it, the synchronous update strategy is
adopted to present a tractable formulation for Problem 1 in Section 3.

3. DMPC algorithm

In the synchronous update scheme, all the agents are permitted to optimize simultaneously at
each time step. An assumed state is introduced to be transmitted among the agents instead of the
real one. Before giving the definitions of the assumed control input and the assumed state, some
notations are summarized in Table 1. The assumed control input of agent i is defined as
( 
ui ðk þ ljk  1Þ; l ¼ 0; 1; …; N  2;
u^ i ðk þ ljkÞ ¼ ð17Þ
κi ðzi ðk  1 þ Njk  1ÞÞ; l ¼ N  1;

where κi ðÞ is the terminal controller to be designed. Correspondingly, the assumed state of agent
i is defined as
( 
zi ðk þ ljk  1Þ; l ¼ 0; 1; …; N  1;
z^ i ðk þ ljkÞ ¼ ð18Þ
zκi ðk þ Njk  1Þ; l ¼ N;

where zκi ðk þ Njk  1Þ is the next state of zi ðk  1 þ Njk  1Þ by using the control input
κi ðzi ðk  1 þ Njk  1ÞÞ. In order to be consistent with z^ i , define z^  i as a vector of the assumed
states of neighbors of agent i.
Note that the introduced assumed states make it possible to optimize in the distributed setting,
despite the deviation between the real predictive trajectory and the assumed one. To keep each
agent operate in accordance with its plan, an additional constraint, namely compatibility
constraint, is imposed to restrict the deviation. To some extent, the compatibility constraint also
represents the agreement among the behaviors of agents.
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2073

3.1. Design of compatibility constraints

The compatibility constraint of each agent is made up of two components, which respectively
imply the position compatibility and the state compatibility. The functions of these two
components are different: one is to realize collision avoidance and the other is to ensure the
stability of the overall system.  
x^ j ðkþljkÞ  x^ i ðkþljkÞ  2R
For agent i, let μij ðk þ ljkÞ ¼ 2 , l ¼ 0; 1; …; N  1, j A Na ⧹fig, and
μi ðk þ ljkÞ ¼ minj A Na ⧹fig μij . The constraint for position compatibility is presented by
 
x^ i ðk þ ljkÞ xi ðk þ ljkÞ ¼ εx ðk þ ljkÞr μi ðk þ ljkÞ; l ¼ 0; 1; …; N  1: ð19Þ
i
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P
Define νi ðkÞ ¼ b2i ðkÞ 4ai ci ðkÞ  bi ðkÞ =2ai , where ai ¼ j A N i ð βi þ 2βj Þ, bi ðkÞ ¼
P   
j A N 2β φ ðkÞ, ci ðkÞ ¼  N γi 1 αi Δzi ðkjkÞ, φji ðkÞ ¼ maxl ¼ 1;2;…;N  1 z^ j ðk þ ljkÞ z^ i ðk þ ljkÞ
 i j ji
 dij , and 0oγ i o1. The constraint for state compatibility is then given as
 
z^ i ðk þ ljkÞ zi ðk þ ljkÞ ¼ εz ðk þ ljkÞr νi ðkÞ; l ¼ 0; 1; …; N  1: ð20Þ
i

Remark 1. Some results are available in the literature on the design of compatibility constraints,
which play an important role to guarantee the closed-loop stability. For the case of continuous-
time systems [17,18], the compatibility constraint is mainly determined by the biases between the
assumed trajectories of different agents, rather than the initial states of the optimizing agents.
However, it is the case that the compatibility constraints of discrete-time systems rely on the
initial states [16].

3.2. Design of collision avoidance constraints

Simply replacing xj with x^ j in Eq. (14) cannot guarantee the satisfaction of the collision
avoidance constraints in real case. Hence, we need to modify Eq. (14) to the form of
   
xi ðk þ ljkÞ  xj ðk þ ljkÞ ¼ xi ðk þ ljkÞ x^ j ðk þ ljkÞ þ x^ j ðk þ ljkÞ xj ðk þ ljkÞ
 
Z xi ðk þ ljkÞ x^ j ðk þ ljkÞ  εxj ðk þ ljkÞ
 
Z xi ðk þ ljkÞ x^ j ðk þ ljkÞ  μij ðk þ ljkÞ: ð21Þ
The above result follows directly from Eq. (19) and the sufficient condition for the satisfaction of
Eq. (14) is then given as
 
xi ðk þ ljkÞ  x^ j ðk þ ljkÞ Z 2R þ μij ðk þ ljkÞ; jA Na ⧹fig: ð22Þ

3.3. Construction of terminal ingredients

The terminal ingredients (including terminal controller, terminal cost, and terminal set) are of
great significance to ensure recursive feasibility and closed-loop stability. The aim of this
subsection is to construct the explicit terminal ingredients for each agent.
First, the terminal controller is designed in the form of
κi ðzi ðkÞÞ ¼ K i1 zi ðkÞ þ K i2 zdi : ð23Þ
2074 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

Substituting it into Eq. (1) and using the definition of Δzi , one obtains the error system

Δzi ðk þ 1Þ ¼ AΔzi ðkÞ þ ðI 2n  AÞzdi  BK i1 zi ðkÞ BK i2 zdi
¼ ðA þ BK i1 ÞΔzi ðkÞ þ ðI 2n  A  BK i1  BK i2 Þzdi :
It can be further rewritten as
Δzi ðk þ 1Þ ¼ ϕi Δzi ðkÞ; ð24Þ
if the following two conditions are satisfied

(i) ϕi ¼ A þ BK i1 is stabilized;
(ii) ðI 2n  A  BK i1  BK i2 Þzdi ¼ 0.

Next let us evaluate the possibility of conditions (i) and (ii). Since (A,B) is controllable, Ki1 can
be easily chosen to satisfy condition (i). For simplicity, ignore the subscript and define
L ¼ K i1 þ K i2 ¼ ½L1 L2 , where L1 ; L2 A Rn
n . Taking into consideration of the specific forms of
A and B, we have
0  In
I 2n  A ¼ ; ð25Þ
0 0
" #
0:5L1 0:5L2
BL ¼ : ð26Þ
L1 L2
T
Further considering the definition of zdi ¼ ½xdi 0T , then condition (ii) holds if L1 ¼ 0, which
means that the first n columns of Ki2 can be taken as the opposite numbers of that of Ki1 and the
other columns are unrestricted.
The terminal cost is defined as
 
Lif zi ; zdi ¼ ‖zdi  zi ‖2Pi ; ð27Þ
and the corresponding terminal set is then given by
n    2 o
Zif ¼ zi ∣Δzi Λi1 þρi zdi Λi2 r 0; ‖Δxi ‖ r Di ;
2
ð28Þ
P
where Λi1 ¼ αi I 2n þ ρi K Ti1 K i1 þ j A N i ð βi þ βj ÞI 2n þ ϕTi Pi ϕi  Pi ,
Λi2 ¼ ðK i1 þ K i2 ÞT ðK i1 þ
  
K i2 Þ, Di ¼ minfDi1 ; Di2 g, Di1 ¼ minj A Na ⧹fig ð‖d ij ‖  2RÞ=2, Di2 ¼ minp A No xop  xdi   2R ,
and Δxi ¼ xdi  xi . Provided that L1 ¼ 0, it easily follows that ρi ‖zdi ‖2Λi2 ¼ 0. Thus the terminal
set (28) can be rewritten as
n     o
Zif ¼ zi ∣Δzi Λi1 r 0; Δxi  r Di :
2
ð29Þ

Before presenting Theorem 1, a standard assumption in MPC approach is given.

Assumption 2. Assume that the terminal set (29) is such that Zif DZi and the terminal controller
(23) is such that for all zi A Zif , κi ðzi ÞA Ui .

Theorem 1. For any agent i A Na , choose Ki1 and Ki2 such that ϕi ¼ A þ BK i1 is stabilized and
L1 ¼ 0. For a given negative definite matrix Qi, if Pi is a positive definite matrix satisfying the
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2075

following Lyapunov equation

X
αi I þ ρi K Ti1 K i1 þ 2ð βi þ βj ÞI þ ϕTi Pi ϕi  Pi ¼ Qi ; ð30Þ
jANi

then the terminal set (29) can be simplified as

Zif ¼ fzi ∣kΔxi k r Di g: ð31Þ
In addition, Zif is a positively invariant set with respect to Eq. (24) and all states in this set
definitely satisfy the collision avoidance constraint (14) and the obstacle avoidance constraint
(15). Furthermore, the terminal cost (27) is a local Lyapunov function on Zif for the system (1)
controlled by Eq. (23), i.e.
X     X  
Lif zi ðk þ 1Þ; zdi  Lif zi ðkÞ; zdi r  Li k; zi ; z  i ; zdi ; κi ðzi Þ ; 8 zi A Zif : ð32Þ
i A Na i A Na

Proof. If Pi is a positive definite solution to the Lyapunov equation (30), it follows immediately
that Λi1 o0. Hence, the first condition in the terminal set (29) can be removed without any effect
and the simplified form of Eq. (29) can be taken as Eq. (31). Due to the contractility of the error
system (24) under the terminal controller (23), the norm-bound on the error zdi  zi is
monotonously non-increasing, which means that the terminal set (31) is positively invariant for
the error system (24).
Moreover, it yields that for zi A Zif and zj A Zjf , ia j,
  
 
‖xi  xj ‖ ¼ xi  xdi  xj  xdj  dij 
Z ‖dij ‖  ‖Δxj ‖  ‖Δxj ‖ Z 2R;
and for zi A Zif and p A No
   
 o   
xp  xi  ¼ xop  xdi  ðxi  xdi Þ
 
 
Z xop  xdi  ‖Δxi ‖ Z 2R:

As a consequence, the collision avoidance constraint (14) and the obstacle avoidance constraint
(15) hold for all the states in the terminal set (31).
Finally, we will show that the terminal cost (27) is a local Lyapunov function in Zif . Rewriting
the terminal controller as κi ðzi ðkÞÞ ¼  K i1 Δzi ðkÞ þ ðK i1 þ K i2 Þzdi and using the triangle
inequality, we obtain that for any zi ðkÞA Zif
X      
Li k; zi ; z  i ; zdi ; κ i ðzi Þ þ Lif zi ðk þ 1Þ; zdi  Lif zi ðkÞ; zdi
i A Na
X X
¼ αi ‖Δzi ðkÞ‖2 þ ρi ‖κ i ðzi ðkÞÞ‖2 þ βi ‖zij ðkÞ‖2
i A Na jANi

þ ‖Δzi ðk þ 1Þ‖2Pi  ‖Δzi ðkÞ‖2Pi

X
r αi ‖Δzi ðkÞ‖2 þ ρi ‖Δzi ðkÞ‖2K T K i1 þ ρi ‖zdi ‖2Λi2
i1
i A Na
2076 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

X
þ 2ðβi þ βj Þ‖Δzi ðkÞ‖2 þ ‖Δzi ðkÞ‖2ϕT Pi ϕ  ‖Δzi ðkÞ‖2Pi
i i
jANi
X 
¼ ‖Δzi ðkÞ‖2Λi1 þ ρi ‖zdi ‖2Λi2 r 0;
i A Na

which completes the proof. □

Remark 2. The evolution of the terminal set Zif from Eq. (28) to Eq. (31) is determined by the
specific forms of the system model and the control objective. Additionally, Zif in Eq. (31) is only
related to the position of each agent, rather than the velocity. The main function of the terminal
set is to ensure the satisfaction of collision avoidance constraints and obstacle avoidance
constraints.

Remark 3. The size of the terminal set Zif depends on the value of Di. If the desired relative
distance between two agents is very close to 2R, Di will become very small, thereby leading to a
small margin for zi ðk þ NjkÞ to satisfy the terminal constraint (31). Such an extreme case may
therefore incur a degree of conservativeness of this result.

3.4. Distributed implementation algorithm

 
For Problem 1, replace the real predictive states of the neighbors zj j A N i with their assumed
 
ones z^ j j A N i , modify the collision avoidance constraint (14) to (22), and add the compatibility
constraint (19) and (20). Then a computationally tractable optimization problem for agent i is
formulated as follows.
Problem 2. At time step k, given zi(k), zdi, and z^  i ðk þ ljkÞ, l ¼ 0; 1; …; N  1, the following
MPC optimization problem is solved.
   
J i k; zi ; z^  i ; zdi ; ui ¼ min J i k; zi ; z^  i ; zdi ; ui
ui ðkþljkÞ;
l ¼ 0;1;…;N  1

subject to for l ¼ 0; 1; …; N  1
zi ðkjkÞ ¼ zi ðkÞ; ð33Þ

zi ðk þ l þ 1jkÞ ¼ Azi ðk þ ljkÞ þ Bui ðk þ ljkÞ; ð34Þ

zi ðk þ ljkÞA Zi ; ð35Þ

ui ðk þ ljkÞA Ui ; ð36Þ
 
x^ j ðk þ ljkÞ xi ðk þ ljkÞ Z 2R þ μij ðk þ ljkÞ; j A Na ⧹fig; ð37Þ
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2077

 
 o 
xp  xi ðk þ ljkÞ Z 2R; p A No ; ð38Þ
(
‖^x i ðk þ ljkÞ xi ðk þ ljkÞ‖ r μi ðk þ ljkÞ;
ð39Þ
‖^z i ðk þ ljkÞ zi ðk þ ljkÞ‖ r νi ðkÞ;

zi ðk þ NjkÞ A Zif ; ð40Þ

where Zif is given in Eq. (31). Note that, in Eq. (37), the terms involving agent jA Na ⧹fig are not
affected by the optimization variable of agent i and appear as fixed values, Problem 2 can
therefore be solved in the distributed way.

Below we propose a DMPC algorithm, which is executed by all the agents synchronously.
Algorithm 1.
Offline Stage : For each agent i, given parameters αi, βi, ρi, γi, dij, zdi , and Qi, determine Ki1, Ki2,
Pi, and Di.
Online Stage: At the initial time step k¼ 0, wait for the initial optimal control inputs ui ðlj0Þ,
l ¼ 0; 1; …; N  1, and the initial optimal states zi ðlj0Þ, l ¼ 0; 1; …; N, for each agent iA Na .
(a) At time step k, k Z 0, for agent i,
 apply ui ðkÞ ¼ ui ð0jkÞ;
 compute the assumed control inputs u^ i ðk þ 1 þ ljk þ 1Þ and the assumed states
z^ i ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N  1, using (17)–(18).
 transmit z^ i ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N  1; to other agents and receive
z^ j ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N  1; from agent jA Na ⧹fig.
(b) At the next time step k þ 1, k Z0, for agent i,
 sample the current state zi ðk þ 1Þ;
 compute μij ðk þ 1 þ ljk þ 1Þ, μi ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N  1, j A Na ⧹fig, and
νi ðk þ 1Þ;
 solve Problem 2 to obtain the optimal control inputs ui ðk þ 1 þ ljk þ 1Þ,
l ¼ 0; 1; …; N  1, and the optimal states zi ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N.
(c) Return to (a).

Remark 4. In addition to the synchronous update, sequential update and iterative update are two
other popular strategies adopted in the DMPC literature. In comparison, the synchronous DMPC
outperforms the sequential DMPC in terms of the optimality of the solution and the scalability of the
network, and outperforms the iterative DMPC in terms of the computation and communication cost.

3.5. Main results

Theorem 2 states that Algorithm 1 guarantees recursive feasibility of the MPC optimization
problem and ensures stability in closed-loop operation.
Theorem 2. At each time step k, all the agents in Eq. (1) are permitted to optimize according to
Algorithm 1. If there exists one feasible solution for each agent at the initial time step k¼ 0, then
all subsequent optimization problems are feasible. Furthermore, all the agents asymptotically
2078 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

converge to their desired positions with the desired formation while satisfying the collision
avoidance constraint (4) and the obstacle avoidance constraint
  i.e., 8iAoNa , as k-1,
(5), 
xi ðkÞ-xi , xj ðkÞ xi ðkÞ-dij , 8 jA N i , and for k Z 0, xj ðkÞ xi ðkÞ Z 2R, xp  xi ðkÞ Z 2R,
d 
8 jA Na ⧹fig, 8 pA No .

Proof. Feasibility: Assume that the optimization problem of agent i is feasible at time step k.
Denote by ui ðk þ ljkÞ, l ¼ 0; 1; …; N  1, the optimal control inputs and by zi ðk þ ljkÞ,
l ¼ 0; 1; …; N, the corresponding optimal states. Then the assumed control inputs u^ i ðk þ 1 þ
ljk þ 1Þ and the assumed states z^ i ðk þ 1 þ ljk þ 1Þ, l ¼ 0; 1; …; N  1, can be defined according
to Eqs. (17)–(18). Let u~ i ðk þ 1 þ ljk þ 1Þ ¼ u^ i ðk þ 1 þ ljk þ 1Þ and z~ i ðk þ 1 þ ljk þ 1Þ
¼ z^ i ðk þ 1 þ ljk þ 1Þ. Next we will show that the input sequence fu~ i ðk þ 1 þljk þ
1Þ; l ¼ 0; 1; …; N  1g is one feasible solution to agent i at time step k þ 1. Applying u~ i ðk þ
1 þ ljk þ 1Þ to agent i, the satisfaction of constraints (33)–(39) follows easily from the definition
of the assumed control inputs and Assumption 2. The positive invariance of the terminal set Zif
further results in the satisfaction of Eq. (40). Therefore, the feasibility of Problem 2 at time step k
implies the feasibility at time step k þ 1. By recursion, Problem 2 is feasible at all future time
instants.
Stability: Consider the summation of optimal costs of all the agents, indicated by
X  
J P ðkÞ ¼ J  k; z ; z^  i ; zd ; u ;
i i i i
i A Na

as a Lyapunov function of the whole system. Then, consider the summation of the costs at time
step k þ 1 associated with the feasible solutions u~ i ðk þ 1 þ ljk þ 1Þ ¼ u^ i ðk þ 1 þ ljk þ 1Þ for all
the agents, indicated by
X  
J~ P ðk þ 1Þ ¼ J i k þ 1; z~ i ; z^  i ; zd ; u~ i :
i
i A Na

The difference between the feasible cost and the optimal cost at two successive time steps is
J~ P ðk þ 1Þ  J P ðkÞ
X    
¼ J i k þ 1; z~ i ; z^  i ; zdi ; u~ i  J i k; zi ; z^  i ; zdi ; ui
i A Na
(
X X
N 1   
¼ Li k þ 1 þ ljk þ 1; z~ i ; z^  i ; zdi ; u~ i
iAN l¼0
a
    
 Li k þ ljk; zi ; z^  i ; zdi ; ui  Lif zi ðk þ Njk ; zdi
)
 
þLif z~ i ðk þ 1 þ Njk þ 1Þ; zi d

(
X X
N 1   
¼ Li k þ ljk; zi ; z i ; zdi ; ui
iAN l¼1
a
   
 Li k þ ljk; zi ; z^  i ; zdi ; ui  Li kjk; zi ; z^  i ; zdi ; ui
    
þ Li k þ Njk; zi ; z i ; zdi ; κi zi  Lif zi ðk þ NjkÞ; zdi
)
 
þ Lif zκi ðk þ 1 þ Njk Þ; zdi : ð41Þ
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2079

By using Eq. (32) in Theorem 1 and the fact that zi ðk þ NjkÞA Zif , we have
    
Li k þ Njk; zi ; z i ; zdi ; κ i zi  Lif zi ðk þ NjkÞ; zdi
 
þ Lif zκi ðk þ 1 þ NjkÞ; zdi r 0: ð42Þ
Furthermore, from the definition of the stage cost in Eq. (9), it follows that
   
Li k þ ljk; zi ; z i ; zdi ; ui  Li k þ ljk; zi ; z^  i ; zdi ; ui
X 
¼ βi ‖zij ðk þ ljkÞ‖2  ‖^z ij ðk þ ljkÞ‖2 ; ð43Þ
jANi

where zij ðk þ ljkÞ ¼ zj ðk þ ljkÞ  zi ðk þ ljkÞ d zij and z^ ij ðk þ ljkÞ ¼ z^ j ðk þ ljkÞ zi ðk þ ljkÞ d zij .
Also, we have
 
Li kjk; zi ; z^  i ; zdi ; ui Z αi ‖Δzi ðkjkÞ‖2 : ð44Þ
Then, substituting Eqs. (42)–(44) into Eq. (41) yields
8
X <NX 1 X 
J~ P ðk þ 1Þ J P ðkÞr βi ‖zij ðk þ ljkÞ‖2  ‖^z ij ðk þ ljkÞ‖2
:
i A Na l ¼ 1 j A N i
)
 αi ‖Δzi ðkjkÞ‖2 : ð45Þ

According to the triangle inequality and the compatibility constraint (20), we have
X NX 1 X  2  2 
βi zij ðk þ ljkÞ  z^ ij ðk þ ljkÞ
i A Na l ¼ 1 j A N i
X NX
1 X   
r βi 2εzj ðk þ ljkÞz^ ij ðk þ ljkÞ
i A Na l ¼ 1 j A N i
 2 
þ εzj k þ ljkÞ
X NX  1 X h 2
r βi εzj ðk þ ljkÞ
i A Na l ¼ 1 j A N i
   i
þ 2εzj ðk þ ljkÞ εzi ðk þ ljkÞ þ φij ðkÞ
X NX
1 X  
r βi 2φij ðkÞεzj ðk þ ljkÞ
i A Na l ¼ 1 j A N i
 2  2 
þ 2εzj ðk þ ljk Þ þ εzi ðk þ ljkÞ
X NX
1 X
 
¼ 2βj φji ðkÞεzi ðk þ ljkÞ
i A Na l ¼ 1 j A N i
   
þ βi þ 2βj εzi ðk þ ljkÞ2
2 3
X X  2 X
r ðN  1Þ4 βi þ 2βj νi ðkÞ þ 2 βj φji ðkÞνi ðkÞ5; ð46Þ
i A Na jANi jANi
2080 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

where εzj ðk þ ljkÞ ¼ ‖zj ðk þ ljkÞ z^ j ðk þ ljkÞ‖. From the definition of νi ðkÞ, it can be concluded
that
X NX 1 X  2  2  X
βi zij ðk þ ljkÞ  z^ ij ðk þ ljkÞ r γ i αi ‖Δzi ðkjkÞ‖2 : ð47Þ
i A Na l ¼ 1 j A N i i A Na

Substituting Eq. (47) into Eq. (45) and considering the optimality of the solution, the following
is derived
X
J P ðk þ 1Þ J P ðkÞ r J~ P ðk þ 1Þ J P ðkÞr  ð1 γ Þαi ‖Δzi ðkjkÞ‖2 ;
i
i A Na

which, as 0oγ i o1 and αi 40, implies that all agents asymptotically converge to their desired
positions with the desired formation. Moreover, collision avoidance and obstacle avoidance can
be achieved by Theorem 1 and (37)–(38). □

4. Simulation results and analysis

Consider a multi-agent system comprising six linear second-order agents. Each agent is
subject to the input constraint ui A Ui ¼ fui ∣ 0:5 r ui;1 ; ui;2 r 0:5g, where ui;j denotes the jth
component of ui. The adjacent matrix E ¼ ½eij , i; j ¼ 1; 2; …; 6, is given by
2 3
0 1 0 0 0 1
61 0 1 0 0 07
6 7
6 7
60 1 0 1 0 07
E¼6 6 0 0 1 0 1 0 7;
7
6 7
6 7
40 0 0 1 0 15
1 0 0 0 1 0
where eij ¼ 1 represents that agent j is the neighbor of agent i. The prediction horizon is chosen
to be N ¼ 10 with the weights αi ¼ 1; βi ¼ 0:1; ρi ¼ 0:1. Choose the parameters offline as
γ i ¼ 0:9; R ¼ 0:25.
The obstacle is located at xo1 ¼ ½ 2  2T . The desired position of each agent is set to
xd1 ¼ ½0 0T ; xd2 ¼ ½1 0T ; xd3 ¼ ½2 1T ;
xd4 ¼ ½1 2T ; xd5 ¼ ½0 2T ; xd6 ¼ ½  1 1T :
The desired relative positions between neighbors are
d 12 ¼ ½1 0T ; d16 ¼ ½  1 1T ;
d 21 ¼ ½  1 0T ; d23 ¼ ½1 1T ;
d 32 ¼ ½  1  1T ; d34 ¼ ½  1 1T ;
d 43 ¼ ½1  1T ; d45 ¼ ½  1 0T ;
d 54 ¼ ½1 0T ; d56 ¼ ½  1  1T ;
d 65 ¼ ½1 1T ; d61 ¼ ½1  1T :
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2081

The initial state of each agent is given as

z1 ð0Þ ¼ ½  3  3  0:5 0:5T ;

z2 ð0Þ ¼ ½  5 2 0:5 0:5T ;
z3 ð0Þ ¼ ½  4 7 0 0:5T ;
z4 ð0Þ ¼ ½  2 5  0:5 0:5T ;
z5 ð0Þ ¼ ½  6 4  0:5  0:5T ;
z6 ð0Þ ¼ ½  2 6 0 0 T :

6
Agent 1
4 Agent 2
Agent 3
Agent 4
2 Agent 5
Agent 6

0
2
x

−2

−4

−6

−8
−8 −6 −4 −2 0 2 4
x1

Fig. 1. Trajectories of all the agents.

9
x21
8 x31
x41
7
x51
x
6 61
Distance

0
2 4 6 8 10 12 14 16 18
T

Fig. 2. The relative distances between agent 1 and other agents.

2082 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

The simulation is implemented in Matlab with the toolboxes ICLOCS [19] and IPOPT [20].
Fig. 1 shows the trajectory of each agent under Algorithm 1 and the black square in this figure
represents the obstacle. All the agents eventually arrive at their desired positions and thereby the
formation of regular hexagon is achieved. Due to space limitations, here we only take agent 1 as
an example to show the guaranteed collision avoidance. As shown in Fig. 2, the relative
distances between agent 1 and other agents are always greater than the safety distance 2R, which
is depicted by the dashed line. It can be shown from Fig. 3 that obstacle avoidance is ensured,
since the relative distances between each agent and the obstacle are always greater than the safety
distance 2R. Figs. 4 and 5 show that two components of the input of each agent lie within the

10
x1o
9 x2o

8 x3o
x4o
7 x5o

6 x6o
Distance

0
2 4 6 8 10 12 14 16 18
T

Fig. 3. The relative distances between each agent and the obstacle.

Agent 1
0.6
Agent 2
Agent 3
Agent 4
0.4
Agent 5
Agent 6
0.2
ui1

−0.2

−0.4

−0.6

2 4 6 8 10 12 14 16 18
T

Fig. 4. The first component of the input of each agent.

 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2083

Agent 1
0.6
Agent 2
Agent 3
Agent 4
0.4
Agent 5
Agent 6
0.2
ui2

−0.2

−0.4

−0.6

2 4 6 8 10 12 14 16 18
T

Fig. 5. The second component of the input of each agent.

1500

1000
−3
x 10
4
J

2
500

0
14 14.5

0
2 4 6 8 10 12 14 16 18
T

Fig. 6. The cost of the whole system.

interval ½ 0:5; 0:5, which means that the input constraints are satisfied. The closed-loop
stability can be illustrated by Fig. 6, in which the cost of the whole system decreases
monotonically to zero. The above simulation results verify the effectiveness of the proposed
DMPC algorithm.

5. Conclusion

In this paper, a DMPC algorithm has been proposed for formation of multiple linear second-
order agents with collision avoidance and obstacle avoidance. The synchronous update strategy
is adopted to formulate a tractable MPC optimization problem in the distributed manner. The
replacement of the real states with the assumed ones requires introducing the compatibility
2084 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085

constraint, which plays an important role to ensure collision avoidance and stability of the whole
system. Furthermore, a terminal controller and a positively invariant terminal set as well as a
corresponding terminal cost are designed with the use of the specific form of the system model
and the control objective. By applying the proposed algorithm, the properties including recursive
feasibility and closed-loop stability of the whole system are guaranteed. Finally, an illustrative
example is presented to show the effectiveness of this algorithm.
On-going research is extending the ideas in this paper to a distributed robust MPC (DRMPC)
strategy for formation of multiple agents with uncertainty. Future research areas include the
event-triggered DMPC scheme and DMPC for multi-agent systems with Markov chain patterns
[21,22].

Acknowledgements

This work was supported by the National Basic Research Program of China (973 Program)
under Grant 2012CB720000, the National Natural Science Foundation of China under Grant
61603041, Grant 61225015, Grant 61105092 and Grant 61422102, the Beijing Natural Science
Foundation under Grant 4161001, and by Foundation for Innovative Research Groups of the
National Natural Science Foundation of China under Grant 61321002.

References

[1] R. Olfati-Saber, A. Fax, R.M. Murray, Consensus and cooperation in networked multi-agent systems, Proc. IEEE 95
(1) (2007) 215–233.
[2] Y. Cao, W. Yu, W. Ren, G. Chen, An overview of recent progress in the study of distributed multi-agent
coordination, IEEE Trans. Ind. Inf. 9 (1) (2013) 427–438.
[3] F. Xiao, L. Wang, J. Chen, Y. Gao, Finite-time formation control for multi-agent systems, Automatica 45 (11)
(2009) 2605–2611.
[4] S. Mastellone, D.M. Stipanović, C.R. Graunke, K.A. Intlekofer, M.W. Spong, Formation control and collision
avoidance for multi-agent non-holonomic systems: theory and experiments, Int. J. Robot. Res. 27 (1) (2008)
107–126.
[5] S.M. Kang, H.S. Ahn, Design and realization of distributed adaptive formation control law for multi-agent systems
with moving leader, IEEE Trans. Ind. Electron. 63 (2) (2016) 1268–1279.
[6] M. Egerstedt, X. Hu, Formation constrained multi-agent control, IEEE Trans. Robot. Autom. 17 (6) (2001)
947–951.
[7] D.Q. Mayne, J.B. Rawlings, C.V. Rao, P.O. Scokaert, Constrained model predictive control: stability and
optimality, Automatica 36 (6) (2000) 789–814.
[8] D.Q. Mayne, Model predictive control: recent developments and future promise, Automatica 50 (12) (2014)
2967–2986.
[9] R.R. Negenborn, J.M. Maestre, Distributed model predictive control: an overview and roadmap of future research
opportunities, IEEE Control Syst. 34 (4) (2014) 87–97.
[10] J.M. Maestre, R.R. Negenborn, Distributed Model Predictive Control Made Easy, Dordrecht, Netherlands Springer,
2014.
[11] W.B. Dunbar, R.M. Murray, Distributed receding horizon control for multi-vehicle formation stabilization,
Automatica 42 (4) (2006) 549–558.
[12] T. Keviczky, F. Borrelli, G.J. Balas, Decentralized receding horizon control for large scale dynamically decoupled
systems, Automatica 42 (12) (2006) 2105–2115.
[13] T. Keviczky, F. Borrelli, K. Fregene, D. Godbole, G.J. Balas, Decentralized receding horizon control and
coordination of autonomous vehicle formations, IEEE Trans. Control Syst. Technol. 16 (1) (2008) 19–33.
[14] A. Richards, J. How, Decentralized model predictive control of cooperating UAVs, in: Proceedings of the 43rd
IEEE Conference on Decision and Control, 2004, pp. 4286–4291.
[15] S.B. Wei, B.C. Ding, G. Chen, Y. Chai, Distributed model predictive control for multi-agent systems with coupling
constraints, Int. J. Model. Ident. Control 10 (3–4) (2010) 238–245.
 L. Dai et al. / Journal of the Franklin Institute 354 (2017) 2068–2085 2085

[16] P. Wang, B.C. Ding, A synthesis approach of distributed model predictive control for homogeneous multi-agent
system with collision avoidance, Int. J. Control 87 (1) (2014) 52–63.
[17] Y. Gao, Y. Xia, L. Dai, Cooperative distributed model predictive control of multiple coupled linear systems, IET
Control Theory Appl. 9 (17) (2015) 2561–2567.
[18] P. Wang, B. Ding, Distributed RHC for tracking and formation of nonholonomic multi-vehicle systems, IEEE Trans.
Autom. Control 59 (6) (2014) 1439–1453.
[19] P. Falugi, E.C. Kerrigan, E.V. Wyk. Imperial College London Optimal Control Software User Guide (ICLOCS),
2010. Available: 〈http://www.ee.ic.ac.uk/ICLOCS/〉.
[20] A. Wächter, L.T. Biegler, On the implementation of an interiorpoint filter line-search algorithm for large-scale
nonlinear programming, Math. Program. 106 (1) (2006) 25–57.
[21] Y. Wei, J. Qiu, H.R. Karimi, Quantized H 1 filtering for continuous-time markovian jump systems with deficient
mode information, Asian J. Control 17 (5) (2015) 1914–1923.
[22] Y. Wei, J. Qiu, H.R. Karimi, M. Wang, Model approximation for two-dimensional Markovian jump systems with
state-delays and imperfect mode information, Multidim. Syst. Signal Process. 26 (3) (2015) 575–597.