2011 American Control Conference

on O'Farrell Street, San Francisco, CA, USA

June 29 - July 01, 2011

Multi-agent Consensus Algorithm with Obstacle Avoidance via

Optimal Control Approach
Jianan Wang and Ming Xin*

Abstract—Multi-agent consensus problem in an obstacle- desired formation, was introduced to achieve the formation
laden environment is addressed in this paper. A novel optimal as well as the collision avoidance between multiple agents.
control approach is proposed for the multi-agent system to In [14], a new distributed robust model predictive control
reach consensus as well as avoid obstacles with a reasonable algorithm was developed for multi-agent trajectory
control effort. An innovative nonquadratic penalty function is
constructed to achieve obstacle avoidance capability from an
optimization utilizing constraint tightening to ensure safety
inverse optimal control perspective. The asymptotic stability in the presence of the environmental changes and generate
and optimality of the consensus algorithm are proven. In an intelligent trajectory around known obstacles. A
addition, the optimal control law of each agent only requires cooperative control law for the individual agent to guarantee
local information from the neighbors to guarantee the proposed collision avoidance in multi-agent systems was proposed in
behaviors, rather than all agents’ information. The consensus [15]. However, it is assumed that every agent knows its
and obstacle avoidance are validated through various
simulations.
desired state and a LQR based optimal control is designed to
track the desired state.
I. INTRODUCTION Most of the obstacle avoidance strategies are designed
either for path planning of the single agent or for multiple
M ulti-agent cooperative missions are becoming
increasingly important and feasible owing to the rapid
advances in computing, communication, sensing, and
agents without considering their interaction topologies and
the information consensus problem. In this paper, we
address both consensus problem and obstacle avoidance in a
actuation. Cooperative control has been recognized to be of unified optimal control framework. A novel avoidance
critically importance to the successful accomplishment of penalty function is constructed based on an inverse optimal
these cooperative missions. control strategy [16, 17] such that an analytical optimal
As a core of multi-agent cooperative control, consensus control law can be obtained. In addition, it can be shown that
problem has been extensively studied in recent years [1-4]. the resultant consensus algorithm is a linear function of the
From the optimization perspective, consensus algorithms Laplacian, and thus only local information from the
have been developed along two lines: 1) fastest convergence communication topology is required to implement the
time: the algorithms were designed to achieve the fastest optimal cooperative control law.
convergence time by finding an optimal weighting matrix The remainder of this paper is organized as follows. The
[5], constructing a proper configuration that maximizes the consensus problem is described in Section II and Section III
second smallest eigenvalue of the Laplacian [6], and presents the main result of this paper. Simulation results and
exploring an optimal interaction graph for the average analysis are shown in Section IV. Some conclusion remarks
consensus problem [7]; 2) Optimal control design: the are given in Section V.
consensus problem was formulated as an optimal control
problem and solved using a linear matrix inequality (LMI) II. PROBLEM STATEMENT
approach [8], a LQR-based optimal linear consensus
algorithm [9], a distributed subgradient method for multi- Consider n agents with double-integrator dynamics:
agent optimization [10], and a locally optimal nonlinear ⎧ pi = vi
⎨ , i = 1,..., n (1a)
consensus strategy by imposing individual objectives [11]. ⎩ vi = ai
In the realistic environment, if obstacles emerge right on or in a matrix form X = AX + BU (1b)
the trajectory, the multiple agents may not be able to safely
achieve desired cooperative behaviors. Therefore, intensive ⎡0 In ⎤ ⎡0 ⎤
A = ⎢ n× n ⎥ ⊗ I m , B = ⎢ n× n ⎥ ⊗ I m
attention has been paid to the cooperative control problem ⎣ 0 n× n 0 n × n ⎦ ⎣ In ⎦
with obstacle/collision avoidance. In [12], three flocking X = [ p1 ,… , pn , v1 ,… , vn ] , U = [a1T ,… , anT ]T
T T T T T

algorithms were proposed to achieve both flocking and

pT vT
obstacle avoidance by adding obstacle avoidance terms to
the group objective. In [13], a constraint force, directly where pi (t ) ∈ R , vi (t ) ∈ R and ai (t ) ∈ R m are, respectively,
m m

converted from the structural distance constraints for a the position, velocity and control input of agent i.
T
X = ⎡ pT v ⎤ and U are the aggregate state and control
T
The authors are both with the Department of Aerospace Engineering, ⎣ ⎦
Mississippi State University, Starkville, MS 39762, USA. (*corresponding
author: 662-325-2139; fax: 662-325-7730; e-mail: [email protected]). input of all agents. ⊗ denotes the Kronecker product.

978-1-4577-0081-1/11/$26.00 ©2011 AACC 2783

Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.
 The consensus problem in this paper is to design a ( L ⊗ I m ) pcs = 0nm×1
distributed control law ai (t ) based on the information (5)
( L ⊗ I m )vcs = 0nm×1
exchange topology such that pi (t ) − p j (t ) → 0 and The final consensus state satisfies the dynamic equation
vi (t ) − v j (t ) → 0 . In addition, each agent is guaranteed to X cs = AX cs + BU cs = AX cs (6)
since U cs = 0nm×1 when the agents reach consensus.
avoid the obstacle along its trajectory.
Fig. 1 shows an example scenario of four agents’ Then, from Eq. (1b) and (6) the error dynamics becomes
consensus problem. R denotes the radius of the obstacle Xˆ = AXˆ + BU (7)
detection region and r denotes the radius of the obstacle. The The consensus is achieved when the system (7) is
dashed line denotes the original consensus trajectory without asymptotically stable.
obstacle. The proposed consensus law will be able to not In this paper, the consensus problem is formulated as an
only drive all the agents along the solid lines to reach optimal control problem with three cost function
consensus but also avoid the obstacle with an optimal components:
control effort. Min : J = J1 + J 2 + J 3
(8)
4
Detection Region
3 S .t. Xˆ = AXˆ + BU
where J1 , J 2 , J 3 represent the consensus cost, obstacle
avoidance cost, and control effort, respectively.
The consensus cost has the form of:
r
R ∞ ∞⎧⎪ ⎛ ⎡ wp 2 L2 0 n× n ⎤ ⎞ ⎫⎪
J1 = ∫ Xˆ T R1 Xˆ dt = ∫ ⎨ Xˆ T ⎜⎢ ⎥ ⊗ I m ⎟⎟ Xˆ ⎬ dt
⎜ wv L − 2 wp wc L ⎥⎦
2 2
⎝ ⎢⎣ 0 n× n
0 0
⎩⎪ ⎠ ⎭⎪
Obstacle (9)
where wp , wv , and wc represent the weights on the position
consensus, velocity consensus, and control effort,
respectively. It is necessary that R1 is positive semi-definite,
1 2
which can be shown in the following proposition.
Fig. 1: Multi-agent consensus scenario with an obstacle
Proposition 4.1: R1 is positive semi-definite if the graph is
III. OPTIMAL CONSENSUS WITH OBSTACLE AVOIDANCE undirected and connected and
wv 2 ei2 − 2wp wc ei ≥ 0 (10)
In this section, we propose a unified inverse optimal
control approach to address the consensus problem with where ei is the eigenvalue of L .
obstacle avoidance capability. For the convenience of Proof: Since L is positive semi-definite and it is
formulation, we define the error state straightforward to show that L2 is also positive semi-definite,
T
Xˆ = ⎡ pˆ T vˆ ⎤ wv 2 L2 − 2 wp wc L = wv 2 QΛ 2 Q −1 − 2w p wc Q ΛQ −1
T
X − X cs (2)
⎣ ⎦
T ⎡ wv 2 e12 − 2 wp wc e1 0 0 ⎤
where X cs = ⎡⎣ pcsT vcsT ⎤⎦ (3) ⎢
0 w e − 2 wp wc e2
2 2
0
⎥
= Q⎢ v 2 ⎥ Q −1
is the final consensus state. For instance, in a planar motion, ⎢ ⎥
T ⎢ ⎥
X cs = ⎡⎣ pcsT
T
vcsT ⎤⎦ = ⎡11× n ⊗ ⎡⎣α x α y ⎤⎦ 11× n ⊗ ⎡⎣ β x β y ⎤⎦ ⎤ (4) ⎣⎢ 0 0 wv 2 en2 − 2 wp wc en ⎦⎥
⎣ ⎦
(11)
where α x , α y are the final consensus position along x axis
where Q is composed of the eigenvectors of L and Λ is the
and y axis, respectively; β x , β y are the final consensus diagonal matrix with the diagonal elements being the
velocity along x axis and y axis, respectively. Note that eigenvalues of L . Since L is positive semi-definite, ei ≥ 0 .
the consensus state X cs is not known a priori. Therefore, wv 2 L2 − 2wp wc L is positive semi-definite if
We follow the standard definitions and concepts from the
wv 2 ei2 − 2wp wc ei ≥ 0 and it follows that R1 is positive semi-
graph theory to describe the interconnection of multi-agent
systems, which can be referred to [18]. In particular, the definite. ■
Laplacian matrix L is commonly used to define the Remark 4.1: The condition (10) is required in Proposition
communication topology among agents. In this paper, the 4.1. One can always find proper weights to satisfy (10). For
information exchange topology is assumed to be undirected instance, a large wv and small enough wp and wc are
and connected. Under this assumption, L is positive semi- applicable due to ei ≥ 0 .
definite and the following property holds when the agents
The obstacle avoidance cost has the form of
reach consensus: [18]

2784
Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.
 ∞
J 2 = ∫ h( Xˆ )dt (12) Proof: Omitted. Refer to [16]. ■
0 The main result of this paper is presented in the following
where h( Xˆ ) will be constructed from an inverse optimal theorem.
control approach in Theorem 4.1. Theorem 4.1: For a multi-agent system (1) with an
The control effort cost has the regular quadratic form of undirected and connected interaction graph, there always
∞ exist a large enough wv , small enough wp and wc , such that
J 3 = ∫ U T R2Udt (13)
0 the feedback control law
where R2 = wc I n ⊗ I m is positive definite and wc is the
2
wp w 1
U =φ(X) = − ( L ⊗ I m ) p − v ( L ⊗ I m )v − 2 g v '( X ) (26)
weighting parameter. wc wc 2wc
The following lemma is introduced to derive our main is an optimal control law for the consensus problem (8) and
result. the closed-loop system is globally asymptotically stable.
Lemma 4.1: [16] Consider the nonlinear controlled
h( Xˆ ) in the obstacle avoidance cost function (12) is
dynamical system
wp
Xˆ (t ) = f ( Xˆ (t ), U (t )), Xˆ (0) = Xˆ , t ≥ 0 (14) 0 h( Xˆ ) = vˆT ( L ⊗ I m )(G p ⊗ I m )( L ⊗ I m ) pˆ
wc
with f (0 , 0 )=0 and a cost functional given by
wv T
J ( Xˆ 0 , U (t ))
∞
T ( Xˆ (t ), U (t ))dt + vˆ ( L ⊗ I m )(G p ⊗ I m )( L ⊗ I m )vˆ − g p 'T ( Xˆ )vˆ (27)
∫
0
(15) wc
where U (t ) is an admissible control. Let D ⊆ R n be an 1 T
+ vˆ ( L ⊗ I m )(G p2 ⊗ I m )( L ⊗ I m )vˆ
open set and Ω ⊆ R . Assume that there exists a
m 4wc 2
continuously differentiable function V : D → R and a where g v '( X ) and g p 'T ( Xˆ ) in (26) and (27) are derived from
control law φ : D → Ω such that the obstacle avoidance potential function defined by
V (0 ) = 0 (16) 1
g ( Xˆ ) = vˆT (G p ⊗ I m )( L ⊗ I m )vˆ (28)
V ( Xˆ ) > 0, Xˆ ∈ D, Xˆ ≠ 0 (17) 2
φ (0 ) = 0 (18) with G p = diag ( m( p1 ) m( p2 ) m( pn ) ) and
V '( Xˆ ) f ( Xˆ , φ ( Xˆ )) < 0, Xˆ ∈ D, Xˆ ≠ 0 (19) ⎧ 0 R < pi − Ob
⎪ 2
H ( Xˆ , φ ( Xˆ )) = 0, Xˆ ∈ D (20) ⎪⎪⎛ R 2 − pi − Ob 2 ⎞
m( pi ) = ⎨⎜ ⎟ r < pi − Ob ≤ R i = 1,..., n (29)
⎪⎜⎝ pi − Ob − r ⎟⎠
2 2
H ( Xˆ , U ) ≥ 0, Xˆ ∈ D, U ∈ Ω (21)
⎪
⎪⎩ not defined pi − Ob ≤ r
where H ( Xˆ , U ) T ( Xˆ , U ) + V '( Xˆ ) f ( Xˆ , U )
is the
T
Hamiltonian function. The superscript ' denotes partial g '( Xˆ ) = ⎡⎣ g p 'T ( Xˆ ) g v 'T ( Xˆ ) ⎤⎦
differentiation with respect to X̂ . T
⎡⎛ ∂g ( Xˆ ) ⎞T ⎛ ∂g ( Xˆ ) ⎞
T T
⎛ ∂g ( Xˆ ) ⎞ ⎤
Then, with the feedback control g p '( Xˆ ) = ⎢⎜⎜ ⎟ ⎥ (30)
⎢⎝ ∂pˆ1 ⎟⎠ ⎜⎜ ∂pˆ ⎟⎟ ⎜⎜ ∂pˆ ⎟⎟ ⎥
U (t ) = φ ( Xˆ (t )) (22) ⎣ ⎝ 2 ⎠ ⎝ n ⎠ ⎦
the solution Xˆ (t ) ≡ 0 of the closed-loop system is locally g v '( Xˆ ) = (G p ⊗ I m )( L ⊗ I m )vˆ = (G p ⊗ I m )( L ⊗ I m )v = g v '( X )
asymptotically stable and there exists a neighborhood of the where g p '( Xˆ ) and g v '( X ) represent the partial
origin D0 ⊆ D such that
differentiation of g ( Xˆ ) with respect to the position error p̂
J ( Xˆ 0 , φ ( Xˆ (t ))) = V ( Xˆ 0 ), Xˆ 0 ∈ D0 (23) and the velocity error v̂ respectively.
In addition, if Xˆ 0 ∈ D0 then the feedback control (22) Proof: Specific to this optimal consensus problem, we have
the following equations corresponding to Lemma 4.1:
minimizes J ( Xˆ , U (t )) in the sense that
T ( Xˆ , U ) = Xˆ T R1 Xˆ + h( Xˆ ) + U T R2U
0
(31)
J ( Xˆ 0 , φ ( Xˆ (t ))) = min J ( Xˆ 0 , U (t )) (24)
U ( t )∈S ( Xˆ 0 ) f ( Xˆ , U ) = AXˆ + BU (32)
where S ( Xˆ 0 ) denotes the set of asymptotically stabilizing ˆ
A candidate Lyapunov function V ( X ) is chosen to be
controllers for each initial condition Xˆ 0 ∈ D . Finally, if V ( Xˆ ) = Xˆ T PXˆ + g ( Xˆ ) (33)
D = R n , Ω = R m , and where P is the solution of a Riccati equation, which will be
V ( Xˆ ) → ∞ as Xˆ → ∞ (25) shown afterwards.
In order for the function V ( Xˆ ) in (33) to be a valid
the solution Xˆ (t ) ≡ 0 of the closed-loop system is globally
Lyapunov function, it must be continuously differentiable
asymptotically stable.

2785
Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.
with respect to X̂ or equivalently g ( Xˆ ) must be Substituting A, B, R1 , R2 in (38) and assuming
continuously differentiable with respect to X̂ . From the ⎡P P2 ⎤
P=⎢ 1 ⊗ I m yields
equations (28) and (29), it suffices to show that m( pi ) is ⎣ P2 P3 ⎥⎦
continuously differentiable in the safety region ⎡ 1 2 1 ⎤
dm( pi ) ⎢ − w 2 P2 P1 −
wc 2
P2 P3 ⎥
⎡ w 2 L2 0 n× n ⎤
{ }
p pi − Ob > r . In fact, this is true if m( pi ) and
dpi
⎢
⎢
c

1 1
⎥+⎢ p
⎥ 0
⎥ = 0 (41)
wv L − 2wp wc L ⎥⎦
2 2

⎢ P1 − 2 P3 P2 2 P2 − 2 P32 ⎥ ⎢⎣ n×n
⎣ wc wc ⎦
are continuous at pi − Ob = R . Since Eq. (29) implies that
Then, P can be solved in the analytical form
lim m( pi ) = 0 = lim m( pi ) , m( pi ) is continuous at
pi − Ob → R − pi − Ob → R + ⎡ w w L2 wp wc L ⎤
P=⎢ p v ⎥ ⊗ Im (42)
pi − Ob = R and thus continuous over the safety region. ⎣⎢ wp wc L wc wv L ⎦⎥
dm( pi ) Next, the cost function term h( Xˆ ) in J 2 is constructed
Similarly, it can be easily shown that is continuous
dpi from solving Eq. (39) and using (42):
over the safety region. Therefore, g ( Xˆ ) and the Lyapunov wp T
h( Xˆ ) = vˆ ( L ⊗ I m )(G p ⊗ I m )( L ⊗ I m ) pˆ
function V ( Xˆ ) are continuously differentiable with respect wc
wv T
to X̂ in the safety region. + vˆ ( L ⊗ I m )(G p ⊗ I m )( L ⊗ I m )vˆ − g p 'T ( Xˆ )vˆ (43)
The Hamiltonian function can be written as: wc
H ( Xˆ , U , V 'T ( Xˆ )) = T ( Xˆ , U ) + V 'T ( Xˆ ) f ( Xˆ , U ) 1 T
(34) + vˆ ( L ⊗ I m )(G p2 ⊗ I m )( L ⊗ I m )vˆ
= Xˆ T R Xˆ + h( Xˆ ) + U T R U + [2 Xˆ T P + g 'T ( Xˆ )][ AXˆ + BU ] 4 wc 2
1 2
which turns out to be (27).
Setting (∂ ∂U ) H ( Xˆ , U ,V 'T ( Xˆ )) = 0 yields the optimal
Using (38) and (39), (36) becomes
control law:
V 'T ( Xˆ ) f ( Xˆ , φ ( Xˆ ))
1 1 (44)
U * = φ ( Xˆ ) = − R2−1 BT V '( Xˆ ) = − R2−1 BT PXˆ − R2−1 BT g '( Xˆ ) (35) 1 1
2 2 = −[ Xˆ T R1 Xˆ + h( Xˆ ) + ( Xˆ T P + g' T ( Xˆ )) S ( PXˆ + g' ( Xˆ ))]
With (35) it follows that 2 2
It can be seen from (51) that the condition (19) can be met
V 'T ( Xˆ ) f ( Xˆ , φ ( Xˆ )) = Xˆ T ( AT P + PA − 2 PSP ) Xˆ
1 (36) if h( Xˆ ) ≥ 0 since Xˆ T R Xˆ is positive semi-definite and
1
− Xˆ T PSg '( Xˆ ) + g 'T ( Xˆ )( A − SP) Xˆ − g 'T ( Xˆ ) Sg '( Xˆ )
2 1 1
( Xˆ T P + g' T ( Xˆ )) S ( PXˆ + g' ( Xˆ )) is positive definite. By
where S BR2−1 BT . Using (35) and (36) into (34) yields 2 2
selecting proper values of the weights wp , wv , and wc , one
H ( Xˆ , φ ( Xˆ ), V 'T ( Xˆ )) = Xˆ T ( AT P + PA + R1 − PSP ) Xˆ
1 T ˆ (37) can always make h( Xˆ ) ≥ 0 . Specifically, if all the agents are
+ g 'T ( Xˆ )( A − SP ) Xˆ + h( Xˆ ) −
g ' ( X ) Sg '( Xˆ )
4 outside the detection region, h( Xˆ ) = 0 by the definition of
In order to prove that the control law (35) is an optimal
G p in (29). h( Xˆ ) > 0 can be guaranteed if one choose a
solution to the consensus problem (8) using the Lemma 4.1,
the conditions (16)-(21) need to be verified. large enough wv , small enough wp and wc such that the
⎡0 ⎤ positive terms vˆ (L ⊗Im )(Gp ⊗Im )(L ⊗Im )vˆ
T
and
Since B = ⎢ n× n ⎥ ⊗ I m , it can be seen that
⎣ In ⎦
vˆ (L ⊗ Im )(G ⊗ Im )(L ⊗ Im )vˆ in (44) are always greater than
T 2
p
1
φ ( Xˆ ) = − R B PXˆ − R2−1 g v '( Xˆ ) . From the form of
−1
2
T
other sign-indefinite terms.
2 Next we will verify the conditions (16) and (17). Note that
ˆ
gv '( X ) in (30), the condition (18), i.e. φ (0 ) = 0 , is satisfied. ⎛ ⎡ w w L2 wp wc L ⎤ ⎞
In order to satisfy the condition (20) in Lemma 4.1 or let Xˆ T PXˆ = Xˆ T ⎜ ⎢ p v ⎥ ⊗ I m ⎟ Xˆ
⎜ ⎢ wp wc L wc wv L ⎥ ⎟
Eq. (37) be zero, we can let ⎝⎣ ⎦ ⎠
AT P + PA + R1 − PSP = 0 (38) = wp wv pˆ ( L ⊗ I m ) pˆ + wc wv vˆ ( L ⊗ I m )vˆ + 2wp wc pˆ T ( L ⊗ I m )vˆ
T 2 T

and require that = wp wv pT ( L2 ⊗ I m ) p + wc wv v T ( L ⊗ I m )v + 2wp wc pT ( L ⊗ I m )v

1 (45)
g 'T ( Xˆ )( A − SP) Xˆ + h( Xˆ ) − g 'T ( Xˆ ) Sg '( Xˆ ) = 0 (39)
4 The last equality is obtained using the property (5).
With (35), (38), and (39), it can be shown that The Lyapunov function finally turns out to be:
H ( Xˆ , U ,V 'T ( Xˆ ))=[U − φ ( Xˆ )]T R2 [U − φ ( Xˆ )] ≥ 0 (40) V ( Xˆ ) = Xˆ T PXˆ + g ( Xˆ )
Therefore, the condition (21) is satisfied.

2786
Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.
 ⎧ Xˆ T PXˆ R < pi − Ob IV. SIMULATION RESULTS AND ANALYSIS
⎪ In this section, two simulation scenarios are used to
⎪ 1
= ⎨ Xˆ T PXˆ + vˆT (G p ⊗ I m )( L ⊗ I m )vˆ r < pi − Ob ≤ R (46) validate the proposed optimal consensus algorithm. Consider
⎪ 2
⎪ not defined pi − Ob ≤ r a planar motion in Fig. 1 with 4 agents and thus m = 2 .
⎩ The initial positions are given by (-2, -2), (2, -2), (2, 2)
It can be seen from (45) and (46) that the condition (16) is and (-2, 2), respectively. The initial velocities are assumed to
satisfied. Moreover, if Xˆ ≠ 0 , i.e. X ≠ X , pT ( L2 ⊗ I ) p
cs m be (0.2, 0.4), (-0.4, 0.2), (-0.2, -0.4), and (0.2, -0.2),
and v ( L ⊗ I m )v will not be equal to zero but positive
T respectively. The weights in the consensus algorithm are set
to wp = 0.04 , wv = 1.2 , and wc = 0.8 .
according to the property of L, i.e. Eq. (5). Note that p = 0
A. Consensus without obstacles on the trajectories
and v =0 that leads to pT ( L2 ⊗ I m ) p = 0 and
In this scenario, an obstacle is assumed to appear on (2,
v T ( L ⊗ I m )v = 0 is a special case of p = pcs and v = vcs 0), which is not on the trajectory of any agent. The radius of
when p = 0 and v = 0 , which implies Xˆ = 0 as well. the obstacle and the detection region are set to
cs cs
r = 0.1 and R = 0.5 . The simulation results of the four
Therefore, the condition (17), V ( Xˆ ) > 0 when Xˆ ≠ 0 , can
agents’ motion are shown in Fig. 2. As can be seen, the
be met by selecting a large enough wv for given wp and wc obstacle avoidance does not take effect since no agent steps
such that the positive terms wp wv pT ( L2 ⊗ I m ) p and into the detection region and the four agents achieve
consensus.
wc wv v T ( L ⊗ I m )v are always greater than the sign-indefinite
terms.
Substituting P and g '( Xˆ ) into (35) leads to
wp wv 1
φ ( Xˆ ) = − ( L ⊗ I m ) pˆ − ( L ⊗ I m )vˆ − 2 g v '( Xˆ ) (47)
wc wc 2 wc
which turns out to be Eq. (26) by substituting pˆ = p − pcs and
vˆ = v − vcs into (47) and using the property of (5). Note that
the optimal control law (26) is only a function of X . This is
desired because X cs is not known a priori.
Now, all the conditions (16)-(21) in Lemma 4.1 can be
satisfied by selecting a large enough wv and small enough
wp and wc . Furthermore, this rule of weight selection also
Fig. 2: Trajectories of the four agents without obstacle
applies to satisfy the condition (10). Therefore, according to Note that h( Xˆ ) in the obstacle avoidance cost function is
Lemma 4.1, the control law (26) is an optimal control law for
the problem (8) in the sense of (23) and (24), and the closed- equal to zero since all the agents are outside the detection
region, which implies that the problem is just a normal
loop system is asymptotically stable. It implies X = X cs and
optimal consensus problem. It has the same form as the
the consensus is achieved. conventional consensus algorithm for networked double
In addition, it can be easily seen from (46) that integrator systems [18] except the weighting parameters.
V ( Xˆ ) → ∞ as Xˆ → ∞ . Therefore, the closed-loop system B. Consensus with multiple obstacles on the trajectories
is globally asymptotically stable. Note that the globally In this scenario, one obstacle with the same radius and
asymptotic stability region excludes the undefined area detection region as Scenario A is assumed to appear on (1,
{p }
pi − Ob ≤ r , which is physically meaningful because 1.3), which is on the trajectory of agent 3. The other obstacle
with r = 0.2 and R = 0.8 is assumed to appear on (0.5, 3.2),
no agent can start from inside the obstacle. ■
which is on the trajectories of agent 1 and agent 4.
Remark 4.2: As can be seen from Theorem 4.1, the optimal
The simulation results are shown in Figs. 3-5. Fig. 3
consensus algorithm is developed from an inverse optimal
demonstrates that all agents avoid the obstacles and reach
control approach since the cost function h( Xˆ ) is not given a the final consensus. Fig. 4 presents the time histories of the
priori but constructed from the optimality condition (39). agents’ positions and velocities. The optimal control inputs
Remark 4.3: From (26) and g v '( X ) in (30), it can be also are shown in Fig. 5. In the bottom two subfigures of Fig. 5,
seen that the optimal control law of each agent only requires the time histories in the first 50 seconds are shown for better
the local information based on the information exchange illustrating the transient responses. The velocity response
topology since it is a linear function of L. and the control response show that the optimal obstacle
avoidance control law does not require large control effort.

2787
Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.
 was constructed from an inverse optimal control approach
such that the optimal control law can be obtained in an
analytical form and was shown to be a linear function of the
Laplacian matrix, which indicates that the control law
requires only the local information and offers a great
implementation advantage. Both globally asymptotic
stability and optimality of this algorithm have been proven.
The simulation results have demonstrated that the proposed
optimal approach is capable of solving the consensus
problem under different obstacle avoidance scenarios.

REFERENCES
[1] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of
mobile autonomous agents using nearest neighbor rules,” IEEE
Transactions on Automatic Control, vol. 48, no. 6, pp. 988-1001, June
2003.
Fig. 3: Trajectories of the four agents with two obstacles [2] J. A. Fax, and R. M. Murray, “Information flow and cooperative
control of vehicle formations,” IEEE Transactions on Automatic
Control, vol. 49, no. 9, pp. 1465-1476, Sep. 2004.
[3] W. Ren, and E. M. Atkins, “Distributed multi-vehicle coordinated
control via local information exchange,” International Journal of
Robust and Nonlinear Control, vol. 17, no. 10, pp. 1002-1033, July
2007.
[4] W. J. Dong, and J. A. Farrell, “Cooperative control of multiple
Nonholonomic mobile agents,” IEEE Transactions on Automatic
Control, vol. 53, no. 6, pp. 1434-1448, July 2008.
[5] X. Lin, and B. Stephen, “Fast linear iterations for distributed
averaging,” Systems and Control Letters, vol. 53, no. 1, pp. 65-78,
2004.
[6] Y. S. Kim, and M. Mesbahi, “On maximizing the second smallest
eigenvalue of a state-dependent graph Laplacian,” IEEE Transactions
on Automatic Control, vol. 51, no. 1, pp. 116-120, Jan. 2006.
[7] J. C., Delvenne, L., Giarre, and S., Zampieri, “Optimal stragies in the
average consensus problem,” in Proceedings of the IEEE Conference
on Decision and Control, pp. 2498-2503, Dec. 2007.
[8] E. S. Kazerooni, and K. Khorasani, “An LMI approach to optimal
consensus seeking in multi-agent systems,” American Control
Fig. 4: Time histories of the four agents’ positions and velocities Conference, June 2009.
[9] Y. Cao and W. Ren, “Optimal linear consensus algorithms: An LQR
perspective,” IEEE Transactions on Systems, Man, and Cybernetics
(Part B), Vol. 40, No. 3, pp. 819-830, Jun. 2010.
[10] A. Dedic, and A. Ozdaglar, “Distributed subgradient methods for
multi-agent optimization,” IEEE Transactions on Automatic Control,
vol. 54, no. 1, Jan. 2009.
[11] D. Bauso, L. Giarre, and R. Pesenti, “Mechanism design for optimal
consensus problems,” Proceedings of the 45th IEEE Conference on
Decision and Control, San Diego, CA, Dec. 2006.
[12] R. Olfati-Saber, “Flocking for multi-agent dynamic systems:
Algorithms and Theory,” IEEE Transactions on Automatic Control,
vol. 51, no. 3, pp. 401-420, 2006.
[13] Y. F. Zou, P. R. Pagilla, and E. Misawa, “Formation of a group of
vehicles with full information using constraint forces,” Journal of
Dynamic Systems, Measurement, and Control, vol. 129, no. 5, pp.
654-661, 2007.
[14] Y. Kuwata, A. Richards, T. Schouwenaars, and P. J. How,
“Distributed robust receding horizon control for multivehicle
guidance,” IEEE Transactions on Control Systems Technology, vol.
15, no. 4, pp. 627-641, 2007.
[15] D. M. Stipanovic, P. F. Hokayem, M. W., Spong, and D. D. Siljak,
Fig. 5: Time histories of the four agents’ optimal control inputs “Cooperative Avoidance Control for Multiagent Systems,” Journal of
Dynamic Systems, Measurement, and Control, vol. 129, pp. 609-707,
2007.
V. CONCLUSION [16] D. S. Bernstein, “Nonquadratic cost and Nonlinear Feedback
In this paper, a novel optimal control law was developed Control,” International Journal of Robust and Nonlinear Control, vol.
3, pp. 211-229, 1993.
for multi-agent consensus with obstacle avoidance. The [17] W. M. Haddad and V. Chellaboina, Nonlinear Dynamical Systems
primary contribution is to formulate the consensus problem and Control: A Lyapunov-Based Approach, Princeton University
and obstacle avoidance in a unified optimal control Press, 2008.
[18] W. Ren and R. W. Beard, Distributed Consensus in Multi-vehicle
framework. A nonquadratic obstacle avoidance cost function Cooperative Control, Springer-Verlag, London, 2008.

2788
Authorized licensed use limited to: QUAID E AZAM UNIVERSITY. Downloaded on July 26,2024 at 15:03:27 UTC from IEEE Xplore. Restrictions apply.