IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO.

4, AUGUST 2007 693

Distributed Coordination Control of Multiagent

Systems While Preserving Connectedness
Meng Ji, Student Member, IEEE, and Magnus Egerstedt, Senior Member, IEEE

Abstract—This paper addresses the connectedness issue in mul- t. Variations in this protocol consist mainly of different weight
tiagent coordination, i.e., the problem of ensuring that a group of factors ki and αij . For example, linear time-varying weights are
mobile agents stays connected while achieving some performance used for αij in continuous time [6], [7] and discrete time [3],
objective. In particular, we study the rendezvous and the formation
control problems over dynamic interaction graphs, and by adding [9]. Nonlinear weights are proposed in [11] and [15]. In addi-
appropriate weights to the edges in the graphs, we guarantee that tion, a robust (in the sense of disturbance rejection) rendezvous
the graphs stay connected. algorithm is presented in [16].
Index Terms—Connected graphs, formation control, graph The formation control problem has also been extensively stud-
Laplacian, multiagent coordination. ied. Generally speaking, there are two kinds of formation control
approaches: the leader-follower approach and the leaderless ap-
I. INTRODUCTION proach. In the leader-follower approach, either an agent [17] or a
HE HISTORY behind this work can be traced back to virtual leader [18]–[21] is chosen as the leader, whose movement
T Reynolds’ “boids” model [1], where each agent only reacts
to its neighboring flock-mates following three adhoc protocols
is constrained by a predefined trajectory. The remaining agents
then track the leader, while obeying some coordination rules to
for autonomous agents, i.e., separation, alignment, and cohe- keep the formation. In contrast, the other approach to forma-
sion. A special case of the “boids” model was studied by Vicsek tion control is the leaderless approach [22], [23]. Here, the con-
et al. [2], where all the agents move at the same constant speed troller is typically given by a mixture of formation-maintenance,
and update their headings according to the nearest neighbor rule. obstacle-avoidance, and trajectory-following terms. Alternative
Velocity cohesion and flocking behavior was observed in both approaches to this problem include, e.g., local navigation func-
cases, and a proof of convergence was provided by Jadbabaie tions, and recent work along these lines can be found in [24]
et al. [3]. and [25].
What makes the multiagent problem challenging is that the In this paper, we will focus on providing solutions to the co-
agents are subjected to limitations on the available information, ordination problem that preserve connectedness in the presence
which has made graph-based models useful and natural tools for of limited sensing and communication ranges. In particular, the
encoding these limitations [4]–[12]. Among several important rendezvous and formation control problems are investigated. It
properties of such graph models, the graph Laplacian stands out, should be noted that these problems have already been solved
and it has been used for proving convergence and characterizing if either connectedness is assumed [5], [7], [15], or connected-
stability. Additional results that rely on such algebraic graph- ness is only required at distinct times [4], [3], [16], [26]–[28]
theoretic tools for graph-based control of mobile agents involve in the sense that the agents sense their environment and then
edge-based control and Lyapunov functions over graphs [3], move in such a way that the network is connected at the sens-
[13]–[15]. ing times, where the agents may be operating synchronously
The agreement problem (or consensus problem) is concerned or asynchronously. In particular, the first solution to the con-
with finding decentralized strategies that achieve convergence nectedness preserving rendezvous problem was given by Ando
to a common value. This problem arises in a number of applica- et al. in [4]; a discrete-time control algorithm was proposed
tions such as swarming, schooling, flocking, or rendezvous. The that evaluated and ensured connectivity, as well as other con-
agreement is typically achieved through a nearest-neighbor-like straints, at each instant of (discrete time). An additional relevant
protocol contribution along these lines can be found in [29], where the

connectivity-maintenance problem for ad hoc networks with
ẋi (t) = −ki (t) αij (t)(xi (t) − xj (t)) discrete-time double-integrator dynamics is considered.
j∈nbhdi (t) In this paper, we show how to make the graph stay connected
where xi is the state vector of agent i, and nbhdi (t) denotes the for all times (thus, removing the separation of the movements
neighborhood set (to be carefully defined later) of agent i at time into sensing and movement phases), and the outline of the paper
is as follows: In Section II, we review some previous results
Manuscript received August 17, 2006; revised January 19, 2007. This paper and recall some basic notions from algebraic graph theory. In
was recommended for publication by Associate Editor L. E. Parker and Editor A.
De Luca upon evaluation of the reviewers’ comments. This work was supported
Section III, we show how to add weights in the static graph
by the U.S. Army Research Office Grant 99838. case in order to solve the rendezvous problem, followed by
The authors are with the School of Electrical and Computer Engineering, the dynamic case in Section IV. The connectedness preserving
Georgia Institute of Technology, Atlanta, GA 30032 USA (e-mail: mengji@
ece.gatech.edu; [email protected]).
control law is extended to the formation control in Section V,
Digital Object Identifier 10.1109/TRO.2007.900638 followed by a collection of simulation results in Section VI.

1552-3098/$25.00 © 2007 IEEE

694 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 4, AUGUST 2007

II. BACKGROUND the node set to be static while allowing the edge-set to undergo
dynamic changes.
The graph Laplacian can be thought of as an encoding of the
discrete topology heat equation [30]–[32], where differential Before we can study the dynamic situation, a few words
should be said about the static case. In this situation, the agents
operators can be defined for functions over graphs. This in-
have established communication links between the predefined
terpretation of the graph Laplacian, defined as a mapping from
graph nodes to graph nodes, tells us that the graph Laplacian de- agents, and these links are assumed to be available through-
out the duration of the maneuver. In fact, by a static interac-
fines a diffusion of information over the network in that a node
value changes based on the values associated with its neigh- tion graph (SIG) G = (V, E), we understand the graph where
boring nodes. And, as already mentioned, pioneering work on the nodes V = {v1 , . . . , vN } are associated with the different
agents, and the static edge set E ⊂ V × V is a set of unordered
consensus problems or agreement problems utilizing the graph
Laplacian can be found in [6] and [11]–[13]. pairs of agents, with (vi , vj ) = (vj , vi ) ∈ E if and only if a
Since few mobile networks have a static network topology communication link exists between agents i and j. We will use
the shorthand V (G) and E(G) to denote the edge and node sets
due to the movements of the individual nodes as well as to the
temporal variations in the available communication channels, associated with a graph G.
interest in networks with changing topologies has been growing Given an agent i, we will associate NG (i) = {j|(vi , vj ) ∈
E(G)} with the neighborhood set to i, i.e., the set of agents
rapidly. In [8] and [33], a dynamic extension of the static graph
theory is proposed as a framework to address network problems adjacent to agent i. Using this terminology, what we understand
by a limited-information time-invariant decentralized control
with time-varying topologies. Ren and Beard [9] find that under
law in (1) is that
a dynamically changing interaction topology, if the union of the

interaction graphs across some time interval contains a spanning ui = f (xi − xj ) (2)
tree at a sufficient frequency as the system evolves, an informa- j∈Nσ (i)
tion consensus is still achievable. An average consensus prob-
lem is solved for switching topology networks in [13], where a where Nσ (i) ⊆ NG (i). The symmetric indicator function
common Lyapunov function is obtained for directed balanced σ(i, j) = σ(j, i) ∈ {0, 1} determines whether or not the infor-
graphs, based on a so-called disagreement function. Moreover, mation available through edge (vi , vj ) should be taken into
similar Lyapunov function arguments were employed in [34], account with
where the stability of coupled nonlinear oscillator networks was j ∈ Nσ (i) ⇔ (vi , vj ) ∈ E(G) ∧ σ(i, j) = 1. (3)
investigated. In this paper, we will draw inspiration from these
results, and we first establish some notation and review some (Using the terminology in [35], just because two nodes are
previous results. “neighbors” it does not follow that they are “friends.”) Along the
N agents, whose positions x1 , . . . , xN take on values in Rn same lines, the decentralized control law f (xi − xj ) is assumed
are given. The problem that we are studying in this paper is to be antisymmetric
that of limited information, decentralized control. As such, we f (xi − xj ) = −f (xj − xi ), ∀(vi , vj ) ∈ E(G). (4)
are focusing our attention on interaction and high-level control
strategies rather than on nonlinear vehicle models, and we as- A few remarks about these particular choices of control laws
sume that the dynamics of each individual agent is given by a and indicator functions should be made. First of all, the fact
single integrator that we only allow f to depend on the relative displacements
between interacting agents is that this is, in general, the only
ẋi = ui , i = 1, . . . , N. (1) type of information available to range-sensor-based information
channels, where agent i simply measures the position of agent
In order to establish what we mean by limited and decentral- j relative to its current position. Secondly, we insist on having
ized control, we follow the standard procedure of associating an the agents be homogeneous in that the same control laws should
interaction graph with the available information flow in that the govern the motion of all agents. This restriction is quite natural
nodes correspond to agents, and edges to available interagent (and arguably necessary) when considering large-scale networks
communication links. Such interaction graphs are, thus, reflec- where it quickly becomes unmanageable to assign and keep
tive of the underlying network topology, and different graphs track of individual control laws.
arise in different applications. Of particular importance to the As a consequence of restricting the permissible control laws
development in this paper are the ∆-disk proximity graphs, of to those given in (2), we obain that the centroid of the system is
connectivity graphs, where edges are established between nodes static. This fact follows directly from the antisymmetry of f in
vi and vj , if and only if the agents are within distance ∆ of each (4), and will be elaborated further in the following sections.
other, i.e., when |xi − xj | ≤ ∆. It sould be noted at this point The type of control terms presented in (2) have appeared
that such graphs are dynamic in nature, i.e., edges may appear repeatedly in the multiagent coordination community, and an
or disappear as agents move in or out of sensing (or communi- intuitive linear control law for solving the rendezvous problem
cation) distance of each other. Moreover, it is conceivable that is given by
agents are added or removed themselves, making not only the
σ(i, j) = 1,
edge set but also the node set a dynamical structure. In this
paper, we will not study this latter situation, and thus, restrict f (xi − xj ) = −(xi − xj ), ∀(vi , vj ) ∈ E(G)
JI AND EGERSTEDT: DISTRIBUTED COORDINATION CONTROL 695

which gives that 1. This, together with the nonnegativity of L(G) and the fact

that span{1} is L(G)-invariant, is sufficient to show that c(x, j)
ẋi = − (xi − xj ), i = 1, . . . , N. (5) approaches span{1} asymptotically.
j∈NG (i)
This result, elegant in its simplicity, can in fact be extended
Under the dynamics in (5), it has been shown that all the agents to dynamic graphs as well. In fact, since c(x, j)T c(x, j) is a
approach the same point asymptotically, provided that the SIG Lyapunov function to the system in (5), for any connected graph
is connected. And, even though this is a well-established result G, the control law
(see, e.g., [7]), we will outline a proof in this paper in order to
establish some needed notation and tools. d
c(x(t), j) = −L(G(t))c(x(t), j) (9)
First, we need to associate an arbitrary orientation to the dt
SIG, G. An orientation is a declaration of direction to each
edge o : E(G) → {−1, 1}such that if (vi , vj ) ∈ E(G), then drives the system to span{1} asymptotically as long as G(t) is
o(vi , vj ) = −o(vj , vi ). Using this orientation, we obtained the connected for all t ≥ 0.
oriented graph G o = (V, E, o). This well-known result is very promising, since dynamic net-
Now, if the total number of edges is equal to M , and we asso- work graphs are frequently occurring in which all real sen-
ciate an index with each edge such that E(G) = {e1 , . . . , eM }, sors and transmitters have finite range. This means that infor-
then the N × M incidence matrix of G o is I(G o ) = [ιij ], where mation exchange links may appear or be lost as the agents
 1, if v is the head of e move around. In fact, if we focus our attention on ∆-disk
i j
proximity graphs, we get the dynamic interaction graph (DIG)
ιij = −1, if vi is the tail of ej . (6)
G(t) = (V, E(t)) where (vi , vj ) = (vj , vi ) ∈ E(t), if and only
0, otherwise
if |xi (t) − xj (t)| ≤ ∆.
Through this incidence matrix, we can now define the graph By applying the control law in (5) to such DIGs, we get a
Laplacian L(G) ∈ RN ×N as system behavior that seemingly solves the rendezvous problem
L(G) = I(G o )I(G o )T (7) quite efficiently. However, the success of the control in (5) hinges
on an assumption that it shares with most graph-based results
where we have removed the orientation dependence in the left- (e.g., [3], [15]), i.e., on the connectedness assumption. Unfor-
hand side of (7). The reason for this is that the Laplacian does not tunately, this property has to be assumed rather than proved,
depend on the particular choice of orientation. In fact, one can and in Fig. 1, an example is shown where connectedness is lost
easily define the Laplacian without any reference to orientation when using (9) to control a system whose network topology is
or incidence matrices, but we follow this definition to ease the a ∆-disk proximity DIG.
notation in future sections. The remainder of this paper will show how this assumption
The graph Laplacian has a number of well-studied proper- can be overcome by modifying the control law in (5) in such a
ties (found, e.g., in [36]); the properties of importance to the way that connectedness holds for all times, while ensuring that
developments in this paper are listed as follows. the control laws are still based solely on local information, in
1) I(G o )I(G o )T = I(G o )I(G o )T for all orientation o, o , the sense of (2).
i.e., the Laplacian is orientation-independent.
2) L(G) is symmetric and positive semidefinite.
3) Let {λi }N III. WEIGHTED GRAPH-BASED FEEDBACK
i=1 be the ordered eigenvalues of L(G).
Then 0 ≤ λ1 ≤ λ2 ≤ · · · ≤ λN . Moreover, λ1 = 0 and In this section, we will restrict the interaction graphs to be
λ2 , . . . , λN > 0 if G is connected. static, i.e., we will only study the SIG-case in which the behav-
4) If G is connected, then the set of eigenvectors √ν1 , . . . , νN ior of the multiagent system is defined through a fixed network
form an orthogonal basis in RN , and ν1 = 1/ N 1, where topology. In particular, we will show how the introduction of
1 denotes the vector with every entry equal to one. In other nonlinear edge-weights can be used to establish certain invari-
words, if G is connected then null(G) = span{1}, where ance properties.
null(·) denotes the null space. To arrive at the desired invariance properties, we will first
If we now let the n-dimensional position of agent i investigate the decentralized control laws of the form
be given by xi = (xi,1 , . . . , xi,n ), i = 1, . . . , N , and let x =
(xT1 , . . . , xTN )T , we can define the componentwise operator as σ(i, j) = 1,
c(x, j) = (x1,j , . . . , xN,j ) ∈ R ,
T N
j = 1, . . . , n. f (xi − xj ) = −w(xi − xj )(xi − xj ), ∀(vi , vj ) ∈ E(G)
Using this notation, together with the observation that (5) can (10)
be decoupled along each dimension, we can in fact rewrite (5)
as where w : Rn → R+ is a positive symmetric weight function
d that associates a strictly positive and bounded weight to each
c(x, j) = −L(G)c(x, j), j = 1, . . . , n. (8)
dt edge in the SIG, based solely on the displacement xi − xj . We
As pointed out in [7] and [36], if G is connected, then the will study in detail as to how to choose the weight w in order to
eigenvector corresponding to the semisimple eigenvalue 0 is maintain connectedness.
696 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 4, AUGUST 2007

Fig. 1. Progression where connectedness is lost even though the initial graph is connected (∆ = 4.5).

This choice of decentralized control law gives We, moreover, define the -interior of a δ-constrained real-

ization of an SIG, G as
ẋi = − w(xi − xj )(xi − xj ) (11)
DG,δ

= {x ∈ RnN | ij ≤ (δ − ), ∀(vi , vj ) ∈ E(G)}.
j∈NG (i)
An edge-tension function Vij , can then be defined as
which can be rewritten as   (x) 2
Vij (δ, x) = δ− i j (x) , if (vi , vj ) ∈ E(G)
ij
d (14)
c(x, j) = −I o W(x)I oT c(x, j), j = 1, . . . , n (12) 0, otherwise
dt with
 2δ− i j (x)
where W(x) =diag(w1 (x), . . . , wM (x)) ∈ RM ×M , where, as ∂Vij (δ, x) (x − xj ), if(vi , vj ) ∈ E(G)
= (δ− i j (x) )2 i (15)
before, M = |E(G)| is the total number of edges, and where we ∂xi 0, otherwise.
have associated a label in {1, . . . , M } with each of the edges. it is to be noted that this edge-tension function (as well as its
We can then define the state-dependent weighted graph derivatives) is infinite when ij (x) = δ for some i, j, and, as
Laplacian as such, it may seem like an odd choice. However, as we will see,
we will actually be able to prevent the energy to reach infinity,
LW (x) = I o W(x)I oT (13)
and instead we will study its behavior on a compact set on which
where, as before, W(x) ∈ RM ×M is a diagonal matrix with it is continuously differentiable.
each element corresponding to a strictly positive edge weight. It The total tension energy of G can now be defined as
is moreover straightforward to establish that as long as the graph 1


N N
is connected, LW (x) is still positive semidefinite, with only one V(δ, x) = Vij (δ, x). (16)
2 i=1 j=1
zero eigenvalue corresponding to the null-space span{1}.
What we would like to show is that, given a critical distance Lemma 3.1: Given an initial position x0 ∈ DG,δ
, for a given
δ, together with the appropriate edge-weights, the edge-lengths  ∈ (0, δ). If the SIG G is connected, then the set Ω(δ, x0 ) :=
never go beyond δ if they start out being less than δ − , for {x|V(δ, x) ≤ V(δ, x0 )} is an invariant set to the system under
some arbitrarily small  ∈ (0, δ). For this, we need to establish the control law
some additional notation, and, given an edge (vi , vj ) ∈ E(G),
2δ − ij (x)
we let ij (x) denote the edge vector between the agents i and j, ẋi = − (xi − xj ). (17)
i.e., ij (x) = xi − xj . (δ − ij (x) )2
j∈NG (i)
JI AND EGERSTEDT: DISTRIBUTED COORDINATION CONTROL 697

Proof: We first note that the control law in (17) can be rewrit- that is
ten as M (δ − )2 2

∂Vij (δ, x) ∂V(δ, x) =
ẋi = − =− = −∇xi V(δ, x).  δ − 
∂xi ∂xi
j∈NG (i) which implies that
This expression may be ill-defined, since it is conceivable that 
ˆ ≤ δ − < δ.
the edge-lengths approach δ and what will be shown is that this M
will not happen. In fact, assume that at time τ we have x(τ ) ∈ Hence,  is bounded away from above from δ, and it is moreover
DG,δ

for some  > 0. Then, the time derivative of V(δ, x(τ )) is bounded from above by a strictly decreasing function in  on
(0, δ). Hence, as V decreases (or at least is nonincreasing), no
V̇(δ, x(τ )) = ∇x V(δ, x(τ ))T ẋ(τ ) edge-distances will tend to δ, which completes the proof.


N The invariance of Ω(δ, x0 ) now leads us to the main SIG
=− ẋi (τ )T ẋi (τ ) theorem.
i=1 Theorem 3.2: Given a connected SIG G with initial condition


n x0 ∈ DG,δ
, for a given  > 0. Then, the the multiagent system
=− c(x(τ ), j)T LW (δ, x(τ ))2 c(x(τ ), j) (18) under the control law in (17) asymptotically converges to the
j=1 static centroid x̄(x0 ).
Proof: The proof of convergence is based on LaSalle’s in-
where LW (δ, x) is given in (13), with weight positive definite
variance theorem. Let DG,δ 
and Ω(δ, x0 ) be defined as before.
(on Ω(δ, x0 )) matrix W(δ, x)
From Lemma 3.1 , we know that Ω(δ, x0 ) is positively invari-
W(δ, x) = diag(wk (δ, x)), k = 1, 2, . . . , M ant with respect to the dynamics in (17). We also note that
span{1} is LW (δ, x)-invariant for all x ∈ Ω(δ, x0 ). Hence, due
2δ − k (x)
wk (δ, x) = (19) to the fact that V̇(δ, x) ≤ 0, with equality only when c(x(t), j) ∈
(δ − k (x) )2
span{1}, ∀j ∈ {1, . . . , n}, convergence to span{1} follows.
where we have arranged the edges such that subscript k cor- Next, we need to show that the agents converge to the centroid.
responds to edge k. We will use this notation interchangeably The centroid is given by
with wij and ij , whenever it is clear from the context.
1

N
We note that for any  bounded away from 0 from below x̄ = xi
and δ from above, and for any x ∈ DG,δ 
, the time derivative N i=1
of the total tension energy is well defined. Moreover, for any
such x, V(δ, x) is nonnegative and V̇(δ, x) is nonpositive (since and the component-wise dynamics of the centroid is
LW (δ, x) is positive semidefinite for all x ∈ Ω(δ, x0 )). Hence, d 1 d 1
in order to establish the invariance of Ω(δ, x0 ), all that needs to c(x, j) = 1T c(x, j) = − 1T LW (δ, x)c(x, j).
dt N dt N
be shown is that as V decreases (or at lest does not increase),
no edge-distances will tend to δ. In fact, since DG,δ 
⊂ DG,δ

if Now, since 1T LW (δ, x) = (LW (δ, x)1)T = 0, ∀x ∈ Ω(x0 ),
 >  , we would have established the invariance of Ω(δ, x0 ) if we directly have x̄˙ = 0, i.e., the centroid is static,
we could find an  > 0 such that whenever the system starts determined entirely by the initial condition x0 . As such, we
from x0 ∈ DG,δ
, we can ensure that it never leaves the superset can denote the centroid by x̄(x0 ). This is in fact just a spe-
cial case of the observation that the centroid is static under any
DG,δ .

control law in (2).
Let
Now, let ξ¯ ∈ RN be any point on span{1} (i.e., ξ¯ =
V̂ := max

V(δ, x). (ξ, . . . , ξ)T for some ξ ∈ R) that is consistent with a static cen-
x∈DG, δ
troid. This implies that
This maximum always exists, and is obtained when all edges
1

N
are at the maximal allowed distance c(x, j) = ξ=ξ
N i=1
M (δ − )2
V̂ =
 and hence, ξ has to be equal to the centroid itself. As a conse-
which is a monotonously decreasing function in  over (0, δ). quence, if xi , i = 1, . . . , n, converged anywhere other than the
We can bound the maximal edge distance that can generate centroid, we would have a contradiction, and the proof follows.

this total tension energy, and the maximal edge-length ˆ ≥ It is to be noted that the construction we have described
δ −  is one where the entire total energy is contributed from corresponds to adding nonlinear state-dependent weights to the
that one single edge. In other words, all other edges have length edges in the graph. One could conceivably also add weights to
0, and the maximal edge length satisfies the nodes as well. Unless these weights were all equal, they
would violate the general assumption in (2), but for the sake of
ˆ2 completeness, we briefly discuss this situation in the next few
V̂ =
δ − ˆ paragraphs.
698 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 4, AUGUST 2007

A node weight would be encoded in the dynamics of the

system through the weight matrix D(x) as

dc(x, j)
= −D(x)LW (x)c(x, j), j = 1, . . . , n.
dt
As long as D(x) is diagonal and positive definite for all x,
(with the diagonal elements bounded away from 0), the null- Fig. 2. Hysteresis protocol for adding interagent tension functions to the total
space remains null(D(x)LW (x)) = span{1}, ∀x ∈ RnN , and tension function only when agents get within a distance ∆ −
of each other,
rather than when they first encouter each other at a distance ∆.
the controller still drives the system to span{1}. However, it
is straightforward to show that in this case, the positions xi ∈
Rn , i = 1, . . . , N approach the same static point x̄D (x0 ) ∈ Rn , In other words, for the ∆-disk proximity DIGs, we propose
given by to let
 −

σ(i, j)[t+ ] = 0 if σ(i, j)[t ] = 0 ∧ ij > ∆ − 
N
1
x̄D (x0 ) = (d−1
i (x0 ))x0,i (20) 1 otherwise
tr(D−1 (x0 )) i=1 
0 if σ(i, j) = 0
f (xi − xj ) = ∂V j (∆,x) (21)
where x0,i ∈ Rn , i = 1, . . . , N is the initial location of agent i, − i ∂x i
otherwise
di (x) is the ith diagonal element of D(x), and tr(D(x)) denotes
the trace of matrix D(x). where we have used the notation σ(i, j)[t+ ] and σ(i, j)[t− ] to
That concludes this section where an SIG was assumed. We denote the value of σ(i, j) before and after the state transition
will show ion Section IV that a similar strategy can be employed in Fig. 2. It is worth noting that if σ(i, j)[t0 ] = 1 for some t0 ,
even if the graph is allowed to change as the agents move around then σ(i, j)[t0 ] = 1 for all t < t0 .
in the environment. Before we can state the rendezvous theorem for dynamic
graphs, we also need to introduce the subgraph Gσ ⊂ G, induced
by the indicator function σ
IV. DYNAMIC GRAPHS
As already pointed out, during a maneuver, the interaction Gσ = (V (G), E(Gσ ))
graph G may change as the different agents move in and out of
where
each others sensory ranges. In this section, we focus on whether
or not an argument, similar to the previous stability result, can E(Gσ ) = {(vi , vj ) ∈ E(G)| σ(i, j) = 1}.
be constructed for the case when (vi , vj ) ∈ E(G) if and only if
xi − xj ≤ ∆. Theorem 4.1: Given an initial position x0 ∈ Dg 0 ,∆ , where
In fact, we intend to reuse the tension energy from the previous  > 0 is the switching threshold in (21), and G 0 is the initial ∆-
section, with the particular choice of δ = ∆. However, since in disk DIG. Assume that the graph Gσ0 is connected, where Gσ0 is
(19) the graph induced by the initial indicator function value. Then,
by using the control law
lim wk (∆, k ) = ∞
k ↑∆
∂Vij (∆, x)
ui = − (22)
we can not directly let the interagent tension energy affect the ∂xi
j∈Nσ (i)
dynamics as soon as two agents form edges in between them,
i.e., as they move within distance ∆ of each other. The reason where σ(i, j) is given in (21), the group of agents asymptotically
for this is that we cannot allow infinite tension energies in the converges to span{1}.
definition of the control laws. To overcome this problem, we Proof: Since, from Lemma 3.1, we know that no edges in
chose to introduce a certain degree of hysteresis into the system Gσ0 will be lost, only two possibilities remain, i.e., that no new
through the indicator function σ. In particular, we let σ(i, j) be edges will be added to the graph during the maneuver, or new
given by the state machine in Fig. 2 edges will in fact be added. If no edges are added, then we know
To elaborate further on the state machine in Fig. 2, we let from Theorem 3.2 that the system will converge to span{1}
the total tension energy be affected by an edge (vi , vj ) that was asymptotically. However, the only graph consistent with x ∈
previously not contributing to the total energy when ij ≤ span{1} is Gσ0 = KN (the complete graph over N nodes), and
(∆ − ), where  > 0 is the predefined switching threshold. hence, no new edges will be added only if the initial, indicator-
Once the edge is allowed to contribute to the total tension en- induced graph is complete. If it is not complete, at least one
ergy, it will keep doing so for all subsequent times. We note new edge will be added. But, since Gσ0 is an arbitrary connected
that the switching threshold can take on any arbitrary value in graph, and connectivity can never be lost by adding new edges,
(0, ∆). The interpretation is simply that a smaller -value cor- we obtain that new edges will be added until the indicator-
responds to a faster inclusion of the inter-robot information into induced graph is complete, at which point the system converges
the decentralized control law. asymptotically to span{1}.

JI AND EGERSTEDT: DISTRIBUTED COORDINATION CONTROL 699

V. FORMATION CONTROL such that

In the previous sections, it was shown that the connectedness- (vi , vj ) ∈ Ed .
preserving control method solves the rendezvous problem. In
this section, we will follow the same methodology to solve the We can then define the displacement from τi at time t as
distributed formation control problem. yi (t) = xi (t) − τi .
We let ij (t) = xi (t) − xj (t) and λij (t) = yi (t) − yj (t), im-
A. Graph Formation
plying that
By formation control, we understand the problem of driving
the collection of mobile agents to some translationally invariant λij (t) = ij (t) − dij .
target geometry, i.e., the control objective is to drive the col- Now, under the assumption that Gd is a connected spanning
lection of autonomous mobile agents to a specific configuration graph of the initial interaction graph, G i.e., V (Gd ) = V (G) and
such that their relative positions satisfy some desired topologi- Ed ⊆ E(G), we propose the following control law:
cal and physical constraints. These constraints can be described
2(∆ − dij ) − ij − dij
by a connected edge-labeled graph Gd = (V, Ed , d), where the ẋi = − (xi − xj − dij ).
subscript d denotes “desired.” Here, Ed encodes the desired (∆ − dij − ij − dij )2
j∈NGd (i)
robot interconnections, i.e., whether or not a desired interagent (23)
distance is specified between two agents or not, and the edge- The reason why this seemingly odd choice makes sense is
labels d : Ed → Rn defines the desired relative interagent dis- because we can again use the edge-tension function V to describe
placements, with dij > ∆ for all i, j such that (vi , vj ) ∈ Ed . this control law. In particular, using the following parameters in
In other words, what we would like is that xi − xj → dij , ∀i, j the edge-tension function
such that (vi , vj ) ∈ Ed . 
λi j 2
Vij (∆ − dij , y) = ∆− di j − λi j , if (vi , vj ) ∈ Ed
One may notice that it is possible that the assignment of gen-
eral edge-labels to a DIG may result in conflicting constraints. 0, otherwise
This is addressed in [37] as the realization problem of connec-
tivity graphs. We will not discuss this problem here and simply (24)
assume that the constraints are compatible. Another issue con- we obtain the decentralized control law
cerning the target formation is that of rigidity, which has been ∂Vij (∆ − dij , y)
discussed in [10], [38], and [39], and will not be discussed fur- σ(i, j) = 1 f (xi − xj ) = − ,
∂yi
ther in this paper. Instead, we assume that the target formation
is chosen in such a way that rigidity is obtained if, in fact, this is ∀(vi , vj ) ∈ Ed . (25)
a desired characteristic of the target formation graph-topology. Note that this control law, in fact, implies something stronger
Given a desired formation, the goal of the distributed forma- than just measurements of displacement. Instead, the agents
tion control is to find a feedback law such that: must also share a common coordinate system. However, they do
F1 dynamic interaction graph G(t) converges to a graph not need to know their exact location in this coordinate system.
that is a supergraph of the desired graph Gd (without Now, along each individual dimension, the dynamics in (25)
labels) in finite time. In other words, what we want is becomes
that Ed ⊂ E(t) for all t ≥ T , for some finite T ≥ 0;
F2 ij (t) = xi (t) − xj (t) converges asymptoti- dc(x, j) dc(y, j)
=
cally to dij for all i, j such that (vi , vj ) ∈ Ed ; and dt dt
F3 feedback law utilizes only local information. = −LW (∆ − d , y)c(y, j), j = 1, 2, . . . n
Here, “F” stands for “formation” and it will be established
where LW (∆ − d , y) is the graph Laplacian associated with
that these properties hold for a particular choice of decentralized
Gd , weighted by W(∆ − d , y), and where we have used the
control law.
convention that the term ∆ − d should be interpreted as
W(∆ − d , y) = diag(wk (∆ − dk , y)), k ∈ {1, |Ed |}
B. Graph-Based Formation Control
2(∆ − dk ) − λk
Analogous to the treatment of the rendezvous problem, we wk (∆, − dk , y) = ). (26)
(∆ − dk − λk )2
first propose a solution to the formation control problem, and
then show that this solution does, in fact, preserve connectedness Here again, the index k runs over the edge set Ed . This con-
as well as guarantee convergence in the sense of F1 and F2 struction allows us to study the evolution of yi , rather than
above. The solution will be based on a variation of the previously xi , i = 1, . . . , N , and we formalize this in the following lemma
derived rendezvous controller. In fact, assume that we have for static interaction graphs.
established a set of arbitrary targets τi ∈ Rn that are consistent Corollary 5.1: Let the total tension energy function be
with the desired interagent displacement
1


N N
V(δ − d , y) = Vij (∆ − dij , y). (27)
dij = τi − τj , ∀i, j 2 i=1 j=1
700 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 4, AUGUST 2007

means that
ij = λij +dij ≤ λij + dij < ∆− dij + dij = ∆
and hence, edges in Ed are never lost under the control law
in (2.3). In other words, Lij (t) < ∆, ∀t ≥ 0, which in turn
implies that connectedness is preserved.

We have, thus, established that if Gd is a spanning graph
of G(x0 ), then it remains a spanning graph of G(x(t)), ∀t > 0
(under certain assumptions on x0 ), even if G(x(t)) is given by
a ∆-disk DIG. And, since the control law in (23) only takes
pairwise interactions in Ed into account, we can view this dy-
namic situation as a static situation, with the SIG being given
by Gd . It still remains to be shown that the system in fact con-
verges in the sense of the formation control properties F1, F2,
and F3, as previously defined. That F3 (decentralized control) is
satisfied follows trivially from the definition of the control law
Fig. 3. State machine describing how the system undergoes transitions from in (23). Moreover, we have already established that F1 (finite
rendezvous (collection of the agents to a tight complete graph) to formation
control. time convergence to the appropriate graph) holds trivially as
long as it holds initially, and what remains to be shown here
is that we can drive the system in finite time to a configuration
Given y0 ∈ DG d ,∆− d , with Gd being a connected spanning in which F1 holds, after which Lemma 5.2 applies. Moreover,
graph, then the set Ω(∆ − d , y0 ) := {y|V(δ − d , y) ≤ we need to establish that the inter-robot displacements (defined
V0 }, where V0 denotes the initial value of the total tension for edges in Ed ) converge asymptotically to the desired relative
energy function, is an invariant set under the control law in (23), displacements (F3), which is the topic of the next theorem.
under the assumption that the interaction graph is static. Theorem 5.3: Under the same assumptions as in Lemma
Proof: By the proposed control law in (23), 5.2 ij (t) = xi (t) − xj (t) converges asymptotically to

dij for all i, j such that (vi , vj ) ∈ Ed .
∂Vij (∆ − dij , y) Proof: Based on the observation that Gd remains a spanning
ẏi = −
∂yi graph to the DIG, together with the observation that
j∈NGd (i)

∂V(∆ − d , y) dc(y, j)
=− = −∇yi V(δ − d , y). = −LW (∆ − d , y)c(y, j), j = 1, 2, . . . n,
∂yi dt
Theorem 3.2 ensures that c(y, j) will converge to span{1}, ∀j ∈
The nonpositivity of V̇ now follows the same argument as in {1, . . . , n}. This implies that all displacements must be the
(18) in the proof of Lemma 3.1. Moreover, for each initial same, i.e., that yi = ζ, ∀i ∈ {1, . . . , N } for some constant
y0 ∈ DG d ,∆− d , the corresponding maximal total tension en- ζ ∈ Rn . But, this simply means that the system converges
ergy induces a maximal possible edge length. Following the asymptotically to a fixed translation away from the target points
same line of reasoning as in the proof of Lemma 3.1, the invari- τi , i = 1, . . . , N , as
ance of Ω(∆ − d , y0 ) thus follows.  
lim yi (t) = lim xi (t) − τi = ζ, i = 1, . . . , N
It is to be noted that Lemma 5.1 says that if we could use t→∞ t→∞
Gd as an SIG, Ω(∆ − d , y0 ) is an invariant set. In fact, it is which, in turn, implies that
straightforward to show that if Gd is a spanning graph to the  
initial proximity ∆-disk DIG, then it remains a spanning graph lim ij (t) = lim (xi (t) − xj (t)
t→∞ t→∞ ¯
to G(x(t))∀t ≥ 0.
Corollary 5.2: Given an initial condition x0 such that y0 = = lim (yi (t) + τi − yj (t) − τj )
t→∞
(x0 − τ0 ) ∈ DG d ,∆− d , with Gd being a connected spanning = ζ + τi − ζ − τj = dij ∀i, j
graph of G(x0 ), the group of autonomous mobile agents adopt-
ing the decentralized control law in (23) can guarantee that such that(vi , vj ) ∈ Ed , which completes the proof.

xi (t) − xj (t) = Lij (t) < ∆, ∀t > 0 and (vi , vj ) ∈ Ed .
Proof: Given two agents i, j that are adjacent in Gd , and C. Hybrid, Rendezvous-To-Formation Control Strategies
suppose that λij = yi − yj approaches ∆ − dij . Since The last property that must be established is that it is possible
Vij ≥ 0, ∀i, j and t > 0, as well as to satisfy F1, i.e., that the initial ∆-disk proximity DIG does,
in fact, converge to a graph that has Gd as a spanning graph
lim Vij = ∞ in finite time. If this was achieved, then Theorem 5.3 would
λi j ↑(∆− di j )
be applicable and F2 (asymptotic convergence to the correct
this would imply that V → ∞, which contradicts Lemma 5.1. interagent displacements) would follow. To achieve this, we
As a consequence, λij is bounded away from ∆ − dij . This propose to use the rendezvous control law developed in Section
JI AND EGERSTEDT: DISTRIBUTED COORDINATION CONTROL 701

Fig. 4. Progression where connectedness is maintained during the rendezvous maneuver, with D = I. Positions of the agents and the edges in the DIG as a
function of time are depicted.

V-B for gathering all agents into a complete graph, of which has N − 1 neighbors, i.e., degree N − 1, all of which are within
any desired graph is a subgraph. Moreover, we need to achieve a distance ε/2, this implies that the maximal separation between
this in such a manner that the assumptions in Theorem 5.3 are two of those neighbors is ε. (This occurs when the agents are
satisfied. polar opposites on an n-sphere of radius ε/2.) Hence, when one
Let KN denote the complete graph over N agents. Moreover, agent detects this condition, it will trigger a switching signal (in-
ε
we will use KN to denote the situation in which the ε-disk volving a 1-bit broadcast communication to all its neighbors),
proximity graph is in fact a complete graph, i.e., a DIG that is a and the transition in Fig. 3 occurs. It is to be noted that first,
complete graph in which no interagent distances are greater than this argument hinges on the fact that the total number of agents
ε. This notation is slightly incorrect in that graphs are inherently N is known to each agent. This could arguably be a concern
combinatorial objects, while interagent distances are geometric, for graphs with time-varying node sets. Second, transition in
and, to be more precise, we will use the notation G = KN ε
to Fig. 3 might actually not occur at the exact moment when G
ε
denote the fact that becomes KN , but rather at a later point. Regardless of the kind
 of transition, we know that this transition will, in fact, occur in
G = KN
finite time in such a way that the initial condition assumptions
ij ≤ ε, ∀(i, j), i = j.
of Theorem 5.3 are satisfied.
The reason for this construction is that, in order for Theo-
rem 5.3 to be applicable, the initial condition has to satisfy
y0 = (x0 − τ0 ) ∈ DG d ,∆− d , which is ensured by making ε VI. EXAMPLES
small enough. Moreover, since the rendezvous controller in (22) In this section, we will show some simulation results that
asymptotically achieves rendezvous, it will consequently drive illustrate the proposed coordination control strategies for differ-
ε
the system to KN in finite time, for all ε bounded below by 0 ent problems. In all of these simulations, the cutoff distance for
and above by ∆. interagent sensing and communication is ∆ = 4, and the switch-
ε
After KN is achieved, the controller switches to the controller ing threshold dictating when to add edges is  = 0.05, i.e., a new
in (23), as depicted in Fig. 3. However, this hybrid control strat- edge is added only when the corresponding interagent distance
egy is only viable if the condition that G = KN ε
is locally veri- is ∆ − .
fiable in the sense that the agents can decide for themselves that The first two simulations show the rendezvous behavior
a synchronous mode switch is triggered [35]. In fact, if an agent under slightly different control laws. In fact, Fig. 4 shows
702 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 4, AUGUST 2007

Fig. 5. Evolution of the formation process. The last graph shows the trajectory of the formation process from t = 0 s to t = 0.5 s, starting from 1, 2, 3, 4, 5 and
ending at 1’, 2’, 3’, 4’, 5’.

the movement of the collection of agents under the weighted [3] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile
Laplacian control law given in (22), under exactly the same autonomous agents using nearest neighbor rules,” IEEE Trans. Autom.
Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003.
initial position as in Fig. 1. What is different here is, as could [4] H. Ando, Y. Oasa, I. Suzuki, and M. Yamashita, “Distributed memo-
be expected, that no links are broken. ryless point convergence algorithm for mobile robots with limited vis-
The third simulation highlights the proposed formation con- ibility,” IEEE Trans. Robot. Autom., vol. 15, no. 5, pp. 818–828, Oct.
1999.
trol strategy, and is implemented based on the formation control [5] J. A. Fax and R. M. Murray, “Graph laplacian and stabilization of vehicle
law in (23). In the simulation, five agents starting from a straight formations,” in Proc. 15th IFAC, 2002, pp. 283–288.
line are to form a pentagonal formation, with Gd = C5 (the [6] J. Fax and R. Murray, “Information flow and cooperative control of vehicle
formations,” IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1465–1476,
cyclic graph with 5 nodes), and the desired interagent distances Sep. 2004.
being dij = 3.2 for all (vi , vj ) ∈ Ed . The movement of the [7] Z. Lin, M. Broucke, and B. Francis, “Local control strategies for groups of
group during the first 0.5 s and the trajectories corresponding to mobile autonomous agents,” IEEE Trans. Autom. Control, vol. 49, no. 4,
pp. 622–629, Apr 2004.
the same time period are shown in Fig. 5. [8] M. Mesbahi, “State-dependent graphs,” in Proc. 42nd IEEE Conf. Deci-
sion Control, Maui, HI, Dec. 2003, pp. 3058–3063.
VII. CONCLUSION [9] W. Ren and R. Beard, “Consensus of information under dynamically
changing interaction topologies,” in Proc. Am. Control Conf., vol. 6, Jun.
In this paper, a collection of graph-based nonlinear feedback 30–Jul. 2, 2004, pp. 4939–4944.
[10] R. Olfati-Saber and R. M. Murray, “Distributed structural stabilization
control laws are studied for distributed multiagent systems. The and tracking for formations of dynamic multi-agents,” Proc. 41st IEEE
nonlinear feedback laws are based on weighted graph Laplacians Conf. Decision Control, vol. 1, pp. 209–215, Dec. 2002.
and they are proved to be able to solve the rendezvous and [11] R. Olfati-Saber and R. M. Murray, “Flocking with obstacle avoidance:
Cooperation with limited communication in mobile networks,” in Proc.
formation-control problems while ensuring connectedness. 42nd IEEE Conf. Decision Control, vol. 2, Maui, HI, Dec. 2003, pp. 2022–
2028.
REFERENCES [12] R. Olfati-Saber and R. M. Murray, “Agreement problems in net-
works with directed graphs and switching toplogy,” in Proc. 42nd
[1] C. Reynolds, “Flocks, herds and schools: A distributed behavioral model,” IEEE Conf. Decision Control, vol. 4, Maui, HI, Dec. 2003, pp. 4126–
in Proc. ACM SIGGRAPH 1987 Conf., 1987, vol. 21, no. 4, pp. 25–34. 4132.
[2] T. Vicsek, A. Czirok, E. B. Jacob, I. Cohen, and O. Schochet, “Novel type [13] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
of phase transitions in a system of self-driven particles,” Phys. Rev. Lett., agents with switching topology and time-delays,” IEEE Trans. Autom.
vol. 75, pp. 1226–1229, 1995. Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004.
JI AND EGERSTEDT: DISTRIBUTED COORDINATION CONTROL 703

[14] J. Desai, J. Ostrowski, and V. Kumar, “Modeling and control of formations [34] A. Jadbabaie, N. Motee, and M. Barahona, “On the stability of the
of nonholonomic mobile robots,” IEEE Trans. Robot. Autom., vol. 17, Kuramoto model of coupled nonlinear oscillators,” in Proc. 2004 Am.
no. 6, pp. 905–908, Dec. 2001. Control Conf., 2004, pp. 4296–4301.
[15] H. Tanner, A. Jadbabaie, and G. Pappas, “Stable flocking of mobile agents, [35] J.-M. McNew and E. Klavins, “Locally interacting hybrid systems with
part II : Dynamic topology,” in Proc. 42nd IEEE Conf. Decision Control, embedded graph grammars,” in Proc. IEEE Conf. Decision Control, San
Maui, HI, Dec.2003, pp. 2016–2021. Diego, CA, Dec. 2006, pp. 6080–6087.
[16] J. Cortés, S. Martı́nez, and F. Bullo, “Robust rendezvous for mobile au- [36] C. Godsil and G. Royle, Algebraic Graph Theory. Berlin, Germany:
tonomous agents via proximity graphs in d dimension,” IEEE Trans. Springer, 2001.
Robot. Autom., vol. 51, no. 8, pp. 1289–1298, Aug. 2006. [37] A. Muhammad and M. Egerstedt, “On the structural complexity of multi-
[17] J. Desai, J. Ostrowski, and V. Kumar, “Controlling formations of multiple agent robot formations,” in Proc. Am. Control Conf. 2004, Boston, MA,
mobile robots,” in Proc. IEEE Int. Conf. Robot. Autom., Leuven, Belgium, pp. 4957–4962.
1998, pp. 2864–2869. [38] R. W. Beard, J. R. Lawton, and F. Y. Hadaegh, “A coordination architecture
[18] M. Egerstedt, X. Hu, and A. Stotsky, “Control of mobile platforms using for spacecraft formation control,” IEEE Trans. Control Syst. Technol.,
a virtual vehicle approach,” IEEE Trans. Autom. Control, vol. 46, no. 46, vol. 9, no. 6, pp. 777–790, Nov. 2001.
pp. 1777–1782, Nov. 2001. [39] T. Eren, P. Belhumeur, B. Anderson, and A. Morse, “A framework for
[19] M. Egerstedt and X. Hu, “Formation constrained multi-agent control,” maintaining formations based on rigidity,” in Proc. 2002 IFAC Congr.,
IEEE Trans. Robot. Autom., vol. 17, no. 6, pp. 947–951, Dec. 2001. pp. 2752–2757.
[20] P. Ögren, M. Egerstedt, and X. Hu, “A control Lyapunov function approach
to multi-agent coordination,” IEEE Trans. Robot. Autom., vol. 18, no. 5,
pp. 847–851, Oct. 2002. Meng Ji (S’06) was born in Tianjin, China. He re-
[21] N. E. Leonard and E. Fiorelli, “Virtual leaders, artificial potentials and ceived the B.S. and M.E. degrees in mechanical engi-
coordinated control of groups,” in Proc. IEEE Conf. Decision Control, neering from Tsinghua University, Beijing, China, in
Orlando, FL, Dec.2001, pp. 2968–2973. 1998 and 2001, respectively. He is currently working
[22] T. Balch and R. C. Arkin, “Behavior-based formation control for multi- toward the Ph.D. degree at the School of Electri-
robot teams,” IEEE Trans. Robot. Autom., vol. 14, no. 6, pp. 926–939, cal and Computer Engineering, Georgia Institute of
Dec. 1998. Technology, Atlanta.
[23] J. Lawton, R. Beard, and B. Young, “A decentralized approach to forma- His research interests include cooperative control
tion maneuvers,” IEEE Trans. Robot. Autom., vol. 19, no. 6, pp. 933–941, of networked systems, distributed sensor networks,
Dec. 2003. multiagent systems, and fault adaptive control.
[24] H. G. Tanner and A. Kumar, “Robotics: Science and systems,”
in Formation Stabilization of Multiple Agents Using Decentralized Navi-
gation Functions. vol. I, Cambridge, MA: MIT Press, 2005, pp. 49–56.
[25] M. C. DeGennaro and A. Jadbabaie, “Formation control for a cooperative Magnus B. Egerstedt (S’99–M’00–SM’05) was
multiagent system with a decentralized navigation function,” in Proc. Am. born in Stockholm, Sweden. He received the B.A.
Control Conf., Minneapolis, MN, 2006, pp. 1346–1351. degree in philosophy from Stockholm University,
[26] J. Lin, A. Morse, and B. Anderson, “The multi-agent rendezvous prob- Stockholm, in 1996, and the M.S. degree in engineer-
lem,” in Proc. 42nd IEEE Conf. Decision Control, Maui, HI, Dec. 2003, ing physics and Ph.D. degree in applied mathematics
pp. 1508–1513. from the Royal Institute of Technology, Stockholm,
[27] L. Moreau, “Stability of multiagent systems with time-dependent commu- in 1996 and 2000, respectively.
nication links,” IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 169–182, From 2000 to 2001, he was a Postdoctoral Fellow
Feb. 2005. at the Division of Engineering and Applied Science,
[28] K. Sugihara and I. Suzuki, “Distributed motion coordination of multiple Harvard University, Cambridge, MA. He is currently
robots,” in Proc. IEEE Int. Symp. Intell. Control, 1990, pp. 138–143. an Associate Professor with the School of Electrical
[29] G. Notarstefano, K. Savla, F. Bullo, and A. Jadbabaie, “Maintaining and Computer Engineering, the Georgia Institute of Technology, Atlanta, where
limited-range connectivity among second-order agents,” in Proc. 2004 he has been on the faculty since 2001. He also holds an adjunct appointment with
Am. Control Conf., Minneapolis, MN, Jun. 2006, pp. 2124–2129. the Division of Interactive and Intelligent Computing, College of Computing at
[30] A. Bensoussan and J. Menaldi, “Difference equations on weighted Georgia Institute of Technology. He is the Chair of the Systems and Controls
graphs,” J. Convex Anal., vol. 12, no. 1, pp. 13–44, 2005. Technical Interest Group at the School of Electrical and Computer Engineer-
[31] G. Ferrari-Trecate, A. Buffa, and M. Gati, “Analysis of coordination in ing and the Associate Director of Research with the Center for Robotics and
multi-agent systems through partial difference equations,” IEEE Trans. Intelligent Machines, Georgia Institute of Technology. His research interests
Autom. Control, vol. 51, no. 6, pp. 1058–1063, Jun. 2006. include optimal control as well as modeling and analysis of hybrid and discrete
[32] G. Ferrari-Trecate, M. Egerstedt, A. Buffa, and M. Ji, “Laplacian sheep: A event systems, with emphasis on motion planning, control, and coordination of
hybrid, stop-go policy for leader-based containment control,” in Proc. 9th mobile robots. He has authored over 100 papers published in several journals
Int. Workshop Hybrid Syst.: Comput. Control, Santa Barbara, CA, Mar. and conference proceedings.
2006, pp. 212–226. Dr. Egerstedt received the ECE/GT Outstanding Junior Faculty Member
[33] M. Mesbahi, “On a dynamic extension of the theory of graphs,” in Proc. Award in 2005, and the CAREER Award from the U.S. National Science Foun-
2002 Am. Control Conf., 2002, vol. 2, pp. 1234–1239. dation in 2003.