Proceedings of the 45th IEEE Conference on Decision & Control FrIP3.

3
Manchester Grand Hyatt Hotel
San Diego, CA, USA, December 13-15, 2006

Distributed Formation Control While Preserving Connectedness∗

Meng Ji† and Magnus Egerstedt†

Abstract— This paper addresses the connectedness issue in II. BACKGROUND

multi-agent coordination, i.e. the problem of ensuring that
the group stays connected while achieving some performance The problem that we are investigating in this paper is how
objective. In particular, we continue our previous work on to implement formation control in a distributive way while
rendezvous control and extend it to the formation control prob- preserving connectedness with only limited information.
lem over dynamic interaction graphs. By adding appropriate Given N agents, whose positions x1 , . . . , xN take on values
weights to the edges in the graphs, we guarantee that the graphs
stay connected, while achieving the desired formations. in Rn , we assume that the dynamics of each individual agent
Index Terms— Multi-Agent Coordination; Formation Con- is given by a single integrator
trol, Graph Laplacian, Connected Graphs.
ẋi = ui , i = 1, . . . , N. (1)
I. I NTRODUCTION We, moreover, associate an interaction graph with the
As a fundamental problem in the distributed multi-agent available information flow in such a way that the nodes
coordination area, formation control has received consider- correspond to agents, and edges to available, inter-agent com-
able attentions. Leader based formation control has been munication links. Such interaction graphs are thus reflective
studied for a long time, where either a real agent [2] or a of the underlying network topology. It is worth of noticing
virtual agent [3], [4], [15] is chosen as the leader. Recently, that such graphs are of dynamic nature in that the existence
as graph theory has been successfully used for solving of an edge between two nodes is affected by their relative
the rendezvous (or agreement) problem [18], [7], [9], [12], positions.
[16], [14], [1], a variety of graph-based formation control By a Static Interaction Graph (SIG) G = (V, E), we
strategies have been proposed [5] and the references therein, understand the graph where the nodes V = {v1 , . . . , vN }
most of which are leaderless. are associated to the different agents and the static edge set
A challenge faced in most multi-agent applications is that E ⊂ V × V is a set of unordered pairs of agents, with
the agents are subjected to limited sensing and commu- (vi , vj ) = (vj , vi ) ∈ E if and only if a communication link
nication capabilities. Therefore, the underlying information exists between agents i and j. We will use the shorthand
network, consisting of sensing and communication links V (G) and E(G) to denote the edge and node sets associated
among the agents, might become disconnected, i.e. broken with a graph G.
up into non-interacting sub-groups, if the control law does Given an agent i, we will associate NG (i) = {j | (vi , vj ) ∈
not take this limitation into account. For example, flocking E(G)} with the neighborhood set to i, i.e. the set of agents
under switching topologies was studied in [14], [19] and the adjacent to agent i. Using this terminology, what we under-
references therein, where artificial potential functions or non- stand by a limited information, time-invariant, decentralized


smooth Lyapunov functions were constructed over the graph control law is that ui = j∈Nσ (i) f (xi −xj ), where Nσ (i) ⊆
structure. In [11], [12], state-dependent dynamic graphs were NG (i). The symmetric indicator function σ(i, j) = σ(j, i) ∈
studied from a combinatoric point-of-view. A measure of {0, 1} determines whether or not the information available
local connectedness of a network was introduced in [17], through edge (vi , vj ) should be taken into account, with
depending entirely on the local interactions. In [20], connec- j ∈ Nσ (i) ⇔ (vi , vj ) ∈ E(G) ∧ σ(i, j) = 1. (Using the
tivity constraints were related to individual agent’s motion by terminology in [10], just because two nodes are ”neighbors”
the construction of a dynamically changing adjacency matrix. it doesn’t follow that they are ”friends”.) Along the same
In our previous work [8], the connectedness problem lines, the decentralized control law f (xi − xj ) is assumed to
was investigated in the context of the rendezvous prob- be anti-symmetric, i.e.
lem. In particular, connectedness was preserved by applying
nonlinear weights on edges, which resulted in a weighted f (xi − xj ) = −f (xj − xi ), ∀ (vi , vj ) ∈ E(G). (2)
graph Laplacian. Based on a variation of these results, this
The type of control terms presented above have appeared
paper extends the weighted graph Laplacian technique to the
repeatedly in the multi-agent coordination community, and
formation control problem.
an intuitive, linear control law for solving the rendezvous
* This work was supported by the U.S. Army Research Office through
the grant # 99838. problem is given by
† Authors are with School of Electrical and Computer 
σ(i, j) = 1
Engineering, Georgia Institute of Technology, Atlanta, GA 30332, ∀ (vi , vj ) ∈ E(G),
{mengji,magnus}@ece.gatech.edu f (xi − xj ) = −(xi − xj )

1-4244-0171-2/06/$20.00 ©2006 IEEE. 5962

45th IEEE CDC, San Diego, USA, Dec. 13-15, 2006 FrIP3.3

which gives that This well-known result is very promising since dynamic
 network graphs are frequently occurring in that all real
ẋi = − (xi − xj ), i = 1, . . . , N. (3) sensors and transmitters have finite range. This means that
j∈NG (i) information exchange links may appear or be lost as the
Under the dynamics in Equation (3), it has been shown that agents move around. In fact, if we focus our attention on ∆-
all agents approach the same point asymptotically, provided disk proximity graphs, where edges are established between
that the SIG is connected. And, even though this is a well- nodes vi and vj if and only if the agents are within distance
established result (see for example [9]), we will here outline ∆ of each other, i.e. when xi − xj ≤ ∆, we get the
a proof in order to establish some needed notation and tools. Dynamic Interaction Graph (DIG) G(t) = (V, E(t)), where
Now, if the total number of edges is equal to M , and (vi , vj ) = (vj , vi ) ∈ E(t) if and only if xi (t)−xj (t) ≤ ∆.
we associate an index with each edge such that E(G) = The success of the control in Equation (3) hinges on
{e1 , . . . , eM }, then the N × M incidence matrix of G o is an assumption that it shares with most graph-based results,
I(G o ) = [ιij ], where e.g. [7], [19], namely on the connectedness assumption.
⎧ Unfortunately, this property has to be assumed rather than
⎨ 1 if vi is the head of ej proved.
ιij = −1 if vi is the tail of ej (4) What we will do for the remainder of this paper is to
⎩
0 otherwise. show how this assumption can be overcome by modifying the
Through this incidence matrix, we can now define the graph control law in Equation (3) in such a way that connectedness
holds for all times, while ensuring that the control laws are
Laplacian L(G) ∈ RN ×N as
still based solely on local information.
L(G) = I(G o )I(G o )T , (5) III. C ONNECTEDNESS P RESERVING R ENDEZVOUS
where we have removed the orientation dependence in the A. Static Graph
left hand side of Equation (5). The reason for this is that First, we will study the behavior of multi-agent system
the Laplacian does not depend on the particular choice of with a fixed network topology. In other words, the interaction
orientation. graph will be of the SIG type, and we will show how
The graph Laplacian has a number of well-studied proper- the introduction of nonlinear edge-weights can be used to
ties, found for example in [6], and we here list the properties establish certain invariance properties.
of importance to the developments in this paper: To arrive at the desired invariance properties, we will first



1) I(G o )I(G o )T = I(G o )I(G o )T for all orientation investigate decentralized control laws of the form
o, o , i.e. the Laplacian is orientation-independent; σ(i, j) = 1
2) L(G) is symmetric and positive semidefinite; ∀(vi , vj ) ∈ E(G),
f (xi − xj ) = −w(xi − xj )(xi − xj )
3) The number of zero eigenvalues of L(G) equals to the (7)
number of connected components in G; where w : Rn → R+ is a positive, symmetric weight func-
4) If G is connected then null(G) = span{1}, where tion that associates a strictly positive and bounded weight to
null(·) denotes the null-space. each edge in the SIG.
If we now let the n-dimensional position of agent i be This choice of decentralized control law gives
given by xi = (xi,1 , . . . , xi,n ), i = 1, . . . , N , and let x = 
ẋi = − w(xi − xj )(xi − xj ), (8)
(xT1 , . . . , xTN )T , we can define the componentwise operator
j∈NG (i)
as c(x, j) = (x1,j , . . . , xN,j )T ∈ RN , j = 1, . . . , n. Using
this notation, together with the observation that Equation which can be rewritten as
(3) can be decoupled along each dimension, we can in fact d
c(x, j) = −I o W(x)I oT c(x, j), j = 1, . . . , n, (9)
rewrite Equation (3) as dt
d where W(x) =diag(w1 (x), . . . , wM (x)) ∈ RM×M , where,
c(x, j) = −L(G)c(x, j), j = 1, . . . , n. (6) as before M = |E(G)|) is the total number of edges, and
dt
where we have associated an identity (1, . . . , M ) to each of
Now, as pointed out in [9] and [6], if G is connected then
the edges.
the eigenvector corresponding to the semi-simple eigenvalue
We can then define the state-dependent, weighted graph
0 is 1. This, together with the non-negativity of L(G) and
Laplacian as
the fact that span{1} is L(G)-invariant, is sufficient to show
LW (x) = I o W(x)I oT , (10)
that c(x, j) approaches span{1} asymptotically.
This result, elegant in its simplicity, can in fact be extended where, as before, W(x) ∈ RM×M is a diagonal matrix with
to dynamic graphs as well. In fact, since c(x, j)T c(x, j) is each element corresponding to a strictly positive edge weight.
a Lyapunov function to the system in (3), for any connected It is moreover straightforward to establish that as long as
graph G, the control law in Equation (6) drives the system the graph is connected, LW (x) is still positive semidefinite,
to span{1} asymptotically as long as G(t) is connected for with only one zero eigenvalue corresponding to the null-
all t ≥ 0. space span{1}.

5963
45th IEEE CDC, San Diego, USA, Dec. 13-15, 2006 FrIP3.3

What we would like to show is that, given a critical where we have arranged the edges such that subscript k cor-
distance δ together with the appropriate edge-weights, the responds to edge k. We will use this notation interchangeably
edge-lengths never goes beyond δ if they start out being less with wij and ij , whenever it is clear from the context.
than δ − , for some arbitrarily small  ∈ (0, δ). For this, Note that for any  bounded away from 0 from below and


we need to establish some additional notation, and, given an δ from above, and for any x ∈ DG,δ , the time derivative of
edge (vi , vj ) ∈ E(G), we let ij (x) denote the edge vector the total tension energy is well-defined. Moreover, for any
between the agents i and j, i.e. ij (x) = xi − xj . such x, V(δ, x) is non-negative and V̇(δ, x) is non-positive
We moreover define the -interior of a δ-constrained (since LW (δ, x) is positive semidefinite for all x ∈ Ω(δ, x0 ).
realization of a SIG, G, as Hence, in order to establish the invariance of Ω(δ, x0 ), all
that needs to be shown is that, as V decreases (or at lest
DG,δ

= {x ∈ RnN | ij ≤ δ −  ∀(vi , vj ) ∈ E(G)}. does not increase), no edge-distances will tend to δ. In fact,


An edge-tension function Vij , can then be defined as since DG,δ

⊂ DG,δ if  >  , we would have established the
 invariance of Ω(δ, x0 ) if we could find an  > 0 such that,
ij (x)2
if (vi , vj ) ∈ E(G) whenever the system starts from x0 ∈ DG,δ 
, we can ensure
Vij (δ, x) = δ−ij (x) (11) 

0 otherwise. that it never leaves the superset DG,δ .
Note that this edge-tension function (as well as its deriva- Let V̂ := max 
V(δ, x). This maximum always exists
x∈DG,δ
tives) is infinite when ij (x) = δ for some i, j, and, as and is obtained when all edges are at the maximal allowed
2
such, it may seem like an odd choice. However, as we will distance δ − , i.e. V̂ = M(δ−)  , which is a monotonously
see, we will actually be able to prevent the energy to reach decreasing function in  over (0, δ).
infinity, and instead we will study its behavior on a compact What we will show next is that we can bound the maximal
set on which it is continuously differentiable. edge distance that can generate this total tension energy, and
The total tension energy of G can now be defined as the maximal edge-length ˆ ≥ δ −  is one where the entire
total energy is contributed from that one single edge. In other
1 
N N
V(δ, x) = Vij (δ, x). (12) words, all other edges have length 0, and the maximal edge
2 i=1 j=1 ˆ2 2
2
length satisfies V̂ = δ−ˆ , i.e. M(δ−) = δ− 
, which

ˆ
implies that  ≤ δ − M < δ. Hence  is bounded away

Lemma 3.1: Given an initial position x0 ∈ DG,δ 
, for a given
from above from δ and it is moreover bounded from above
 ∈ (0, δ). If the SIG G is connected then the set Ω(δ, x0 ) :=
by a strictly decreasing function in  on (0, δ). Hence, as V
{x | V(δ, x) ≤ V(δ, x0 } is an invariant set to the system
decreases (or at least is non-increasing), no edge-distances
under the control law
will tend to δ, which completes the proof.
 2δ − ij (x)
ẋi = − (xi − xj ). The invariance of Ω(δ, x0 ) now leads us to the main SIG
(δ − ij (x) )2 (13)
j∈NG (i) theorem.
Proof: We first note that the control law in Equation (13) Theorem 3.2: Given a connected SIG G with initial con-
can be rewritten as dition x0 ∈ DG,δ

, for a given  > 0. Then the multi-agent
 ∂Vij (δ, x) ∂V(δ, x) system under the control law in Equation (13) asymptotically
ẋi = − =− = −∇xi V(δ, x).
∂xi ∂xi converges to a common point.
j∈NG (i)
Proof: The proof of convergence is based on LaSalle’s
This expression may be ill-defined since it is conceivable invariance theorem. Let DG,δ
and Ω(δ, x0 ) be defined as
that the edge-lengths approach δ and what will be shown is before. From Lemma 3.1, we know that Ω(δ, x0 ) is positive
that this will not happen. In fact, assume that at time τ we invariant with respect to the dynamic in Equation (13).


have x(τ ) ∈ DG,δ for some  > 0. Then the time derivative We also note that span{1} is LW (δ, x)-invariant for all
of V(δ, x(τ )) is x ∈ Ω(δ, x0 ). Hence, due to the fact that V̇(δ, x) ≤ 0, with

N equality only when c(x(t), j) ∈ span{1}, ∀j ∈ {1, . . . , n},
V̇(δ, x(τ )) = ∇x V(δ, x(τ ))T ẋ(τ ) = − ẋi (τ )T ẋi (τ ) convergence to span{1} follows.
i=1
B. Dynamic Graphs

n
2
=− c(x(τ ), j) LW (δ, x(τ )) c(x(τ ), j),
T
As already pointed out, during a maneuver, the interaction
j=1 graph G may change as the different agents move in and
(14) out of each others sensory ranges. What we focus on in
this section is whether or not an argument, similar to the
where LW (δ, x) is given in Equation (10), with weight previous stability result, can be constructed for the case when
positive definite (on Ω(δ, x0 )) matrix W(δ, x) (vi , vj ) ∈ E(G) if and only if xi − xj ≤ ∆.
W(δ, x) = diag(wk (δ, x)), k = 1, 2, . . . , M, In fact, we intend to reuse the tension energy from the
previous section, with the particular choice of δ = ∆.
2δ − k (x) (15)
wk (δ, x) = , However, since in Equation (15) lim wk (∆, k ) = ∞,
(δ − k (x) )2 k ↑∆

5964
45th IEEE CDC, San Diego, USA, Dec. 13-15, 2006 FrIP3.3

we can not directly let the inter-agent tension energy affect IV. F ORMATION C ONTROL
the dynamics as soon as two agents form edges in between
them, i.e. as they move within distance ∆ of each other. In the previous section, the connectedness-preserving con-
The reason for this is that we can not allow infinite tension trol method solves the rendezvous problem. In what follows,
energies in the definition of the control laws. To overcome we will follow the same methodology to solve the distributed
this problem, we chose to introduce a certain degree of formation control problem.
hysteresis into the system, through the indicator function σ.
In particular, let the total tension energy be affected by an A. Graph Formation
edge (vi , vj ) that was previously not contributing to the total By formation control, we understand the problem of driv-
energy, when ij ≤ (∆ − ), where  > 0 is the predefined ing the collection of mobile agents to some translationally
switching threshold. Once the edge is allowed to contribute invariant target geometry, such that their relative positions
to the total tension energy, it will keep doing so for all satisfy some desired topological and physical constraints.
subsequent times. Note that the switching threshold can take These constraints can be described by a connected, edge-
on any arbitrary value in (0, ∆). The interpretation is simply labelled graph Gd = (V, Ed , d), where the subscript d
that a smaller -value corresponds to a faster inclusion of the denotes “desired”. Here, Ed encodes the desired robot inter-
inter-robot information into the decentralized control law. In connections, i.e. whether or not a desired inter-agent distance
other words, what we propose for the ∆-disk proximity DIGs is specified between two agents or not, and the edge-
is thus to let labels d : Ed → Rn defines the desired relative inter-
 agent displacements, with dij < ∆ for all i, j such that
0 if σ(i, j)[t− ] = 0 ∧ ij > ∆ −  (vi , vj ) ∈ Ed . In other words, what we would like is that
σ(i, j)[t+ ] =
 1 otherwise xi − xj → dij ∀i, j such that (vi , vj ) ∈ Ed .
0 if σ(i, j) = 0 One may notice that it is possible that the assignment
f (xi − xj ) = ∂V (∆,x)
− ij∂xi otherwise, of general edge-labels to a DIG may result in conflicting
(16) constraints. This is addressed in [13] as the realization
where we have used the notation σ(i, j)[t+ ] and σ(i, j)[t− ] problem of connectivity graphs. We will not discuss this
o denote σ(i, j)’s value before and after the state transition. problem here and simply assume that the constraints are
Before we can state the rendezvous theorem for dynamic compatible.
graphs, we also need o introduce the subgraph Gσ ⊂ G, Given a desired formation, the goal of the distributed
induced by the indicator function σ: Gσ = (V (G), E(Gσ )), formation control is to find a feedback law such that:
where E(Gσ ) = {(vi , vj ) ∈ E(G) | σ(i, j) = 1}. F1) The dynamic interaction graph G(t) converges to a
graph that i s a supergraph of the desired graph Gd (without
Theorem 3.3: Given an initial position x0 ∈ DG 0 ,∆ , where labels) in finite time. In other words, what we want is that
 > 0 is the switching threshold in Equation (16), and where Ed ⊂ E(t) for all t ≥ T , for some finite T ≥ 0;
G 0 is the initial ∆-disk DIG. Assume that the graph Gσ0 F2) ij (t) = xi (t) − xj (t) converges asymptotically
is connected, where Gσ0 is the graph induced by the initial to dij for all i, j such that (vi , vj ) ∈ Ed ; and
indicator function value. Then, by using the control law F3) The feedback law utilizes only local information.
Here “F” stands for “formation” and what will be estab-
 ∂Vij (∆, x)
ui = − , (17) lished is in fact that these properties hold for a particular
∂xi choice of decentralized control law.
j∈Nσ (i)

where σ(i, j) is given in Equation (16), the group of agents B. Graph-Based Formation Control
asymptotically converges to span{1}. Analogous to the treatment of the rendezvous problem, we
Proof: Since, from Lemma 3.1, we know that no edges in first propose a solution to the formation control problem, and
Gσ0 ill be lost, only two possibilities remain, namely that no then show that this solution does in fact preserve connect-
new edges will be added to the graph during the maneuver, or edness as well as guarantee convergence in the sense of F1
new edges will in fact be added. If no edges are added, then and F2 above. The solution will be based on a variation of
we know from Theorem 3.2 that the system will converge to the previously derived rendezvous controller. In fact, assume
span{1} asymptotically. However, the only graph consistent that we have established a set of arbitrary targets τi ∈ Rn
with x ∈ span{1} is Gσ0 = KN (the complete graph over that are consistent with the desired inter-agent displacement,
N nodes), and hence no new edges will be added only if i.e.dij = τi −τj , ∀ i, j s.t. (vi , vj ) ∈ Ed . We can then define
the initial, indicator induced graph is complete. If it is not the displacement from τi at time t as yi (t) = xi (t) − τi . As
complete, at least one new edge will be added. But, since Gσ0 before, we let ij (t) = xi (t) − xj (t) and we moreover let
is an arbitrary connected graph, and connectivity can never λij (t) = yi (t) − yj (t), implying that λij (t) = ij (t) − dij .
be lost by adding new edges, we get that new edges will be Now, under the assumption that Gd is a connected span-
added until the indicator induced graph is complete, at which ning graph of the initial interaction graph G, i.e. V (Gd ) =
point the system converges asymptotically to span{1}. V (G) and Ed ⊆ E(G), we propose the following control

5965
45th IEEE CDC, San Diego, USA, Dec. 13-15, 2006 FrIP3.3

law: Note that Lemma 4.1 says that if we could use Gd as

 2(∆ − dij ) − ij − dij a SIG, Ω(∆ − d , y0 ) is an invariant set. In fact, it is
ẋi = − (xi − xj − dij ). straightforward to show that if Gd is a spanning graph to
(∆ − dij − ij − dij )2
j∈NGd (i) the initial proximity ∆-disk DIG, then it remains a spanning
(18) graph to G(x(t)) ∀t ≥ 0.
The reason why this seemingly odd choice makes sense
is because we can again use the edge-tension function V to Lemma 4.2: Given an initial condition x0 such that y0 =
describe this control law. In particular, using the following (x0 − τ0 ) ∈ DG d ,∆−d , with Gd being a connected spanning
parameters in the edge-tension function graph of G(x0 ), the group of autonomous mobile agents
 adopting the decentralized control law in Equation (18) can
λij 2
if (vi , vj ) ∈ Ed guarantee that xi (t) − xj (t) = lij (t) < ∆, ∀t >
Vij (δ − dij , y) = ∆−d ij −λij 
0 otherwise, 0 and (vi , vj ) ∈ Ed .
(19) Proof: Given two agents i, j that are adjacent in Gd , and
we obtain the decentralized control law suppose that λij = yi − yj approaches ∆ − dij . Since
 Vij ≥ 0, ∀i, j and t > 0, as well as lim Vij = ∞,
σ(i, j) = 1 λij ↑(∆−dij )
∂V (∆−d ,y) ∀ (vi , vj ) ∈ Ed . this would imply that V → ∞, which contradicts Lemma 4.1.
f (xi − xj ) = − ij ∂yi ij
As a consequence, λij is bounded away from ∆ − dij .
As such, along each individual dimension, the dynamics This means that
becomes dc(x,j)
dt = dc(y,j)
dt = −LW (∆ − d , y)c(y, j), ij = λij +dij ≤ λij + dij < ∆− dij + dij = ∆,
where j = 1, 2, . . . n and LW (∆ − d , y) is the graph
Laplacian associated with Gd , weighted by W (∆ − d , y), and hence edges in Ed are never lost under the control law
and where we have used the convention that the term ∆− d in Equation (18). In other words, lij (t) < ∆, ∀t ≥ 0,
should be interpreted in the following manner: which in turn implies that connectedness is preserved.
We have thus established that if Gd is a spanning graph of
W (∆ − d , y) = diag(wk (∆ − dk , y)), G(x0 ) then it remains a spanning graph of G(x(t)), ∀t > 0
2(∆ − dk ) − λk (20) (under certain assumptions on x0 ), even if G(x(t)) is given
wk (∆, − dk , y) = ).
(∆ − dk − λk )2 by a ∆-disk DIG. And, since the control law in Equation
(18) only takes pairwise interactions in Ed into account, we
where k = 1, 2, . . . |Ed |. Here, again, the index k runs over
can view this dynamic situation as a static situation, with
the edge set Ed . Note that this construction allows us to study
the SIG being given by Gd . Now we need to verify the
the evolution of yi , rather than xi , i = 1, . . . , N , and we
properties of F1, F2, and F3. That F3 (decentralized control)
formalize this in the following lemma for static interaction
is satisfied follows trivially from the definition of the control
graphs:
law in Equation (18). Moreover, we have already established
Lemma 4.1: Let the total tension energy function be that F1 (finite time convergence to the appropriate graph)
holds trivially as long as it holds initially, and what remains
1 
N N
V(∆ − d , y) = Vij (∆ − dij , y). (21) to be shown here is thus that we can drive the system in
2 i=1 j=1 finite time to a configuration in which F1 holds, after which
Lemma 4.2 applies. Moreover, we need to establish that the
Given y0 ∈ DG d ,∆−d , with Gd being a connected spanning inter-robot displacements (defined for edges in Ed ) converge
graph, then the set Ω(∆ − d , y0 ) := {y | V(∆ − d , y) ≤ asymptotically to the desired, relative displacements (F3),
V0 }, where V0 denotes the initial value of the total tension which is the topic of the next theorem.
energy function, is an invariant set under the control law
in Equation (18), under the assumption that the interaction Theorem 4.3: Under the same assumptions as in Lemma
graph is static. 4.2, ij (t) = xi (t) − xj (t) converges asymptotically to
Proof: By the proposed control law in Equation (18), dij for all i, j such that (vi , vj ) ∈ Ed .
 ∂Vij (∆ − dij , y) Proof: Based on the observation that Gd remains a spanning
ẏi = − (22) graph to the DIG, together with the observation that
∂yi
j∈NGd (i)
dc(y, j)
∂V(∆ − d , y) = −LW (∆ − d , y)c(y, j), j = 1, 2, . . . n,
= − = −∇yi V(∆ − d , y). dt
∂yi
Theorem 3.2 ensures that c(y, j) will converge to
The non-positivity of V̇ now follows the same argument as span{1}, ∀j ∈ {1, . . . , n}. What this implies is that all
in Equation (14) in the proof of Lemma 3.1. Moreover, for displacements must be the same, i.e. that yi = ζ, ∀i ∈
each initial y0 ∈ DG d ,∆−d , the corresponding maximal, {1, . . . , N } for some constant ζ ∈ Rn . But, this simply
total tension-energy induces a maximal possible edge length. means that the system converges asymptotically to a fixed
Following the same line of reasoning as in the proof of translation away from the target points τi , i = 1, . . . , N , i.e.
Lemma 3.1, the invariance of Ω(∆ − d , y0 ) thus follows. 
lim yi (t) = lim xi (t) − τi = ζ, i = 1, . . . , N,
t→∞ t→∞

5966
45th IEEE CDC, San Diego, USA, Dec. 13-15, 2006 FrIP3.3

ij
≤ ∆ − 
which in turn implies that
 σ(i, j) = 1
lim ij (t) = lim (xi (t) − xj (t)
σ(i, j) = 0
(23) f (xi − xj ) = 0 ∂V (∆,x)
f (xi − xj ) = − ij∂xi
t→∞ t→∞ ¯

= lim yi (t) + τi − yj (t) − τj
t→∞ (vi, vj ) ∈ Ed
= ζ + τi − ζ − τj = dij , and
G = KNε
(vi, vj ) ∈ Ed
and
G = KNε
∀i, j s.t. (vi , vj ) ∈ Ed , which completes the proof.

C. Hybrid, Rendezvous-to-Formation Control Strategies σ(i, j) = 0 σ(i, j) = 1

∂V (∆−
d
,y)
f (xi − xj ) = 0 f (xi − xj ) = − ij ∂yi ij
The last property that must be established is that it is
possible to satisfy F1, i.e. that the initial ∆-disk proximity
DIG does in fact converge to a graph that has Gd as a span- Fig. 1. The figure shows a state machine describing how the system
ning graph in finite time. If this was achieved then Theorem undergoes transitions from rendezvous (collection of the agents to a tight,
4.3 would be applicable and F2 (asymptotic convergence complete graph), to formation control.
[2] J. Desai, J. Ostrowski, and V. Kumar. Controlling formations of
to the correct inter-agent displacements) would follow. To multiple mobile robots. In Proc.IEEE Int. Conf. Robot. Automat.,
achieve this, we propose to use the rendezvous control law pages 2864–2869, Leuven, Belgium, May 1998.
developed in the previous section for gathering all agents [3] M. Egerstedt and X. Hu. Formation constrained multi-agent control.
IEEE Trans. Robot. Automat., 17(6):947–951, December 2001.
into a complete graph, of which trivially any desired graph [4] M. Egerstedt, X. Hu, and A. Stotsky. Control of mobile platforms
is a subgraph. Moreover, we need to achieve this in such a using a virtual vehicle approach. IEEE Trans. Automat. Contr.,
manner that the assumptions in Theorem 4.3 are satisfied. 46(46):1777–1782, November 2001.
[5] J.A. Fax and R.M. Murray. Information flow and cooperative control
Let KN denote the complete graph over N agents. More- of vehicle formations. IEEE Trans. Automat. Contr., 49:1465–1476,
ε
over, we will use KN to denote the DIG that is a complete Sept 2004.
graph in which no inter-agent distances are greater than [6] C. Godsil and G. Royle. Algebraic graph theory. Springer, 2001.
[7] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of
ε. This notation is slightly incorrect in that graphs are mobile autonomous agents using nearest neighbor rules. IEEE Trans.
inherently combinatorial objects, while inter-agent distances Automat. Contr., 48(6):988–1001, June 2003.
are geometric, and, to be more precise, we will use the [8] M. Ji and M. Egerstedt. Connectedness preserving distributed coordi-
nation control over dynamic graphs. In Proceedings of the American
notation G = KN ε
to denote the fact that G = KN and Control Conference 2005, pages 93–98, Portland, OR, June 2005.
ij ≤ ε, ∀ (i, j), i = j. The reason for this construction [9] Z. Lin, M. Broucke, and B. Francis. Local control strategies for
is that, in order for Theorem 4.3 to be applicable, the initial groups of mobile autonomous agents. IEEE Trans. Automat. Contr.,
49(4):622–629, 2004.
condition has to satisfy y0 = (x0 − τ0 ) ∈ DG d ,∆−d , [10] John-Michael McNew and Eric Klavins. Locally interacting hybrid
which is ensured by making ε small enough. Moreover, since systems with embedded graph grammars. In Conference on Decision
the rendezvous controller in Equation (17) asymptotically and Control, 2006. To Appear.
[11] M. Mesbahi. On a dynamic extension of the theory of graphs. In
achieves rendezvous, it will consequently drive the system Proceedings of the 2002 American Control Conference, volume 2,
to KN ε
in finite time, ∀ε ∈ (0, ∆). pages 1234–1239, May 2002.
ε [12] M. Mesbahi. State-dependent graphs. In Proceedings of the 42nd
After KN is achieved, the controller switches to the
IEEE Conference on Decision and Control, pages 3058–3063, Maui,
controller in Equation (18), as depicted in Figure 1. However, Hawaii USA, December 2003.
this hybrid control strategy is only viable if the condition that [13] A. Muhammad and M. Egerstedt. On the structural complexity of
G = KN ε
is locally verifiable in the sense that the agents multi-agent robot formations. In American Control Conference 2004,
Boston, MA, June 2004.
can decide for themselves that a synchronous mode switch [14] R. Olfati-Saber. Flocking for multi-agent dynamic systems: Algo-
is triggered [10]. In fact, if an agent has N − 1 neighbors, rithms and theory. IEEE Trans. Automat. Contr., 51(3):401–420,
i.e. degree N − 1, all of which are within a distance ε/2. March 2006.
[15] P. Orgen, M. Egerstedt, and X. Hu. A control lyapunov function
Hence, when one agent detects this condition, it will trigger approach to multi-agent coordination. IEEE Trans. Robot. Automat.,
a switching signal, and the transition in Figure 1 occurs. 18(5):847–851, October 2002.
Regardless of which, we know that this transition will in fact [16] W. Ren and R.W. Beard. Consensus of information under dynamically
changing interaction topologies. In Proceedings of the American
occur in finite time in such a way that the initial condition Control Conference 2004, volume 6, pages 4939–4944, June 30-July
assumptions of Theorem 4.3 are satisfied. 2 2004.
V. C ONCLUSION [17] D. Spanos and R. Murray. Robust connectivity of networked vehicles.
In Proceeding of the 43rd IEEE Conference on Decision and Control
In this paper, a graph-based nonlinear feedback control law 2004, volume 3, pages 2893–2898, December 2004.
[18] K. Sugihara and I. Suzuki. Distributed motion coordination of multiple
is studied for distributed formation control. The nonlinear robots. In Proceedings of IEEE Int. Symp. on Intelligent Control, pages
feedback law is based on weighted graph Laplacians and it 138–143, 1990.
is proved to be able to solve the formation control problem [19] H. Tanner, A. Jadbabaie, and G.J. Pappas. Stable flocking of mobile
agents, part II : Dynamic topology. In Proceedings of the 42nd IEEE
while preserving connectedness. Conference on Decision and Control, pages 2016–2021, 2003.
R EFERENCES [20] M. Zavlanos and G. Pappas. Controlling connectivity of dynamic
graphs. In Proceeding of the 44th IEEE Conference on Decision
[1] J. Cortés, S. Martı́nez, and F. Bullo. Robust rendezvous for mobile and Control 2005 and 2005 European Control Conference (CDC-ECC
autonomous agents via proximity graphs in d dimension. IEEE Trans. ’05), pages 6388–6393, December 2005.
Robot. Automat., 51(8):1289–1298, 2006.

5967