Decentralized Cohesive Motion Control of

Multi-Agent Formations*
Srikumar Sandeep

Bar Fidan and Changbin Yu

University of Auckland
Auckland, New Zealand

National ICT Australia and

the Australian National University
Locked Bag 8001, Canberra ACT 2601, Australia

Abstract - This paper presents a set of decentralized

control laws for the cohesive motion of 2-dimensional multiagent formations. We consider rigid and constraint-consistent
formations that can be modeled by directed graphs, where
each agent is represented by a vertex and the inter-agent
distance constraints are represented by directed edges. We
analyze both of the two main hierarchical structures for such
2-dimensional formations: The leader-follower and the threecoleader structures. For each structure, we derive a control
scheme that moves a given rigid and constraint-consistent
formation whose initial position and orientation are specified
to a new desired position and orientation cohesively, i.e.,
without deforming the shape of the formation during the
motion. We elaborate our designs considering the path
smoothness, chattering, and agent dynamics issues and
demonstrate their effectiveness via a set of simulation results.
Index Terms rigid formations, multi-agent systems,
decentralized control, cohesive motion, persistence.

I. INTRODUCTION
In recent years, the topic of cooperative control and
control of multi-agent formations has gained a lot of
attention [1-4] in parallel with the increase of interest in
real-life multi-agent system applications such as the ones
involving groups of combat robots, unmanned aerial
vehicles, ground vehicles, underwater vehicles, etc. [5-7].
Beside the usual requirement of assigning decentralized
decision-making mechanisms, the foci and approaches of
the studies on formation control are diverse depending on
the agent dynamics, control goals, control strategies, and the
inter-agent control and information structure considered.
In this paper, our focus is on the cohesive motion of the
entire formation rather than the individual agent dynamic
behaviors. Hence, we assume a point-agent system model
and represent each multi-agent formation of interest by a
directed graph where each agent Ai is represented with a
vertex i and each neighbor agent pair

(A , A ) ,
i

the

distance between which is required to be explicitly

maintained at a desired value d ij , is represented by a
directed edge, say (i, j ) . The direction of the edge (i, j )
from i to j implies that agent Ai is responsible to keep the

distance between Ai , A j at the desired value d ij , i.e., Ai

has the constraint of staying at a distance d ij from A j .

A formation is called rigid, if the distance between each

pair of agents remains constant, i.e., the formation shape is
maintained during any continuous motion provided that
each agent satisfies all the distance constraints on it [8]. For
a rigid formation, some of the distances between the
neighbor agents (which are connected by edges) are
explicitly maintained. Because of this explicit maintenance,
the remaining inter-agent distances are also (implicitly)
maintained. Thus the formation retains its shape during any
continuous motion. If each agent in a given formation is
able to satisfy all the constraints on it once all the other
constraints within the formation are satisfied, then the
formation is called constraint-consistent. If a formation is
both rigid and constraint-consistent, then it is called
persistent [9]. Rigorous definitions and elaborated analysis
of rigid, constraint-consistent and persistent formations can
be found in [8,9].
The main contribution of this paper is providing a
decentralized control strategy for cohesive motion of
persistent formations. The aim of the control strategy is to
move a given persistent formation whose initial position and
orientation are specified to a new desired position and
orientation cohesively, i.e., without violating the persistence
of the formation during the motion. Note that this is a
decentralized task shared by the individual autonomous
agents of the formation and each of these agents make
decisions based only on its own observations and state. To
our best knowledge, none of the existing papers have
addressed this problem. For simplicity, we consider only the
case of 2-dimensional multi-agent formations, and the
number of agents is kept small for clarity of illustration.
Based on the distribution of the distance constraints
(and hence the degrees of freedom) of the agents, persistent
formations can be categorized into two groups [8]: The
formations with the leader-follower structure and the ones
with the three-coleader structure. The notion of degrees of
freedom and the two formation categories above shall be
elaborated more in Section II. In our decentralized motion
control design and analysis we consider both of these
categories separately.
For simplicity we present our control schemes assuming
a simple integrator model for the agents, which is used in
many other works in the field [5,6]. However, our designs
can be generalized for other (more realistic) dynamic
models as well, as demonstrated for the nonholonomic
unicycle model [10-13] in Section V. We demonstrate the

This work is supported by National ICT Australia, which is funded by the Australian Governments Department of Communications, Information
Technology and the Arts and the Australian Research Council through the Backing Australias Ability Initiative. Corresponding author: Bar Fidan (e-mail:
Baris.Fidan@anu.edu.au)

effectiveness of the control laws developed for each case via

simulation results.
The rest of the paper is organized as follows. In Section
II, we elaborate the notions of rigidity, persistence, and
degrees of freedom, and formulation of the cohesive
formation motion problem. Our control design for the
formations with the leader-follower and the 3-coleader
structures are presented, respectively, in Sections III and IV.
In Section V, we discuss the effects of assuming an agent
model other than the simple integrator one and redesign our
schemes in Sections III and IV for the nonholonomic
unicycle model. The paper ends with concluding remarks
and future work in Section VI.
II. PROBLEM DEFINITION
In this paper, we consider 2-dimensional pointagent
formations whose shapes are required to be maintained
during any motion by explicitly maintaining the distances
between certain agent pairs. We call each such agent pair a
neighbor agent pair. We model each multi-agent formation
using a directed graph, which is also called the directed
underlying graph of the formation. In the directed
underlying graph of a given multi-agent formation, each
agent Ai is represented with a vertex i and each neighbor

agent pair Ai , A j , the distance between which is required

to be explicitly maintained at a desired value d ij , is
represented by a directed edge, say (i, j ) . The direction of
the edge (i, j ) from i to j implies that agent Ai is

responsible to keep the distance between Ai , A j

) at the

desired value d ij , i.e., Ai has the constraint of staying at a

distance d ij from A j . In this case we also say that Ai
follows A j or Ai is a follower of A j .
Definitions of rigid, constraint-consistent and
persistent formations were given in Section I. For a given
persistent formation F, if by removal of any single edge (in
the directed underlying graph) would make F nonpersistent then F is further called minimally persistent, i.e.,
a minimally persistent formation is a persistent formation
which preserves its persistence with a minimal number of
edges. Next we provide some fundamental characteristics of
persistent formations relevant to our work, whose details
can be found in [8,9]. A 2-dimensional formation is always
constraint consistent, if the out-degree of each vertex i (the

number edges (i, j ) originating at i ) in its directed

underlying graph is less than or equal to 2. Hence a 2dimensional rigid formation is always persistent, if the outdegree of each vertex i in its directed underlying graph is
less than or equal to 2.
It can be easily observed that the out-degree of a vertex
in a directed underlying graph is equal to the number of
distance constraints on the corresponding agent. Hence the
degree of freedom (DOF) of an agent (for its motion in 2 )
can be defined based on the out-degree of its representative
vertex, i.e., the corresponding vertex in the underlying

graph. Given an agent Ai , if the out-degree of its

representative vertex i is 0 then the agent has 2-DOF, i.e., it
can move freely in 2 . If the out-degree is 1 then Ai can
only rotate around the agent it follows in order to meet its
distance constraint and hence Ai has 1-DOF. If the outdegree is 2 or more then the motion of Ai completely
depends on the agents it follows, i.e. it has 0-DOF.
For a persistent formation, it has been shown in [8,9]
that the sum of DOFs of individual agents is at most 3 (and
for minimally persistent formation, exactly 3), which is the
same as the DOF of a free rigid (non-vertex) object in 2
(2 for translation and 1 for rotation). Based on the
distribution of these 3 DOFs, persistent formations can be
divided into two categories: Formations with the leaderfollower structure where one agent has 2-DOF, another has
1-DOF and the rest have 0-DOF, and the 3-coleader
structure where three agents have 1-DOF and the rest have
0-DOF. In the leader-follower structure, the 2-DOF agent is
called the leader and the 1-DOF is called the first follower.
In the 3-coleader structure, the 1-DOF agents are called the
coleaders. In both structures the 0-DOF agents are called
the (ordinary) followers.
From [8], it can be seen that the underlying graph of a
formation with leader-follower structure is always acyclic
(cycle-free). Some stability properties of leader-follower
structures are investigated in [11]. Any formation with the
three-coleader structure has at least one cycles. Due to the
presence of a cycle, the motion of the three coleaders are
cyclically dependent on each other and hence the motion
control for the formation requires a more implicit strategy.
To our best of knowledge, there is no existing literature that
deals with motion control of persistent (or rigid) formations
with cycles.
In our control design and analysis, we consider both of
the two categories above. One can verify via simple
examples that there exist persistent formations with the
leader-follower structure where the first follower does not
directly follow the leader but another agent, and there exist
persistent formations with the 3-coleader structure where
the coleaders do not directly follow each other but some
other agents. However for simplicity, in our work we
assume that the first follower directly follows the leader in a
leader-follower structured formation and the three co
leaders follow each other in a 3-coleader structured
formation.
Our control task is to move a given persistent formation
F with say m 3 agents A1 , K , Am whose initial position
and orientation are specified, to a new desired position and
orientation in 2 (the xy-plane) cohesively, i.e., without
violating the persistence of F during motion, using a
decentralized strategy. The initial and final positions and
orientations of the formation F are defined by the initial
positions pi 0 and final positions pif of the individual
agents Ai for i = 1, K , m , respectively. We assume that the
final positions pif are consistent with the desired inter-

agent distances d ij between neighbor agent pairs Ai , A j

in the formation F .
For each agent, we assume a velocity integrator
dynamics, i.e., for each agent Ai we assume that its
dynamics is given by

p& i (t ) = vi (t )

(1)

pi (t ) = (xi (t ), yi (t ) ) 2

where

and

vi (t ) = (v xi (t ), v yi (t ) ) denote the position and velocity

2

of Ai at time t . The velocity v i (t ) is considered as the

control signal of the agent Ai to be generated by the
individual controller of Ai . It is required that v i (t ) is
continuous and satisfies vi (t ) v for some constant
maximum speed limit v > 0 at any t 0 for any
i {1, K , m}. We assume that each agent Ai knows its final
desired position pif and can sense its own position p i (t )
and velocity v i (t ) as well as the position p j (t ) of each
agent A j it follows at any time t 0 . It is also assumed
that the distance sensing range for a neighbor agent pair
Ai , A j is sufficiently larger than the desired distance d ij

to be maintained.
For each 1-DOF agent, the control laws below are
chosen so that meeting the distance constraint (i.e.,
formation maintenance) has a higher priority than reaching
to the final desired position, i.e., a 1-DOF agent can
undergo its degree of freedom movement, only when its
distance constraint is satisfied within a certain error bound.
Based on the agent dynamics (1) and the assumptions
above, in the next two sections we develop control schemes
to address cohesive motion task defined above for persistent
formation with both the leader-follower and the 3-coleader
structures. Later we will discuss extension of these control
schemes for agent dynamics more complex than (1).
III. CONTROL FOR THE LEADER-FOLLOWER STRUCTURE

Fig. 1.

The directed underlying graph of a persistent graph

with the leader-follower structure.
For optimality considerations and to cope with constant
velocity requirements in certain UAV and other flight
formation cases, we assert the following two guidelines in
our control design: (i) Any agent will move at the constant
maximum speed v > 0 at any time instant t 0 unless it
is impossible to do that at that particular instant t due to,
e.g., initial and final zero-velocity constraints, etc. (ii) Any
agent with distance constraints moves through a path of
shortest distance in order to satisfy these constraints.
Based on these two assertions and the earlier
assumptions, the derivation of our control law to address the
cohesive motion task in Section II uses basic vector analysis
and borrows ideas from the virtual vector field concept,
details of which can found in [6,14,15]. The main idea in
the virtual vector field approach is obtaining the overall
velocity vector (for each agent) as the superposition of the
vectors defining each of the separate motion tasks (of this
agent). In our case the two separate motion vector types of
an agent is (i) to maintain a distance constraint with each of
the agents it follows and (ii) to move towards a final
destination. Nevertheless, note that the virtual vector field
approaches described in [6,14,15] are somewhat different
from our approach, as in these works, the inter-agent
distance constraints are not considered hence do not
constitute a vector field.
A. Control Law for the Ordinary Followers
Consider an (ordinary) follower agent Ai (for example

A4 in Fig. 1) and the two agents A j and Ak it follows ( A2

and A3 for the example) . Due to the distance constraints of
keeping pi (t ) p j (t ) , pi (t ) p k (t ) at the desired values
of d ij , d ik respectively, the motion of Ai is dictated by the
motion of Aj and Ak . Hence at each time t 0 , the
desired position p id (t ) of Ai is the point whose distances
to p j (t ) and p k (t ) are d ij , d ik respectively; as a function
of time p id (t ) is continuous. Assuming p i (t ) pid (t ) is

In this section, we derive control laws for the cohesive

motion of a persistent formation with the leader-follower
structure. We illustrate our designs on a formation with one
leader A1 , one first-follower A2 and two (ordinary)

sufficiently small, p id (t ) can be explicitly determined as

followers A3 and

where p jk (t , p 0 ) for any p0 2 denotes the intersection

A4 depicted in Fig. 1.

of the circles C ( p j (t ), d ij ) and C ( p k (t ), d ik ) that is

A2

A4

A1

A3

pid (t ) = p jk (t , pi (t )) = (x jk (t , p i (t )), y jk (t , pi (t )) )

closest to p 0 , and in the notion C (,) the first argument

denotes the center and the second denotes the radius. Based
on this observation, we propose the following control law
for the follower agents:

v i ( t ) = v i ( t ) id ( t ) id ( t )

id ( t ) = p id ( t ) p i ( t ) = p jk ( t , p i ( t )) p i ( t )

i ( t ) = id

0,
(t ) k

(2)

id ( t ) < k

the dot product operation. In (3), via the switching term

i ( t ) , the controller switches between a translational

k id ( t ) < 2 k

action (4) to keep the distance Ai A j close to desired value

id ( t ) 2 k

d ij and a rotational action (5) (around A j ) to move the

1,

agent Ai towards its desired final position pif . Note that

where v > 0 is the constant maximum speed of the agents

and k > 0 is a small design constant. In (2), the switching
term i ( t ) is introduced to avoid chattering due to small
but acceptable errors in the desired inter-agent distances.
B. Control Law for the First-Follower
Let agent Ai be the first-follower and Aj be the leader.
First, observe that once the first-follower satisfies its
distance constraint with the leader, it is free to rotate in a
clockwise or counter-clockwise direction around the leader.
For the example in Fig. 1, the first-follower can move on the
circle with center A1 (the leader) and radius d 21 provided
that it does not need to use the whole of its velocity capacity
to satisfy the distance constraint A1 A2 = d 21 .
Based on the above observation and the assumption in
Section II that for each 1-DOF agent, meeting the distance
constraint (i.e., formation maintenance) has a higher priority
than reaching to its final desired position, we propose the
following control scheme for the first-follower agent Ai :

v i ( t ) = i ( t ) v i1 ( t ) + 1 i2 ( t ) v i 2 ( t )

ji ( t ) = (

jix

( t ),

jiy

and k , f > 0 are small design constants and , denotes

(t ) ) = p j (t ) p i (t )

sufficiently close to d ij .
In (5),

ji

(t ) is the unit vector perpendicular to the

distance vector ji ( t ) = p j ( t ) p i ( t ) with clockwise

orientation with respect to the circle C ( p j (t ), d ij ) , and the

term sgn

if

( t ),

ji

(t )

) determines the orientation of

motion that would move Ai towards Aif . The switching

term i (t ) is for avoiding chattering due to small but
acceptable errors in the final position of Ai .
C. Control Law for the Leader
If a given agent Ai is the leader of the formation (e.g.,

A1 in Figure 1), since it does not have any distance

constraint to satisfy, it can use its full velocity capacity only
to move towards its desired final position pif . Hence the
velocity input at each time t can be simply designed as a
vector with magnitude v and the same direction as
p if ( t ) p i ( t ) :

v i ( t ) = v i ( t ) if ( t ) if ( t )

ji ( t ) = ij ( t ) d ij

0,

ji ( t ) k
i (t ) =
k

1,

the rotational action (5) can take place only when Ai A j is

(3)

ji ( t ) < k

if
i (t ) =

k ji ( t ) < 2 k
ji ( t ) 2 k

(6)

if ( t ) = p if ( t ) p i ( t )
if ( t ) <

0,
(t )

if ( t ) < 2

if ( t ) 2

1,

where

v i1 ( t ) = v sgn ( ji ( t ) ) ji ( t ) ji ( t )

v i 2 ( t ) = v i ( t ) sgn if ( t ),

if ( t ) = p if ( t ) p i ( t )

ji ( t ) = (

jiy

( t ),

0,

if ( t )
i (t ) =
f

1,

jix

ji

(t )

ji

if ( t ) < 2

if ( t ) 2

The switching term i (t ) is defined again for

avoiding chattering due to small but acceptable errors in the
final position of Ai .

(5)

D. Simulation Results
The control laws (2)-(6) have been successfully tested
via a number of simulations on various formations with
leader-follower structure for various initial and desired final
settings. In this subsection we present the results for the
formation depicted in Fig.1 for a sample initial and desired
final setting pair.
We assume that all the desired distance between each
neighbor agent pair Ai , A j whose representative vertices

(t )

( t ) ) ji ( t )

if ( t ) <

(4)

are connected with the edge (i, j ) is d ij = 1 m. The design

parameters are chosen as v = 1 m/s and k = 0.01 . Figure

2 shows the motion of the formation for the following initial
and desired final positions of the agents (units are all in m):
p1 (0) = (0,0), p2 (0) = ( 3 2 ,0.5), p3 (0) = ( 3 2 ,0.5),
p 4 (0) = ( 3 ,0),

p1 f = (5,2),

A4

p 2 f = (5 3 2 ,2.5),

A2

A1

p 3 f = ( 5 3 2 ,1.5), p 4 f = (5 3,2) .

A3
Fig. 3. The directed underlying graph of a persistent graph
with the 3-coleader structure.

(a)

We use the same assumptions and the same

methodology as in Section IV to design one control scheme
to be used in each of the co-leader and another scheme to be
used in the followers. Note that in Section IV the control
law for each agent is designed based on the individual task
of the agent, which is satisfying its two distance constraints
for a 0-DOF agent, satisfying its single distance constraint
and then moving to its desired final position for a 1-DOF
agent, and moving to its desired final position for a 2-DOF
agent. The corresponding control laws are derived as (2),
(3)-(5), and (6), respectively.
Using the same argument and noting that each co-leader
has 1-DOF and each follower has 0-DOF in a persistent
formation with the 3-coleader structure, we simply select (2)
as the control scheme for the followers and (3)-(5) as the
control law for each of the three coleaders.
B. Simulation Results
Similarly to the tests reported in Section III, we have
successfully tested the control scheme designed in Section
IV.A via a number of simulations on various formation
settings with the 3-coleader structure. In this subsection we
present the results for the formation depicted in Fig. 3 for a
sample initial and desired final setting pair.
We assume that all the desired distances between each
neighbor agent pair Ai , A j whose representative vertices

are connected with the edge (i, j ) is d ij = 1 m. Fig. 4

(b)
Fig. 2. Cohesive motion of a formation with the leaderfollower structure: (a) The path. (b) Inter-agent distances
during motion.
IV. CONTROL FOR THE THREE-COLEADER STRUCTURE
A. Control Law Design
In this section, we derive the counterparts of the control
laws designed in Section III for persistent formations with
the 3-coleader structure. Again to illustrate our designs, we
use a simple formation with three co-leaders A1 , A2 , A3 and
one (ordinary) follower A4 as depicted in Fig. 3.

shows the motion of the formation for the same design

parameters and the same initial and desired final positions of
the agents as in the example in Section III.D.
V. REDESIGN FOR NONHOLONOMIC AGENT DYNAMICS
In Sections III and IV, we have designed control
schemes for cohesive motion of persistent formations based
on the integrator model (1) for the agent dynamics. In this
section we discuss the effects of assuming an agent model
other than (1) and redesign our schemes in Sections III and
IV for the nonholonomic unicycle model [7]. For each agent
Ai , we consider the following kinematic model for unicycle
nonholonomic robots:

( x& i , y& i ) = (v i cos i , v i sin i )

& =
i

(7)

where p i (t ) = ( xi (t ), y i (t ) ) denotes the position of Ai (as

in the previous sections) and i (t ) denotes the orientation

of the agent with the convention depicted in Fig. 5 (a). The
control inputs are the scalar translational velocity vi and
the angular velocity i .
Next we redesign the control schemes presented in
previous sections based on the agent dynamics (7). Due to
space limitations, we only present our redesign for
formations with the 3-coleader structure. We illustrate the
use of our control design again on the formation depicted in
Fig. 3. In the redesign of the control scheme of Section III,
we employ the separation-separation control idea that was
published in [7]. We first describe the idea for a virtual
setting where one agent with dynamics (7) follows two
other agents keeping a specific distance from each. Then we
present our redesign based on this idea.

( x& i , y& i )

yi

Ai
x

xi
(a)

il
Al
Ai

( x& l , y& l )

(b)

Fig. 5. Conventions: (a) The parameters in (7) for agent Ai .

(b) The parameters in separation-separation control for
agent Ai following agents A j and Ak .

A. Redesign Using a Separation-Separation Scheme

Consider an agent Ai that follows two other agents, A j
and Ak (all with dynamics (7)). Suppose that it is desired
for Ai to keep distances

d ij and d ik from A j and Ak ,

respectively. The separation-separation control law for

this task is given as follows [7]:
vi =

i =

(a)

sik sin ij sij sin ik + v k cos ik sin ij v j cos ij sin ik

sin( ij ik )
s ik cos ij + sij cos ik v k cos ik cos ij + v j cos ij cos ik

(8)

k 2 sin( ij ik )

s il = k1 ( d il pi pl ), il = l + il i

for l { j , k }

where il is the separation angle defined in Fig. 5 (b) and

k1 , k 2 > 0 are some design coefficients.

In a persistent formation with the 3-coleader structure,
we can use (8) directly for a follower agent Ai that follows
agents A j and Ak since the task of Ai as a follower is
exactly the same as the above task tackled by separationseparation control.
However, if Ai is a co-leader following agent A j , one
cannot use (8) directly. In this case we consider a virtual
agent Ak (with speed vk (t ) ) which lies on the line segment

[ pi (t ) pif ], t and a corresponding time-varying desired

(b)
Fig.4. Cohesive motion of a formation with the three-coleader structure (a) The path. (b) Inter-agent distances
during motion.

distance d ik (t ) which are explicitly defined as follows:

A closer look at (8) reveals that due to the sine function in

the denominator, the angular and translational velocity
inputs are unbounded. In our work, this problem is
eliminated in the implementation of the controller using a
velocity limiter.

 p (t ) + ( pif pi (t ) ) pif pi (t ) if pif pi (t ) > 1

p k (t ) = i
pif
otherwise

k (t ) = pif pi (t ) pi (t ),
pi (t ) = pi (t ) + (1,0)

(9)

1 ( m / s ) if pif pi (t ) > 1
v k (t ) =
otherwise
pif pi (t )
1 (m)
if pif pi (t ) > 1
d ik (0) =
(
)

p
p
t
otherwise
if
i

In summary, we use the control law (8) for a follower

Ai following agents A j , Ak , and the control law (8),(9) for

a co-leader Ai following agent A j .
B. Simulation Results
We have tested the control scheme of Section V.A via a
number of simulations. The result for the example in
Section IV.B with nonholonomic unicycle agent dynamics
(7) and the control scheme of Section V.A with
k1 = 1, k 2 = 0.2 , which is shown in Fig. 7.

As it can be easily observed in Fig. 7, although the

control goal of reaching the final positions is achieved, the
performance is poor compared to the performance seen in
Section IV in terms of both the path length and the distance
constraints.
One reason for the above phenomenon is (design-)
parametric sensitivity and certain nonlinear characteristics
of the separation-separation control law (8), particularly
the denominator term whose magnitude may become very
small time to time. Another reason is that the control laws
for the coleaders are not derived based directly on their
specicific duties but rather adapted from the separationseparation idea which is mainly for an agent following two
other agents.
Design of an enhanced control scheme to obtain a
performance level that is comparable to the results in
Sections III and IV is currently being investigated by the
authors. Nevertheless, the simulation results demonstrate the
feasibility of cohesive motion control with agent dynamics
models that are more complex then the velocity integrator
model.
VI. CONCLUSION

(a)

(b)
Fig. 7. Cohesive motion of a formation with three co-leader
structure with agent dynamics (7): (a) The path. (b) Interagent distances during motion.

In this paper, we have formulated and presented a

solution for the problem of controlling the cohesive motion
of persistent (rigid and constraint consistent) autonomous
formations, where the rigidity of the formation is maintained
via a set of constraints on the agents to keep their distances
from certain other agents at some predefined values. We
have considered both of the two possible hierarchical
(underlying directed graph) structures for such formations.
For both of these structures, we have designed decentralized
control laws to move the formation having an arbitrary
initial position and orientation to a desired final position and
orientation. The designs have been initially performed
assuming a simple integrator agent dynamics, and later
extended for a nonholonomic unicycle agent model.
Directed graph notions, basic vector analysis, and switching
are the main tools used in system modeling and control
design. The effectiveness of the designed control schemes
have been demonstrated via a set of simulation results.
Since the main goal of this paper was formulating the
problem of cohesive motion of persistent formations and
finding a set of solutions to this general problem and due
to space limitations, there exist a number of issues which
have not been covered in the paper. A formal stability and
convergence analysis for the proposed control designs is
underway for an extended version of this paper. Beside that,
the proposed control laws and strategies can be enhanced to
make the paths of the agents smoother and/or more optimal
in the sense of total displacement of all agents, etc. The
control schemes can be re-designed for agent dynamics
other than the velocity integrator and nonholonomic
unicycle dynamics models used in this paper. Another
possible redesign task would Consideration of obstacles in
the region of motion and addition of the obstacle avoidance
task to the cohesive motion problem is another future
research aspect.

REFERENCES
[1] R. Olfati-Saber and R. Murray, Distributed cooperative control of
multiple vehicle formations using structural potential functions, in
Proc. IFAC World Congress, Barcelona, Spain, 2002.
[2] J. Baillieul and A. Suri, Information patterns and hedging Brockett's
theorem in controlling vehicle formations, in Proc. of the 42nd IEEE
Conf. on Decision and Control, vol. 1, pp. 556-563, December 2003.
[3] R. Olfati-Saber and R. M. Murray, Consensus problems in networks
of agents with switching topology and time-delays, IEEE Trans. on
Automatic Control, vol. 49, no. 9, pp. 520 - 1533, September 2004.
[4] V. Gazi and K.M. Passino, Stability analysis of swarms, IEEE Trans.
on Automatic Control, vol. 48, no. 4, pp. 692-697, April 2003.
[5] H.G. Tanner, A. Jadbabaie and G.J. Pappas, Flocking in teams of
nonholonomic agents in Cooperative Control, V. Kumar, N. Leonard
and A. S. Morse (Eds.) Springer-Verlag, pp. 229-239, 2005.
[6] S. Drake, K. Brown, J. Fazackerley and A. Finn, Autonomous Control
of Multiple UAVs for the Passive Location of Radars, in Proc. 2nd
Int. Conf. on Intelligent Sensors, Sensor Networks and Information
Processing (ISSNIP), pp.403-409, December 2005.
[7] R. Fierro, P. Song, A. Das, and V. Kumar, Cooperative control of
robot formations, in Cooperative Control and Optimization, R.
Murphey and P. Pardalos (eds.), Kluwer Academic Press, pp. 73-94,
2002.
[8] C. Yu, J.M. Hendrickx, B. Fidan, B.D.O. Anderson, and V.D. Blondel,
Three and higher dimensional autonomous formations: Rigidity,
persistence and structural persistence, to appear in Automatica, 2006.
[9] J.M. Hendrickx, B. Fidan, C. Yu, B.D.O. Anderson, and V.D. Blondel,
Rigidity and persistence of three and higher dimensional formations,
in Proc.1st Int. Workshop on Multi-Agent Robotic Systems (MARS
2005), pp. 3946, September 2005.
[10]Z. Lin, B. Francis, and M. Maggiore, Necessary and sufficient
graphical conditions for formation control of unicycles, IEEE
Transactions on Automatic Control, vol. 50, no.1, pp. 121-127, January
2005.
[11]H.G. Tanner, G.J. Pappas, and V. Kumar, Leader-to-formation
stability, IEEE Trans. on Robotics and Automation, vol. 20, no.3, pp.
443-455, June 2004.
[12]A.K. Das, R. Fierro, and V. Kumar, Control graphs for robot
networks, in Cooperative Control: Models, Applications and
Algorithms, S. Butenko, R. Murphey and P. Pardalos (eds.), pp. 55-73,
Kluwer Academic Press, 2003.
[13]A.K. Das, R. Fierro, V. Kumar, J.P. Ostrowski, J. Spletzer, and C.J.
Taylor, A vision-based formation control framework, IEEE Trans. on
Robotics and Automation, vol. 18, no. 5, pp. 813825, October 2002.
[14]J. Borenstein and Y. Koren, Teleautonomous guidance for mobile
robots, in Proc. IEEE Trans. on Systems, Man and Cybernetics, vol.
20, no. 6, pp. 1437 1443, Nov.-Dec. 1990.
[15]M. Kohji and M. Yoshiki, Control of shape and internal movement of
a homogenous autonomous mobile robot herd employing simple virtual
interactive forces, in SICE 2003 Annual Conference, vol. 3, 2003, pp.
2912-2915.