IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 18, NO.

5, OCTOBER 2002 847

A Control Lyapunov Function Approach to

Multiagent Coordination
Petter Ögren, Magnus Egerstedt, Member, IEEE, and Xiaoming Hu, Member, IEEE

Abstract—In this paper, the multiagent coordination problem is into the category of “virtual structures.” The formation function
studied. This problem is addressed for a class of robots for which we introduce has similarities with the task function of [13], but
control Lyapunov functions can be found. The main result is a there is no connection with Lyapunov theory in that approach.
suite of theorems about formation maintenance, task completion
time, and formation velocity. It is also shown how to moderate The motivation for studying this type of multiagent coordi-
the requirement that, for each individual robot, there exists a con- nation problem mainly stems from the observation that there is
trol Lyapunov function. An example is provided that illustrates the robustness and strength in numbers. If more than one agent is
soundness of the method. asked to carry out a given task, e.g., search a disaster area, the
Index Terms—Coordinated control, Lyapunov methods, mobile likelihood of success increases as more agents are included in
robots, multirobot system, robot formation control. the mission. In other situations, cost and energy efficiency in-
dicates that using many small robots might be more beneficial
than using one big robot.
I. INTRODUCTION
The outline of this paper is as follows. In Section II, we de-

I N THIS PAPER, we investigate the problem of how to

coordinate a collection of robots in such a way that they
maintain a given formation relative to each other. The main
fine what we mean by a formation, and show how this can be
formalized in terms of a formation function. We define a sub-
class called Lyapunov formation functions, and show how these
assumption about the dynamics of the individual robots that we can be constructed. Next, in Section III, we add error feedback
initially make in this paper is that they have control Lyapunov to the time evolution of the formation. In Section IV, we then
functions (CLFs). Based on this assumption, an abstract and prove a suite of theorems about bounded formation errors, task
theoretically sound coordination strategy can be developed. completion times, and group velocities. We conclude, in Section
Multiagent formation control problems have been extensively V, with an example, illustrating the usefulness of our proposed
studied in the literature, and our main contribution is that we use method.
CLFs to define the formation. By doing this, we convert the for-
mation control problem, typically a constrained motion control
problem of multiple systems, into a stabilization problem for II. FORMATION FUNCTIONS
one single system. By this approach, we neither cast the problem Our primary object of study is a collection of robots, whose
without real dynamics [2], nor with an explicit nonlinear robot dynamics can be described by the following set of controlled
model [5]. Instead, we believe that by requiring the existence of differential equations:
CLFs, we can capture at least some aspects of the platform dy-
namics, while not having to spend our main effort on nonlinear
robot control. Thus, we can focus on the coordination problem (1)
at a higher level.
In addition to the CLF approach, we use the idea of virtual ve- where , and .
hicles discussed in [6]. Concepts of similar flavor are the “action Now, a desired formation in is simply a set
reference” suggested by Kang et al. [7] and the “dynamic coor- , and we define this set implicitly
dination variable” proposed by Beard et al. [3]. Furthermore, in through the null set of a so-called formation function.
the terminology of Beard et al. in [9], our approach would fall Definition II.1—Formation Function: Given a formation
parameterized by a scalar .
We say that a positive definite, continuously differentiable
Manuscript received June 11, 2001; revised February 7, 2002. This paper map is a formation function to the set
was recommended for publication by Associate Editor E. Pagello and Editor if gives a
S. Hutchinson upon evaluation of the reviewers’ comments. This work was sup-
ported in part by the SSF through the Centre for Autonomous Systems at the unique for each choice of .
Royal Institute of Technology (KTH), Stockholm, Sweden. This paper was pre- Example II.1: One choice is
sented in part at IEEE CDC’01, Orlando, FL, December, 2001. . Clearly, implies a unique
P. Ögren and X. Hu are with the Division of Optimization and Systems
Theory, Royal Institute of Technology (KTH) SE-100 44 Stockholm, Sweden for each , as demanded in the definition. But as we will see,
(e-mail: [email protected]; [email protected]). robot dynamics (Definition II.2) can be taken into account
M. Egerstedt is with the School of Electrical and Computer Engi- when choosing .
neering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail:
[email protected]). In order to focus on the high-level coordination issues, we
Digital Object Identifier 10.1109/TRA.2002.804500 connect the formation function with the concept of CLFs.
1042-296X/02$17.00 © 2002 IEEE
848 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 18, NO. 5, OCTOBER 2002

Definition II.2—Lyapunov Formation Function: A forma- tically important systems [8], including feedback linearizable
tion function is a Lyapunov formation function if there systems (as will be seen in Section V). Note also the Artstein
exists a class function and a control , such that Sonntag theorem on existence of CLFs [1].
We now go on to state and prove the main existence theorem
of Lyapunov formation functions.
Theorem II.1—Lyapunov Formation Function: If the vehi-
cles have translationally invariant (in position coordinates) dy-
Furthermore, the left-hand side of this expression goes to in-
namics and LPD CLFs , with locally negative definite time
finity as approaches zero.
derivatives, then we can form a Lyapunov formation function
A class function is a function , continuous,
by a weighted sum of the parameterized CLFs
strictly increasing, and satisfying
[11].
To set the stage for our main existence theorem on Lyapunov (2)
formation functions, we now state two lemmas.
Lemma II.1—Locally Positive Definite (LPD) Implica- Proof: By Lemma II.1, the bound of Definition II.2 holds,
tions: If the Lyapunov function is LPD and decresent and pointwise. The hypothesis that the vehicle dynamics are in-
is LPD, then the bound of Definition II.2 is fulfilled. variant with respect to position translations lets us parameterize
Proof: The conditions on implies [11] that there exist them with respect to , and we have that
class functions and , such that

where the are given by as in Lemma II.1.

Since the functions are of class , they are all invertible and
We have that
positive, and we have and

Note that this makes a class function. Now we have

and we can choose . Furthermore, if

the property only holds locally, i.e., when , then this
condition can be replaced by , since
where
.
Remark II.1: These are the standard Lyapunov assumptions
for showing that a system is asymptotically stable [11]. The case
of a semidefinite Lyapunov function for the whole formation is
The second inequality above follows, since there exists a such
discussed in [12].
that , and thus
Lemma II.2—Limit Property: Given a such that

The limit property follows from Lemma II.2.

The limit property of Definition II.2 will be fulfilled if we Remark II.4: Note that the different vehicles in the forma-
choose a new . tion can have completely different dynamics [and therefore
Proof: Since the bound still holds. Furthermore Lyapunov functions in (2)]. Furthermore, the choice
of coefficients in (2) reflects how large deviations from zero
are allowed for each .

III. COORDINATED CONTROL

By establishing these observations about the Lyapunov for-
which clearly approaches infinity as . mation functions derived from the individual CLFs, we can
Remark II.2: The local stability property of Definition II.2 now shift our attention to actually controlling the evolution of
in combination with the feedback of (3) in Section III will be the formation. The one parameter that we can control is the
shown to guarantee successful traversal of the whole trajectory. parameter, i.e., the parameterization of the time evolution of
Remark II.3: In general, finding a CLF is an open problem, the desired positions. We do this by specifying the trajectory
however, it is known that CLFs exist for a large class of prac- that we want the so-called virtual leader, , to follow.
ÖGREN et al.: A CONTROL LYAPUNOV FUNCTION APPROACH TO MULTIAGENT COORDINATION 849

This nonphysical leader is a reference point in the state space and

with respect to which we can define the rest of the formation.
We denote the trajectory executed by the virtual leader by
. Intuitively, one might want to set .
But, due to robustness considerations, we incorporate error This directly gives that if , then
feedback into the time evolution of (see, for example, [6]) along trajectories. Thus, the set
and let be given by is invariant, i.e., if , then
for all .
Remark IV.1: In most cases, a careful choice of can guar-
(3) antee collision avoidance.
Theorem IV.2—Finite Completion Time: If the trajectory is
Here, is a small positive constant that prevents from given as an interval, , then by using the con-
becoming singular, and is the bound of Definition II.2 or troller in (3) we can find an upper bound , such that
something smaller chosen by the user. It will be shown to be
an upper bound on the Lyapunov formation function .
The idea is to say that the formation is being respected as In other words, there is an upper bound on the completion time.
long as . is the class function of Definition Proof: Let the completion time be defined in such a way
II.2. Furthermore, is the nominal velocity that we want that
the formation to move with, and as we will see later, it holds
that when is small.

IV. THEORETICAL PROPERTIES If there exists a constant such that , then we

In the following paragraphs we will investigate what the- obviously have
oretical properties the evolution of the multi-agent formation
exhibits when letting be given by (3). We will show that if
, then remains bounded by along
trajectories for all times greater than . We will also show that The proof thus consists of finding such a lower bound on . Let
if and if is governed by (3), then reaches
in finite time. We will conclude our theoretical investiga-
tions by showing that if the formation function
is small enough.
Theorem IV.1—Error Bound: If , then
We have that

i.e., the Lyapunov formation function will never exceed .

Proof: We directly have

since . Thus

which concludes the proof.

To show that the set is in- Theorem IV.3—Formation Velocity: If the formation error is
variant, we will note that for any such that small, and , then
, we have

i.e., the formation velocity is .

For such a , it holds that Proof: We will start by showing that the right-hand term
in the brackets of (3) grows to infinity as approaches zero

as
850 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 18, NO. 5, OCTOBER 2002

by Assumption II.2. Thus, the left-hand term governs when

is small. In that case, we have

(a)

V. SIMULATION EXAMPLES
We will now go on to illustrate Theorem IV.1 and Theorem (b)
IV.3, as well as the effect of measurement noise on our proposed
approach. In the following example, we model the robots using
the standard unicycle model (see, for example, [4] and [5]). Such
a model is applicable to the Nomadic Scout, which is the robot
we work with at the Royal Institute of Technology in Stock-
holm. It is, furthermore, adequate for most all-terrain caterpillar
vehicles as well.
(c)
The equations of motion are
Fig. 1. (a) Top view of the robot trajectories. (b) The formation function.
(c) The formation velocity. The horizontal axis of the two lower plots refer to
the horizontal position of the topmost robot.
where is the center of the wheel axis, is the directional
angle, and are forward and angular velocities. The controls the successful simulation does indicate some robustness of the
are the applied force and torque. We choose the output to approach.
be the position of an off-axis point,
, perhaps the center of gravity. VI. CONCLUSION
It was shown in [9] that this model can be feedback linearized
to a two-dimensional double integrator . (This In this paper, we propose a stable coordination strategy for
property was also used implicitly in [5]). A parameterized CLF a team of formation constrained autonomous agents. A Lya-
and feedback control of a one-dimensional double integrator is punov formation function defined under standard assumptions
can be constructed from individual-CLFs. The Lyapunov for-
mation function is used to prove properties such as formation
maintenance, task completion time, and formation velocity. Fi-
nally, we present an example that illustrates the soundness of
yielding . We note that there is some our method.
flexibility in choosing the second term in , since the only hard
formation constraint is . ACKNOWLEDGMENT
Now we can choose , as in (2), and
The authors would like to thank the three anonymous re-
it is straightforward to check that
viewers for their insightful comments.
fulfills the bound and limit property of Definition II.2.
A formation of three feedback linearized dynamic unicycles
thus gives a 12–dimensional system (Fig. 1). In the first part REFERENCES
of the simulation, the three desired trajectories meet to form a [1] Z. Arstein, “Stabilization with relaxed controls,” Nonlinear Anal., vol.
7, no. 11, pp. 1163–1173, 1983.
side-by-side formation. When the robots are close to horizontal [2] T. Balch and R. C. Arkin, “Behavior-based formation control for multi-
coordinate 6 m, we increase to . This will robot teams,” IEEE Trans. Robot. Automat., vol. 14, pp. 926–939, Dec.
drive the formation function close to, but not above, the upper 1998.
[3] R. W. Beard, J. Lawton, and F. Y. Hadaegh, “A coordination architecture
limit of . When this happens, the velocity is decreased to for spacecraft formation control,” IEEE Trans. Control Syst. Technol., to
a value below . Finally, in the last part, when passing be published.
the 12–m mark, we introduce a stochastic measurement error [4] C. Canudas de Wit, B. Siciliano, and G. Bastin, Theory of Robot Con-
trol. Berlin, Germany: Springer-Verlag, 1996.
in the control of the topmost robot. Since we are already close [5] J. Desai, J. Ostrowski, and V. Kumar, “Control of formations for mul-
to the upper bound , the disturbance makes the whole tiple robots,” Proc. IEEE Int. Conf. Robotics and Automation, vol. 4, pp.
formation slow down when needed (as seen in the lowest plot) 2864–2869, May 1998.
[6] M. Egerstedt, X. Hu, and A. Stotsky, “Control of a car-like robot using
to respect the bound. This is not guaranteed by Theorem IV.1, a virtual vehicle approach,” presented at the 37th IEEE Conf. Decision
since the proof is only valid in a deterministic setting. However, and Control, Tampa, FL, Dec. 1998.
ÖGREN et al.: A CONTROL LYAPUNOV FUNCTION APPROACH TO MULTIAGENT COORDINATION 851

[7] W. Kang, N. Xi, and A. Sparks, “Formation control of autonomous Magnus Egerstedt (S’99–M’00) was born in
agents in 3D workspace,” presented at the IEEE Int. Conf. Robotics Stockholm, Sweden, in 1971. He received the B.A.
and Automation, San Francisco, CA, Apr. 2000. degree in philosophy from Stockholm University,
[8] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adap- Stockholm, Sweden, in 1996, and the M.S. degree
tive Control Design. New York: Wiley, 1995. in engineering physics and Ph.D. degree in applied
[9] J. R. T. Lawton, B. J. Young, and R. W. Beard, “A decentralized approach mathematics from the Royal Institute of Technology
to elementary formation maneuvers,” IEEE Trans. Robot. Automat., to (KTH), Stockholm, Sweden, in 1996 and 2000,
be published. respectively.
[10] P. Ögren, M. Egerstedt, and X. Hu, “A control Lyapunov function ap- He is currently an Assistant Professor in Electrical
proach to multiagent coordination,” presented at the IEEE Conf. Deci- and Computer Engineering with the Georgia Institute
sion and Control, Orlando, FL, Dec. 2001. of Technology, Atlanta. In 2000–2001, he was a Post-
[11] H. K. Khalil, Nonlinear Systems. Englewood Cliffs, NJ: Prentice-Hall, doctoral Fellow at the Division of Engineering and Applied Science, Harvard
1996. University, Cambridge, MA. His research interests include optimal control as
[12] P. Ögren, E. Fiorelli, and N. Leonard, “Formations with a mission: Stable well as modeling and analysis of hybrid and discrete event systems, with em-
coordination of vehicle group maneuvers,” presented at the 15th Int. phasis on motion planning and control of mobile robots.
Symp. Mathematical Theory of Networks and Systems, South Bend, IN,
Aug. 2002.
[13] C. Samson, M. Le Borgne, and B. Espiau, Robot Control, the Task Func-
tion Approach, ser. Oxford Engineering Science Series. Oxford, U.K.:
Clarendon, 1991. Xiaoming Hu (M’90) was born in Chengdu, China
in 1961. He received the B.S. degree from the
University of Science and Technology, Hefei, China,
in 1983, and the M.S. and Ph.D. degrees from
Arizona State University, Tempe, in 1986 and 1989,
Petter Ögren was born in Stockholm, Sweden, in respectively.
1974. He received the M.S. degree in engineering He was a Research Assistant with the Institute
physics in 1998 from the Royal Institute of Tech- of Automation, Academia Sinica, Beijing, China,
nology (KTH), Stockholm, Sweden, where he is from 1983 to 1984. From 1989 to 1990, he was
currently working toward the Ph.D. degree in the a Gustafsson Postdoctoral Fellow at the Royal
Division of Optimization and Systems Theory. Institute of Technology, Stockholm, Sweden, where
His research interests include multirobot systems, he is currently an Associate Professor. His current research interests are
navigation, and obstacle avoidance. in nonlinear feedback stabilization, nonlinear observer design, sensing and
active perseption, motion planning and control of mobile robots, and mobile
manipulation.