Stable Flocking of Mobile Agents,

Part I: Fixed Topology

Herbert G. Tanner Ali Jadbabaie and George J. Pappas
Mechanical Engineering Dept Electrical and Systems Engineering Dept
University of New Mexico University of Pennsylvania

Abstract—This is the first of a two-part paper that in- for example [15], [2], [22], [24], [6], [17], [8], [29], [12],
vestigates the stability properties of a system of multi- [20], [4], [13], [36], [23].
ple mobile agents with double integrator dynamics. In
this first part we generate stable flocking motion for the
group using a coordination control scheme which gives The animal aggregation model of [25] aimed at generat-
rise to smooth control laws for the agents. These control ing computer animation of the motion of bird flocks and
laws are a combination of attractive/repulsive and align- fish schools. It was based on three dimensional compu-
ment forces, ensuring collision avoidance and cohesion
of the group and an aggregate motion along a common tational geometry of the sort normally used in com-
heading direction. In this control scheme the topology puter animation or computer aided design. This flock-
of the control interconnections is fixed and time invari- ing model consists of three steering behaviors which
ant. The control policy ensures that all agents eventually
align with each other and have a common heading direc-
describe how an individual agent maneuvers based on
tion while at the same time avoid collisions and group the positions and velocities its nearby flockmates:
into a tight formation.
• Separation: steer to avoid crowding local flock-
mates.
I. Introduction • Alignment: steer towards the average heading of
local flockmates.
Over the last years, the problem of coordinating the • Cohesion: steer to move toward the average po-
motion of multiple autonomous agents, has attracted sition of local flockmates.
significant attention. Besides the links of this issue to
problems in biology, social behavior, statistical physics, The superposition of these three rules results in all
and computer graphics, to name a few, research was agents moving in a formation, with a common head-
partly motivated by recent advances in communication ing while avoiding collisions.
and computation. Considerable effort has been directed
in trying to understand how a group of autonomous Generalizations of this model include a leader follower
moving creatures such as flocks of birds, schools of strategy, in which one agent acted as a group leader
fish, crowds of people [34], [18], or man-made mobile and the other agents would just follow the aforemen-
autonomous agents, can cluster in formations without tioned cohesion/separation/alignment rules, resulting
centralized coordination. in leader following. Vicsek et al. [35] proposed such
a model in 1995. Although developed independently,
Similar problems have been studied in ecology and the- Vicsek’s model turns out to be a special case of [25], in
oretical biology, in the context of animal aggregation which all agents move with the same speed (no dynam-
and social cohesion in animal groups [1], [21], [37], [10], ics), and only follow an alignment rule. In [35], each
[7]. A computer model mimicking animal aggregation agent heading is updated as the average of the head-
was proposed in [25]. Following the work in [25] sev- ings of agent itself with its nearest neighbors plus some
eral other computer models have appeared in the lit- additive noise. Numerical simulations in [35] indicate a
erature (cf. [11] and the references therein), and led coherent collective motion, in which the headings of all
to creation of a new area in computer graphics known agents converge to a common value. This was quite a
as artificial life [25], [31]. At the same time, several surprising result in the physics community and was fol-
researchers in the area of statistical physics and com- lowed by a series of papers [3], [33], [32], [27], [19]. The
plexity theory have addressed flocking and schooling first rigorous proof of convergence for Vicsek’s model
behavior in the context of non-equilibrium phenomena (in the noise-free case) was given in [12].
in many-degree-of-freedom dynamical systems and self
organization in systems of self-propelled particles [35], Inspired by the results of [25], this paper introduces a
[33], [32], [19], [16], [28], [5], [14]. Related problems set of control laws that give rise to flocking behavior
have become a major thrust in systems and control the- and provides a system theoretic justification by com-
ory, in the context of cooperative control, distributed bining results from classical control theory, mechanics
control of multiple vehicles and formation control; see and algebraic graph theory. In this first part of the
paper, we consider the case where the topology of the vi
control interactions between the agents is fixed. Each
agent regulates its position and orientation based on a αi θi
fixed set of “neighbors”. In this case, the control in-
puts for the agent are smooth. The case where the set
of neighbors may change in time, depending on the rel- ŷ
ative distances between the agent and its flockmates, is
treated separately in Part II [30]. Here we show that
under fixed control interconnection topology, the sys- ri ai
tem of mobile agents is capable of coordinating itself so
that all agents attain a common heading, they cluster to
a tight formation. Collision free fashion can be guaran-
ẑ x̂
teed under sufficient network connectivity assumptions.
The control laws that ensure cohesion and separation
Fig. 1. Control forces acting on agent i.
can be decoupled from alignment.

This paper is organized as follows: in section II we de- III. Graph Theory Preliminaries
fine the problem addressed in this paper and sketch the
solution approach. In section III we give a brief in-
troduction on algebraic graph theory. The purpose of The following is a brief and selective introduction to al-
section IV is to introduce the control scheme that trig- gebraic graph theory. For more information, the reader
gers flocking and analyze the stability of the closed loop is referred to [9].
system. Results are verified in section V via numerical
simulations. Section VI summarizes and highlights new A graph G consists of a vertex set V, and an edge set E,
research directions. where an edge is an unordered pair of distinct vertices
in V.. If x, y ∈ V, and (x, y) ∈ E, then x and y are adja-
II. Problem Description cent, or neighbors and we denote this by x ∼ y. A graph
is called complete if any two vertices are neighbors. A
Consider N agents, moving on the plane with dynamics path of length r from vertex x to vertex y is a sequence
described by: of r+1 distinct vertices starting with x and ending with
y, such that consecutive vertices are adjacent. If there
ṙi = vi (1a) is a path between any two vertices of a graph G, then
G is said to be connected. An orientation in a graph
v̇i = ui i = 1, . . . , N , (1b)
is the assignment of a direction to each edge, so that
edge (i, j) is an arc from vertex i to vertex j. We de-
where ri = (xi , yi )T is the position vector of agent i,
note G σ the graph G with orientation σ. The incidence
vi = (ẋi , ẏi )T is its velocity vector and ui = (uxi , uyi )T
matrix B(G σ ) of a graph G σ is the matrix whose rows
its control (acceleration) input. The heading of agent
and columns are indexed by the vertices and edges of
i, θi , is defined as:
G respectively, such that the i, j entry of B(G) is equal
to 1 if the edge j is incoming to vertex i, −1 if edge j
θi = arctan(ẏi , ẋi ) . (2)
is outcoming from vertex i, and 0 otherwise.
Relative position vector between agents i and j is de-
noted rij = ri − rj . The symmetric matrix defined as:

The objective is for the whole group to move at a com- L(G) = B(G σ )B(G σ )T
mon speed and direction and maintain constant dis-
tances between agents. The control input for agent i is is called the Laplacian of G and is independent of the
a combination of two components (Figure 1): choice of orientation σ. It is known that the Lapla-
cian matrix captures many topological properties of the
ui = ai + αi . (3) graph. Among those, it is the fact that L is always
positive semidefinite and the algebraic multiplicity of
The first component, ai , is derived from the field pro- its zero eigenvalue is equal to the number of connected
duced by an artificial potential function, Vi , that de- components in the graph. For a connected graph, L has
pends on the relative distances between agent i and its a single zero eigenvalue, and the associated eigenvector
flockmates. This term is responsible for collision avoid- is the n-dimensional vector of ones, 1n . The second
ance and cohesion in the group. The second compo- smallest eigenvalue of L, denoted λ2 is known as the al-
nent, αi regulates the velocity vector of agent i to the gebraic connectivity of the graph because it is directly
weighted average of that of its flockmates. related with the way the nodes are interconnected.
 IV. Control Law with Fixed Topology Vij

In this section we will refine the acceleration input of

(3) into specific expressions for the components ai and
αi . To represent the control interconnections between
the agents we use a graph with a vertex correspond-
ing to each agent. The edges capture the dependence
of agent controllers on the state of other agents. Adja-
cency in the graph will thus induce a (logical) neighbor-
ing relation between agents. In Part II, this neighboring
relation will also be associated with physical adjacency.
|| rij ||
Definition IV.1 (Neighboring graph) The neigh-
boring graph, G = {V, E}, is an undirected graph con- Fig. 2. Example of an inter-agent potential function.
sisting of:

• a set of vertices (nodes), V = {n1 , . . . , nN }, indexed Consider the following positive semi-definite function
by the agents in the group, and N
1

W = (Vi + viT vi ).
• a set of edges, E = {(ni , nj ) ∈ V × V | ni ∼ nj }, con- 2 i=1
taining unordered pairs of nodes that represent neigh-
boring relations.
The level sets of W ,
Assumption IV.2 Graph G is connected.
Ω = {(vi , rij ) | W ≤ c} (6)
Since G is constant with respect to time, the above define compact sets in the space of agent velocities and
assumption ensures that G will remain connected for relative distances. This is because the set {rij , vi } such
all time. The set of all neighbors of agent i is called that W ≤ c, for c > 0 is closed by continuity. Bound-
the neighboring set, denoted: Ni
{j | i ∼ j} ⊆ edness, on the other hand, follows from connectivity:
{1, . . . , N } \ {i}. Cohesion and separation is achieved from W ≤ c we have that Vij ≤ c. Connectivity en-
using artificial potential fields [26]. In fact, although sures that a path connecting nodes i and j has length
cohesion is ensured for a connected graph, the fixed at most N − 1. Thus rij  ≤ Vij−1 c(N − 1) . Similarly,
topology of the graph cannot guarantee collision avoid- √
viT vi ≤ c yielding vi ≤ c.
ance unless the neighboring graph is complete - when
two agents are not linked, they cannot be aware of be- Due to Vi being symmetric with respect to rij and the
ing close to each other. Cohesion and separation forces fact that rij = −rji ,
exerted to a pair of neighboring agents are generated
by a potential function Vij (Figure 2) which satisfies: ∂Vij ∂Vij ∂Vij
= =− , (7)
∂rij ∂ri ∂rj
Definition IV.3 (Potential function) Potential Vij
is a differentiable, nonnegative, radially unbounded and therefore it follows:
function of the distance rij  between agents i and j, N N
d
1

such that Vi = ∇ri Vi · vi .
dt i=1 2 i=1
1. Vij (rij ) → ∞ as rij  → 0,
Theorem IV.4 (Flocking in a fixed network)
2. Vij attains its unique minimum when agents i Consider a system of N mobile agents with dynamics
and j are located at a desired distance. (1), each steered by control law (5) and assume that the
neighboring graph is connected. Then all agent velocity
Having defined Vij we can express the total potential vectors become asymptotically the same, collisions
of agent i as
between interconnected agents are avoided and the
Vi = Vij (rij ), (4) system approaches a configuration that minimizes all
j∈Ni agent potentials.

The control law ui can then be defined as: Proof: Taking the time derivative of W , we have:


 
ui = − (vi − vj ) − ∇ri Vij . (5) N N
1




j∈Ni
  
j∈Ni
  Ẇ = V̇i − viT  (vi − vj ) + ∇ri Vi  (8)
2 i=1 i=1 j∼i
αi ai
which due to the symmetric nature of Vij , simplifies to V. Simulations
 

N N


In this Section we verify numerically the stability re-
Ẇ = viT ∇ri Vi − viT  (vi − vj ) + ∇ri Vi 
sults obtained in Section IV. In the simulation example,
i=1 i=1 j∼i
the group consists of ten mobile agents with identical
N


second order dynamics. Initial positions were generated
=− viT (vi − vj ) = −v T (L ⊗ I2 )v randomly within a ball of radius R0 = 2.5[m] centered
i=1 j∼i at the origin. Initial velocities were also selected ran-
where v is the stack vector of all agent (three dimen- domly with arbitrary directions and magnitude in the
sional) velocity vectors, L is the Laplacian of the neigh- (0, 1)[m/s] range. The interconnection graph was also
boring graph and ⊗ denotes the Kronecker matrix prod- generated in random, with the only requirement being
uct. Writing the quadratic form explicitly, that it is connected.

Ẇ = −vxT Lvx − vyT Lvy (9) Figures 3-8 depict snapshots of the system’s evolution
within a time frame of 100 simulation seconds. The
where vx and vy are the stack vectors of the compo-
corresponding time instant is given below each Figure.
nents of the agent velocities along x̂ and ŷ directions
The position of each agent is represented by a small dot
(Figure 1), respectively.
and the neighboring relations by line segments connect-
For a connected graph G, L is positive semidefinite and ing them. Velocity vectors are depicted as arrows, with
the eigenvector associated with the single zero eigen- their base point being the position of the corresponding
value is 1. Thus Ẇ = 0 implies that both vx and vy agent. Dotted lines show the trajectory trails for each
belong to span{1}. This means that all agent velocities agent. Simulation verifies that the system converges to
have the same components and are therefore equal. It an invariant set that corresponds to a tight formation
follows immediately that ṙij = 0, ∀ (i, j) ∈ N × N . and a common heading direction. The shape of the for-
mation which the group converges to, is determined by
Application of LaSalle’s invariant principle establishes the artificial potential functions.
convergence of system trajectories to S = {v | Ẇ = 0}.
In S, the agent velocity dynamics become:
 
 ∇r V1 ..
1
 . 
v̇ = − .. = −(B ⊗ I2 )  ∇rij Vij  (10)
. ..
∇rN VN
.
which can be expanded to
v̇x = −B[∇rij Vij ]x , v̇y = −B[∇rij Vij ]y .
Thus, v̇x and v̇y belong in the range of the incidence ma-
trix B. For a connected graph, range(B) = span{1}⊥
and therefore
v̇x , v̇y ∈ span{1}⊥ . (11)
In an invariant set within S, 0

vx , vy ∈ span{1} ⇒ v̇x , v̇y ∈ span{1}. (12)

Fig. 3. Initial configuration.
Combining (11) and (12),
VI. Conclusions
v̇x , v̇y ∈ span{1} ∩ span{1}⊥ ≡ {0}.
Thus, in steady state agent velocities must not not In this paper we demonstrate how a group of au-
change. Furthermore, from (10) it follows that in steady tonomous mobile agents, can cooperate to exhibit a
state the potential Vi of each agent i is minimized. In- flocking behavior. Flocking requires all the agents to
terconnected agents cannot collide since this will result have a common heading and stay close to each other
in Vi → ∞ and the system departing Ω, which is a while avoiding collisions. We model flocking and in-
contradiction since Ω is positively invariant. troduce local controllers that establish a stable, coor-
dinated flocking motion. These local controllers rely
Remark IV.5 Collision avoidance between all agents on a fixed communication (or sensing) network, that
can only be guaranteed with this control scheme when all allows the exchange of state information between inter-
agents are interconnected to each other. This requires connected agents. The stability of the group motion fol-
the neighboring graph to be complete. lows from the connectivity properties of the underlying
 2.22842 22.8931

Fig. 4. Initial maneuvering may bring agents close. Fig. 6. Velocity vectors converge.

5.21464 55.0766

Fig. 5. Potential forces ensure cohesion. Fig. 7. The group moves in the same direction.

network topology. The case where the interconnection [4] J. P. Desai, J. P. Ostrowski, and V. Kumar. Mod-
topology is dynamic, is treated separately in [30]. eling and control of formations of nonholonomic
mobile robots. IEEE Transactions on Robotics and
Acknowledgment Automation, 17(6):905–908, 2001.
[5] M.R. Evans. Spontaneous symmetry breaking in a
The authors would like to thank Glenn Vinnicombe for one dimensional driven diffusive system. Physical
his help in the mathematical formulation. Review Letters, 74:208–211, 1995.
[6] J. Alexander Fax and Richard M. Murray. Graph
References Laplacians and stabilization of vehicle formations.
15th IFAC Congress, Barcelona, Spain, 2002.
[1] C. M. Breder. Equations descriptive of fish schools [7] G. Flierl, D. Grunbaum, S. Levin, and D. Olson.
and other animal aggergations. Ecology, 35:361– From individuals to aggregations: the interplay be-
370, 1954. tween behavior and physics. Journal of Theoretical
[2] Jorge Cortes, Sonia Martinez, Timur Karatas, and Biology, 196:397–454, 1999.
Francesco Bullo. Coverage control for mobile sens- [8] V. Gazi and K. M. Passino. Stability analysis of
ing networks. IEEE Transactions on Robotics and swarms. IEEE Transactions on Automatic Con-
Automation, 2002. (submitted). trol, 48(4):692–696, April 2003.
[3] A. Czirok, A. L. Barabasi, and T. Vicsek. Collec- [9] C. Godsil and G. Royle. Algebraic Graph Theory.
tive motion of self-propelled particles:kinetic phase Springer Graduate Texts in Mathematics # 207,
transition in one dimension. Physical Review Let- New York, 2001.
ters, 82:209–212, 1999. [10] D. Grunbaum and A. Okubo. Modeling social ani-
 dynamic multi-agents. In IEEE Conference on De-
cision and Control, Las Vegas, NV, 2002.
[23] A. Pant, P. Seiler, T. J. Koo, and K. Hedrick.
Mesh stability of unmanned aerial vehicle clusters.
Proceedings of the American Control Conference,
1:62–68, June 2001.
[24] John H. Reif and Hongyang Wang. Social poten-
tial fields: A distributed behavioral control for au-
tonomous robots. Robotics and Autonomous Sys-
tems, 27:171–194, 1999.
[25] C. Reynolds. Flocks, birds, and schools: a dis-
tributed behavioral model. Computer Graphics,
21:25–34, 1987.
[26] E. Rimon and D. Koditschek. Exact robot
100
navigation using artificial potential functions.
IEEE Transactions on Robotics and Automation,
Fig. 8. Steady state. 8(5):501–518, October 1992.
[27] C. P. Schenk, P Schutz, M. Bode, and H.G. Pur-
mal aggregations. Frontiers in Theoretical Biology, wins. Interaction of self organized quasi parti-
100 of Lecture Notes in Biomathematics:296–325, cles in a two dimensional reaction-diffusion system:
1994. The formation of molecules. Physical Review E,
58:6480–6486, 1998.
[11] http://www.red3d.com/cwr/boids/.
[28] N. Shimoyama, K. Sugawa, T. Mizuguchi,
[12] A. Jadbabaie, J. Lin, and A. S. Morse. Coordina-
Y. Hayakawa, and M. Sano. Collective motion in
tion of groups of mobile autonomous agents using
a system of motile elements. Physical Review Let-
nearest neighbor rules. IEEE Transactions on Au-
ters, 76:3879–3873, 1996.
tomatic Control, 48(6):988–1001, July 2002.
[29] P. Tabuada, G. J. Pappas, and P. Lima. Feasi-
[13] E.W. Justh and P.S. Krishnaprasad. A simple con- ble formations of multi-agent systems. In Proceed-
trol law for UAV formation flying. Technical Re- ings of the American Control Conference, Arling-
port 2002-38, Institute for Systems Research, 2002. ton, VA, 2001.
[14] M. Kardar, G. Parisi, and Y.-C. Zhang. Dynamic [30] Herbert G. Tanner, Ali Jadbabaie, and George J.
scaling of growing interfaces. Physical Review Let- Pappas. Stable flocking of mobile agents, Part II:
ters, 56:889–892, 1986. Dynamic topology. In Proceedings of the IEEE
[15] N. Leonard and E. Friorelli. Virtual leaders, arti- Conference on Decision and Control, 2003. (to ap-
ficial potentials and coordinated control of groups. pear).
IEEE Conference on Decision and Control, 2001. [31] D. Terzopoulos. Artificial life for computer graph-
[16] H. Levine and W. J. Rappel. Self organization in ics. Communications of the ACM, 42(8):32–42,
systems of self-propelled particles. Physical Review 1999.
E, 63:208–211, 2001. [32] J. Toner and Y. Tu. Long range order in a two di-
[17] Y. Liu, K. M. Passino, and M. M. Polycar- mensional xy model: How birds fly together. Phys-
pou. Stability analysis of m-dimensional asyn- ical Review Letters, 75:4326–4329, 1995.
chronous swarms with a fixed communication [33] J. Toner and Y. Tu. Flocks, herds, and schools:
topology. IEEE Transactions on Automatic Con- A quantitative theory of flocking. Physical Review
trol, 48(1):76–95, February 2003. E., 58:4828–4858, 1998.
[18] D. J. Low. Following the crowd. Nature, 407:465– [34] T. Vicsek. A question of scale. Nature, 411:421,
466, 2000. May 24 2001.
[19] A. S. Mikhailov and D. Zannette. Noise in- [35] T. Vicsek, A. Czirok, E. Ben Jacob, I. Cohen, and
duced breakdown of collective coherent motion in O. Schochet. Novel type of phase transitions in
swarms. Physical Review E, 60:4571–4575, 1999. a system of self-driven particles. Physical Review
[20] P. Ögren, M. Egerstedt, and X. Hu. A control Letters, 75:1226–1229, 1995.
Lyapunov function approach to multi-agent coor- [36] R. Vidal, O. Shakernia, and S. Sastry. Dis-
dination. IEEE Transactions on Robotics and Au- tributed formation control with omnidirectional vi-
tomation, 18, October 2002. sion based motion segmentation and visual servo-
[21] A Okubo. Dynamical aspects of animal grouping: ing. Robotics and Automation Magazine, 2003. (to
swarms, schools, flocks, and herds. Advances in appear).
Biophysics, 22:1–94, 1986. [37] K. Warburton and J. Lazarus. Tendency-distance
[22] R. Olfati and R. M. Murray. Distributed struc- models of social cohesion in animal groups. Journal
tural stabilization and tracking for formations of of Theoretical Biology, 150:473–488, 1991.