2008 IEEE International Conference on

Robotics and Automation

Pasadena, CA, USA, May 19-23, 2008

Control of Swarms Based on Hydrodynamic Models

Luciano C. A. Pimenta, Nathan Michael, Renato C. Mesquita, Guilherme A. S. Pereira, Vijay Kumar

Abstract— We address the problem of pattern generation formation control are applied in [6] and [7] but in these cases
in obstacle-filled environments by a swarm of mobile robots. proofs of stability of the given formation are not discussed.
Decentralized controllers are devised by using the Smoothed The use of artificial potential fields to address the pattern
Particle Hydrodynamics (SPH) method. The swarm is modelled
as an incompressible fluid subjected to external forces. Actual generation task in the absence of obstacles is discussed in [8]
robot issues such as finite size and nonholonomic constraints are and [9]. In these works, scalable approaches are derived and
also addressed. Collision avoidance guarantees are discussed. proofs of convergence and stability are given. The problem
Finally, in the absence of obstacles, we prove for the first time of staying connected during the task is also considered in [9].
stability and convergence of controllers based on the SPH. In contrast to the others, in this work we use the Smo-
I. INTRODUCTION othed Particle Hydrodynamics method to model the swarm
emulating an incompressible fluid in an obstacle-filled envi-
The focus of this work is on using analogies with fluid ronment. We deal with static obstacles by computing a global
dynamics models to control swarms of robots. The main potential function and using virtual particles. The incompres-
motivation stems from the fact that a great variety of charac- sible fluid model allows for a loose way of controlling the
teristics desirable for a large group of robots may be observed connectivity of the swarm. We provide a decentralized and
in fluids. Some examples of such characteristics are: (i) fluids scalable approach which relies only on local information.
are easily deformed, (ii) fluids can easily flow around objects, Moreover, there is no need to label the agents in the swarm.
and (iii) the flow field variables and also the fluid phase can
be easily manipulated in order to design desired behaviors. II. SMOOTHED PARTICLE HYDRODYNAMICS
Other works have already mimicked fluid behaviors to The Smoothed Particle Hydrodynamics (SPH) is a mesh-
control large groups of robots. The authors of [1] used free particle numerical method. It is a particle numerical
Stokesian Dynamics. By using this technique the robots method since it employs a set of finite number of disordered
had the behavior of particles suspended in a fluid and the discrete particles to represent the state of the simulated
group shape could be controlled without losing the group system. It is mesh-free due to the fact that it is not necessary
coherence. In [2] the kinetic theory of gases was used to to generate a mesh to provide connectivity of the particles.
sweep a group of robots through a bounded region. Additionally, SPH is considered a Lagrangian method, which
In this work we are interested in the so-called pattern means that the individual particles are modelled in space and
generation problem, which may be stated as follows: in time in contrast to Eulerian methods that model the flux
Given N robots and any initial spatial distribution, the of material through a control volume fixed in space.
geometry of the environment with static obstacles defining a The continuum governing equations of fluid dynamics
compact domain Ω ⊂ R2 , and a curve Γ : I → Ω, where capture three salient phenomena: (i) conservation of mass;
I ⊂ R, find a controller which enables the robots, without (ii) conservation of momentum; and (iii) conservation of
colliding with static obstacles and each other, to form Γ. energy. For inviscid compressible fluids, in the absence of
Possible applications of an efficient solution to this task heat flux, the Lagrangian description is given by:
are surveillance and cordoning off of hazardous areas. We Dρ
are interested in a decentralized solution which scales from = −ρ∇ · v , (1)
Dt
tens to hundreds of robots. Some approaches as in [3] Dv ∇P
assume that each robot knows the positions of all the others. =− , (2)
Dt ρ
To achieve scalability, it is more interesting to have an

approach that relies only on local information. Approaches De P
=− ∇·v, (3)
that depend on labelling the robots are also hard to scale Dt ρ
since it may be difficult to uniquely identify agents in the where ρ is density, v is velocity, e is the internal energy per
swarm. Approaches based on leader-follower controllers are unit of mass, P is the hydrostatic pressure, ∇ is the gradient
examples of this class [4], [5]. Behavior-based techniques for operator, ∇ · () is the divergence, and D/Dt is the total
time derivative which is physically the time rate of change
L. C. A. Pimenta, R. C. Mesquita, and G. A. S. Pereira are with
the Department of Electrical Engineering, Universidade Federal de Minas following a moving fluid element.
Gerais, Belo Horizonte, MG, 31270-901, Brazil. {lucpim, renato, In the SPH method, the continuum equations of fluid
gpereira}@cpdee.ufmg.br dynamics are converted to a set of ordinary differential
N. Michael and V. Kumar are with the GRASP Laboratory, Uni-
versity of Pennsylvania, Philadelphia, PA 19104, USA. {nmichael, equations, where each one controls the evolution of an attri-
kumar}@seas.upenn.edu bute of a specific particle. This conversion is performed by

978-1-4244-1647-9/08/$25.00 ©2008 IEEE. 1948

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.
 1.8

1.6 be done by using the equation of state [11]:

1.4

γ
ρi
Pi = Bi −1 ,
1.2

(10)
1

ρ0

W
0.8

where ρ0 is the reference density (1000kg/m3 in the case

0.6

0.4

0.2 of water), γ is the ratio of specific heats, and Bi is the bulk

0
−2.5 −2 −1.5 −1 −0.5 0
q
0.5 1 1.5 2 2.5 modulus. The bulk modulus is computed to guarantee a small
Mach number, M , (typically 0.1 − 0.01 ). The following
Fig. 1. One-dimensional kernel function. The solid, dashed, and dashdot expression may be used [11]:
lines correspond respectively to h = 1, h = 0.75, and h = 0.5.
2
vmax
Bi = ρi , (11)
M
using differentiable interpolation kernels, which approximate

a delta function. The kernel is chosen to satisfy Υ W (q − where vmax is the maximum velocity of the flow. For
q
, h)dq
= 1 and limh→0 W (q − q
, h) = δ(q − q
), liquids, the speed of sound of a particle i, which represents
where Υ is the volume that contains q, h is a parameter the speed at which sound travels through the fluid element
that controls the influence area of W , and δ(x − x
) is the represented by the particle, is given by ci = Bi /ρi .
Dirac delta function. The error in approximating the integral The authors of [12] proposed decentralized controllers for
representation of a function by summations of the function large groups of robots based on the SPH method. In [12], the
evaluated at particle locations weighted by interpolation SPH equations for compressible fluids were used to mimick
kernels is O(h2 ) [10]. In this work, we use the cubic spline: the behavior of air at 20o C. In our previous work [13] the
 SPH method was used to provide decentralized controllers
1 − 3 κ2 + 3 κ3 if 0 ≤ κ ≤ 1,
10  1 2 3 4 in a pattern generation task with static obstacles. In the next
W (q, h) = (2 − κ) if 1 ≤ κ ≤ 2,
7πh2  4 section, we present extensions of the technique in [13].
0 otherwise ,
(4) III. CONTROL POLICY
where κ = q/h. It can be observed that the function We assume that each robot of the team is a SPH particle
support is determined by 2h (see Fig. 1). subjected to an external force, and since we use kernels with
The SPH conservation equations for a particle i are: compact support, it is possible to derive decentralized control

ρi = mj W (qij , h) , (5) laws based on the SPH equations. The resulting controllers
j
are decentralized in the sense that only local information is
 necessary: the gradient of a potential function at the location
dvi  Pi Pj of the robot i and position and velocity of the robot i itself
=− mj 2 + 2 + Πij ∇i Wij + fi , (6) and of the robots in the neighborhood of i. For a robot i
dt ρi ρj
j with configuration qi = [xi , yi ]T we define Ni as the set of
 robots in the neighborhood of robot i:
dei 1 Pi Pj
= mj 2 + 2 + Πij vij · ∇i Wij , (7) Ni = {j = i | qj − qi  < D}.
dt 2 j ρi ρj (12)
The distance D is determined by the kernel support size,
where qi is the particle vector position [xi , yi ]T , qij = qi −
which in the case of the kernel in (4) is given by D = 2h.
qj , Wij = W (qij ), vij = vi − vj , fi is the sum of external
forces normalized by the mass mi , and Πij is a dissipative A. Global Potential Functions
term called artificial viscosity added to handle shocks. Our approach relies on the computation of a global po-
There are several variants for the viscosity term, the most tential function. In this subsection we present two examples
common of which is given by [10]: of such functions: harmonic functions [13] and shape func-
1 tions [9]. Harmonic functions can be efficiently computed in
ρij (−ξ1 cij µij + ξ2 µ2ij ) if vij · qij < 0 ,
Πij = obstacle-filled environments. To numerically compute these
0 if vij · qij > 0 , functions, we use the Finite Element Method (FEM). The
(8) efficiency of such a method is due to its ability to work
where properly with unstructured meshes which are used to exactly
hvij · qij
µij = . (9) decompose the solution domain.
qij 2 + η 2
If a safety factor, , is defined such that the desired pattern
In (8), ρij is the average between the densities of i and j, is represented by a region between two curves Γ1 and Γ2 ,
ξ1 and ξ2 are positive viscosity constants, cij is the average we can define a harmonic function which drives the robots
speed of sound, and η 2 is a term added to avoid singularities. toward the goal region and at the same time drives the
The motion of incompressible fluids can also be simulated robots away from the obstacles. If the desired pattern, Γ,
using the SPH method. The key idea is to make a compres- is parameterized by a function s(x, y) = 0, then Γ1 is
sible fluid behave like a nearly incompressible one. This can such that s(x, y) =  and Γ2 is such that s(x, y) = −.

1949

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.
 (a) (b)

Fig. 2. Domain example.

Figure 2 presents an example of a domain with a circular

pattern. Harmonic functions are solutions to the Laplace’s
equation. In order to guarantee uniqueness in the solution, we (c) (d)
must define boundary conditions. We use constant Dirichlet
boundary conditions such that a maximum value is obtained Fig. 3. Simulation with 121 point robots from a starting configuration
at the boundaries of the configuration space and a minimum (Fig. 3(a)) to the goal (Fig. 3(d)), with intermediate configurations (Figs.
value is obtained at the desired pattern. This boundary value 3(b) and 3(c)).
problem (BVP) is given by:
 2 damping to stabilize the system. It is important to mention
 ∇ φ = 0, that ui (q) in (15) can be computed by taking into account
Γ1 = Γ2 = 0 , (13) only robots in the neighborhood Ni defined in (12) due to

∂Ω1 = ∂Ω2 = P = Vc , the compact support of the kernel, W . Figure 3 presents
where φ is the harmonic function, Vc is a positive constant, a simulation with 121 point robots with control law given
and P is a point inside desired closed curves (see Fig. 2). by (15). Additional simulations can be found in [13].
In obstacle-free environments with desired smooth star C. Finite-size, Nonholonomic Robots
shapes, we can use shape functions. For a desired curve
Now, we will describe how our approach may be adapted
parameterized by a function s(x, y) = 0, a shape function, φ,
to take into account actual robot issues. The first issue we
is defined such that φ is a positive semi-definite function with
address is the finite size of actual robots. The static obstacles
a minimum value equal to zero at the boundary Γ. In this
are directly taken into account since we plan our potential
case, the pattern may be described in polar coordinates by
functions in the robots configuration space. We also assume
r = (α) and the following shape function may be applied:
that our robots are circular in shape with radius R. Given
φ = ((r − (α)) ◦ f )2 , (14) two robots, we guarantee that the robots do not collide with
each other if qij  ≥ 2R + ε, where ε is a safety factor.
where f : (x, y) → (α, r). The collision avoidance of our approach is performed by the
B. Holonomic Point Robot Abstraction artificial viscosity term in (8) with
Our controller is derived by considering each robot as a hvij · qij
µij = . (18)
SPH particle at qi = [xi , yi ]T subjected to an external force (qij  − (2R + ε))2
given by the descent gradient of a global potential function. This adaptation guarantees a repulsive term in (16)
Under the assumption of fully-actuated, holonomic, point between robots which are moving toward each other. This
robots, each robot’s acceleration is given by q̈i = ui (q, t), term is repulsive since Πij ≥ 0 and ∇i Wij points in the
where q = [qT1 , . . . , qTN ]T . The controller for each robot is: direction of −qij . Note that Πij → ∞ when qij  →
ui (q) = bi − ζvi + kfi , (15) (2R + ε), i.e., when the robots are about to collide.
Besides the robot size, motion constraints are also an im-
where portant consideration. In our experiments we use differential

 Pi Pj
drive, kinematically controlled robots. To control such robots
bi = − mj 2 + 2 + Πij ∇i Wij , (16) we use feedback linearization:
ρi ρj
j
v cos(θ) sin(θ) ẋd
= · , (19)
k and ζ are positive constants and fi is given by a vector ω − sin(θ)
d
cos(θ)
d
ẏd
defined by −∇φ. In fact, we use vector fields of the form:
where v and ω are linear and angular velocities, respectively,
∇φ(qi ) and θ is the robot orientation. The parameter d defines a
− ∇φ(q if ∇φ(qi ) = 0
fi = i )
β
, (17) point, [xd , yd ]T , in the global frame, which corresponds to
0 if ∇φ(qi ) = 0
[d, 0]T in the robot frame. Now, each robot is represented in
where β is a non-negative integer number. In (15) we include its configuration space by the point [xdi , ydi ]T such that the
a dissipative term proportional to vi , which represents a physical extent of the robot lies within the circle of center

1950

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.
 [xdi , ydi ]T and radius R
= R + d. The SPH particles are where e
i is the part of the internal energy related to
also placed at the points [xdi , ydi ]T . Since our controllers conservative forces such that:
were devised for robots fully actuatedin their acceleration 
we compute the integral [ẋdi , ẏdi ]T = ui (q)dt. de
i 1 Pi Pj
= m + 2 vij · ∇i Wij (23)
dt 2 j ρ2i ρj
D. Virtual Particles
When controlling multiple robots, due to the presence of and e0 is the minimum internal energy which is obtained
inter-particle forces, bi , the external force, fi , may be not when vi = 0 and ρi = ρ0 for all i. Consider also the set
enough to avoid collisions. We add virtual particles right Ωc = {x ∈ X|V (q, v) ≤ c}, where X is the state space
at the boundaries of the configuration space such that we defined by x = [qT1 v1T . . . qTN vN ] . Given the set S with
T T

can guarantee collision avoidance. In fact, we want to take boundary determined by the desired pattern Γ and the system
advantage of the collision avoidance property provided by the of robots defined in Proposition 1 with initial conditions x0 ∈
artificial viscosity. There are several ways of implementing Ωc , the system converges to an invariant set, ΩI ⊂ Ωc , such
this virtual particle idea. One option is to create a virtual that the points in ΩI minimize the measure function φS .
particle at the closest boundary point, p. Then we adapt the
term bi in (16) such that: Proof: Since V is continuous, we conclude that  Ωc is
 closed for some c > 0. Also, due to the fact that φS + i (e
i −
 Pi Pj e0 ) ≤ c and vT v ≤ c we conclude that Ωc is compact.


bi =− mj 2 + ρ2 + Πij ∇i Wij (h)
ρ i j (20) We have that
j
  de

− λΠip ∇i Wip (h
) , V̇ = (k∇φTi q̇i + viT v̇i ) + i
. (24)
dt
where λ is a positive constant, j iterates only through the N i i
particles that represent real robots, and p refers to the virtual By using (15) and (23), and the fact that ∇i Wij =
particle. Due to the fact that the size of the robot is already −∇j Wji and Πij = Πji
taken into account in the configuration space, we use R = 0  1
in (18). Note that the virtual particle does not change the V̇ = − ζviT vi − mΠij vij
T
∇i Wij ≤ 0.
density ρi and also does not have its own density. The other i i
2 j
terms necessary to compute Πip are ρip = ρi and cip = ci .
By using the LaSalle’s Invariance Principle, we conclude
Another option is to assign virtual particles to each cell
that for any x0 ∈ Ωc the system converges asymptotically
with obstacle in a local occupancy grid. This option was
to the largest invariant set ΩI = {x ∈ X|V̇ = 0}, which
found to be the most robust during experiments.
corresponds to vi = 0 ∀i with ΩI ⊂ Ωc . Since ΩI contains
IV. ANALYSIS all equilibrium points in Ωc and based on the Proposition 1
Our stability and convergence analysis is built upon the we conclude that all points in ΩI minimize φS .
results in [9]. The next four results correspond to our version
of the Results 4.1 to 4.4 in [9]. We assume obstacle-free Proposition 3 Consider the set ΩS defined by
environments and fi in (15) is given by −∇φ, where φ is
ΩS = {x|φ(qi ) = 0, vi = 0, ρi = ρ0 ∀i} , (25)
the shape function in (14). We also assume that the robots
are represented by identical SPH particles with mass m. where φ is a shape function. Given the system of N robots
Before presenting our results we define the function φS (q) defined in Proposition 1, the set ΩS is a stable submanifold
as a measure of performance that should be minimized: and ΩS ⊂ ΩI .

φS (q) = k φ(qi ). (21) Proof: Since vi = 0 for all i, ΩS ⊂ ΩI . The potential
i
energy of the system is given by U = φS + i e
i . We need
Proposition 1 Given a system of N point robots with dyna- to show that the hessian of U , HU = HφS + H i e
i , is
mics q̈i = ui (q, t) and a control law determined by (15), positive semi-definite when qi ∈ Γ and ρi = ρ0 ∀i.
where fi = −∇φ and φ is a shape function, the system It is proved in [9] that the 2N ×2N matrix HφS is positive
equilibrium points are at an extremum of φS . semi-definite when φ(qi ) = 0. Therefore, we need to prove
that H i e
i ≥ 0 when ρi = ρ0 . By using ∇i Wij = −∇j Wji
Proof: Since the system is in equilibrium we have q̈ = in (23) we can write after some algebra
0
 and q̇ = 0. We have also Πij = 0 and ui = 0. Therefore,
  
∂ n e
n 
i i = 0. Since ∇i Wij =
u −∇j Wij , we have i ui = Pi Pj
 = m 2 + ρ2 ∇i Wij . (26)
k i ∇φi = 0. However, k i ∇φi = 0 is the necessary ∂qi ρi j
j
condition for φS to be at an extremum.
After using the state equation (10), computing the second
Proposition 2 Consider the positive semi-definite function: derivatives, and using ρi = ρ0 , we can obtain
 1
V = φS + (e
i − e0 ) + vT v , (22) m2 γσ
2 H i e
i = AAT ≥ 0 , (27)
i ρ20

1951

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.
 where σ is a positive constant and
  W1k 
k ∂ ∂x1 ∂W 12
... ∂W 1N

  ∂ W1k ∂ W12 . . .
∂x1 ∂x1
W1N 
 k ∂y1 ∂y1 ∂ ∂y1 
 
A= 
..
.
..
.
..
.
..
.
.

  WN k 
 ∂W N1
∂ WN 2
. . . ∂ 
∂xN ∂xN k W ∂xN
∂W∂yN
N1
∂ WN 2
∂yN . . . k ∂ ∂yN
Nk

Fig. 4. The 20 × 13.5 × 22.2 cm3 differential drive robotics platform.

Proposition 4 For any smooth star shape, Γ, the system of

N robots defined in Proposition 1, with ρi = ρ0 ∀i during In the experiments we used a team of seven differen-
all instants of time t, converges to Γ for any x0 ∈ Ωc . tial drive, kinematically controlled robots called Scarabs
(see Fig. 4). Details of the Scarabs are described in [14]. The
Proof: If ρi = ρ0 , ∀i, during all instants of time t, team of robots were provided with a map of the environment
then the equilibrium of the system is given by ∇i φ(qi ) = 0, which was defined by the boundaries of the experimental
∀i. Proposition 2 guarantees the system converges to ΩI . area and static obstacles. A vector field based on harmonic
As proven in [9] we have that ∇i φ(qi ) = 0 if and only if functions was computed off-line. Each robot computed its
qi ∈ Γ. Therefore ΩI ≡ ΩS ≡ Γ. location in the map based on localization information from
Next, we present results concerning collision avoidance. an overhead tracking system and current velocity from its
motor controller. This information was broadcasted over
Proposition 5 Given a pair of robots, i and j, with radius the network to the other robots. To emulate the notion of
R, dynamics q̈i = ui (q, t), and control law (15), where µij neighborhoods in the smaller experimental area, each robot
is given by (18), the robots will never collide with each other. ignored messages from robots a distance greater than 2 m. At
every update of the control algorithm, each robot computed
Proof: The worst case happens when the robots i and j its current SPH state based on its local information and the
drive with opposite maximum finite velocities vm and −vm , information received over the network. Additionally, each
respectively, and bi and fi assume constant maximum finite robot incorporated virtual particles based on the map within
values in the direction qji , and bj and fj assume constant a region 2h
× 2h
, with h
= 0.3 m.
maximum finite values in the direction qij . Since −∇i Wij A vignette of one trial run of the implementation is shown
points in the direction of qij and Πij ≥ 0 we conclude that in Fig. 5. The control law was executing at an update rate of
the term given by the artificial viscosity is a repulsive term. 10 Hz while the inter-agent network communications were
Due to the symmetry we have in this case vi = −vj . Further, executing at a higher rate of 20 Hz to accommodate the
by checking the expression of µij in (18) one should notice asynchronous system design. One can see that two robots
that the artificial viscosity term corresponds to a nonlinear did not converge to the goal. This was due to limitations of
damping −F (qi , qj )vi −G(qi , qj )vi2 , where F ≥ 0 and G ≥ the minimum allowed velocities of the actual robots. Since
0. The artificial viscosity is active in the interval qij  < 2h, one of the robots that had already reached the goal was not
and since F → +∞ and G → +∞ when qij  → (2R + ε) moving, the two other ones could not converge.
we can guarantee that vi → 0 somewhere in the interval We implemented simulations using the 3D environment
2R + ε ≤ qij  < 2h. Due to symmetry we can guarantee GAZEBO [15], which allows for simulating conditions that
that vij · qij ≥ 0 when qij  → 2R + ε. are very close to the real world ones. We used robot models
that capture the geometry, kinematics, and dynamics of the
Proposition 6 Given a robot, i, with dynamics q̈i = Scarab robots. Moreover, we considered a virtual world with
ui (q, t), and the control law determined by (15), with virtual the same dimensions of the area where the experiments were
particles defined as in Subsection III-D, the robot will never performed. The obstacles and the goal were also the same.
collide with static obstacles. A simulation trial with 15 robots is presented in Fig. 6.
Proof: We just need to guarantee that vi · (qi − p) ≥ 0
VI. CONCLUSIONS
when qi − p − ε → 0, where p is the closest point in the
boundary of the configuration space. We have that vp = 0 In this work we extended the results of a novel decen-
and b
i is determined by (20). The arguments of the proof tralized approach to control swarms of robots presented in
of Proposition 5 complete the proof. [13] to the case of finite size and differential drive robots.
We used the SPH technique to model the swarm as an
V. EXPERIMENTAL AND SIMULATION RESULTS incompressible fluid and also a global potential function
In this section, we present experimental and simulation that allows for solving the pattern generation task in the
results that verify the proposed approach for finite size and presence of obstacles. We chose the potential function to
nonholonomic robots. Movies can be seen on the web page be a harmonic function. The Finite Element Method was
http://kumar.cis.upenn.edu/∼lucpim/icra2008. used to compute this function which guarantees efficiency in

1952

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.
 2.5

1.5

0.5

y (m)
0

−0.5

−1

−1.5

−2

−2.5
−3 −2 −1 0 1 2 3 4
x (m)

(a) (b)
(a) (b)
2.5

1.5

0.5
y (m)

−0.5

−1

−1.5

−2

−2.5
−3 −2 −1 0 1 2 3 4
x (m)

(c) (d)
(c) (d)
2.5

1.5

1
Fig. 6. Simulation with 15 robots from a starting configuration (Fig. 6(a))
0.5 to the goal (Fig. 6(d)), with intermediate configurations (Figs. 6(b) and 6(c)).
y(m)

−0.5

−1

−1.5

−2
W911NF-05-1-0219, and the NSF ITR Grant 0324977.
−2.5
−3 −2 −1 0 1 2 3 4

R EFERENCES
x (m)

(e) (f)
[1] M. Shimizu, A. Ishiguro, T. Kawakatsu, Y. Masubuchi, and M. Doi,
2.5
“Coherent swarming from local interaction by exploiting molecular
2 dynamics and stokesian dynamics methods,” in Proc. IEEE/RSJ Int.
1.5

1
Conf. Intel. Robots and Systems, vol. 2, 2003, pp. 1614 – 1619.
0.5 [2] W. Kerr and D. F. Spears, “Robotic simulation of gases for a
y (m)

−0.5
surveillance task,” in Proc. IEEE/RSJ Int. Conf. Intel. Robots and
−1

−1.5
Systems, 2005, pp. 2905–2910.
−2 [3] K. Sugihara and I. Suzuki, “Distributed algorithms for formation of
−2.5
−3 −2 −1 0
x (m)
1 2 3 4 geometric patterns with many mobile robots,” Journal of Robotic
Systems, vol. 13, no. 3, pp. 127–139, 1996.
(g) (h) [4] J. P. Desai, J. P. Ostrowski, and V. Kumar, “Modeling and control of
formations of nonholonomic mobile robots,” IEEE Trans. Robot. and
2.5
Automat., vol. 17, no. 6, pp. 905–908, 2001.
2

1.5
[5] H. C.-H. Hsu and A. Liu, “Multiple teams for mobile robot formation
1

0.5
control,” in Proceedings of the IEEE International Symposium on
Intelligent Control, 2004, pp. 168–173.
y (m)

−0.5

−1
[6] T. Balch and R. C. Arkin, “Behavior-based formation control for
−1.5 multirobot teams,” IEEE Trans. Robot. and Automat., vol. 14, no. 6,
−2

−2.5
pp. 926–939, 1998.
−3 −2 −1 0 1 2 3 4
x (m)
[7] Z. Cao, M. Tan, S.Wang, Y. Fan, and B. Zhang, “The optimization
(i) (j) research of formation control for multiple mobile robots,” in Proc.
of the 4th World Congress on Intel. Control and Automat., 2002, pp.
1270–1274.
Fig. 5. Experimental Results. A team of seven robots, starting from an [8] L. Chaimowicz, N. Michael, and V. Kumar, “Controlling swarms of
initial configuration (Fig. 5(a)), control around an obstacle (Figs. 5(c), 5(e), robots using interpolated implicit functions,” in Proc. IEEE Int. Conf.
and 5(g)) to a goal circular formation (Fig. 5(i)). The results in the Robot. Automat., Barcelona, Spain, April 2005, pp. 2498–2503.
configuration space are also shown in Figs. 5(b), 5(d), 5(f), 5(h), and 5(j). [9] M. A. Hsieh and V. Kumar, “Pattern generation with multiple robots,”
in Proc. IEEE Int. Conf. Robot. Automat., Orlando, USA, May 2006,
geometrically complicate domains. By adapting the artificial pp. 2442–2447.
viscosity term and also by including virtual particles we [10] J. J. Monaghan, “Smoothed particle hydrodynamics,” Annual Review
of Astronomy and Astrophysics, vol. 30, pp. 543–574, 1992.
were able to guarantee that the robots are driven safely. [11] ——, “Simulating free surface flow with SPH,” Journal of Computa-
For obstacle-free environments we used shape functions and tional Physics, vol. 110, pp. 399–406, 1994.
proved stability and convergence of our controllers. [12] J. R. Perkinson and B. Shafai, “A decentralized control algorithm for
scalable robotic swarms based on mesh-free particle hydrodynamics,”
Our approach is scalable and there is no need for labelling in Proc. of the IASTED Int. Conf. on Robot. and Applications, 2005,
the robots. The approach is also robust to dynamic deletion pp. 1–6.
and addition of agents. Since the robots try to keep the [13] L. C. A. Pimenta, M. L. Mendes, R. C. Mesquita, and G. A. S. Pereira,
“Fluids in electrostatic fields: An analogy for multi-robot control,”
density of the fluid model constant, this approach allows for IEEE Trans. on Magnetics, vol. 43, no. 4, pp. 1765 – 1768, 2007.
a loose way of controlling the group connectivity. [14] N. Michael, J. Fink, and V. Kumar, “Experimental testbed for large
multi-robot teams: Verification and validation,” IEEE Robotics and
VII. ACKNOWLEDGMENTS Automation Magazine, Mar. 2008, to Appear.
[15] B. Gerkey, R. T. Vaughan, and A. Howard, “The Player/Stage project:
The authors would like to acknowledge the financial Tools for multi-robot and distributed sensor systems,” in Proc. of the
support of CNPq – Brazil, ARO MURI SWARMS Grant Int. Conf. on Advanced Robotics, 2003, pp. 317–323.

1953

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on August 19, 2009 at 12:09 from IEEE Xplore. Restrictions apply.