2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Congress Center Hamburg

Sept 28 - Oct 2, 2015. Hamburg, Germany

Segregating Multiple Groups of Heterogeneous Units in Robot Swarms

using Abstractions
Edson B. F. Filho1 and Luciano C. A. Pimenta1

Abstract— This paper addresses the problem of segregation one aerial robot work together and Pimenta [14] proposed
of groups of heterogeneous units in robot swarms. We propose the collaboration of robots with heterogeneous sensing ca-
a controller that can drive robots in a way that each group pabilities.
composed of robots of a similar type will form clusters while
maintaining segregation from other groups. The approach is In the case of heterogeneous swarms an important ability
based on abstractions created to represent each group of of the system which might be useful in several applications is
robots and an artificial potential function used to segregate the the capacity of autonomous segregation. This is the ability of
groups. Different from previous works on swarm segregation, forming groups, each one composed solely of robots of the
we can mathematically guarantee that by using our approach same kind. In order to provide this capacity to the system
the system will always converge to a state where multiple
dissimilar groups are segregated. Moreover, in some situations, one must design individual control laws that make robots
our controller does not require all robots to have information of the same group form clusters while maintaining distance
about all the other robots in the system. We demonstrate the from other groups.
effectiveness of our controller with simulations with different Few works directly tackle the segregation problem. Studies
types of robots and varying number of robots and groups. relevant in the solution of the segregation problems are [15],
[16], [17] and [18].
I. INTRODUCTION
Groß [15] developed a centralized algorithm that can
Swarms of robots are systems composed of a number of segregate robots based on the Brazilian nut effect, in which
autonomous agents that need to interact and cooperate to nuts are segregated based on its granulometry. This algorithm
achieve a common goal [1]. These systems are characterized is then implemented [17] in e-puck robots.
by decentralized control, limited communication between Kumar et al. [16] used an artificial potential function based
robots, use of local information, and emergence of global in the differential adhesion model for biological cells. They
behavior [2]. The first researcher to reproduce computa- showed a proof of asymptotic convergence to segregation and
tionally the behavior of animal swarms was Reynolds [3], stability analysis of a robotic swarm with only two groups.
the goal was the automation of those behaviors in graphics Kumar’s work [16] was then extended in [18] with a
computation. Since then, different works have addressed the similar potential function that is capable of segregating more
problem of controlling swarms of robots [4], [5], [6], [7]. than two groups. Santos’s [18] and Kumar’s [16] controllers
There are multiple advantages inherent to swarms such as are distributed, although in both cases each robot needs
fault tolerance due to the redundancy in its construction. information from every other robot in the system during all
Some recent works are now focusing on applications of the time.
swarms of robots, such as perimeter surveillance [8], spill This paper presents a controller that differs from previous
detection [9], interactions with humans [10], [11] among works [16], [18]. The controller is based on the use of
others [12], [13]. abstractions [4] to represent each group of robots and an
Swarms of heterogeneous robots are those composed of artificial potential function [19] to create the artificial force
different types of robots, either in its design or in its role that segregate the groups represented by the abstractions.
in the task to be performed. For example, one can design Our controller has two clear advantages. The first one is
a system of heterogeneous robots to be used in perimeter with respect to the convergence to segregation with multiple
surveillance where some robots have cameras and are re- groups. In [16], it is shown the convergence to segregation
sponsible ford the surveillance while some other robots are with only two groups of heterogeneous robots and in [18]
designed to warn humans if there is a breach in the perimeter. it is only shown stability but not convergence for multiple
Some works address the use of heterogeneous robots groups. In this work we show convergence to segregation
in different contexts. For example, Dorigo [2] proposed a with multiple groups of heterogeneous robots. The second
scheme in which two different types of ground robots and advantage is that our controller might not require that each
robot receive information from all the other robots in the
*This work has been supported by the Brazilian agency CAPES. This system, during all the time. This property appears particularly
work has also been supported by the Brazilian agencies CNPq and
FAPEMIG. when there is a great number of groups in the system, in that
1 Edson Bernardes Ferreira Filho and Luciano Cunha de Araújo Pi- case, each robot will only need information from the robots
menta are with Graduate Program in Electrical Engineering - Federal of its own group and from groups called neighbors most of
University of Minas Gerais - Av. Antônio Carlos 6627, 31270-901, Belo
Horizonte, MG, Brazil. emails: [email protected], the time.
[email protected] This paper is organized in five sections. Section II presents

978-1-4799-9993-4/15/$31.00 ©2015 IEEE 401

the segregation problem formulation. Section III is com-
posed of four parts that leads to the proposed controller and
its convergence proof. In section IV, a simulation is shown
and the results are discussed. Section V concludes this paper
with final observations and future work perspective.

II. PROBLEM FORMULATION

Consider N holonomic robots moving freely in a two
dimensional Euclidean space, as in the other works [18] and
[19]. The dynamics of each robot is given by the double
integrator

q˙i = vi , v˙i = ui i = 1, 2, ...N, (1) Fig. 1. N = 16 point robots unevenly distributed into M = 3 groups.
Robots of the same type have the same color and are inside the same
where the position vector of each robot is given by qi = abstraction. Asterisks represent abstractions center (µj ).
[xi , yi ]T , the velocity vector by vi = [ẋi , y˙i ]T and the control
input by ui = [uxi , uyi ]T . Each robot is assigned to a group
Nj , j ∈ M = {1, 2, ...M } and M is the number of groups. mean and covariance of the positions of all robots in a group.
Therefore, the system is composed of N robots divided into The mean of each group is given by
the groups N1 + N2 + ... + NM . Robots of the same group  x n
are considered to be robots of the same type. µ 1X
µj = jy = qi , (3)
In this paper, we are interested in the segregation problem µj n i=1
[16], [18]. This is the problem of designing a control law that
drives the system to a state in which robots of the same type where n is the number of robots in the group associated with
or group form clusters separated from the robots of other the abstraction φj . In the two dimensional Euclidean space,
types. When this state is reached, the system is said to be physically, the abstraction is a circle with center given by
segregated. In this work we will assume that each group of (3) and “size” given by
robots of the same type is represented by an abstraction [4], n n
1X 1X 2
which is invariant to robots permutations and with dimension σj = ((xi − µxj )2 + (yi − µyj )2 ) = kqi − µj k .
independent of the number of robots. More specifically, each n i=1 n i=1
abstraction φj , j ∈ 1, ..., M will be defined by a circle with (4)
√
mean µj and radius Rj . Now, we can formally state our The radius of the abstraction circle is Rj = nσj . As
version of the segregation problem to be solved: shown in [20] it is easy to see that by definition, the
Problem Statement 1: Given N robots with dynamics robots associated with φj remainsPn inside the2 circle with
2
given by (1) of M types, where N ≥ M , design individual Rj . Note that kqi − µj k ≤ i=1 kqi − µj k = nσj ⇒
√
control laws ui that guarantee that each robot i remains in the kqi − µj k ≤ nσj . It is important to mention that if
interior of the abstraction φj that represents the robots of the n = 1, then the circle is degenerated to a point given by the
same type of robot i and at the same time each abstraction position vector qi . Formally, the abstraction φj is a surjective
φj converges to a state where: submersion φj = R2n → R3 mapping from the original
\ configuration space to a lower dimensional space:
φj = ∅. (2)
y T
φj (q̂j ) = µxj µj σj ,

j={1,...,M } (5)
Figure 1 shows a system segregated according to our defini-
in which q̂j is a vector composed of the position of all the
tion.
robots of a given group: q̂j = [x1 , y1 , x2 , y2 , ..., xn , yn ]T .
III. METHODOLOGY In order to design our individual controllers it is important
to relate the motion of the abstraction with the motion of the
We propose a strategy to guarantee the segregative behav- robots. Thus, differentiating (5)
ior of the system. This strategy consists mainly in coupling
two ideas: the use of abstractions to represent each group φ˙j = dφj q̂˙j . (6)
and an artificial potential function to segregate those groups.
Then we encapsulate both ideas into a control law which is By using (3), (4) and (5) we can obtain dφj :
used by each robot. T
1 0 2(x1 − µxj )

y
0 1 2(y1 − µj ) 
A. Abstractions
1 
dφj =  ... ... ..
 . (7)
 
The abstraction considered in this work is the same one n . 
x
1 0 2(xn − µj )
defined in [4], in the context of motion planning for large
multi-robot systems. Each abstraction is defined using the 0 1 2(yn − µyj )

402
 (a) (b)
1000 0.015 where  is a parameter larger than zero to guarantee that
0.01 kzkσ is differentiable everywhere. For the sake of simplicity,
800

0.005
in this work we use  = 1.
600 Now, considering the means of the abstractions as the high
ψα

γα
0
400
level agents to be guided, the collective potential function
−0.005
defined in [19] is given by:
200
−0.01
1 XX
0 −0.015
V (µ) = ψα (kµj − µi kσ ), (14)
0 5 10 15 0 5 10 15 2 i
||µ − µ ||
j i σ
||µ − µ ||
j i σ j6=i

Fig. 2. Parameters: h = 0.3, c = 0.01, dα = 10 and rα = 1.8dα = 18 where Z z

(a) Example of artificial potential function interaction between two agents ψα (z) = γα (s)ds, (15)
versus the distance between them. (b) Gradient based force between two dα
agents versus the distance between them.
c(z − dα )
γα = ρh (z/rα ) p , (16)
1 + (z − dα )2
Since we have agents with double integrator dynamics, we
need a relation between the abstraction motion and the robots and function ρh (z) is a bump function, that smoothly varies
acceleration. By differentiating (4) twice we have from 1 to 0:
T 

x1 − µxj 
 h 1, i z ∈ [0, h)
 y1 − µyj  1
ρh (z) = 2 1 + cos(π( z−h1−h )) z ∈ [h, 1] (17)
 x2 − µxj 
  
0, otherwise.

2 y
σ̈j =  y2 − µj  q̂¨j + 2σj0 , (8)

n . Figures 2(a) and 2(b) show examples of functions ψα and
..




 γα respectively. Parameter c is a gain of the function, and
xn − µx  parameter h acts in the smoothness of the corresponding
j
yn − µyj gradient. The parameters rα and dα are the finite cut-off
where, rα = krkσ and the global minimum of ψα , dα = kdkσ ,
n respectively (see Figure 2).
1X
σj0 = (ẋi − µ̇xj )2 + (y˙i − µ̇yj )2 . (9) By using the artificial potential function previously shown,
n i=1
we can define artificial forces composed of two terms:
Now, we can write: X X
  Fi = γα (kµij kσ )nij + ρh (kµij kσ /rα )(µ̇j − µ̇i ),
0 j∈Bi j∈Bi
φ̈j = dφj q̂¨j +  0  . (10) (18)
2σj0 where nij is a vector pointing in the direction µj − µi :
In section III-C we are going to propose a control law for (µj − µi )
each robot of the group so that the corresponding abstraction nij = q , (19)
2
dynamics in (10) can be simplified to: (1 + kµij k )

φ¨j = wj , (11) where µij = µi −µj and Bi is the neighborhood of group i,

i.e. these are the other groups so that the distance kµi − µj k
where wj is a virtual input for the abstraction which will be
is less than r, which is the parameter that defines the finite
given by:
µ T cut-off of the potential function.
wj = kµ Uj Ujσ ,

(12) First term of (18) is a gradient term based on (14), and the
where Ujµ is an artificial force that guides the motion of the second term acts as a velocity damping, where µ̇i and µ̇j are
group mean, Ujσ determines the evolution of the abstraction velocities of the centers of abstractions i and j, respectively.
size and kµ is a positive gain. Our choice of Ujµ and Ujσ In the case where the parameters r and d are such that
will determine the success of our strategy. d < r < 2d, the important Lemma, which is proved in [19],
holds:
B. Potential function Lemma 1: (Lemma 3 in [19]) Every local minima of V (µ)
This section describes an artificial potential function with is an α−lattice and vice-versa.
a finite cutoff inspired by the one first presented in [19] with An α−lattice is a formation such that the following set of
the aim of generating a proper artificial force Ujµ . algebraic constraints hold:
Before showing the potential function we need to define
the σ-norm of a vector. This is a map Rm → R≥0 given by kµj − µi k = d, ∀j ∈ Bi . (20)
[19]:
1
q
2
This Lemma will be useful in the proof of convergence of
kzkσ = [ 1 +  kzk − 1], (13) the proposed controller in the next section.


403
C. Control law in which δ is a positive small value to guarantee (25).
We can now show the individual control laws: Each abstraction can have a different number of robots
    n. Parameters kµ , k1 , k2 and δ are fixed and equal to all
0
  abstractions.
ui = dφTj (dφj dφTj )−1 −  0  + wj , (21)

2σj0
 Individual control law (28) is dependent on the number of
robots in the abstraction, the state of the robot itself (qj , q˙j ),
where φj is the abstraction of robot i and wj has to the state of the abstraction φj and the state of the neighbor
be designed to control the state of the abstraction. Note abstractions.
(2σ )
that det(dφj dφTj ) = n3j , then as long as σj 6= 0, the
determinant is different from zero which means that the
inverse always exist. From (10) by applying this control law D. Controller analysis
in every robot, each abstraction will move according to wj , In this section we formally analyze the proposed controller
as follows: to demonstrate its effectiveness to solve the problem of
kµ Ujµ
 
φ̈j = wj = . (22) segregation.
Ujσ
Theorem 1: Applying individual control law (28) in the
We design wj by using two components, Ujµ and Ujσ . system with M groups and N robots with dynamics given
Component Ujµ guides the motion of the mean of abstraction by (1), and assuming we do not have a situation where all
j and it is defined by the artificial forces in (18): robots of the same group are placed at the same position
at the same time and the system does not start at a local
Ujµ = Fj . (23)
maximum or saddle point of function V (µ) in (14), the
This choice of artificial force will guide the system to form system will converge to segregation i.e, the problem defined
an α−lattice with parameter d. in the Problem Statement 1 will be solved.
We design component Ujσ in order to reach the desired size Proof: Our method was constructed to guarantee the
for each abstraction. To guarantee that the segregation will be solution of the problem, which means that our proof is
obtained according to our definition in (2), we have to specify straight forward. The analysis is conducted in two parts.
the proper desired size of each abstraction. This desired size First, we have to prove that all robots in an abstraction will
has to be specified so that the distance between the means of stay inside it and the abstraction state will converge to the
two abstractions will be greater than the sum of the radius desired size. The second part is to show that the abstractions
of both abstractions when the system reaches convergence, a will end separated apart without intersections.
suficient condition is that, the radius Rj of each abstraction Given the assumption that the robots are not at the same
is less than half the distance d in the α−lattice. We know position at the same time, the determinant of dφj dφTj is
√
that the radius of each abstraction is Rj = nσj [20], and different from zero and the inverse in (21) always exist, then
we now impose that Rj < (d/2), to guarantee segregation the motion of the abstraction will be given by (22). From
as t → ∞. Thus, we define the parameters so that: (26) it should be clear that if k1 , k2 are properly designed the
q d dynamics given by σ¨j = Ujσ will be such that σj converges to
nσjdes < , (24) σjdes exponentially [21]. Since the radius is defined according
2 √
or to Rj = nσj , we know from section III-A that the robots
(d2 ) of φj will remain inside the abstraction during all the time.
σjdes < , (25)
4n For the second part of the proof, we consider the proof
where σjdes is the desired value for σj . of Theorem 1 in [19]. In this theorem, LaSalle’s invariance
Now, we propose the following dynamics for Ujσ : principle is used to show that a set of agents with double
integrator dynamics subject to the artificial potential force in
Ujσ = σ̈jdes + k1 (σ̇jdes − σ˙j ) + k2 (σjdes − σj ), (26) (18) (see Algorithm 1 in [19]) asymptotically converges to a
where k1 and k2 are properly designed positive gains, and configuration which is an equilibrium of function V . Since
we assume that the system does not start at a local maximum
n
2X i or at a saddle point of V and these are unstable equilibria
σ̇j = (x − µxj )ẋij + (yji − µyj )ẏji . (27)
n i=1 j we can guarantee that the system asymptotically converges
to a local minimum of V . By using Lemma 1 (see III-B) we
We can set σjdes as a constant, σ̈jdes and σ̇jdes to zero, so can conclude that the system asymptotically converges to an
that the abstraction size will have zero velocity and zero α−lattice formation.
acceleration when t → ∞. Thus, each individual robot will As the abstractions reach the desired size, with all the
be guided by the control law composed of (21), (23), (25) robots of the abstraction inside, together with the fact that
and (26), as follows: the other parameters (see (25)) were specified to guarantee
(qi − µj ) d2 absence of intersections among abstractions when forming
ui = kµ Ujµ + [2σj0 − k1 σ˙j + k2 ( − δ − σj )], the α−lattice, then the problem of segregation as defined in
σj 4n
(28) the Problem Statement 1 will be solved as t → ∞.

404
Fig. 3. Simulations in MATLAB, each group has n = 10 robots. From top to bottom: (a) M = 5 groups. (b) M = 10 groups. (c) M = 20 groups.
From left to right, 4 snapshots of initial to final iterations. Last snapshot of each simulation also highlight the abstraction size and the formation of the
α−lattices.

Fig. 4. Snapshots of simulation in ROS/Stage with 20 robots distributed unevenly in 4 groups. Each group is represented by a different color. From left
to right we have the initial simulation step, a intermediary step and the final step.

IV. NUMERICAL EXPERIMENTS parameters, such as the desired distance between groups and
We tested the proposed controller in two different plat- the normal distribution of robots were set in a way that we
forms ROS/Stage and MATLAB. In ROS/Stage we used can better visually evaluate our approach and are dependent
differential drive robots and in MATLAB we used a model of on the number of groups and robots. The simulations were
holonomic robots. In this section we present one simulation stopped as soon as segregation was reached according to our
run in ROS/Stage and three runs in MATLAB. In addition, definition in (2). Simulations with 5, 10 and 20 groups of
we discuss the results and the advantages and limitations of robots are shown in Figure 3.
the method. A video of all four simulations is found on the In order to better depict the “local” property of our
web: http://youtu.be/mFxI3YQrhSk. controller in comparison to the works in [16] and [18], Figure
5 shows the average number of groups in neighborhood
A. Simulations Bi versus the iterations, that is, the average amount of
In all simulations we assume that all robots start with information needed for robots from initial time to the time
zero velocity and the robots were positioned according to a segregation was reached.
normal distribution. The simulations were performed in two In ROS/Stage we performed simulations with a different
environments. In MATLAB, we performed simulations to purpose. Figure 4 shows an example of this simulations.
test the feasibility of our approach with a varying number of We aim to show the applicability of our controller with
robots and groups. In ROS/Stage we are interested in analyze differential drive robots in unbalanced groups. Groups are
our controller in a more realistic simulation environment. composed of 20 robots divided in 9, 5, 3 and 3 robots
We performed extensive simulation in MATLAB, using a per group. We made use of the feedback linearization [6]
holonomic robot model. Potential function parameters were approach in order to use the designed controller with the
assumed: r = 1.5d, h = 0.1, c = 50, δ = 0.01. Gains differential drive robot. The simulation was stopped when,
kµ , k1 and k2 were set to 10, 5, 50 respectively. Other visually, the robots were segregated.

405
 16

5 groups
10 groups
R EFERENCES

Number of groups in neighborhood Bi

14
20 groups
[1] V. Trianni, Evolutionary Swarm Robotics: Evolving Self-Organising
12 Behaviours in Groups of Autonomous Robots. Vol. 108. Ed. Springer
Science & Business Media, 2008.
10 [2] M. Dorigo, et al., Swarmanoid: A Novel Concept for the Study
of Heterogeneous Robotic Swarms, IEEE Robotics & Automation
8
Magazine. Dez. 2013, pp. 60-71.
[3] C. Reynolds, Flocks, herds and schools: A distributed behavioral
model, Computer Graphics, 1987, pp. 25-34.
6
[4] C. Belta and V. Kumar, Abstraction and Control for Groups of Robots,
IEEE Transactions on Robotics, vol. 20, no. 5, Oct. 2004 pp. 865875.
4
[5] T. Balch and R. Alkin, Behavior-based formation control for multi-
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 robot teams, IEEE Transactions on Robotics and Automation, v. 14,
Iterations n. 6, 1998, 926-939.
[6] L. Pimenta, G. Pereira, N. Michael, R. Mesquita, M. Bosque, L.
Fig. 5. Information of how many groups each robot needs (average) versus Chaimowicz and V. Kumar, Swarm Coordination based on Smoothed
iterations. e.g. With 10 and 20 groups, after the iteration 2000, each robot Particle Hydrodynamics technique, IEEE Transactions on Robotics,
needs information only of less than 6 neighbor abstractions. vol. 29, no. 2, 2013, pp. 4372-4277.
[7] V. Santos and L. Chaimowicz, Hierarchical Congestion Control for
Robotic Swarms, Proceedings of the 2011 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), 2011, pp. 4372-
B. Discussion 4277.
After the simulations it is easy to see that our approach is [8] L. Pimenta, G. Pereira, M. Gonalves, N. Michael, M. Turpin, and V.
Kumar, Decentralized controllers for perimeter surveillance with teams
different from Kumar’s and Santos’s works. Our experiments of aerial robots, Advanced Robotics, v. 27, n. 9, 2013, 697-709.
showed segregative behavior independent of the total number [9] G. Zhang, G. Fricke and D. Garg, Spill Detection and Perimeter
of groups and robots and how robots are distributed into Surveillance via Distributed Swarming Agents, IEEE/ASME Trans-
actions on Mechatronics, v. 18, n. 1, Feb. 2013, pp. 121-129.
groups. The work in [18] verified problems in segregation in [10] P. Walker, S. Amraii, N. Chakraborty, M. Lewis and K. Sycara, Human
the case of unbalanced groups. Control of Robot Swarms with Dynamic Leaders, Proceedings of the
The main advantage in our approach is that we can 2014 IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), Sept. 2014, pp. 1108-1113.
formally guarantee that segregation will be always achieved [11] J. Nagi, A. Giusti, L. Gambardella and G. Di Caro, Human-Swarm
when t → ∞. Another advantage of our controller is that Interaction Using Spatial Gestures, Proceedings of the 2014 IEEE/RSJ
in many situations robots only need local information to International Conference on Intelligent Robots and Systems (IROS),
Sept. 2014, pp. 3834-3841.
segregate. As can be seen in Figure 5, as the clusters begin [12] I. O’Hara, J. Paulos, J. Davey, N. Eckenstein, N. Doshi, T. Tosun, J.
to be formed, the amount of information needed by each Greco, J. Seo, M. Turpin, V. Kumar and M. Yim, Self-Assembly of a
group decreases. Even with 10 and 20 groups, after some Swarm of Autonomous Boats into Floating Structures, Proceedings of
the 2014 IEEE International Conference on Robotics and Automation
time, robots only need the states of the robots of its own (ICRA), 2014. p. 1234-1240.
group and the state of, in average, 5 other abstractions. In [13] B. Remes, D. Hensen, F. van Tienen, C. Wagter, E. van der Horst, G.
Figure 5 we can also see that if we increase the number of de Croon, Paparazzi: how to make a swarm of Parrot AR Drones
groups in the system, the amount of information needed by fly autonomously based on GPS, International Micro Air Vehicle
Conference and Flight Competition (IMAV2013), Toulouse, France,
each group is not proportionally increased. 17-20 Sept. 2013.
Our controller is also robust to adding or subtracting robots [14] L. Pimenta, V. Kumar, R. Mesquita, G. Pereira, Sensing and Coverage
in the system as long as the condition in (25) is always for a Network of Heterogeneous Robots. Sensing and coverage for
a network of heterogeneous robots. Proceedings of the 47th IEEE
satisfied. Conference on Decision and Control, Dec. 2008, pp. 3947-3952
A current limitation of our approach is the lack of a col- [15] R. Groß, S. Magnenat and F. Mondada, Segregation in swarms of
lision avoidance strategy. In practical situations, a low-level mobile robots based on the Brazil nut effect, Proceedings of the 2009
IEEE/RSJ International Conference on Intelligent Robots and Systems
collision avoidance scheme can be implemented, although (IROS), Oct. 2009, pp. 4349-4355.
that would imply losing the guarantee of segregation. [16] M. Kumar, D. Garg, and V. Kumar, Segregation of heterogeneous
units in a swarm of robotic agents, IEEE Transactions on Automatic
V. CONCLUSIONS Control, vol. 55, no. 3, Mar 2010, pp. 743-748.
[17] J. Chen, M. Gauci, M. J. Price and R. Groß, Segregation in swarms
We have presented an approach to the problem of segrega- of e-puck robots based on the brazil nut effect, Proceedings of the
tion of multiple heterogeneous units in a robotic swarm based 11th International Conference on Autonomous Agents and Multiagent
on the use of abstractions guided by potential functions. Systems, vol. 1, 2012, pp. 163-170.
[18] V. Santos, L. Pimenta and L. Chaimowicz, Segregation of Multiple
In contrast to previous work, we have shown a method Heterogeneous Units in a Robotic Swarm, Proceedings of the 2014
with guaranteed convergence to segregation with multiple IEEE International Conference on Robotics and Automation (ICRA),
groups of robots. Moreover, our approach may use only local 2014, pp. 1112-1117.
[19] R. Olfati-Saber, Flocking for multi-agent dynamic systems: algorithms
information during part of the time to segregate groups in and theory, IEEE Transactions on Automatic Control, vol. 51, no. 3,
swarms of robots. 2006, pp. 401-420.
Future work will focus on collision avoidance strategies [20] C. Belta and V. Kumar, Towards Abstraction and Control for Large
Groups of Robots, Control Problems in Robotics, Ed. Springer-Verlag
integrated with our controller. Also, we will focus on strate- Berlin Heidelberg, 2003, pp. 169-182.
gies to segregate robots using only local information in any [21] J. Slotine and W. Li, Applied nonlinear control. Vol. 60. NJ: Prentice-
situation. Hall, 1991.

406