COLLECTIVE PERCEPTION IN A SWARM

OF AUTONOMOUS ROBOTS

Giuseppe Morlino, Vito Trianni and Elio Tuci

Institute of Cognitive Sciences and Technology, National Research Council

Via San Martino della Battaglia 44, 00185 Rome, Italy

Keywords:

Swarm robotics, Collective behaviour, Perceptual discrimination, Evolutionary algorithms, Artiﬁcial neural

networks.

Abstract:

We present a study that aims at understanding how perception can be the result of a collective, self-organising

process. A group of robots is placed in an environment characterized by black spots painted on the ground.

The density of the spots can vary from trial to trial, and robots have to collectively encode such density into a

coherent ﬂashing activity. Overall, robots should prove capable of perceiving the global density by exploiting

only local information and robot-robot interactions. We show how we can synthesize individual controllers

that allow collective perception by exploiting evolutionary robotics techniques. This work is a ﬁrst attempt to

study cognitive abilities such as perception, decision-making, or attention in a synthetic setup as result of a

collective, self-organising process.

1 INTRODUCTION

How can a distributed system collectively encode the

magnitude of a macroscopic variable? This ques-

tion holds over multiple domains, and at different

scales. First and foremost, in the context of cognitive

neuroscience, this question can be reformulated as:

what are the neural mechanisms underlying percep-

tion? This is a fundamental question, which must be

answered ﬁrst in order to lay the foundations for fur-

ther investigations on other cognitive processes, such

as decision-making, attention or learning. For this

reason, the literature abounds of models about neural

coding of every sort of stimuli, from the basic ones—

e.g., vibro-tactile or visual stimuli (Romo and Sali-

nas, 2003; Lofﬂer, 2008)—to more complex percep-

tual conditions—e.g., multi-stability, face recognition

or numbers (Leopold and Logothetis, 1996; Rubin,

2003; Grill-Spector, 2003; Dehaene, 2003).

The problem of suitably encoding environmental

stimuli, however, does not pertain exclusively indi-

vidual animals, but is of fundamental importance also

for collective systems, such as bird ﬂocks and honey-

bee swarms. Similar systems are often considered as

super-organisms, due to their high cohesion and inte-

grated functioning (H¨olldobler and Wilson, 2008; De-

train and Deneubourg, 2006). It is therefore interest-

ing to look at how super-organisms can achieve a co-

herent perception of macroscopic features of the envi-

ronment they inhabit. For instance, while searching a

new nesting site, honeybees explore the environment

thanks to scouts that report their discoveries to the

nest. In there, a collective perception and a decision-

making process is carried on, which results in the

recognition and selection of the best site among the

discovered choices (Passino et al., 2008). In this pro-

cess, no single bee has the full picture. However, the

partial information of many bees is aggregated in the

nest and through a self-organising process decision-

making is successfully performed. (Passino et al.,

2008) recognise strong similarities between honey-

bees behaviour and the mechanisms that support per-

ception and decision-making in neural systems. In

particular, cross-inhibition within neural populations

is functionally similar to negative feedback between

bee workers committed to different nesting sites. The

parallel between cognitive systems and swarm be-

haviour goes beyond qualitative considerations. (Mar-

shall et al., 2009) compared the nest site selection

behaviour in ants and honeybees and the brain dy-

namics during decision making in a perceptual choice

task. They show that the swarm behaviour can be de-

scribed by the same model that was proposed for de-

cision making in (Ratcliff and Smith, 2004). As a

Morlino G., Trianni V. and Tuci E..

COLLECTIVE PERCEPTION IN A SWARM OF AUTONOMOUS ROBOTS.

DOI: 10.5220/0003075400510059

In Proceedings of the International Conference on Evolutionary Computation (ICEC-2010), pages 51-59

ISBN: 978-989-8425-31-7

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

consequence, the two decision processes can be di-

rectly compared, and similarities can be drawn be-

tween cognition in the brain and in the swarm.

In this paper, we aim at studying collective per-

ception in a robotic swarm. The goal of this study is

understanding which are the self-organising processes

underlying the collective perception of a macroscopic

environmental feature, which is not accessible to the

individual robots due to their limited perceptual abil-

ities and due to the nature of their individual explo-

ration strategies. Therefore, multiple robots need to

interact in order to give a collective response that cor-

relates with the macroscopic variable. It is worth

noticing that the perceptual discrimination task em-

ployed could in principle be solved by a single robot,

given an effective exploration strategy and enough

time to accomplish it. The reason why we let a group

of robots to ﬁnd a collective solution is because we

believe that the study of successful collective discrim-

ination strategies in this particular artiﬁcial scenario

may shed a light on the mechanisms of collective per-

ception in natural organisms.

In this robotic model, we synthesise the robot

neural controllers through evolutionary robotic tech-

niques, and we afterwards analyse the obtained results

in order to uncover the mechanisms that support the

collective perception process.

The usage of evolutionary techniques for collec-

tive and swarm robotics has been demonstrated in var-

ious recent studies. For instance, (Trianni and Nolﬁ,

2009) evolved self-organising synchronisation for a

group of robots that presented an individual periodic

behaviour, while (Sperati et al., 2010) showed how a

robotic swarm evolved through an evolutionary pro-

cess managed to collectively explore the environment

and forming a path to navigate between two target ar-

eas, which were too distant to be perceivedby a single

agent at the same time. (Hauert et al., 2009) exploited

artiﬁcial evolution to synthesise Swarming Micro Air

Vehicles (SAMVs) able to organise autonomously,re-

laying only on local informations, to estabilish a wire-

less communication network between users located

on the ground.

Our working hypothesis is that the evolutionary

process can produce optimal solutions to the given

task. Therefore, by analysing these solutions, we can

discover general mechanisms for collective percep-

tion, which are adapted to the experimental conditions

we have devised. This allows us to discuss the dis-

covered mechanisms with respect to known processes

performed by individuals and collectives.

2 EXPERIMENTAL SETUP

As mentioned above, in this paper we study how a

swarm of robots can collectively encode a macro-

scopic feature of the environment. We have set up

an experimental arena in which black circular spots

are painted on a grey background. The macroscopic

feature that must be encoded by the robotic swarm is

the density of black spots, which may vary from trial

to trial in the range d ∈ [0,1]. Robots can perceive the

colour of the ground only locally, through a noisy in-

frared sensor placed under their chassis. Robots can

emit ﬂashing signals, which can be perceived by all

other robots. By combining the locally acquired infor-

mation through this kind of simple communication,

the group should encode the global density through

the frequency of the emitted signals: the higher the

density, the higher the frequency of the collective

ﬂashing signal. In the following, we give the details

of the experimental setup and of the evolutionary al-

gorithm we used to synthesise the robot neural con-

trollers.

2.1 The robots and the Environment

Figure 1: Two snapshots of the simulated arena are shown.

The black disks spots are painted on a grey ﬂoor. The spots

are positioned on a grid of 40 × 40 cells. The density (i.e.

the probability to ﬁnd a spot in a cell) varies in the range

[0,1] (left: d = 0.26, right: d = 0.66).

The experimental arena is square (side l = 2m) and

surrounded by walls. Circular black spots are painted

on the ground in order to probabilistically obtain a

desired global density. The spots are homogeneous in

colour and size (radius r = 2.5cm), and are aligned to

a square grid of 40 × 40 cells (see Fig. 1). The den-

sity d represents the probability that each cell of the

grid is ﬁlled with a black spot. Therefore, when the

density is 0, no spot is present and the arena ground

is completely grey; when the density is 1, the arena

is completely ﬁlled with black circular spots. In this

way, we can control the black spot density with a sin-

gle parameter, and we can create multiple instances

for the same macroscopic value.

ICEC 2010 - International Conference on Evolutionary Computation

Ten robots (radius 3.75cm) are randomly placed in

the environment. Each robot is equipped with two

wheels that provide a differential drive motion (maxi-

mum linear speed: v

max

= 8.2cm/s). Robots can per-

ceive walls and other obstacles by means of eight in-

frared sensors placed around the turret (see Fig. 2).

The infrared sensors can be exploited for obstacle

avoidance. The ground colour is perceived through an

infrared sensor placed under the chassis of the robot,

in the front part (indicated by ‘G’ in Fig. 2). In the

absence of noise, the ground sensor returns 0 when is

over a black spot, and 0.5 when is over the grey back-

ground. Additionally, we make this sensor very unre-

liable by adding 30% white noise to the absolute sen-

sor reading. Finally, each robot r can emit a ﬂashing

signal S

(t) switching on for a time-step (∆t = 0.1s)

the LEDs placed around its turret. This signal can be

perceived by all the other robots present in the envi-

ronment in a binary way: s(t) = 1 if there is at least

one robot r emitting a signal, otherwise s(t) = 0. A

robot can perceive the ﬂashing signals through the

omni-directional camera, including the signals emit-

ted by the robot itself.

O,3

O,2

H,1

H,4

I,10

I,1

I,8

O,1

I,9

H,3

H,2

to actuators

from sensors

Figure 2: The robot and the neural controller. Left: a

schema of the simulated robot. Eight proximity sensors

(IR

0−7

) are positioned at 3.25cm from the ground pointing

horizontally. The sensors detect obstacles at a maximum

distance of ≈ 5cm. The ﬂoor colour is perceived through

an infrared sensor (G) positioned on the robots’ front and

pointing the ﬂoor. Right: the agents’ controller is a CTRNN

with 10 sensory neurons, 4 hidden and three motor units.

2.2 The Controller and the

Evolutionary Algorithm

Each robot is controlled by a continuous time recur-

rent neural network (CTRNN) (Beer, 1995). The

neural network has a multi-layer topology, as shown

in Fig. 2: neurons N

I,1

to N

I,10

take input from the

robot’s sensory apparatus, neurons N

O,1

to N

O,3

con-

trol the robot’s actuators, and neurons N

H,1

to N

H,4

form a fully recurrent continuous time hidden layer.

The input neurons are simple relay units, while the

output neurons are governed by the following equa-

tions:

= σ(O

+ β

), (1)

∑

i=1

σ(H

+ β

), (2)

σ(z) = (1+ e

−z

)

−1

, (3)

where, using terms derived from an analogy with real

neurons, O

and H

are the cell potentials of respec-

tively output neuron j and hidden neuron i, β

and β

are bias terms,W

is the strength of the synaptic con-

nection from hidden neuron i to output neuron j, and

and h

= σ(H

+ β

) are the ﬁring rates. The hidden

units are governed by the following equation:

= −H

∑

i=1

σ(H

+ β

) +

∑

i=1

, (4)

where τ

is the decay constant, W

is the strength

of the synaptic connection from hidden neuron i to

hidden neuron j, W

is the strength of the connec-

tion from input neuron i to hidden neuron j, and I

the intensity of the sensory perturbation on neuron i.

The weights of the connection between neurons, the

bias terms and the decay constants are genetically en-

coded parameters. Cell potentials are set to 0 each

time a network is initialised or reset. State equations

are integrated using the forward Euler method with an

integration step-size of 0.1 seconds.

Eight input neurons—N

I,1

to N

I,8

—are set from

the infrared sensors. Input neuron N

I,9

is set from the

ground sensor. Finally, input neuron N

I,10

is a binary

input set by the perception of the ﬂashing signal s(t).

The neurons N

O,1

and N

O,2

are used to set the speed of

the robot’s wheels. Neuron N

O,3

is used to switch on

the LEDs. In order to emit a ﬂashing signal that lasts a

single time-step, the LEDs are switched on only when

the neuron activation surpasses the threshold 0.5:

(t) = 1 ⇐⇒ o

(t) ≥ 0.5∧ o

(t − 1) < 0.5. (5)

This means that in order to ﬂash again, the activation

of neuron N

O,3

must go belowthe threshold, and up

again. The minimum period for oscillations is there-

fore 2 time-steps, that is, 0.2s.

The free parameters of the robot’s neural con-

troller are encoded in a binary genotype, using 8 bits

for each real number. Evolution works on a popu-

lation of 100 randomly generated genotypes. After

evaluation of the ﬁtness, the 20 best genotypes sur-

vive in the next generation (elitism), and reproduce

COLLECTIVE PERCEPTION IN A SWARM OF AUTONOMOUS ROBOTS

by generating four copies of their genes with a 2%

mutation probability of ﬂipping each bit. The evolu-

tionary process lasts 5000 generations. During evolu-

tion, genotype parameters are constrained to remain

within the range [0,1]. They are mapped to produce

CTRNN parameters with the following ranges: con-

nection weights W

∈ [−4,4]; biases β ∈ [−4, 4]; con-

cerning decay constants, the genetically encoded pa-

rameters are ﬁrstly mapped onto the range [−1,2] and

then exponentially mapped onto τ ∈ [10

−1

,10

]. The

lower bound of τ corresponds to the integration step

size used to update the controller; the upper bound is

arbitrarily chosen.

2.3 The Fitness Function

A genotype is translated into N = 10 identical neu-

ral controllers which are downloaded onto N iden-

tical robots (i.e., the group is homogeneous). Each

group of robots is tested for 20 trials, which last either

1000 or 2000 time-steps (one time-step corresponds

to 0.1s). The density is varied systematically, mak-

ing the group experience 20 different values, equally

distributed in [0, 1]. The robots’ neural controllers

are not reset from trial to trial, therefore the order in

which trials are presented is relevant. At each ﬁtness

evaluation, we randomly shufﬂe the sequence of en-

vironments experienced by the same group, in order

to remove regularities that could be exploited by spu-

rious behaviours. In order to evaluate the ﬁtness of

s(t)

T T T

i+2i i+1

flashing event i flashing event i+1

Figure 3: Schematic representation of the collective ﬂashing

signal, through which the group of robots encodes the black

spot density.

a genotype, we measure how well the corresponding

robotic group encodes the black spot density d. To

do so, we demand that robots as a group display a

periodic ﬂashing activity with a frequency that corre-

lates with the black spot density. The group ﬂashing

activity is measured on the global signal s(t) that re-

sults from the coupled activity of each robot. When

robots ﬂash in subsequent time-steps, their signals are

perceived as a single ﬂashing event (i.e., a sequence

of consecutive ﬂashes is perceived as a square signal,

see Fig. 3). We measure the period T

as the time be-

tween the start of two subsequent events. In this way,

we obtain a series of inter-ﬂash periods that we use to

compute the ﬁtness. First of all, we compute through

an exponential moving average the average period

and the average difference between two consecutive

periods ∆T:

T = α

T + (1− α)T

, (6)

∆T = α∆T + (1− α)|T

− T

i−1

|, (7)

where α = 0.9 is the time constant of the moving av-

erage. At the end of the trial θ,

T should encode the

density. We measure the encoded density by linearly

scaling the average period:

enc

−

− T

, (8)

where T

= 5s and T

= 1s are respectively the max-

imum and minimum periods, arbitrarily chosen. Fi-

nally, the two ﬁtness components are computed: F

rewards the group for suitably encoding the black spot

density:

= Φ(1.0− |d − d

enc

|), (9)

where Φ(x) is a piecewise linear function that simply

constrains the ﬁtness value in the interval [0,1]. This

component therefore rewards the group for minimis-

ing the difference between the black spot density and

the group encoded density. However, it does not as-

sure that the system converges towards a periodic sig-

nalling. For this purpose, a second ﬁtness component

is computed, that minimises the difference between

consecutive periods:

∆

= Φ(1.0−

∆T

), (10)

where ∆T

= 2s is the maximum difference allowed.

By minimising the difference among consecutive pe-

riods, the system is rewarded to produce periodic sig-

nals. Finally, the ﬁtness of a genotype is the product

of the two ﬁtness components, averaged over multiple

trials:

F =

∑

θ=1

· F

∆

. (11)

A trial is stopped and the ﬁtness is zero when no ﬂash-

ing event is detected within the last 10s, therefore pro-

moting a sustained ﬂashing activity during the whole

trial. Similarly, a trial is stopped if any robot collides

with another robot or with a wall, and the ﬁtness is

zero for that trial. This indirect selective pressure al-

lows to evolve obstacle avoidance.

3 RESULTS

We have performed 20 evolutionary runs for 5000

generations. For each evolutionary run, we selected

ICEC 2010 - International Conference on Evolutionary Computation

0 0.04 0.1 0.14 0.2 0.24 0.3 0.34 0.4 0.44 0.5 0.54 0.6 0.64 0.7 0.74 0.8 0.84 0.9 0.94

0.0 0.2 0.4 0.6 0.8 1.0

real density

performance / encoded density

Fitness

Comp1: Density

Comp2: Regularity

Encoded Density

Figure 4: Generalisation test for the excitatory strategy. Boxes represent the inter-quartile range of the data, while the hori-

zontal lines inside the boxes mark the median values. The whiskers extend to the most extreme data points within 1.5 times

the inter-quartile range from the box. The empty circles mark the outliers.

a single genotype to be further analysed. To this

aim, we evaluated the performance of the 20 best

individuals of the last generation, measuring the ﬁt-

ness over 100 trials for each of the 20 density val-

ues (2000 trials in total), and we selected the geno-

type with the highest mean performance to represent

the evolutionary run. Among the selected genotypes,

9 out of 20 resulted in a good collective behaviour

while the remaining ones resulted in sub-optimal so-

lutions, in which the group always converged to a

ﬁxed signalling frequency, therefore failing to suit-

ably encode the black spot density. The performance

of the best genotypes is presented in Table 1. Despite

Table 1: Performance of the genotypes that result in a good

collective perception behaviour. Data are sorted in decreas-

ing order and, for each column, the mean and standard de-

viation are shown. The columns represent the ﬁtness F and

the the two components F

and F

∆

run F F

∆

4 0.87±0.06 0.92±0.06 0.95±0.02

19 0.85±0.08 0.92±0.08 0.93±0.04

14 0.84±0.07 0.89±0.08 0.94±0.03

9 0.83±0.07 0.92±0.06 0.91±0.05

20 0.83±0.08 0.92±0.07 0.90±0.05

10 0.82±0.06 0.91±0.07 0.90±0.03

15 0.81±0.07 0.88±0.08 0.92±0.03

3 0.80±0.10 0.85±0.11 0.94±0.03

13 0.80±0.07 0.89±0.08 0.89±0.04

the variability in performance, the behaviours evolved

in different evolutionary runs are qualitatively simi-

lar: robots mainly rotate on the spot, in some cases

slightly moving away from the initial position. While

rotating on the spot, the ground sensor positioned on

the robot front gives a very local and noisy estimate

of the ground colour. The ground information is in-

tegrated over time, and modulates an internal oscil-

lator that allows to tune the frequency of a periodic

signalling. However, this frequency is related just to

the local density perceived by the robot, which may

be signiﬁcantly different from the global density: in

fact, an individual robot rotating in one place can per-

ceive only a limited number of different ground pat-

terns, which do not represent well the global density,

above all for intermediate density values. Moreover,

the 30% white noise of the ground sensor makes it dif-

ﬁcult to have even a good and stable local perception.

For these reasons, robots have to coordinate to bet-

ter estimate the global density, and to do so, they can

exploit the ﬂashing signals. By analysing the com-

munication strategies evolved in the different evolu-

tionary runs, we found that they can be grouped into

two classes. In some cases, the ﬂashing signals are

excitatory, that is, signal reception anticipates or pro-

vokes the signal production. This is the case for the

behaviour evolved in runs 4, 9, 10 and 15. In the

other cases—namely, runs 3, 13, 14, 19 and 20—

ﬂashing signals are inhibitory, that is, signal reception

prevents or delays the signal production. In order to

understand the mechanisms behind these two strate-

gies, we analyse the best performing genotype of each

class, namely the one obtained in run 4 for the exci-

tatory strategy, and the one obtained in run 19 for the

inhibitory one.

3.1 Analysis of the Excitatory Strategy

To better understand the properties of the evolved be-

haviour, we ﬁrst perform a generalisation test that

aims at revealing how well the system behaves with

varying densities. For this purpose, we have recorded

COLLECTIVE PERCEPTION IN A SWARM OF AUTONOMOUS ROBOTS

the performance of the system over 50 different black

spot densities uniformly distributed in the range [0,1]

(200 trials per density value). The results are dis-

played in Fig. 4. Here, we plot the ﬁtness F and the

two components F

and F

∆

for each density. More-

over, we also plot the encoded density d

enc

. The ﬁg-

ure reveals that the system shows a good behaviour

for almost all densities. In average, performance F

oscillates in the interval [0.8,1.0]. Moreover, it is pos-

sible to observe that the component F

∆

is very high,

especially for high densities. This indicates that the

group is able to converge to a very regular ﬂashing

activity, while for smaller values the period is nois-

ier. However, for small density values the component

is higher, revealing that the system performs better

in this conditions. The actual abilities of the robotic

group can be discussed looking at the encoded den-

sity d

enc

, which is plotted against the ideal case y = x.

We note that for densities up to d = 0.6, the encoded

density nicely follows the real one. For larger values,

however, a sort of phase transition occurs, in which

the robots present a fast signalling behaviour that en-

codes a density around 0.84.

0.5

600 700 800

0.5

600 700 800

time-step (t)

Figure 5: Excitatory strategy: neural activation and sig-

nalling status for two of the ten robots. The bold line in-

dicates the activation of the neuron N

O,3

, which controls the

ﬂashing signal. The vertical grey bands indicate the per-

ceived signal s(t). The small dark vertical lines within the

grey band indicate the time-step in which the robot itself is

signalling. The plot refers to a density d = 0.2.

How can the robots produce such behaviour? We

answer this question by analysing the behavioural and

communication strategy. In this case, robots rotate on

the spot without changing position. By observing an

isolated robot, we noticed that the ﬂashing activity is

regulated by the locally perceived density: the higher

the density, the higher the ﬂashing frequency. How-

ever, the individual robot always underestimates the

real density, in average. Therefore a collective mech-

anism must be in place. As mentioned above, in the

excitatory strategy the reception of a signal provokes

the emission of a signal. The dynamics of the os-

0.5

600 700 800

0.5

600 700 800

time-step (t)

Figure 6: Excitatory strategy: neural activation for d = 0.8.

0.25

0.5

0.75

0 0.2 0.4 0.6 0.8 1

encoded density

density

1 robot

2 robots

3 robots

4 robots

10 robots

Figure 7: Encoded density for varying group size. Each

line represents the average of 100 trials, performed for 100

density values in [0, 1]. Data for 1, 2, 3, 4 and 10 robots are

shown.

cillation for a low density d = 0.2 can be observed

looking at Fig. 5. In correspondence of a perceived

signal, the activation o

of the neuron controlling the

signal output increases until it goes beyond the 0.5

threshold, making the robot itself signal. Upon the

sustained perception of a signal, the activation o

re-

mains high, therefore delaying the following ﬂash.

For instance, in the ﬁrst signalling event in Fig. 5, the

top robot ﬂashes the earliest, and the persistence of

the signal afterwards delays the following ﬂash. In-

stead, if more than one ﬂash is required for o

to over-

come the signalling threshold, the following ﬂash is

anticipated: the bottom plot reveals that in correspon-

dence of a very delayed ﬂash—during the second sig-

nalling event—theactivation o

is just over the thresh-

old and goes immediately down afterwards, allowing

the robot to anticipate the following ﬂash. The se-

quence of perceived ﬂashes functions both as a pos-

itive and negative feedback mechanism: robots com-

pete in emitting the ﬁrst ﬂash, and consequentlymutu-

ally accelerate their rhythm. This acceleration is how-

ever limited by the presence of multiple signals that

slow the ﬂashing frequency down. The same mech-

anism is in place for larger frequencies (see Fig. 6).

However, in this case the system converges into a dif-

ferent dynamic regime, in which robots differentiate

in two groups that alternately signal. This is evident in

the dynamics of the activation o

shown in Fig. 6: the

ICEC 2010 - International Conference on Evolutionary Computation

0 0.04 0.1 0.14 0.2 0.24 0.3 0.34 0.4 0.44 0.5 0.54 0.6 0.64 0.7 0.74 0.8 0.84 0.9 0.94

0.0 0.2 0.4 0.6 0.8 1.0

real density

performance / encoded density

Fitness

Comp1: Density

Comp2: Regularity

Encoded Density

Figure 8: Generalisation test for the inhibitory strategy.

asymmetric oscillations indicate that robots engage in

a sort of turn-taking, achieving the maximum ﬂashing

frequency. This also justiﬁes the phase transition we

observed in Fig. 4: for high densities the probability

of converging into this fast ﬂashing regime is higher.

In order to further test the hypothesis that robots com-

pete to emit the ﬁrst signal, we run a series of experi-

ments varying the number of robots in the arena. The

results plotted in Fig. 7 show that the average encoded

density increases with the number of robots, thus sug-

gesting that robots are able to collectively accelerate

their ﬂashing rhythm.

3.2 Analysis of the Inhibitory Strategy

In the case of the inhibitory strategy, we performed

similar analyses. The results of the generalisation test

are plotted in Fig. 8. It is possible to notice that the

system has a very similar performance with respect to

the excitatory case: the performance F

is very high

for each density, while F

∆

slightly increases for large

d. Therefore, also in this case the group converges

towards very regular and precise ﬂashing activity, es-

pecially for high densities. Looking at d

enc

, it is pos-

sible to notice that the system presents a phase tran-

sition similar to the one discussed for the excitatory

strategy.

All these similarities, however, result from rad-

ically different mechanisms. As we already men-

tioned, in this case signals are inhibitory: when a

robot perceives a ﬂash, the neural activity o

that con-

trols the ﬂashing signal is reset, whatever its value

is. This means that there is normally only one robot

ﬂashing at any time, that is, the one the reaches the

signalling threshold the earliest. This behaviour is ev-

ident looking at Fig. 9, in which the dynamics of the

neural activity of two different robots are plotted for a

density d = 0.2: the bottom plot reveals that the corre-

sponding robot ﬂashes the earliest in the ﬁrst four sig-

nalling events, preventing other robots to ﬂash them-

selves. The situation is similar in the case of d = 0.8,

shown in Fig. 10, in which we observe that robots

compete in order to ﬂash the earliest, similarly to what

happens for the excitatory strategy. However, in this

case the inhibitory signal does not allow a negative

feedback mechanism. In fact, if a robot ﬂashes with

an individual frequency higher than the other robots

(e.g., the robot locally perceives a higher density), it

would impose its frequency to the group by inhibiting

all other robots. If this is the case, the group system-

atically overestimates the black spot density due to

those robots that locally perceive a high value. There-

fore there must exist another mechanism that serves

as negative feedback to control the frequency of the

group. By looking at the behaviour of the robots, we

notice that at the beginning of the trial robots slightly

move from their initial position while rotating on the

spot. This allows robots to explore the neighbour-

hood for a short time. In order to understand the

role of these movements, we tested the robotic system

ﬁxing the motor outputs to constant values (o

= 1

and o

= 0), forcing the robots to turn on the spot

without changing position. The results are shown in

Fig. 11, for varying density and varying number of

robots. From these tests, we infer that the slight mo-

tion of the robots is an adaptive mechanisms, given

that the system without motion performs worse. As

predicted, when robots cannot search the neighbour-

hood of their initial position, they slightly overesti-

mate the density. We also observe that robots exploit

the information coming from other robots. In fact,

there is no difference between the performance of a

single robot when it can and when it cannot move.

This means that the motion alone does not allow a

single robot to better estimate the global density. We

therefore believe that the initial motion of the robots is

COLLECTIVE PERCEPTION IN A SWARM OF AUTONOMOUS ROBOTS

0.5

600 700 800

0.5

600 700 800

time-step (t)

Figure 9: Inhibitory strategy: neural activation for d = 0.2.

performed when there are discrepancies between the

locally perceived density and the global ﬂashing ac-

tivity. In other words, a robot moves in search of a

local density that corresponds to the globally encoded

density. This is brought forth only for a short time

at the beginning of the trial. After this short time,

the robot stops in place, whatever the local density

is. With this mechanism, robots can average out the

global density.

0.5

600 700 800

0.5

600 700 800

time-step (t)

Figure 10: Inhibitory strategy: neural activation for d = 0.8.

0.25

0.5

0.75

0 0.2 0.4 0.6 0.8 1

encoded density

density

10 robots

10 robots NM

1 robot

1 robot NM

Figure 11: Encoded density for varying group size and for

blocked . Each line represents the average of 100 trials,

performed for 100 density values in [0,1]. Data for 1, 2, 3,

4 and 10 robots are shown.

4 DISCUSSIONS AND

CONCLUSION

In this paper, we have analysed how a group of robots

can collectively encode a macroscopic variable that is

not accessible to the single individuals in the group.

By evolving the collective perception behaviour, we

have found two possible strategies that use the com-

munication channel in a opposite way: signals are

either excitatory or inhibitory. In both cases, robots

compete in ﬂashing the earliest. In doing so, they

share the information gathered locally, allowing to

collectively encode an average value close to the ac-

tual density.

It is important to remark the fact that, besides the

excitatory or inhibitory communicative interaction, a

second mechanism is necessary to regulate the ac-

tivities of the group. On the one hand, this mecha-

nism has been found in the length of the signalling

event, which limits robots in producing the ﬁrst ﬂash

for many times consecutively. On the other hand, we

observed that robots move from their initial position,

therefore exploring the neighbourhood. In both cases,

these regulatory mechanisms allow multiple robots to

participate in the collective perception in order to have

a better estimate of the macroscopic variable. It is

therefore possible to identify a general strategy that

supports the collective perception in our system: indi-

vidually, robots encode the local density in a ﬂashing

frequency and compete in producing the ﬁrst ﬂash,

which is globally perceived and inﬂuences the whole

group. At the same time, robots try to hand on the

leader role and to listen to the other robots. The bal-

ancing of these two tendencies leads to the correct en-

coding of the global density.

The presence of two counteracting mechanisms

that regulate the activity of the group is common to

systems as diverse as brains and swarms. The positive

feedback loop allows to amplify small perturbations

and quickly spread information in a system, while

the negative feedback loop controls the competition

between different options and modulate the informa-

tion spreading. The relevance of the negative feed-

back is recognised in neural systems—in which it is

provided by specialised inhibitory inter-neurons and

mediated by glycine and gamma-aminobutyric acid

(GABA) transmitters (Jonas and Buzsaki, 2007)—

and in super-organisms—in which it may results

from speciﬁc stop signals issued by some individu-

als (Nieh, 2010). In our system, we have not pro-

vided a speciﬁc interaction modality different from

the ﬂashing signals. Despite this limitation, evolu-

tion could synthesise other mechanisms that resulted

in regulatory processes.

ICEC 2010 - International Conference on Evolutionary Computation

In future work, we plan to continue the study

of cognitive abilities displayed by collective sys-

tems (Trianni and Tuci, 2009). The experimental

scenario we have presented here has been conceived

to investigate the collective perception and decision

making. We plan to study whether groups of robots

can select the most dense environment among two

or more possibilities presented sequentially or segre-

gated in space. By comparing the results obtained in

different artiﬁcial setups, we aim at discovering gen-

eral principles about collective perception and deci-

sion making that could be generalised also to the bio-

logical reality.

ACKNOWLEDGEMENTS

The authors thank the Institute of Cognitive Sci-

ences and Technology of the Italian National Re-

search Council for having funded the research work

presented in this paper. The authors also thank the

members of LARAL group for the constructive com-

ments during the early preparation of this research

work.

REFERENCES

Beer, R. D. (1995). A dynamical systems perspective on

agent-environment interaction. Art. Intell., 72:173–

215.

Dehaene, S. (2003). The neural basis of the weber-fechner

law: a logarithmic mental number line. Trends in Cog-

nitive Sciences, 7(4):145–147.

Detrain, C. and Deneubourg, J.-L. (2006). Self-organized

structures in a superorganism: do ants “behave” like

molecules? Physics of Life Reviews, 3:162–187.

Grill-Spector, K. (2003). The neural basis of object percep-

tion. Current Opinion in Neurobiology, 13(2):159–

166.

Hauert, S., Zufferey, J.-C., and Floreano, D. (2009).

Evolved swarming without positioning information:

an application in aerial communication relay. Au-

tonomous Robots, 26(1):21–32.

H¨olldobler, B. and Wilson, E. O. (2008). The Superorgan-

ism: The Beauty, Elegance, and Strangeness of Insect

Societies. W. W. Norton & Company, New York, NY.

Jonas, P. and Buzsaki, G. (2007). Neural inhibition. Schol-

arpedia, 2(9):3286.

Leopold, D. A. and Logothetis, N. K. (1996). Activity

changes in early visual cortex reﬂect monkeys per-

cept during binocular rivalry. Nature, 379:549–553.

Lofﬂer, G. (2008). Perception of contours and shapes: Low

and intermediate stage mechanisms. Vision Research,

48:2106–2127.

Marshall, J. A. R., Bogacz, R., Dornhaus, A., Planqu´e, R.,

Kovacs, T., and Franks, N. R. (2009). On optimal

decision-making in brains and social insect colonies.

Journal of the Royal Society Interface, 6(40):1065–

1074.

Nieh, J. C. (2010). Negative feedback signal that is trig-

gered by peril curbs honey bee recruitment. Current

Biology, 20:310–315.

Passino, K., Seeley, T., and Visscher, P. (2008). Swarm

cognition in honey bees. Behavioral Ecology and So-

ciobiology, 62:401–414.

Ratcliff, R. and Smith, P. L. (2004). A comparison of

sequential sampling models for two-choice reaction

time. Psychological Review, 111:333–367.

Romo, R. and Salinas, E. (2003). Flutter discrimination:

neural codes, perception, memory and decision mak-

ing. Nature Reviews Neuroscience, 4:203–218.

Rubin, N. (2003). Binocular rivalry and perceptual multi-

stability. Trends in Neurosciences, 26(6):289–291.

Sperati, V., Trianni, V., and Nolﬁ, S. (2010). Evolution of

self-organised path formation in a swarm of robots.

In Proceedings of the 7th International Conference on

Swarm Intelligence (ANTS 2010). In press.

Trianni, V. and Nolﬁ, S. (2009). Self-organising sync in a

robotic swarm. a dynamical system view. IEEE Trans-

actions on Evolutionary Computation, Special Issue

on Swarm Intelligennce, 13(4):722–741.

Trianni, V. and Tuci, E. (2009). Swarm cognition and ar-

tiﬁcial life. In Advances in Artiﬁcial Life. Proceed-

ings of the 10th European Conference on Artiﬁcial

Life (ECAL 2009).

COLLECTIVE PERCEPTION IN A SWARM OF AUTONOMOUS ROBOTS