A Bio-inspired Approach in Decision-making of Multiple Robots Applied

on Partitioned Surveillance Task

Bruno Massaki Emori and Rodrigo Calvo

Department of Computer Science, State University of Maringa, 5790 Colombo Avenue, 87020-900, Maringa-PR, Brazil

Keywords:

Bio-inspired Computing, Autonomous Navigation of Multiple Robots, Exploration and Surveillance of

Environment, Collective Intelligence.

Abstract:

This paper proposes a robot coordination strategy for surveillance task execution. The strategy is based on

artiﬁcial ants behavior, both for non explored regions and areas already discovered by the robots. In a strategy

already known, the robots are not capable to distinguish its own pheromone from the pheromone of the others.

In the proposed strategy, the ability to distinct the substances causes the environment’s partition. Experimental

results shows the typical behavior of each strategy applied to different environments and show the superiority

of the proposed strategy due to the environment partitioning.

1 INTRODUCTION

The use of multi-agent systems to achieve an ob-

jective can be justiﬁed by many aspects, for example:

the inviability to develop a coordination system that

involves all available and necessary resources to solve

a complex problem; the possibility to decompose the

problem into several other subproblems, which are

solved simultaneously by individual agents; and the

possibility to use agents of small complexity to per-

form simple tasks and when grouped are able to solve

a more complex problem.

The use of multiple agents is adopted in applica-

tion which the individual behavior of robots leads into

the solution as a whole. In this case, the robots act

cooperatively. Many approaches use this kind of sy-

stem in real and virtual world applications. Examples

of these applications are: surveillance tasks and moni-

toring (Krishnan, 2015), cloud computing or automa-

tion (Sun et al., 2013), scheduling problems (Adhau

et al., 2013), optimal solution search problems (Liem-

hetcharat et al., 2015), rescue (Eoh et al., 2013) and

tasks allocation (Robu et al., 2011).

Regarding the surveillance problem, some propo-

sed systems of multiple robots are based on mathema-

tical models of the environment, speciﬁcally, a graph.

In this case, the environment’s structure is known, as

well as the positioning of each robot in the scene in

any given moment (Portugal, 2013). Generally, this

model is adopted in (Fazli et al., 2013) to develop an

off-line surveillance system. Initially, strategic points

from the environment are previously deﬁned in order

to maximize the covering area of the scene. Then,

a sequence of algorithms is executed to reduce the

graph’s size and also the trajectory established by the

robots. A model based on graphs is also adopted in

(Anisi et al., 2010). The robots’ trajectory is determi-

ned using heuristic algorithms. The authors present a

mathematical analysis and prove that the surveillance

problem is NP-complete. In (Wallar et al., 2015), air

robots perform the surveillance task.

A coordination strategy has an important role in

multi-agent systems, since it creates conditions for a

synergistic behavior to emerge. If this kind of col-

laborative behaviors exists, then the agents achieve

better results interacting with each other rather than

acting alone. Without synergisms, the system is just

a mere group of robots. In this sense, bio-inspired

approaches based on stigmergic interactions (e.g., ar-

tiﬁcial ants system) has drawn the scientiﬁc commu-

nity’s attention (Xiang and Lee, 2008).

The multiple robots system investigated here was

developed to deal with the surveillance task that dif-

fers from the techniques usually presented in the lite-

rature. In this case, the requirements that deﬁne the

task are kept (the environment must be sensed repea-

tedly and indeﬁnitely) and other, added: the environ-

ment must be virtually partitioned in disjointed regi-

ons (sectors) of same extension in quantity equivalent

to the number of robots; and each robot must perform

the surveillance task in one of these sectors.

The proposed coordination strategy, named as Par-

Massaki Emori, B. and Calvo, R.

A Bio-inspired Approach in Decision-making of Multiple Robots Applied on Partitioned Surveillance Task.

DOI: 10.5220/0006708803110318

In Proceedings of the 20th International Conference on Enterprise Information Systems (ICEIS 2018), pages 311-318

ISBN: 978-989-758-298-1

311

titioned Surveillance System (PSS), is based on a mo-

diﬁed artiﬁcial ants algorithm (Dorigo, 1992). The

logic presented differs from the traditional one, the

pheromone deposited by an agent has a repulsive pro-

perty. This way, the agents tend to disperse throug-

hout the environment, avoiding that two or more

agents perform the surveillance task in the same area

or in very close areas. Besides that, the strategy tends

to guide the robots into areas with no recent visitation,

due to the absence of the substance.

The strategy must partition the environment in

sectors that will be occupied by each one of the ro-

bots. The robots, individually, will be able to deﬁne

autonomously their sectors due to the adopted beha-

vior when the pheromone is detected. Hence, the

number of sectors in an environment is equal to the

number of robots. The environment’s partition favors

the surveillance task fulﬁllment more efﬁciently.

Each robot is equipped with two kinds of sensors:

one of them detects the pheromone released by the

robot itself; the second one detects the pheromone re-

leased by the other robots.

The coordination is not dependent of the parame-

ters that deﬁne an environment and no information

about the position and direction of the robots. Experi-

ments performed in different scenarios shows a gene-

ral behavior of the robots, consisting in two phases:

ﬁrst, the robots develop a dispersion behavior, and

then, each agent deﬁnes a restrictive sector to cover

and sensor. Lastly, o system reaches a stability stage.

In this instant, the surveillance task is performed fol-

lowing the desired requirements: the environment is

totally and repeatedly sensed and virtually partitioned

into small disjointed sectors, which each one is moni-

tored by a robot.

The paper is structured as it follows. The proposed

strategy is described in Section 2 as well as the models

of the sensors used. In Section 3, the experiments and

discussion about the obtained results in simulations

are presented. Main contributions, relevant aspects

and future works are presented in Section 4.

2 PARTITIONED

SURVEILLANCE SYSTEM

The PSS strategy was developed to deal with explo-

ration and surveillance tasks of unknown environ-

ments, according to the principles of the ants algo-

rithm (Dorigo, 1992). Essentially, the system is based

on multiple mobile agents able to take decisions of

movement adjustment, individually, according to the

stimuli from the environment.

The mentioned tasks are performed due to the

robot’s ability to deposit a substance, named phero-

mone, on the environment, to mark the regions which

they visited (or monitored). This substance has a

repulsive feature, i.e, the robots tend to avoid areas

with pheromone concentration. That behavior cau-

ses the robots spreading out in the environment. Alt-

hough the robots deposit only one kind of pheromone,

each one is able to distinguish its pheromone, (own

pheromone) from the one deposited by the others (odd

pheromone).

The proposed strategy is distributed, reactive and

independent of the parameters usually considered in

multiple agent systems: environment structure and di-

mension, number of robots and robots’ position. The

robots play a sequence of actions (Figure 1) in order

to execute the surveillance task, they are: pheromone

dispersion, pheromone detection and direction adjus-

tment. That sequence determines an iteration to exe-

cute the surveillance task. For all actions of the se-

quence, the robots keep a constant speed. The robot’s

angular speed is changed according to the external sti-

muli detected by the sensors. Each action is described

in next sections.

Figure 1: Task diagram for one robot.

2.1 Robot and Sensors Models

The robots are equipped with two types of phero-

mone sensors. One of them allows the detection of its

own pheromone and the other detects only odd phero-

mone. Both of them has the same physical structure.

Consider the index s, s ∈ {own,odd}, refers to the

pheromone sensors, own and odd, respectively. The

sensor ﬁeld is a sector of a circle C

deﬁned accor-

ding to a radius R

centralized on the robot’s frontal

part. The sensors covers an area of β

degrees from

the left to β

degrees from the right of the robot’s mo-

vement direction, β

∈ [0

◦

,180

◦

] (Figure 2). The total

cover area of 2β

degrees is divided in identical circu-

lar sectors C

, each one measuring α

degrees.

Every iteration, the pheromone sensors detect a set

of stimuli from the environment at a speciﬁc and pa-

rameterizable distance. The detection only occurs at

the boundary of the sensors (detection limit).

The robots also possess a proximity sensor. Its

model is similar to the pheromone sensors. At each

iteration, this sensor detects, in each circular sector,

a set of stimuli corresponding to the distance of ob-

stacles. The amount of pheromone accumulated close

to obstacles generates not attractive areas for robots.

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

312

Figure 2: Robot’s sensor model.

Then, the pheromone releasing is enough to robots

guide over safe trajectories. However, a speciﬁc al-

gorithm for obstacle avoidance is triggered at instant

that robots face imminent collision situations (for ex-

ample, if the distance between the obstacle and the

robot is too low). That algorithm is activated rarely

due to robot’s dimensions.

2.2 Direction Adjustment Mechanism

To determine the direction angle, it is proposed a me-

chanism that combines the information obtained by

the pheromone sensors. Consider a set P of circular

sectors C

, in which the i−th element of the set corre-

sponds to the quantity of own and odd pheromone of

the i−th circular sector. From this set, only the sectors

with odd pheromone quantity lower than ψ, ψ > 0 are

selected and inserted to the set Q. From this selection,

it is possible to achieve two scenarios:

1) If the quantity of elements in Q is smaller than

η, then the robot is located in an area with high con-

centration of odd pheromone. In this case, the own

pheromone becomes attractive (only for this robot),

in order to maintain the robot on its own area and

not monitor areas visited by the others. Therefore,

the content of Q is removed and the elements of P

with the highest quantity of own pheromone are in-

serted in Q until its size is equal to half of P’s size. A

probabilistic value is attributed to each circular sector

in P directly proportional to the quantity of own

pheromone deposited on the respective angular inter-

val. Speciﬁcally, a probability P(r) attributed to the

circular sector C

is:

P(r) =

∑

i∈{r|C

∈Q}

(τ

)

(1)

where τ

is the quantity of pheromone correspondent

to the circular sector C

2) If the quantity of elements in Q is higher than

η, it means there is low or none quantity of odd phero-

mone surrounding the robot. So, these areas have

been for a long time without visits (or never have been

visited yet). The advance of a robot into areas that are

not of its domain is only possible if the own phero-

mone remains repulsive. Hence, priority is given to

the exploration behavior. This way, elements in Q

with quantity of own pheromone higher than ψ are

removed from the set in such a way that the length

of Q will be reduced by up to half. Similarly to the

previous scenario, a probabilistic value is attributed

to each circular sector, however, inversely proporti-

onal to the quantity of own pheromone deposited in

the respective angular interval. The probability P(r)

is given as:

P(r) =

1 − τ

∑

i∈{r|A

∈Q}

(1 − τ

)

(2)

2.3 Pheromone Dispersion and

Evaporation

The pheromone is dispersed on a wide frontal area

of the robot, corresponding to the area covered by

the sensors. The amount of pheromone released in

a position changes according to the distance between

the position and the agent. Up next, the model of

pheromone releasing is described. Consider L

and

Q the coverage area of the sensor in iteration t and the

complete space of the environment, respectively, such

that L(t) ⊂ Q ⊂ R

. The concentration of pheromone

∆

(t) which the k−th robot deposits on the position

q ∈ Q in iteration t is given as:

∆

(t) = (τ

max

− τ

(t − 1))Γ

(t), and (3)

(t) =

(

δ e

−(q−q

)

,if q ∈ L

0,otherwise

(4)

where q

is the position of the k−th robot; τ

max

the maximum limit of saturation of pheromone’s con-

centration; λ is the dispersion rate of pheromone; and

δ ∈ (0,1).

The pheromone is a volatile substance. Hence,

it evaporates at the end of each iteration, i.e, when

all robots complete a cycle of tasks. The amount of

evaporated pheromone, given by the equation (5) de-

pends on the speciﬁc rate, a parameterizable constant

A Bio-inspired Approach in Decision-making of Multiple Robots Applied on Partitioned Surveillance Task

313

and the amount of pheromone on the area at the given

instant.

(t) = φτ

(t) (5)

where ε, 0 ≤ ε ≤ 1, is the evaporation rate and τ

(t)

is the total amount of pheromone on the position q in

the instant t.

3 EXPERIMENTAL RESULTS

This section presents the experiments to validate the

proposed strategy (PSS). The results are analyzed and

compared with a coordination strategy already known

with results proved in previous researches for the

exploration task and surveillance of unknown envi-

ronments, called in this paper as traditional strategy

(Calvo et al., 2015; Calvo et al., 2012). The perfor-

mance of the PSS strategy is satisfactory if it is simi-

lar or superior than the performance of the traditional

strategy. The surveillance task is executed if the entire

of the environment are sensed.

The platform Morse Simulator is used to execute

the experiments. The model adopted for the robots

is the ATRV, equipped with the sensor SICK LMS

500. For both strategies, was adopted the following

parameters: (s ∈ {own,odd}); R

= 10m; β

= 90

◦

;

γ = 0.4; robots’ linear speed = 0.5m/s; S = 360;

φ = 0.01; τ

(0) = 0.5 (quantity of pheromone at ite-

ration t = 0 for each q ∈ Q). These parameters were

deﬁned according to the numerous tests executed pre-

viously.

The experiments are performed in environments

with dimensions of 80×50m (Figure 3). The environ-

ments are divided in connected regions called rooms.

Each one of them are divided in small squared areas

called cells. The cells are portions of the environment

used to check if the entire of the environment is sen-

sed. At instant when robots visit all cells at least once,

then it is said that a cycle of surveillance was comple-

ted. After that, a new cycle os surveillance is started.

Two criteria of evaluation are adopted to the experi-

ments: number of cycles of surveillance completed

through the simulation and the average of iterations

between two consecutive completed cycles. Each si-

mulation is executed 30 times with 2000 iterations.

The PSS strategy is validated by means of the ana-

lysis of its performance when compared to the traditi-

onal strategy. The maps correspond to the simulation

whose performance is closer to the average perfor-

mance obtained in the simulations. Next, is presented

the performance of each strategy of coordination.

(a)

(b)

Figure 3: Environment models: (a) #1; (b) #2.

3.1 Traditional Strategy

In the traditional strategy, the direction adjustment

occurs differently from the PSS strategy. Here, the ro-

bots do not distinguish the types of pheromones. Only

one kind of pheromone is detected, and it has only the

repulsive property. The direction adjustment is also

based on a probabilistic model, but the best circular

sectors to be chosen are those which have low phero-

mone. Others sectors are chosen randomly to favor

the exploratory behavior in the environment. In this

strategy, the robots tend to monitor areas with low or

none pheromone quantity, regardless of the agent that

deposited it.

Considering the environment #1 (Figure 3(a)), the

traditional strategy is performed, using ﬁve robots, all

of them starting on the room of number 1. In 30 simu-

lations, the robots completed an average of 20.5 cy-

cles (standard deviation of 1.1670), with an average

of 95.28 iterations per cycle (standard deviation of

5.2373). The Figures 4 and 5 show the trajectory and

the average of pheromone deposited on the environ-

ment for each robot, respectively.

Note that, in Figure 4, the robots traveled throug-

hout the environment, each one occupying, practi-

cally, all rooms on a same simulation. The Figure

5 shows the average of pheromone deposited by the

robots in one simulation. This information indicates

the most visited areas by the robots and that the robots

were too close to each other. The occurrence of this

situation leads to the waste of resources like sensor

and traveled path. The concentration of close robots

also causes the increase of time in which a room re-

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

314

(a) Robot 1 (b) Robot 2

(e) Robot 5

Figure 4: Trajectory of the robots on the environment #1

using the traditional strategy.

(a) Robot 1 (b) Robot 2

(e) Robot 5

Figure 5: Average of pheromone dispersion on the environ-

ment #1 using the traditional strategy.

mains unvisited.

For the environment #2 (Figure 3(b)), six robots

are employed. All of them started on the room of

number 1. The average amount of surveillance cy-

cles completed was of 26.16 (standard deviation of

1.7827), with an average of 74.81 iterations per cy-

(a) Robot 1 (b) Robot 2

(e) Robot 5 (f) Robot 6

Figure 6: Trajectory of the robots on the environment #2

using the traditional strategy.

(a) Robot 1 (b) Robot 2

(e) Robot 5 (f) Robot 6

Figure 7: Average of pheromone dispersion on the environ-

ment #2 using the traditional strategy.

cle (standard deviation of 4.6290). Similarly to the

previous experiment, the robots are dispersed throug-

hout the environment, alternating the visits among the

rooms. The Figure 6 shows that the robots occu-

pied all rooms. According to the Figure 7, it can be

A Bio-inspired Approach in Decision-making of Multiple Robots Applied on Partitioned Surveillance Task

315

seen that many robots remained on the same rooms,

in most part of the simulation, indicating proximity

with each other. This means that another rooms of the

scene remained without visits for a long period.

3.2 Partitioned Strategy

The robots’ behavior on the partitioned strategy con-

sists of the dispersion throughout the environment in

a way that the monitored region by a robot do not get

occupied by any other robot. After a period in which

each robot is on a region, it is said that the strategy sta-

bilized itself. At this instant, the robots deﬁned their

own partition on the environment. The following ex-

periments show this behavior.

For the environment #1, ﬁve robots are used, star-

ting on the room number 1. The number of completed

surveillance cycles is, on average, 26.36 (standard de-

viation of 2.5795), with an average of 75.23 iterations

per cycle (standard deviation of 7.5414). Compared

to the traditional strategy, the partitioned strategy in-

creased about 28.59% the number of average comple-

ted cycles. This is due to the better surveillance task

distribution among the robots.

The Figures 8 and 9 show the environment’s par-

tition, where each robot remains on its own area indi-

vidually.

On the environment #2, for six robots starting on

room number 1, the average of completed cycles was

31.83 (standard deviation of 3.0522), with an average

of 62.59 (standard deviation of 5.7205) iterations per

cycle, an increase of 21.67% of number of completed

cycles. Figures 10 and 11 show that, even in a diffe-

rent environment , the strategy’s behavior still keeps

the map partitioned among the agents.

The Tables 1 and 2 summarize the results for both

experiments. The S.D. column stands for standard de-

viation of its previous column.

The main reason for the better performance of the

PSS strategy is highly tied to the partition of the en-

vironment whose partitions are occupied by only one

robot. The presence of odd pheromone avoid a robot

move to occupied partition. That behavior induces a

robot stay in areas with its own pheromone, i.e, areas

where that the robot visited recently.

If each robot patrols its own partition, there will

not redundancy of covered areas, improving the per-

formance of the surveillance task. Therefore, the exe-

cution of the task is distributive and more effective

than the strategy where the robots do not able to split

virtually the environment in small partitions.

(a) Robot 1 (b) Robot 2

(e) Robot 5

Figure 8: Trajectory of the robots on the environment #1

using the partitioned strategy.

(a) Robot 1 (b) Robot 2

(e) Robot 5

Figure 9: Average of pheromone dispersion on the environ-

ment #1 using the partitioned strategy.

4 CONCLUSIONS

The PSS strategy presented a higher performance than

the traditional strategy. The fundamental feature of

this strategy consists on the ability to partition the

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

316

Table 1: Results of the experiments on environment. #1.

Strategy Average cycles S.D. Average iterations per cycle S.D.

Traditional 20,5 1,1670 95,28 5,2373

Partitioned 26,36 2,5795 75,23 7,5414

Increase 28,59% 121,04% -21.04% 43,99%

Table 2: Results of the experiments on environment. #2.

Strategy Average cycles S.D. Average iterations per cycle S.D.

Traditional 26.16 1,7827 74.81 4.6290

Partitioned 31.83 3.0522 62.59 5.7205

Increase 21,67% 71,21% -16.33% 23,58%

(a) Robot 1 (b) Robot 2

(e) Robot 5 (f) Robot 5

Figure 10: Trajectory of the robots on the environment #1

using the partitioned strategy.

scene autonomously into smaller sectors, in which

just one robot remains on the sector, performing the

surveillance task. The fact that a robot does not in-

vade other’s areas avoids the redundant monitoring of

the same region, distributing the surveillance equally

on the environment.

As future works, it is intended to verify the inﬂu-

ence of the pheromone evaporation and dispersion ra-

tes on the number of robots and the dimensions of the

environment. Investigate the behavior of the agents

with scenes without obstacles and develop a strategy

for this scenario that can simulate, for example, a dis-

tribution task of a mobile phone operator’s antennas.

Also it is intended to make feasible the PSS strategy

on real robots, along with mapping and location algo-

rithms needed for the surveillance task.

(a) Robot 1 (b) Robot 2

(e) Robot 5 (f) Robot 5

Figure 11: Average of pheromone dispersion on the envi-

ronment #1 using the partitioned strategy.

REFERENCES

Adhau, S., Mittal, M., and Mittal, A. (2013). A multi-

agent system for decentralized multi-project schedu-

ling with resource transfers. International Journal of

Production Economics, 146(2):646 – 661.

Anisi, D. A., Ogren, P., and Hu, X. (2010). Coopera-

tive minimum time surveillance with multiple ground

vehicles. IEEE Transactions on Automatic Control,

55(12):2679–2691.

Calvo, R., Constantino, A., and Figueiredo, M. (2015). A

multi-pheromone stigmergic distributed robot coordi-

nation strategy for fast surveillance task execution in

unknown environments. In Neural Networks (IJCNN),

2015 International Joint Conference on, pages 1–8.

Calvo, R., de Oliveira, J., Figueiredo, M., and Francelin Ro-

mero, R. (2012). A bioinspired coordination strategy

A Bio-inspired Approach in Decision-making of Multiple Robots Applied on Partitioned Surveillance Task

317

for controlling of multiple robots in surveillance tasks.

In International Journal on Advances in Software, vo-

lume 5, pages 146–165. IARIA.

Dorigo, M. (1992). Optimization, Learning and Natural

Algorithms. PhD thesis, Politecnico di Milano, Italy.

Eoh, G., Choi, J. S., and Lee, B. H. (2013). Faulty

robot rescue by multi-robot cooperation. Robotica,

31(8):12391249.

Fazli, P., Davoodi, A., and Mackworth, A. K. (2013). Multi-

robot repeated area coverage. Autonomous Robots,

34(4):251–276.

Krishnan, D. (2015). A distributed self-adaptive intrusion

detection system for mobile ad-hoc networks using

tamper evident mobile agents. Procedia Computer

Science, 46:1203 – 1208. Proceedings of the Inter-

national Conference on Information and Communica-

tion Technologies, ICICT 2014, 3-5 December 2014

at Bolgatty Palace & Island Resort, Kochi, India.

Liemhetcharat, S., Yan, R., Tee, K. P., and Lee, M. (2015).

Multi-robot item delivery and foraging: Two sides of

a coin. Robotics, 4(3):365–397.

Robu, V., Noot, H., Poutr, H. L., and van Schijndel, W.-

J. (2011). A multi-agent platform for auction-based

allocation of loads in transportation logistics. Expert

Systems with Applications, 38(4):3483 – 3491.

Sun, Q., Yu, W., Kochurov, N., Hao, Q., and Hu, F. (2013).

A multi-agent-based intelligent sensor and actuator

network design for smart house and home automation.

Journal of Sensor and Actuator Networks, 2(3):557–

588.

Wallar, A., Plaku, E., and Sofge, D. A. (2015). Reactive

motion planning for unmanned aerial surveillance of

risk-sensitive areas. IEEE Transactions on Automa-

tion Science and Engineering, 12(3):969–980.

Xiang, W. and Lee, H. (2008). Ant colony intelligence

in multi-agent dynamic manufacturing scheduling.

Engineering Applications of Artiﬁcial Intelligence,

21(1):73 – 85.

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

318