Draft version.

Cite this work as:

A.L. Alfeo, M.G.C.A. Cimino, N. De Francesco, A. Lazzeri, M. Lega, G. Vaglini, "Swarm coordination of mini-UAVs for target search using imperfect
sensors", Intelligent Decision Technologies, IOS Press, Vol. 12, Issue 2, Pages 149-162, 2018
The final publication is available at IOS Press through http://dx.doi.org/10.3233/IDT-170317

Swarm coordination of mini-UAVs for target

search using imperfect sensors
Antonio L. Alfeo1, Mario G. C. A. Cimino1,*, Nicoletta De Francesco1, Alessandro
Lazzeri1, Massimiliano Lega2, Gigliola Vaglini1
1
Department of Information Engineering, University of Pisa, Largo L. Lazzarino 1, 56127 Pisa, Italy
2
Department of Engineering, University of Naples "Parthenope", 80143 Naples, Italy

[email protected], [email protected], [email protected],

[email protected], [email protected], [email protected]

* Corresponding author:

Mario G. C. A. Cimino, [email protected]

Tel: +39 050 2217 455; Fax: +39 050 2217 600.

Abstract Unmanned Aerial Vehicles (UAVs) have a great potential to support search tasks

in unstructured environments. Small, lightweight, low speed and agile UAVs, such as

multirotors platforms can incorporate many kinds of sensors that are suitable for detecting

object of interests in cluttered outdoor areas. However, due to their limited endurance,

moderate computing power, and imperfect sensing, mini-UAVs should be into groups using

swarm coordination algorithms to perform tasks in a scalable, reliable and robust manner. In

this paper a biologically-inspired mechanisms is adopted to coordinate drones performing

target search with imperfect sensors. In essence, coordination can be achieved by combining

stigmergic and flocking behaviors. Stigmergy occurs when a drone releases digital

pheromone upon sensing of a potential target. Such pheromones can be aggregated and

diffused between flocking drones, creating a spatiotemporal attractive potential field.

Flocking occurs, as an emergent effect of alignment, separation and cohesion, where drones

self organise with similar heading and dynamic arrangement as a group. The emergent

coordination of drones relies on the alignment of stigmergy and flocking strategies. This
 2
paper reports on the design of the novel swarming algorithm, reviewing different strategies

and measuring their performance on a number of synthetic and real-world scenarios.

Keywords Swarm intelligence, mini-UAV, Stigmergy, Flocking, Target search

1 Introduction and Problem Statement

In recent years, several research groups are working on new procedures and

technologies to operate and monitor complex scenarios. Two specific areas include search

and rescue and environmental monitoring. Both these topics require solutions to critical

issues related to the mission requirements and the mission profile. The choice of a specific

aerial platform for the monitoring of complex scenarios should be made by examining

particular correspondence to the needs of the mission at the same time, and the multiplying

effect of what is measurable by sensors positioned on the ground as fixed configuration.

Advanced aerial platforms such as Unmanned Aerial Vehicles (UAVs), often called

drones, are today the most frequent response to the needs of different missions. In

particular, the specific category of mini-

are generall

have recently received a strong technological acceleration thanks to recent advances in

miniaturization of battery, of communication, processing and sensing technology [1].

The remote/proximal sensing data obtained using mini-UAVs were validated in

several environmental monitoring missions with complex scenarios as reported in previous

research; these include: fusion of optical data with synthetic aperture radar data to detect

environmental hazards [2,3], use of thermal imagery to monitor landfills [4], surface

waters contamination [5] and to detect illegal dumping [6,7] and to identify other illegal

activities [8]. In addition, remote sensing data can be strategically combined with other

data layers in geographic information systems to monitor the vulnerability of cultural sites

[9] and anticipate environmental violations [10,11].

3
The use of a range of aerial platforms and advanced sensors to detect the illegal

activities was validated in several real missions in Italy [12,13]. These are the first known

use of these methods in both the fields of environmental research and law

enforcement/environmental forensics. They also represent an example of collaboration

between law enforcement and university teams on developing enhanced environmental

protection methods.

In the operational surveillance for successful identification and prosecution of

environmental pollution culprits it is required an integrated system based on data from

several sources. The surveillance service must also include geospatially tagged forensic

data analysis (information arising from navigation/positional systems).

The detection, identification and localisation of a target are key elements in all the

above operations. Groups of mini-UAVs equipped with self-localisation and sensing

capabilities offer new opportunities; indeed, groups of mini-UAVs can explore cluttered

outdoor environments, where access to conventional platforms is inefficient, limited,

impossible, or dangerous. In brief, the main motivations for adopting the UAV technology

in the survey process are the following: reduction of risk of human falling, reduction of

safety costs for plant stoppage, improved data density and quality due to a better

proximity, accessibility to locations where people or vehicles have no access, faster and

cheaper data acquisition due to the involvement of less workforce and equipment.

The coordinated swarming drones could be also considered as single array of sensors

configured to the measure of a host of environmental parameters. In search and rescue

tasks, for example, a more effective appr to

identify key locations as quick as possible. This exclusion process enables organisers to

rescan the key locations that provided some circumstantial evidence. In this context, the

quality of the sensing has also a direct impact on the overall mission performance [14].
 4
Therefore, an important aspect of the swarm coordination is the possibility to require a

sufficient number of redundant samples of the target to reliably classify it

A cooperative approach that exploits drones sensing, minimizes the error in target

recognition [15]. In contrast, to use a unique drone implies costly structure and design, as

well as vulnerability. Hence, a number of considerations support the use of coordinated

swarming drones. An important requirement of the coordination strategy is to avoid

centralized control approaches, leading to exponential increases in communication

bandwidth and software complexity [16]. Swarm intelligence methodologies can be

investigated to solve problems cooperatively while maintaining scalability. The main

inspiration for swarming drones comes from the observation of social animals, such as

insects, winged animals, and fish, that exhibit a collective intelligence which appears to

achieve complex goal through simple rules and local interactions [17]. The main benefits of

a swarm drones includes: robustness (for the ability to cope with the loss of individuals);

scalability (due to the ability to perform well with different group size); and flexibility

(thanks to the capability to manage a broad spectrum of different environments and tasks).

To this aim, each individual of the swarm: acts with a certain level of autonomy; performs

only local sensing and communication; operates without centralized control or global

knowledge, and cooperates to achieve a global task [17].

In this paper, different coordination strategies are reviewed and tested empirically with

both synthetic and real-world scenarios, with obstacles having irregular complex shapes.

For this purpose, it is adopted a multi-agent simulation platform with the possibility of

importing environments with obstacles and targets sampled from real landscapes.

The paper is structured as follows. In Section 2 early requirements and coordination

strategies are reported. Section 3 briefly characterizes the related work. In Section 4, the
5
analysis and the integration of the emergent schemes is covered. Section 5 reports on the

design of the algorithm. Experimental studies are detailed in Section 6. Section 7 draws

conclusions and future work.

2 Early requirements and coordination strategies

From a structural standpoint, it is assumed that each mini-UAV is provided with the

following capabilities: wireless communication capability for sending and receiving

information from the ground station; self-location capability based on Global Position

System (GPS) and inertial technology, returning the coordinates of its current location; one

or more target sensing technology, capable of acquire data in the area over which it flies;

processor with limited computing capability; obstacle avoidance capability, that is, locally

managed detection and steering to avoid flying towards surrounding barriers and drones.

Moreover, it is assumed that a certain level of uncertainty comes from noisy of faulty sensor

measurements.

Marker-based stigmergy is a fundamental swarm coordination mechanism, based on

the release of information in the environment in the form of pheromones [18,19]. The

pheromone is a volatile substance diffused locally and staying temporarily for other

individuals that can properly react and modify their behavior [20]. Simulated (that is,

digital) pheromones can be used to coordinate groups of drones for various tasks. In a

distributed environment, a pheromone map of the search space can be maintained and made

[21].

When the sensing system of a drone determines a potential target, it tries to trigger the

cooperation of its swarm to achieve reliable sensing and target detection.

Reliable sensing. Sensors on mini-UAV can generate faulty measurements for a number

of reasons, such as power loss, software failure, small bias, miscalibration, slow-drifting,

loss of accuracy, temporary freezing, to name a few [22]. In the literature, some
 6
approaches to fault recognition assume the fault types can be described by a static

parameterized model. If parameterized models for the fault types are available, a fault

recognition algorithm can be applied. However, without health monitoring a static fault

model is often not known [22]. In practice, there is likely to be some deviation between

what the actual faults look like and what their models predict. This residual may be

irrelevant for a single sensor system. However, in a conventional distributed approach

even small residuals can have a significant impact on the overall effectiveness, due to the

high number of occurrences potentially involved. In contrast, swarm systems could

exhibit much higher levels of robustness, in the sense of tolerance to individual (or few)

residual(s), than in conventional distributed systems. Nevertheless, a simplistic

modelling approach may make incorrect assumptions, because the question of how many

agents are needed to guarantee a required emergent behavior in a particular swarm and

for a particular behavior is not straightforward [23]. This potential tolerance cannot be

natively assumed without special analysis, design, and test, since swarm systems can

exhibit a number of unexpected behaviors. Therefore, the proposed coordination

algorithm needs to incorporate some mechanism able to exploits the inherent collective

influence between measures, in order to verify its effectiveness under assumption of

uncertainty in individual sensing. To this aim, this study tries to achieve a control on the

number of redundant measures of the targets that are sufficient to ensure a sufficient

level of reliability.

Target detection. For a distributed target, the detection process is the identification of

any parts of it, with sufficient detail to permit the intended action. For example: to detect

a landmine means to find the location of it to avoid being maimed or killed. To detect

radioactive substance means to trace perimeters were radioactivity levels are considered

dangerous. To detect gas leak means to identify the area were natural gas seeping from
7
the ground implies fire and explosion hazards, and so on. The search problem is

formulated by discretizing the environment into a set of cells. Each target is stationary

and usually covers many cells. The objective is to determine in which cells the targets

lie. Due to the distribution, the task requires that drones are dynamically arranged so as

to be efficiently engaged when some member detects a part of the possible target.

For this purpose, the drone releases a particular amount of pheromone on the cell of the

sensed possible target, whose diffusion acts as an attractive potential on neighboring drones.

To be attracted by pheromone trails, the available drones should be spatially organized into

flocks. Flocking is a strategy to allow the self-organization of drones into a number of

flocks. Flocking behavior is an emergent effect of individual rules based on alignment,

separation and cohesion [24]. With alignment rules the drones tends to move in the same

direction that nearby drones. With separation rules, the drone keeps a minimum distance

able to provide the drone with flexibility when moving in the swarm, and for a better

exploration. Finally, with cohesion rules the drone tends to move towards the swarm.

As a result of flocking, each member of a flock has approximately the same heading of

the other members, and attempts to remain in range with them. For this purpose, the

structural dimensions of the pheromone should take into account the average size of a

swarm (or vice versa). Otherwise, a highly diffused or poorly evaporated pheromone could

attract disproportionate resources on a single target, thus interfering with the progressive

development of the emergent behavior. In contrast, a poorly diffused or highly evaporated

pheromone could not be sensed at all.

As an effect of pheromone attraction, other drones can confirm the possible target

through repeated sensing, and can surround the detected location in order to map the whole

distribution. Thus a considerable amount of pheromone is aggregated for each possible

target. Once a predefined number of drones confirmed the sensing of the possible target, it
 8
is definitively considered to be a true target and then its sensing cannot activate additional

pheromone. Since pheromone evaporates over time, after a certain time the pheromone

intensity cannot be reinforced in a fully explored region, and in practice disappears.

In the presented approach, stigmergy and flocking are two emergent behavioral

patterns which should work in conjunction with other basic behavioral patterns of the drone,

such as obstacle and boundary avoidance. The process of designing a combination strategy

is bottom-up and consists in finding the right setting at the micro-level (agent-level) in order

to obtain a coherent emergent behavior at macro-level (swarm-level) [25].

3 Related Work

The goal of this section is to briefly characterize the main approaches and results in the

literature on stigmergic mechanisms coordinating swarms of small robots to perform target

search or similar tasks. The published works in the field can be distinguished into three

categories: using a physical substance as a pheromone, which is necessarily transmitted in

an indirect way between robots, by means of the physical environment; using a digital

pheromone, transmitted via direct communication between robots; using a digital

pheromone, transmitted via an indirect communication between robots. The latter is the

category of our approach.

Kuyucu et al. in [26] use a swarm of robots releasing physical substance as a repulsive

pheromone, for environment exploration. In particular, robots act combining three basic

behaviors, with decreasing priority: wall avoiding, pheromone coordination, and random

walk. Actually there are various approaches in the literature using physical pheromones,

because they do not require a computational structure. Although real pheromones are not

usable with aerial vehicle, they can be simulated. Thus, this type of research can be

interesting to model new types of digital stigmergy.

An example of stigmergic coordination between drones using direct communication is

9
presented by Dasgupta [27], where he focuses on automatic target recognition. Potential

target are marked by drones, which also communicate the gossiped pheromone to nearby

drones, with probability inversely proportional to the distance from the source. The

proposed stigmergic schema employs also repulsive pheromone, as a negative feedback,

when a predefined number of drones identify the same target. A disadvantage of such

scheme is that the bandwidth required goes into an exponential explosion as the population

grows. To avoid redundancy in target evaluation each UAV has to maintain in memory the

state of each potential and confirmed target. In this way, the direct communication in the

swarm should be strongly limited [28].

A swarm coordination schema with indirect coordination is proposed by Sauter et al.

[18]. Here the coordination of a swarm of vehicles is based on digital pheromones

maintained in an artificial space called pheromone map and composed by an arbitrary graph

of place agents, that is, intermediate control nodes. There are two classes of agents which

deposit, withdraw, and read pheromones, that is, walkers and avatars. A walker agent aims

to make movements and action decisions, whereas avatars collect location information to

make estimates when sensor information is not available. The schema has been applied to a

range of scenarios, among which target acquisition. An important problem of this approach

is that the exploration depends on the initial position of the swarm. This model does not

consider complex targets but only simple targets without structure.

To handle the unreliability in sensing, a certain number of drones must be attracted on

a potential target. To achieve this goal a spatial organization of the available drones is

required in order to sense the pheromone deposit released during a survey leaded by a peer

of the same group. This result can be achieved keeping flocking formation. Flocking

behavior is exhibited during the group flight. It is an emergent effect caused by the

observance of three rules: preserving heading alignment with flock-mates, while

 10
maintaining separation with respect to the nearest one and cohesion with the entire group, as

described by Reynolds [24]. This flocking behavior formalization have been extensively

used in swarming robots and drones coordination. Bouraqadi et al. [29] accomplish an

unknown environment survey via a group of robots which has to stay close enough to

maintaining the ability to communicate with each other. This objective is reached using

Reynolds rules to organize the robots distribution and movements. Hauert et al. [30] apply

flocking rules for the management of a drones swarm in order to keep an ad-hoc network

during their flight and to coordinate their task. However, this application is based on the

assumption of well-known search field, and then it is not applicable to unstructured

environments, which is one of our requirements.

4 Behavioral specification of the proposed approach

This section aims to characterize the emergent behavior of the coordination algorithm,

via the integration of a variety of mechanisms. This purpose is achieved using the Tropos

agent-oriented methodology [31]. Tropos is based on the notion of agent, which in this

context is a drone, with related notions such as goals and plans. It allows a clear modeling

of the operating environment and of the interactions that should occur between drones.

Figure 1a shows a legend of the main concepts: actor, goal, plan, resource, capability, and

social dependency between actors for goal achievement. Actors may be further specialized

based on roles (circle with a bottom line) or agents (circle with upper line). A software

agent represents a physical instance (human, hardware or software) of an actor that

performs the assigned activities. A role represents a speci c function that, in different

circumstances, may be played by the agents. Edges Edges between nodes form

dependencies of the form: actor goal/task/resource actor . In additional to hard

goals, soft goals are also used when having no clear-cut definition and/or criteria as to

whether they are satisfied, for example for modeling goal/plan qualities and non-functional
11
requirements [32]. A detailed account of modeling activities can be found in [31].

Figure 1c represents a top view of the proposed algorithm. More specifically, on the

bottom, a Physical Environment is a resource modeling the search field, which contains all

the physical elements interacting with drones, whereas a Virtual Environment is a resource

managing virtual pheromones and the targets (cells discovered or not). In the middle, Drone

is the main actor, supporting the primary goal look for target, collectively attained by two

levels of organization: the flock, that is, the organism consisting in locally coordinated

drones, and the swarm, that is, the organism consisting in globally coordinated flocks.

Conversely, a drone depends on the swarm for saving fly time, since the coordinated search

is purposely organized to reduce the overall time. This purpose is based on the resource

accomplishment time, managed via update plans of the virtual environment (on the bottom

right of the figure): count target found and time unit. Other update plans of the virtual

environment are diffuse and evaporate pheromones. The basic needs of a drone consist in

the sensing procedures, carried out via both the physical and virtual environments, whereas

the basic soft goals of a drone consist of: to cover the search space, collective flight, to

point towards targets, and to follow obstacle-free paths. Such soft goals are attained via

related roles (in ascending order of priority): obstacle avoider, tracker, flockmate, and

explorer [33]. Above all, Fig. 1d shows how a drone reacts to local conditions. Each role is

further detailed in Fig. 1b. Fig. 1e represents the obstacle avoider role, with the first priority

level. At the second priority level, Fig. 1f represents the tracker role. Fig. 1b shows the third

priority level, flockmate. Finally, Fig. 1g represents the minimum priority level, explorer.

The above specification is a mixed actor-dependency model in which

explained. As a result, the drone task and goals and its precedencies between roles have

been detailed. The next Section focuses on the system design, to show how to implement
 12
and integrate the main functional and architectural components.

(a)

(b) (c)

(d) (e) (f) (g)

Fig. 1. Behavioral representation of the proposed approach

5 Architectural and functional Design of the main subsystems

This section is devoted to the modeling of environment and drones.

13
4.1 The design of the environment: pheromones and error dynamics

It is assumed that the environment is constrained to a specific area. Without loss of

generality, the area is discretized through a grid consisting of C2 cells, each identified by a

pair (x,y) of coordinates, with x,y [ C]. The actual size of the area and the number of

discretized squares depend on the specific application domain. Figure 2 shows a basic

scenario of the pheromone dynamics, focused to the most significant stages of diffusion and

evaporation. The levels of pheromone intensity are represented by different grey gradations:

the darker the gradation is, the higher the intensity.

(a) t = 1 (b) t = 2 (c) t = 3 (d) t = 5

(e) t = 20 (f) t = 25

Fig. 2. Basic scenario of pheromone dynamics: (a) releasing; (b) mainly diffusing; (c-d)

diffusing and evaporating; (e-f) mainly evaporating.

More specifically, in Fig. 2: (a) a single pheromone intensity I is released; (b) at the first

steps, the pheromone is mainly diffusing (moving) to the nearby cells, with a constant

diffusion rate [0,1] , StigDiffusion; (c-d) the pheromone is diffusing and evaporating; by

evaporating, the pheromone decreases its intensity over time; it is ruled by the constant rate

[0,1] , StigEvaporation; (e-f) the pheromone is mainly evaporating. More formally, the

pheromone intensity p released at the instant t on the cell (x,y) is then characterized by the
 14
following dynamics:

px, y (t) (1 ) px, y (t 1) px, y (t 1, t) d x, y (t 1, t) (1)

where (1- ) px,y(t-1) represents the amount remaining after diffusion to nearby cells, px,y(t-

1,t) the additional deposits made within the interval (t-1,t], and dx,y(t-1,t) the input

pheromone diffused from all the nearby cells. The latter can be formally calculated as:

1 1
d x , y (t 1, t ) px i,y j (t 1) (2)
8i 1j 1
( i , j ) (0,0)

since each of the 8 neighbor cells propagates the portion of its pheromone to the cell (x,y)

at each update cycle. The total amount in (1) is also multiplied by (StigEvaporation) to

take the evaporation into account.

The Environment supports also the management of the target detection with imperfect

sensors. It is assumed that each target sensing can provide both false positive and false

negative. However, this occurs, with a certain error probability of , only while checking

the target cell or the cells adjacent to the target, as represented in Fig. 3.

Fig. 3. Imperfect sensor model

In essence, it is introduced the notion of degradation of the sensing quality as a function

of the proximity to the target: as the proximity increases, the sensing may generate an

altered measure resulting in a wrong detection. This assumption implies that the

discontinuities represented in Fig.1 should be small with respect to the source signal.

4.2 The design of the drone behavior

15
The drone behavior is structured into a prioritized logic, where each priority level

implements one basic behavior, or role. At each update cycle, or tick, the role assumed by

the drone is a consequence of the environmental sensing. In descending order or priority,

the roles are: obstacle avoider, tracker, flockmate and explorer.

Figure 4 shows an overall representation of the drone behavior, using a UML (Unified

Modeling Language) activity diagram. Here, every tick period, represented by the hourglass

icon on bottom left, the environment updates his status, whereas the drone performs in

parallel: the target detection, in which case it releases pheromone controlled by

StigDiffusion and StigEvaporationRate parameters; the obstacle avoider. If a close object is

detected, within the ObstacleVision radius, the drone points toward a free direction, when

available, and moves forward. Otherwise, if there are no close objects detected, the drones

play the tracker role: it tries to sense pheromone within the Olfaction radius and, if

detected, points toward the pheromone peak. Alternatively, if pheromone is not detected,

the drone plays the flockmate role: it tries to detect surrounding drones within the

FlockVision radius, in order to point toward the flock. Finally, if there are no surrounding

drones, as an explorer it performs a random turn within the WiggleVar angle, and then

moves forward. Figure 5 represents a detailed modeling of the main procedures and roles

played by a drone.

Figure 5a models the basic drone behavior consisting in releasing attractive pheromone

with StigIntensity intensity, upon target detection and moving forward according to a given

velocity set to DroneVel. In Fig. 5b, Fig. 5c, and Fig. 5d the obstacle avoider, the tracker

and the flockmate roles are modelled, respectively. Figure 6 shows the main procedures of

[24]. In the flocking

behavior, the drone takes into account only drones within a FlockVision. Figure 6a

represents the separation behavior: drones close to others have to separate for better
 16
exploration; thus, if a drone senses another drone closer than the MinimumSeparation, it

turns by an angle MaxSeparateTurn. Figure 6b shows the alignment behavior: the drone

calculates the average direction of the drones in the flock vision and turns by an angle

MaxAlignTurn to conform its direction to the flock direction. Figure 6c illustrates the

cohesion behavior: the drone calculates the barycenter of the drones in the Flock vision and

turns by an angle MaxCohereTurn towards the barycenter.

Fig. 4. Overall modeling of the drone behavior modularized in roles.

17

(a) Drone behavior (b) Obstacle avoider

(c) Tracker (d) Flockmate

Fig. 5. Detailed modeling of the main roles played by a drone

(a) Separation (b) Alignment (f) Cohesion

Fig. 6. Illustration of the procedures of the flocking behavior
 18

6 Experimental studies

The proposed model has been implemented on NetLogo1, a leading simulation platform

for swarm intelligence. The output of the system is the total time needed to find the 95% of

targets. According to Fig. 4, the model requires 12 parameters, to be tuned via three-phases:

early analysis, under the assumption of reliable sensing (that is, sensing error probability

and sensing redundancy set to 0.1 and 1, respectively); parameter sensitivity analysis on

representative scenarios, by evaluating the uncertainty in the output for each parameter;

finally, accurate setting on each of the most sensitive parameters, via a bisection method to

find the value minimizing the output.

main structural and behavioral parameters of the model, with their range and their value set.

Table 1 Structural and behavioral parameters.

Name Description (unit measure) Range Set v.

DroneVel Drone horizontal speed (m/s) (0,15) 1
WiggleVar Drone max rand- (0,180) 150
ObstacleVision Drone object sensing distance (m) (0, 5) 2
FlockVision Flock visibility radius (m) [0, 50] 7
MinimumSeparation Flock mobility distance (m) [0,5] 3
MaxSeparateTurn (0,180) 30
MaxAlignTurn Fl (0,180) 20
MaxCohereTurn (0,180) 5
Olfaction Pheromone sensing distance (m) (0, ) 1
StigIntensity Pheromone release intensity (0, ) 40K
StigDiffusion Pheromone diffusion rate (%) [0,1] 0.85
StigEvaporation Pheromone evaporation rate (%) [0,1] 0.05
SensingError Sensing error probability (%) (0, 100) [0.1,1]
Redundancy Sensing Redundancy (0, ) {1,3,5}

The algorithm has been tested on four different scenarios, such as Field, Dumps, Urban

and Mines. The Field scenario is made by 5 groups of targets scattered over the area, with

about 10 targets per group. There are no obstacles. Figure 7 shows a snapshot with the
19
spatial arrangement of flocks of different forms and sizes, together with four stigmergic

formations on corresponding groups of targets. Here, it can be observed that stigmergic

formations attracted flocks of drones.

Fig. 7. A snapshot of the Field scenario with flocks and stigmergic formations

An initial configuration of the Field scenario is shown in Fig. 8a. Here, 80 total drones

(represented by triangular forms) are arranged into four dense flocks, placed at the

antipodes of the area, whereas the targets are represented by clusters of black dots. The

second scenario, called Dumps (Fig. 8b) represents a synthetic reconstruction of woodland

with three abusive garbage aggregations, modelled by three groups of targets. Here, 30

targets and 100 trees are represented by gray and black dots, respectively. 80 total drones,

arranged into 4 flocks, are initially placed at the antipodes of the area. The third scenario,
 20
that is, Urban (Fig. 8c) is characterized by two cluster of 110 total targets placed on two

sides of corresponding buildings. Overall, 7 total buildings are located. 40 drones, arranged

into four flocks, are placed at the antipodes of the area, with no trees at all. Finally, the

Urban Mines (Fig. 8d) scenario is derived from real-world examples of areas near Sarajevo,

in Bosnia-Herzegovina, with landmine objects, selected from publicly available maps 2.

Recently, some authors actually proposed the use of mini-UAVfor detecting landmines

[34]. Drones have been initially placed on the boundaries of the area. With respect to the

map of the first three scenarios, whose area is 200 square meters, in the last scenario the

area is 400 square meters.

To carry out the experiments under the requirement of imperfect sensor model, a sensing

error probability in the interval [0.1, 1] percent with uniform distribution has been added.

Then, the system output has been evaluated by requiring a prefixed number of repeated

measures of the targets in the termination criterion, that is, sensing redundancy values 3 and

5.

To assess the effectiveness of the proposed approach, the performance of the model has

been evaluated on three approaches: tigmergic

Stigmergic and Flocking aproach 10 trials have been carried

out. It has been determined that the resulting performance indicator samples are well-

modeled by a normal distribution, using a graphical normality test. Hence, the 95%

confidence intervals have been calculated. Table 2 summarizes, for each scenario, the

that the use of stigmergy speeds up the target search process in any scenario. Moreover,

results become even better in combination with flocking. It can be remarked that all

scenarios have been processed by using a general purpose parameterization determined a

21
with reliable sensor model. Indeed, a parameterization ad initialization adapted to types of

scenario might produce better results.

Table 2. Characteristics and numerical results (mean confidence interval) of each

scenario.

Field Dumps Urban Urban

Mines
# targets 50 30 110 40
# clusters 5 3 2 40
# trees 0 100 0 54
# buildings 0 0 7 28
# drones 80 80 40 25
R (1) 2,604 248 2,252 212 2,340 229 651 55
Algorithm (redundancy)

S (1) 1,383 126 1,297 102 1,748 188 560 49

S+F (1) 1,078 106 1,009 141 1,259 102 487 29
R (3) 4,161 269 3,993 266 3,688 286 944 55
S (3) 1,758 151 1,513 116 2,089 197 707 84
S+F (3) 1,484 147 1,289 135 1,861 166 594 34
R (5) 6,173 361 6,163 399 4,647 271 1,167 51
S (5) 2,109 246 2,208 208 2,488 280 770 93
S+F (5) 1,591 136 1,823 233 2,102 151 726 32

(a) (b)
 22

(c) (d)
Fig. 8. Models of three synthetic and one real-world scenarios: (a) Field; (b) Dumps; (c)

Urban; (d) Urban Mines.

To better highlight the scalability of our approach against redundancy, Fig. 9a-d shows

the completion time for redundancy 1, 3 and 5, for each scenario. Here, it is apparent that

Stigmergy introduces a significant improvement of trend over Random Fly, both alone and

combined with flocking behavior.

(a) (b) (d)

23
(c)

Fig. 9. Completion time against redundancy, for each scenario and with different
approaches: Random Fly (dotted line), Stigmergy (dashed), and Stigmergy + Flocking
(solid).

7 Conclusions and Future Work

In this paper, it is presented a novel swarm approach for coordinating mini-UAVs to

perform target search with imperfect sensors. The approach uses a stigmergic behavior,

consisting in the release of information in the environment in the form of attractive digital

pheromones, in areas where potential targets are sensed. Moreover, the approach employs

flocking behavior, resulting in a flexible arrangement of drones according to the stigmergic

potential field. The paper illustrates the approach from the behavioral and architectural

point of views, and then discusses the experimental studies. Results on synthetic and real-

world scenarios prove the benefits of both stigmergy and flocking, in terms of tolerance to

errors and scalability for increasing redundancy requirements.

The overall mechanism can be better enabled if structural parameters are correctly tuned

for the given scenario. Determining such correct parameters is not a simple task since

different areas have different features. Thus, an appropriate tuning to adapt parameters to

the specific search area is desirable to make the search more effective. For this purpose, to

use a parameter optimization strategy is considered a key investigation activity for future

work.

Acknowledgements

This work was carried out in the framework of the SCIADRO project, cofunded by the

Tuscany Region (Italy) under the Regional Implementation Programme for Underutilized

Areas Fund (PAR FAS 2007-2013) and the Research Facilitation Fund (FAR) of the

Ministry of Education, University and Research (MIUR)

 24

References

1. K. Whitehead, C. H. Hugenholtz, S. Myshak, O. Brown, A. LeClair, A. Tamminga, T.

E. Barchyn, B. Moorman, B. Eaton, Remote sensing of the environment with small

unmanned aircraft systems (UASs), part 2: scientific and commercial applications,

Journal of Unmanned Vehicle Systems, 02 (2014), 86-102.

2. A. Errico, C. V. Angelino, L. Cicala, G. Persechino, C. Ferrara, M. Lega, A. Vallario,

C. Parente, G. Masi, R. Gaetano, G. Scarpa, D. Amitrano, G. Ruello, L. Verdoliva, G.

Poggi, Detection of environmental hazards through the feature-based fusion of optical

and SAR data: a case study in southern Italy, International Journal of Remote Sensing,

36 (2015), 3345-3367.

3. A. Errico, C. V. Angelino, L. Cicala, D. P. Podobinski, G. Persechino, C. Ferrara, M.

Lega, A. Vallario, C. Parente, G. Masi, R. Gaetano, G. Scarpa, D. Amitrano, G. Ruello,

L. Verdoliva, G. Poggi, SAR/multispectral image fusion for the detection of

environmental hazards with a GIS, Proceedings of SPIE - The International Society for

Optical Engineering, 9245 (2014), doi: 10.1117/12.2066476.

4. M. Lega, R. M. A. Napoli, A new approach to solid waste landfills aerial monitoring,

WIT Transactions on Ecology and the Environment, 109 (2008), 193-199.

5. M. Lega, R. M. A. Napoli, Aerial infrared thermography in the surface waters

contamination monitoring, Desalination and Water Treatment, 23 (2010), 141-151.

6. G. Persechino, P. Schiano, M. Lega, R. M. A. Napoli, C. Ferrara, J. Kosmatka,

Aerospace-based support systems and interoperability: The solution to fight illegal

dumping, WIT Transactions on Ecology and the Environment, 140 (2010), 203-214.
25
7. G. Persechino, M. Lega, G. Romano, F. Gargiulo, L. Cicala, IDES project: An

advanced tool to investigate illegal dumping, WIT Transactions on Ecology and the

Environment, 173 (2013), 603-614.

8. M. Lega, C. Ferrara, G. Persechino, P. Bishop, Remote sensing in environmental police

investigations: Aerial platforms and an innovative application of thermography to

detect several illegal activities, Environmental Monitoring and Assessment, 186 (2014),

8291-8301.

9. M. Lega, L. D'Antonio, R. M. A. Napoli, Cultural Heritage and Waste Heritage:

Advanced techniques to preserve cultural heritage, exploring just in time the ruins

produced by disasters and natural calamities, WIT Transactions on Ecology and the

Environment, 140 (2010), 123-134.

10. M. Lega, G. Persechino, GIS and infrared aerial view: Advanced tools for the early

detection of environmental violations, WIT Transactions on Ecology and the

Environment, 180 (2014), 225-235.

11. F. Gargiulo, G. Persechino, M. Lega, A. Errico, IDES project: A new effective tool for

safety and security in the environment, Lecture Notes in Computer Science (LNCS),

8286 (2013), 201-208.

12. M. Lega, D. Ceglie, G. Persechino, C. Ferrara, R. M. A. Napoli, Illegal dumping

investigation: A new challenge for forensic environmental engineering, WIT

Transactions on Ecology and the Environment, 163 (2012), 3-11.

13. M. Lega, J. Kosmatka, C. Ferrara, F. Russo, R. M. A. Napoli, G. Persechino, Using

Advanced Aerial Platforms and Infrared Thermography to Track Environmental

Contamination, Environmental Forensics, 13 (2012), 332-338.

 26
14. L. F. Bertuccelli, J. P. How, Robust UAV search for environments with imprecise

probability maps, Proceedings of the 44th IEEE Conference on Decision and Control,

(2005), 5680-5685.

15. B. Bethke, M. Valenti, J. How, Cooperative vision based estimation and tracking using

multiple UAVs, In Advances in Cooperative Control and Optimization, Springer,

Berlin Heidelberg, (2007), 179-189.

16. R. McCune, R. Purta, M. Dobski, A. Jaworski, G. Madey, Y. Wei, A. Madey and M.B.

Blake, Investigations of DDDAS for command and control of uav swarms with agent-

based modeling, Proceedings of the 2013 Winter Simulation Conference: Simulation:

Making Decisions in a Complex World, (2013), 1467-1478.

17. M. Brambilla, E. Ferrante, M. Birattari, M. Dorigo, Swarm robotics: a review from the

swarm engineering perspective, Swarm Intelligence, 7 (2013), 1-41.

18. J. A. Sauter, R. Matthews, H. Van Dyke Parunak, S. A. Brueckner, Performance of

digital pheromones for swarming vehicle control, Proceedings of the fourth

international joint conference on Autonomous agents and multiagent systems, (2005),

903-910.

19. M.G.C.A. Cimino, A. Lazzeri, G. Vaglini, Combining stigmergic and flocking

behaviors to coordinate swarms of drones performing target search, Proceedings of the

Sixth International Conference on Information, Intelligence, Systems and Applications

(2015), 1-6.

20. H. Van Dyke Parunak, M. Purcell, R. O'Connell, Digital pheromones for autonomous

coordination of swarming UAV's, Proceedings of the 1st Technical Conference and

Workshop on Unmanned Aerospace Vehicles, (2002), 48105-1579.

21. G. Ermacora, A. Toma, B. Bona, M. Chiaberge, M. Silvagni, A cloud robotics

architecture for an emergency management and monitoring service in a smart city

27
environment, Proceedings of the International Conference of Intelligent Robots and

Systems, (2013).

22. S. Reece, S. Roberts, C. Claxton, D. S. Nicholson, Multi-sensor fault recovery in the

presence of known and unknown fault types, Proceedings of the 12th International

Conference on Information Fusion, (2009).

23. A. F. T. Winfield, C. J. Harper and J. Nembrini, Towards the Application of Swarm

Intelligence in Safety Critical Systems, Proceedings of the International Conference on

System Safety, (2006), 89-95.

24. C. W. Reynolds, Flocks, herds and schools: A distributed behavioral model,

Proceedings of the ACM Siggraph Conference on Computer Graphics, (1987), 25-34 .

25. C. Bernon, M. P. Gleizes, G. Picard, Enhancing self-organising emergent systems

design with simulation, Engineering Societies in the Agents World VII, Springer Berlin

Heidelberg, (2007), 284-299.

26. T. Kuyucu, I. Tanev, K. Shimohara, Superadditive effect of multi-robot coordination in

the exploration of unknown environments via stigmergy, Neurocomputing, 148 (2015),

83-90.

27. P. Dasgupta, A multiagent swarming system for distributed automatic target

recognition using unmanned aerial vehicles, IEEE Transactions on Systems, Man, and

Cybernetics Part A: Systems and Humans, 38 (2008) 549-563.

28. Y. Tan, Z. Zhong-Yang, Research advance in swarm robotics, Defence Technology 9

(2013), 18-39.

29. N. Bouraqadi, A. Doniec, E. M. de Douai. Flocking-based multi-robot exploration.

Proceedings of National conference on control architectures of robots, (2009).

30. S. Hauert, S. Leven, M. Varga, F. Ruini, A. Cangelosi, J.-C. Zufferey, D. Floreano,

Reynolds flocking in reality with fixed-wing robots: communication range vs.

 28
maximum turning rate, Proceedings of the International Conference on Intelligent

Robots and Systems, (2011).

31. P. Bresciani, A. Perini, P. Giorgini, F. Giunchiglia, J. Mylopoulos, Tropos: An agent-

oriented software development methodology, Autonomous Agents and Multi-Agent

Systems, 8 (2004) , 203 236.

32. F. Pagliarecci, L. Penserini, L. Spalazzi, From a Goal-Oriented methodology to a BDI

agent language: the case of Tropos and Alan, Proceedings of Workshop on Agents,

Web Services and Ontologies Merging, (2007).

33. P. Giorgini, J. Mylopoulos, A. Perini, A. Susi, The Tropos Metamodel and its Use,

Informatica, 29 (2005), 401.408.

34. J. Rodriguez, C. Castiblanco, I. Mondragon, J. Colorado, Low-cost quadrotor applied

for visual detection of landmine-like objects, Proceedings of the International

Conference on Unmanned Aircraft Systems, (2014).