L. Garattoni and M. Birattari (2016). Swarm robotics. In J.G. Webster (Eds.

) Wiley
Encyclopedia of Electrical and Electronics Engineering. John Wiley & Sons, Hoboken, NJ.

Swarm Robotics
Lorenzo Garattoni and Mauro Birattari
IRIDIA, Université Libre de Bruxelles, Belgium

Abstract
Swarm robotics is an approach to coordinating a highly redundant group
of robots. A robot swarm is an autonomous entity that acts in a self-
organized way: the complexity of its collective behaviors is the result of
the local interactions between the individual robots. A robot swarm nei-
ther has a leader nor any other centralized entity that is responsible for its
coordination. Self-organization, high redundancy, and the lack of single
points of failure promote fault tolerance, scalability and flexibility. These
are desired properties for systems deemed to successfully function in the
real world. However, these properties also pose a challenging engineering
problem: the behavior of the individual robot cannot be conceived indi-
vidually;it must be conceived by considering the collective behavior that
it produces when executed by a large number of robots. Designing the
robot-robot and the robot-environment interactions that would result in
the desired collective behavior is a difficult endeavor. Research towards
the definition of an engineering methodology for designing, analyzing, and
maintaining robot swarms is currently ongoing. In this article, we present
swarm robotics from an engineering perspective: we describe works that
contribute to the advancement of swarm robotics as an engineering field
and to its forthcoming uptake in real-world applications.

Keywords—swarm robotics, collective robotics, robotics, self-organization

1 Introduction
Swarm robotics is an approach to robotics in which a mission is entrusted to a
large group of robots, a so called robot swarm.
A robot swarm operates in an autonomous and self-organized way. A swarm
does not rely on any centralized entity for making decisions and for coordinating
its activities. In particular, a swarm does not rely on a leader robot or on
external infrastructures: the collective behavior of the swarm is the result of the
interactions between the individual robots and between robots and environment.
A robot swarm might comprise only robots that are endowed with identical
hardware and control software. In this case, the swarm is called a homogeneous
swarm. Alternatively, a swarm might comprise robots endowed with hardware
and/or control software that belong to a number of classes. In this case, the
swarm is called a heterogeneous swarm. What homogeneous and heterogeneous
swarms have in common is that they are highly redundant: there are multiple

1
robots in the swarm that are able to perform each of the individual actions that
are required to accomplish the given mission. In other words, no single robot is
indispensable.
A characteristic of robot swarms is locality of interaction: each robot has
a limited range of communication and perception. As a consequence, at any
moment in time, each robot directly interacts only with the other (relatively
few) robots that happen to be in its neighborhood. The main implication of this
is that each robot is unaware of the overall size of the swarm and is unaffected
by it.
In a typical swarm robotics application, robots operate in parallel on multiple
tasks. They switch from task to task according to contingencies. Coherently
with what has been stated above, task allocation is autonomous, self-organized,
and based only on locally available information.
Autonomy, self-organization, redundancy, locality, and parallel execution
are highly appreciated characteristics of a robot swarm as they are commonly
deemed to promote fault tolerance, scalability, and flexibility.
Fault tolerance: high redundancy and the lack of a single point of failure
(no leader robot, no external infrastructure) promote the realization of a system
that is robust to failures of individual robots.
Scalability: locality of interaction promotes the realization of a system in
which the addition (or removal) of robots does not qualitatively change the
behavior of the system and therefore does not require modifying the behavior of
the individual robots—provided that robot density is not dramatically altered.
Flexibility: parallel execution and autonomous task allocation promote the
realization of a system that reacts and adapts to contingencies, modifications of
the environment, and variations of the working conditions.
Swarm robotics juxtaposes itself to the single-robot approach [1] and to
classical multi-robot approaches [2, 3, 4].
In the single-robot approach, a mission to be performed is entrusted to a
single, monolithic robot, for example, a humanoid robot. With respect to the
classical single robot approach, swarm robotics appears to be more promising
in applications in which fault tolerance, scalability, and flexibility are partic-
ularly desirable. Moreover, each of the individual robots composing a swarm
is mechanically simpler than a single monolithic robot whose capabilities are
comparable to the one of the swarm. As a consequence, it should be expected
that the cost of hardware design is reduced in the case of swarm robotics. On
the other hand, designing the individual robot behavior that, through robot-
robot and robot-environment interactions, would produce the desired collective
behavior is more complex than designing the behavior of a single, monolithic
robot.
In the classical multi-robot approaches, the mission of interest is entrusted
to a relatively small team of robots [5, 2], smaller than a typical swarm. Usually,
the team behaviors are tailored to the specific team size and thus need to be
adjusted as the team size varies. In the classical multi-robot approaches, each
team member has a role that is defined at design time. Also the patterns of
interaction are defined at design time and are typically more rigid than those
that characterize a robot swarm. As a consequence, a classical multi-robot
system is not as fault tolerant, scalable, and flexible as a robot swarm. On
the other hand, the interaction protocols of a classical multi-robot system are
typically simpler to define than those of a robot swarm, as interactions are well

2
defined and predictable and all relevant information is available at design time.
Beside being a promising engineering approach to the development of com-
plex robotics systems, swarm robotics can be a powerful tool for studying social
behaviors in biology as it is attested by a significant body of literature [6]. When
swarm robotics is used as a tool to study social behaviors, the robots are pro-
grammed to reproduce as faithfully as possible the behavior observed in the
biological system under analysis.
This article presents swarm robotics from an engineering perspective. The
focus is on methods for designing and analyzing robot swarms with the ultimate
goal of adopting them in real-world applications. When swarm robotics is in-
tended as a field of engineering, social behaviors of insects, birds, and mammals
are often a valuable source of inspiration for the designer. Nonetheless, as the
goal of a designer is the pragmatic one of producing a system that accomplishes
a given mission, the source of inspiration is loose and the designer is ready to de-
part from the biological system should this be needed to meet the requirements.
In an engineering perspective, the biological plausibility of the final result is not
a value in itself.
This article is an introduction to swarm robotics and covers the most impor-
tant contributions to date, although it is not meant to be a comprehensive and
exhaustive account of the swarm robotics literature. Contributions are classified
into design methods, modeling methods, and collective behaviors. This classifi-
cation is inspired by the one proposed by Brambilla et al. [7]. Additionally, the
article describes few particularly notable robot swarms that serve as concrete
examples of the recent achievements in the field of swarm robotics. Finally, we
cover some promising prospective applications of swarm robotics. Other articles
have previously reviewed the swarm robotics literature: The already mentioned
Brambilla et al. [7] recently reviewed the swarm robotics literature from an
engineering perspective. Şahin [8] was the first to formally define the basic
concepts of swarm robotics and to provide a survey of the literature. Bayindir
and Şahin [9] presented the literature via five taxonomies: modeling, behav-
ior design, communication, analytical studies and problems. Iocchi, Nardi, and
Salerno [4] classified multi-robot systems depending on their degree of aware-
ness, coordination, and decentralization and dedicates a section to applications
of multi-robot systems. Finally, Gazi and Fidan [10] surveyed the literature from
a control-theory perspective, focusing on the problems of modeling the dynamics
of a robot swarm, and presenting approaches for its control and coordination.
The rest of the article is organized as follows: Section 2 presents the most
common approaches used to design collective behaviors of robot swarms; Sec-
tion 3 presents methods for modeling and analyzing the behavior of robot
swarms; Section 4 describes a number of collective behaviors that have been
realized and discussed in the literature; Section 5 describes six notable robot
swarms that have been demonstrated; Section 6 discusses some promising ap-
plication areas; and Section 7 concludes the article.

2 Design
Designing a robot swarm that is able to accomplish a given mission is a diffi-
cult endeavor. Usually, requirements on the mission are expressed at the swarm
level, the so called macroscopic level, while the designer works at a lower level

3
by implementing the behavior of the individual robots that compose the swarm,
the so called microscopic level. The interaction of the individuals gives rise to a
collective behavior that should satisfy the swarm-level requirements. In partic-
ular, the resulting collective behavior should allow the swarm to accomplish the
given mission. To date, no general formal method exists to derive the individual
behavior from the swarm-level requirements. The problem of designing robot
swarms is tackled either manually or via automatic methods.

2.1 Manual design

In manual design, the designer of the swarm develops, by hand, the behavior
of the individual robots that yields the desired collective behavior. In swarm
robotics, the behavior of the individual robots is typically reactive—that is,
robots act in response to contingencies (possibly influenced by their memory),
without planning their future actions nor reasoning on their effects. The soft-
ware architecture that is most broadly adopted is a particular class of finite
state machines, the probabilistic finite state machine [11].
Although most robot swarms are still developed through trial and error, in
recent years principled design approaches have been proposed. The following
two sections are devoted to the trial-and-error design approach and to some of
the most promising principled design approaches.

Trial-and-error design
Designing a robot swarm by trial and error is more of an art than a science. The
designer operates in an unstructured way with little scientific basis and technical
tools: the designer searches for an individual-level behavior that, through the
complex interaction of a large number of robots, would result in the desired
collective behavior. The search process is performed via educated guesses that
rely solely on the expertise and the ingenuity of the designer. The designer starts
by defining a first implementation of the individual robot behavior. The designer
then tests the behavior, usually by means of computer-based simulations, and
iteratively adjusts it until the resulting collective behavior meets the swarm-level
requirements. Often, the designer takes inspiration from biological systems:
when the goal is to design a robot swarm whose swarm-level behavior is similar
to the one of a biological system (e.g., a swarm of insects, a flock of birds, or
a herd of mammals), the designer might find convenient to design the behavior
of the individual robot by mimicking the one of the individual member of the
biological system.
The relationship between the microscopic and the macroscopic levels poses
challenging issues to the trial-and-error design approach. In particular, the
behavior of the individual robot cannot be evaluated directly and per se: it
must be evaluated indirectly by observing the collective behavior of a swarm
composed by a large number of individual robots that execute the behavior
under analysis.
Notwithstanding its limitations, the trial-and-error approach has been suc-
cessfully used to develop several collective behaviors, including aggregation [8],
chain formation [12], and task allocation [13]. These behaviors are described in
more detail in Section 4.

4
Principled design
Although a general engineering framework for designing robot swarms is not
available yet, a number of promising principled design methods have been pro-
posed. These methods borrow concepts and tools from different disciplines and
address different issues.
In virtual physics-based design [14] each robot is considered as a virtual parti-
cle that exerts forces on other particles—that is, other robots. Each robot is thus
immersed in a field of forces that depends on the presence and distance
Pk of neigh-
boring robots. The virtual force acting on each robot is f = i=1 fi (di )ejθi ,
where j denotes the imaginary unit, di and θi are the distance and the direc-
tion of the ith neighboring robot, and the function fi (di )is the derivative of
an artificial potential function. The Lennard-Jones potential [15] (Figure 1(a))
is commonly used in this context [e.g., 14, 16, 17]. Figure 1(b), (c), and (d)
show three examples of the virtual force that acts on a robot depending on the
position of its neighboring peers. The designer can associate virtual repulsive
forces to obstacles and other objects in the environment to prevent collisions.
Each robot estimates the virtual forces that act on it and translates them into
motion commands. The main benefit of virtual physics-based design is that it
allows the designer to formally prove properties of swarm-level behaviors includ-
ing stability, convergence, and robustness. An extension of virtual physics-based
design is the Hamiltonian method [18]. Starting from a mathematical descrip-
tion of the swarm at the macroscopic level, the Hamiltonian method derives the
microscopic behavior that minimizes or maximizes the value of a relevant quan-
tity (e.g., the virtual potential energy of the state of the swarm). The major
drawback of virtual physics-based design and of the Hamiltonian method is that
they are suitable only for designing spatially-organizing collective behaviors (see
Section 4.1).
Control theory is the theoretical framework of a few principled design meth-
ods that have been proposed. Some of these methods combine virtual physics
with sliding mode control to design robot swarms that perform aggregation, for-
aging, and pattern formation [19, 20] (see Section 4). Other methods use kine-
matic equations to model the motion of robots and a set of control-Lyapunov
functions to develop an individual behavior for pattern formation [21, 22] (see
Section 4). The main advantage of methods based on control theory is that
some properties of the resulting robot swarms (e.g., stability and robustness)
can be proved using theoretical tools such as Lyapunov stability theory. How-
ever, the application of control theory typically relies on assumptions—such as
deterministic behavior, global communication, and full synchronization—that
are often unrealistic in swarm robotics.
Defining a general link between the desired swarm-level (macroscopic) be-
havior and the individual (microscopic) behavior is the key issue in the prin-
cipled design of robot swarms. Some studies showed that, under a series of
assumptions, a microscopic implementation can be derived from a macroscopic
model [23, 24]. Through analytical means, the parameters of a macroscopic
model described by a set of advection-diffusion-reaction partial differential equa-
tions are mapped onto the individual behavior. This method was successfully
used to design robot swarms that perform task allocation and area coverage.
However, the underlying assumptions—such as infinite number of robots and
global communication—are often violated in swarm robotics. As a result, the

5
 Lennard-Jones potential function
v(d) (potential)

d (distance)

(a) (b)

(c) (d)

Figure 1: Virtual physics-based design. (a) The Lennard-Jones potential func-

tion. The potential v depends on the current distance d between two robots. σ is
the desired distance between the robots.  is a parameter called well depth and
corresponds to the depth of the potential function. In this example, σ = 0.2 and
 = 2.5. (b), (c), and (d) are examples of the virtual force that acts on a robot
(gray-filled circle) depending on the position of two neighboring robots (white-
filled circles). In (b), the two neighboring robots are farther than the desired
distance σ. Therefore, the robot is attracted by the two neighbors with forces
that are determined by the derivative of the Lennard-Jones potential function
at the point given by the distance of the neighbors. The resulting force is the
sum of the individual forces. In (c), the two neighboring robots are closer than
the desired distance σ and hence exert repulsive forces on the robot. In (d), one
neighbor is at a distance d1 < σ and thus exert a repulsive force on the robot,
while the other neighbor is at a distance d2 > σ and thus attracts the robot.

6
macroscopic model often fails in predicting the performance of the final system,
as shown by a comparative experimental analysis [25]. Another study proposed
a method to design the individual switching probabilities in task allocation un-
der soft deadlines [26]. The total amount of work performed by the swarm is
described as a Poisson process. Via formal means, the proposed method derives
off-line the switching probabilities—that is, the parameters of a Markov chain
that describes the individual behavior. The method is based on the assumption
that the size and the deadline of the target tasks are all known at design time.
Recently, a formal method borrowed from supervisory control theory has
been used to design a segregation behavior [27]. The method is intended to
be platform independent: the behavior produced was shown to successfully
accomplish the mission both when executed by a swarm of e-pucks [28] and
when executed by a swarm of kilobots [29]. Supervisory control theory is a
method largely applied in manufacturing for automatically synthesizing control
software that drives the behavior of a plant so that specifications are met. In its
adaptation to swarm robotics, the method uses a description of the swarm-level
requirements and a set of specifications for the behavior of individual robots to
generate the individual control logic. In particular, the method requires that the
designer specifies the set of possible events that may occur and the corresponding
responses of the individual robots necessary to generate the desired collective
behavior. In other terms, the method does not support the designer in the most
crucial step: devising the appropriate individual behavior that generates the
desired swarm-level behavior.
Property-driven design [30] was introduced with the ultimate goal of deriving
an individual behavior from swarm-level requirements. The method is based on
prescriptive modeling and model checking. The design process is composed
of four phases: First, the designer defines a set of desired properties that the
swarm should meet. Second, the designer produces a prescriptive model of
the swarm and uses model checking to verify that the model complies with
the specified properties. Third, the designer implements a simulated version
of the robot swarm using the prescriptive model as a blueprint. Fourth, the
designer implements the final robot swarm and validates the previous steps.
Models are described by means of Markov chains and properties are defined
by statements in probabilistic computation tree logic, a probabilistic temporal
logic that captures well both temporal and stochastic aspects. Property-driven
design is a structured design process supported by formal tools. However, the
step from the prescriptive macroscopic model and the correspondent microscopic
implementation is not yet automatic, and it is still reliant on the intuition of
the designer.
Rather defining a unifying framework to obtain any possible collective be-
havior, a number of works have proposed the idea of defining a catalogue of
design patterns [31, 32, 33]. In the context of swarm robotics, a design pat-
tern is a collection of guidelines to obtaining a specific collective behavior. It
must provide: i) a macroscopic model that describes the swarm-level require-
ments,ii) a description of the microscopic behavior, and iii) a mapping from the
parameters of macroscopic model to those of the microscopic behavior. A first
example of a design pattern for collective decision-making has been proposed
and successfully used to design a collective foraging behavior [34].
An interesting design method has been proposed to obtain self-assembly [35]
(see also Section 4) and construction [36] (see also Section 4 and Section 5). The

7
method promotes the decomposition of the design problem [37]: First, the user
provides a global description of the desired aggregate/structure (e.g., the shape
of the aggregate, the height of the structure). Second, the method defines a set
of steps necessary to build the desired aggregate/structure. Third, the method
maps these steps onto individual rules (e.g., rules of motion along the border
of the aggregate or over the structure). The method enables the validation of
some properties of the final system, such as correctness and convergence. It has
been applied successfully to self-assembly and construction. Unfortunately, its
applicability is limited to a restricted set of missions.
Finally, although programming and scripting languages cannot be consid-
ered as design methods on their own, they can significantly ease the principled
design of robot swarms. Two prominent examples are Protoswarm [38] and
Buzz [39]. Protoswarm [38] is a scripting language based on the abstraction
of an amorphous computational medium [40].The amorphous computational
medium assumes that the environment is filled with entities that can compute
and communicate locally with each other. Protoswarm enables the definition of
behaviors for the individual robots by writing scripts at the level of the swarm.
The scripting language features swarm-level primitives that deal both with space
and time. These primitives are translated approximately into individual robot
behaviors by a runtime library. Recently, the idea of languages based on ma-
nipulations of computational mediums has received increasing attentions. New
languages have been proposed for the creation of swarm-level programs with a
sound mapping between the swarm-level and the individual-level primitives [41,
42]. On the other hand, Buzz [39] is a programming language for heterogeneous
robot swarms. Buzz offers primitives to work either at the individual level or
at the swarm level. For the time being, the swarm-level primitives mostly serve
to create different teams and assign robots to each team. Moreover, Buzz pro-
vides mechanisms to share information locally and globally, thanks to a virtual
stigmergy mechanism based on a distributed tuple space. One of the prominent
features of Buzz is its modularity: primitives can be combined or defined anew
to create modules that can be tested, compared, and reused.

2.2 Automatic design

In automatic design methods, the individual behavior of the robots that com-
pose the swarm is generated automatically through an optimization process.
The burden of searching for the individual behavior that results in the desired
collective behavior moves from the designer to a computer program. The major-
ity of work on automatic design of control software for robot swarms has been
produced within the evolutionary robotics domain. A few alternative methods
have been recently proposed.

Evolutionary robotics
Evolutionary robotics [43] takes inspiration from the Darwinian principles of
natural selection and evolution to automatically design control software for sin-
gle and multi-robot systems. In evolutionary robotics, the design process starts
typically from a population of behaviors generated at random. At each itera-
tion, the behaviors are evaluated over a set of experiments via computer-based
simulations. The same behavior is used as control software for all the robots of

8
the swarm. The evaluation is performed by a fitness function that measures the
performance of the swarm. The best scoring behaviors are used to produce the
next generation by means of genetic operators:cross-over, which mixes treats
from parent solutions to produce a child behavior from them, and/or mutation,
which alters small treats of single behaviors. The process terminates when a
time limit or a certain performance threshold are attained or when the fitness
function stops improving. The individual behavior can be represented in several
ways, but the most common one is via an artificial neural network. The evolu-
tionary process searches the parameters of a neural network that, when used as
control software on all the robots, maximizes the performance of the swarm.
Despite the large number of works that showed the effectiveness of evolution-
ary techniques, an engineering methodology for the application of evolutionary
robotics is still unavailable [44]. The main issues are that the evolutionary
process does not give any guarantee of convergence and the neural networks
that result from the design process are black-boxes: they are difficult to an-
alyze and understand. Moreover, most of the behaviors produced so far via
evolutionary robotics are relatively simple and thus easily obtainable via man-
ual design. Some promising ideas have been proposed that could contribute
to the development of an engineering methodology for evolutionary robotics.
Multi-objectivization is deemed to improve the effectiveness of the design pro-
cess by guiding the evolutionary search in rugged fitness landscapes [45]. Nov-
elty search [46] is deemed to promote diversity among candidate behaviors and
improve the exploration of the search space [47]. Finally, the hierarchical de-
composition of the control software into modules is deemed to ease the design
process [48, 49]. For a recent review and critical discussion of the evolutionary
robotics literature, see [50].

Other methods
Because of the limitations of evolutionary robotics [44, 50], other automatic
design methods for robot swarms have been proposed in the recent years.
Reinforcement learning is widely adopted in robotics. It has been elegantly
defined and successfully used in single-robot scenarios [51, 52]. The multi-robot
case has been considered only by few works with limited scope [53, 54]. Swarm
robotics appears to pose major problems to reinforcement learning and only a
very limited number of studies have been proposed [55, 56]. The results are
limited to specific tasks and have been demonstrated in experiments with only
few robots.
A number of studies focused on on-line adaptation in multi-robot systems.
In these studies, the execution of population-based algorithms is distributed
over a group of robots [57]. In this form of embodied evolution the robots are
used as computation nodes. Several works have tested the feasibility of this ap-
proach, proposing different solutions, including open-ended and task-dependent
evolution and the use of finite state machines [58, 59, 60]. The implementation
of distributed evolutionary algorithms in robot swarms has been tested in other
variants: some study explored the idea of cultural evolution in robot swarms
using an imitation-based algorithm [61]. The particle swarm optimization algo-
rithm was compared to genetic algorithms for on-line adaptation and proven to
provide a higher degree of diversity in the robot swarm [62, 63].
Another promising and effective approach that has been proposed adopts

9
a fixed control architecture and focuses on tuning only a small set of parame-
ters. Genetic algorithms and evolutionary strategies were used to optimize the
parameters of finite state machines for a cooperative foraging and object clus-
tering [64, 65]. Exhaustive search was used to determine the optimal parameters
for self-organized aggregation [66]. A similar approach combines evolutionary
computation with virtual-physics based design to learn off-line the parameters
for the Lennard-Jones potential function in a navigation task [67].
A recently proposed novel approach to the automatic design of control soft-
ware for robot swarms is AutoMoDe [68]. In AutoMoDe, the control software is
automatically designed in the form of a probabilistic finite state machine. The
design process works by combining and fine-tuning preexisting modules, which
have parameters that regulate their functioning. A search algorithm optimizes
these parameters along with the topology of the probabilistic finite state ma-
chines to maximize a task-dependent performance measure. This design method
was proven effective in overcoming the reality-gap and in subsequent studies was
shown to outperform human designers in designing control software for five dif-
ferent missions [69].

3 Modeling
A robot swarm can be modeled at two levels: the microscopic level, which
is the level of the single individuals and the interactions among them, or the
macroscopic level, which is the level of the swarm and its collective dynamics.

3.1 Microscopic models

Modeling a robot swarm at the microscopic level involves creating a detailed
representation of each individual. A microscopic model can be defined with
different levels of abstraction: simple models consider robots as point-masses,
while the most complex ones include detailed representations of each sensor and
actuator of the robots. Often, microscopic models are inspired by biochemical
systems and chemical reactions [70, 71, 24]. The behavior of each individual
robot is also an element of microscopic models and, because of its stochastic
nature, it is usually represented by probabilistic finite state machines or Markov
chains [72].
The large number of robots and interactions involved make microscopic
modeling problematic. For this reason, microscopic modeling often relies on
computer-based simulations [73, 74]. Simulations are among the most used tools
for validation and analysis of robot swarms. The vast majority of the collective
behaviors presented in Section 4 have been studied by means of computer-based
simulations.

3.2 Macroscopic models

Macroscopic models consider a robot swarm as a whole. The details of the
individuals and their behavior are neglected in favor of a model of the system
at a higher level. The remainder of this section is devoted to the description of
the most important techniques for modeling robot swarms at the macroscopic
level.

10
Rate and differential equations
Rate equations describe the evolution in time of the portion of robots that are
in a certain state over the total number of robots. Deriving the rate equations
of the swarm from the probabilistic finite state machine that describes the indi-
vidual behavior is straightforward [75]. First, a variable is defined for each state
of the state machine. These variables count the fraction of robots that are in the
corresponding states. Second, for each variable a rate equation is defined that
describe the time evolution of the variable—that is, the time evolution of the
portion of robots in the corresponding state. A rate equation contains a set of
parameters, one for each input and output transition of the corresponding state.
In a seminal work [75], rate equations were used to model an object clustering
behavior. In subsequent works, rate equations have been used to model several
other tasks, among which stick pulling [76, 72], wireless network formation [77],
aggregation [78], and foraging [13]. The advantage of rate equations is that
they allow the derivation of macroscopic models directly from microscopic ones.
The main limitations are that the time is assumed to be discrete and motion
patterns or complex spatial aspects are difficult to model.
A modeling technique for robot swarms that considers spatiality, stochas-
ticity and noise is based on differential equations [79]. This technique uses
Langevin and Fokker-Planck equations. The Langevin equation describes the
motion of a robot in an environment populated by other peers. It consists
of two components: a deterministic one, which expresses the laws of motion
of the robot (microscopic component); and a stochastic one, which describes
the effects of the interactions with its peers (macroscopic component). Be-
cause of this double nature, the Langevin equation is a mesoscopic model—an
intermediate level between microscopic and macroscopic. From the Langevin
equation, it is possible to derive a Fokker-Planck equation that describe the
dynamics of the entire swarm. The derivation is possible by means of tools
of statistical mechanics and problem-dependent intuition. This technique was
applied to analyze coordinated motion, aggregation, and foraging [80]. In fur-
ther studies, the predictions of four models based on Fokker-Planck equations
were compared with the results of computer-based simulations and real-robot
experiments [81]. The comparison revealed that spatial models are more accu-
rate than non-spatial ones for short time spans, but the difference is very small
for long time spans. In principle, the modeling technique based on Langevin
and Fokker-Planck equations can be used to model any robot swarm and any
collective behavior. However, this technique still depends on the intuition of
the designer, who must properly model communication aspects between robots.
Moreover, the Fokker-Planck equation is difficult to solve analytically and often
requires demanding numerical algorithms. A similar modeling technique uses
a set of advection-diffusion-reaction partial differential equations [23]. These
equations were used to model the behavior of robot swarms performing task
allocation and area coverage [24]. The Gillespie algorithm [82] has been used
to numerically approximate stochastic differential equations that model robot
swarms of finite size [83, 33, 84]. These equations present non-linearities that
prevent the use of analytical methods for their resolution.

11
Control and stability theory
Classical control and stability theory can be used to verify whether a collective
behavior will eventually drive the swarm to a desired state. The first works
that adopted control and stability theory in swarm robotics modeled swarms
in a one-dimension space, using discrete time and discrete event dynamical
systems [85]. By using Lyapunov stability theory, a collective behavior was
shown able of social foraging in presence of noise [86, 20]. Other works used
delay differential equations to model task allocation and to prove the stability
of the configuration obtained [87]. The modeling techniques based on control
and stability theory have the advantage of a strong mathematical formulation.
However, these techniques rely on several assumptions that are often violated
in swarm robotics because of noise, stochasticity, and asynchronism.

Model checking
Model checking is a method for formally and automatically verifying whether
a system meets its specifications. In swarm robotics, model checking involves
encoding the collective behavior by means of a mathematical model and checking
whether the model possesses the desired properties. Properties of a system are
expressed as temporal logic formulas.
In one of the first applications of model checking to swarm robotics, re-
searchers used linear temporal logic to define and prove two properties of a
robot swarm: safety, which is a property verified if the swarm does not display
undesirable behaviors, and liveness, which is a property verified if the dynamics
of the swarm evolves over time [88]. A recent work proposes the use of proba-
bilistic computation tree logic to describe desired properties of models expressed
by Markov chains [30]. Probabilistic computation tree logic is well suited for
swarm robotics because it captures well both temporal and stochastic aspects.
In other studies, a high level modeling language called Bio-PEPA was used to
analyze the properties of robot swarms [70]. From a description in Bio-PEPA,
it is possible to derive different models and analyze them by means of stochastic
simulations and model checking.
Recent studies have proposed novel solutions to the main problem of model
checking [89, 90]—that is, the state-space explosion problem when the number
of robots is higher than few units. These studies present a robust methodology
for identifying cutoffs with respect to expressive temporal-epistemic specifica-
tions. This methodology enables the verification of properties of robot swarms
independently of the number of robots in the swarm.

Markov chains
Markov chains are stochastic processes that undergo transitions between states
in a given state space. Markov chains are memoryless: the following state of the
process depends only on the current state, and not on the past history. Markov
chains are applied as statistical models of many real-world processes [91].
Because of the ability to model stochastic processes, Markov chains are well
suited for swarm robotics. Several collective behaviors have been modeled using
Markov chains. A first example is aggregation [92, 93]; the predictions of the
models of aggregation were then validated using computer-based simulations.
More recently, Markov chains have been used to model a protocol of spatially

12
targeted communication between aerial and ground robots of a swarm [94].
The protocol allows robots to open communication links with target robots
depending on their location in space. This enables spatial coordination (i.e.,
pattern formation, morphogenesis) in the swarm. The model was validated
using computer-based simulations and real-robot experiments.
Markov chains enabled the analysis of several collective decision-making in-
stances. For example, the designers could gain insights into the distribution
of the number of individual decisions (by the majority rule) necessary to reach
consensus [95], and into the effects of a dynamic neighborhood size on the deci-
sion dynamics [96]. In other studies, urn models and Markov chains have been
used to study collective decision making [97, 96]. The advantage of urn models
is their ability to capture qualitatively key features of a system notwithstanding
their simplicity. Finally, Markov chains and death-birth processes have been
used to study a scenario in which, at any moment in time, each robot can be
either inactive or engaged in a task [98]. Death-birth processes enable the esti-
mation of several properties of the swarm, such as the energy consumption, the
amount of work accomplished, the time required to complete the task, and the
expected cost-reward.

4 Collective behaviors
Collective behaviors are basic behavioral units of robot swarms that can be com-
bined to create complex collective behaviors. Here we describe the main collec-
tive behaviors studied in the literature and we divide them into five categories:
spatially-organizing behaviors, navigation behaviors, collective decision-making,
interaction with humans, and other behaviors.

4.1 Spatially-organizing behaviors

Spatially-organizing behaviors are collective behaviors that focus on how the
robots distribute and organize in space.

Aggregation
The goal of aggregation is to group the robots in a region of the environment.
Aggregation is a useful building block for many complex behaviors as it allows
robots to gather and thus to interact with each other. The implementation of
aggregation in robot swarms is often inspired by similar behaviors observed in
natural systems such as bacteria, bees, and cockroaches. Aggregation has been
obtained with either manual or automatic design methods.
Manual design methods typically adopt a simple probabilistic finite state ma-
chine: the robots wander in the environment and, when they find other robots,
they decide stochastically whether to stay in their proximity or depart from
them. Typically, robots join an aggregate (or leave it) with a probability that is
a function of the size of the aggregate itself: the larger the aggregate, the higher
the probability of staying. This favors the formation of a single, large aggregate,
as small aggregates tend to disband. This basic behavior can be adapted and
tuned to obtain either static or moving aggregates [99, 100]. Aggregation has
been obtained also via a principled design method based on control theory [19]

13
(see Section 2). Automatic design methods mostly use artificial evolution to
find the parameters of a neural network that produces the desired aggregation
behavior. Either static or moving aggregates can be obtained with this ap-
proach [101, 92]. Other automatic design approaches work on a fixed control
architecture and tune a small set of parameters. This approach successfully
produced an aggregation behavior with memoryless robots that are equipped
only with a single binary sensor [66].
Aggregation can be modeled using different modeling techniques. Rate equa-
tions are particularly suited because of their ability to describe the evolution in
time of the portion of robots in a particular state (the aggregate) [78]. Other
modeling methods used in the literature are based on Langevin and Fokker-
Planck equations [79, 80], on Markov chains [92, 93], and on control and stability
theory [102, 103].

Pattern formation
Pattern formation is a behavior that aims at positioning robots in space ac-
cording to a certain, well defined, pattern. Pattern formation can be useful for
a number of purposes such as covering an area with a fixed number of robots,
achieving a certain network topology and forming the initial configuration for
coordinated motion (see Section 4.2). Examples of pattern formation that of-
ten inspire research in swarm robotics can be found both in biology (e.g., the
chromatic patterns on some animal’s coat) and physics (e.g., crystal formation
and Bénard cells).
Pattern formation in robot swarms is typically obtained using virtual
physics-based design. As already mentioned in Section 2, in virtual physics-
based design robots are considered immersed in the virtual potential field
generated by the neighboring robots. Motion commands are computed by each
robot based on the sum of the virtual forces exerted by its neighbors. If all
the robots exert the same force, this simple mechanism yields an hexagonal
lattice [14]. By dividing the swarm in two groups with different attraction/re-
pulsion thresholds, it is possible to obtain a square lattice [16, 17]. Virtual
physics can be combined with tools borrowed from control theory. In this case,
the stability of the resulting formation can be proved analytically [19, 22] (see
Section 2). Virtual springs can be used alternatively to compute the forces of
attraction and repulsion. Combined with different interaction rules (e.g., full
connectivity, nearest neighbor, K-nearest neighbors), they can produce different
patterns [104, 105].
Recently, a pattern formation behavior with a thousand robots has been
demonstrated [35]. Few robots act as the seed of the pattern and define the
origin and orientation of the coordinate system that is used to build the desired
shape. Starting from the seed robots and using also an internal representation
of the target pattern, other robots of the swarm gradually join the pattern.
Robots localize themselves with respect to the initial seed using an information
gradient. The thousand robots have been shown to successfully form different
shapes.
An important application of pattern formation is area coverage: when the
number of robots is limited, a lattice formation of equally-spaced robots op-
timizes the coverage of the space [106]. Area coverage is often modeled us-
ing differential equations: two examples of differential equations used to model

14
area coverage are a set of advection–diffusion–reaction partial differential equa-
tions [24] and the Fokker–Planck equations [81]. In the latter example, the accu-
racy of four models based on Fokker-Planck equations was tested by comparing
their predictions with the results of computer-based simulations and real-robot
experiments.

Chain formation
In chain formation, robots arrange themselves in the environment to create a
chain that connects two locations. The chain is then used by other robots as a
navigation aid (see Section 4.2). This behavior is inspired by Argentine ants,
which form chains of individuals that connect their nest to foraging sites [107].
Chain formation can be developed using different design methods. Typically
it is obtained by manually designing control software in the form of a proba-
bilistic finite state machine. The chain is built incrementally from the starting
location. The robots that find a growing chain follow it until the end and join
it in the last position with a certain probability. The last robot in the chain can
always leave the chain with a certain probability. This prevents the chain from
becoming entrapped in dead ends and allows an effective exploration of the en-
vironment. When the chain reaches the target location, it becomes stable. The
robots in the chain might use a tricolor-pattern to indicate the direction of the
chain [108, 12]. A variant of this solution is based on a probabilistic finite state
machine and network routing. The result is a chain of moving robots [109].
Virtual physics-based design and automatic design methods can also be used
to design chain formation. In virtual physics, virtual forces are used to maintain
a desired distance between robots in the chain and between robots and walls
in order to create chains that strongly depends on the shape of the environ-
ment [110]. Concerning automatic design, artificial evolution has been shown
able to produce chains of moving robots [111].

Self-assembly and morphogenesis

Self-assembly is the process in which robots physically connect to each other.
Self-assembly can be useful, for example, to increase mechanical stability and
ease navigation on rough terrains. When the connected robots form a particular
pattern or shape, the process is called morphogenesis. Morphogenesis is used
when a particular structure allows the swarm to perform a specific task. For
instance, a line of connected robots can navigate over a hole whereas a single
robot would fall into it. Several natural systems show self-assembly and mor-
phogenesis behaviors: ants are able to create bridges, rafts and walls to perform
specific tasks; cells self-organize structures to form tissues and organs.
Self-assembly and morphogenesis can be designed in several ways. These be-
haviors pose many challenges to the design process: when and how the assembly
should start, which robots should connect to each other, which shape should be
formed. Each of this problems can be addressed in different ways.
Robots can trigger the self-assembly process when they encounter obstacles
or adversities that they are not able to overcome on their own. Empirical studies
showed that connected robots are able to navigate in hazardous terrains better
than individual robots [112], they can overcome obstacles that a single robot
cannot overcome [113], and they can transport heavy objects faster and for

15
longer distances [114]. Robot swarms have also been demonstrate capable of
creating 3D structures through self-assembly [115].
Homogeneous robots can self-organize the assembly process by signaling the
docking points in different locations of their bodies. Other robots can then
connect stochastically to those docking points. In this way the robots can form
different structures, such as lines, stars and circles [116]. Alternatively, the
capabilities of heterogeneous robots can ease the process of self-assembly. For
example, an aerial robot can recognize the task to perform and indicate to the
ground robots which robots should self-assemble and what structure they should
create to perform the task [117].
Self-assembly and morphogenesis have not been modeled often in the lit-
erature. A study showed that a self-assembly behavior that allows robots to
form lines can be modeled using a set of chemical reactions [118]. This set of
chemical reactions was then abstracted by a set of differential equations, solved
approximately by means of stochastic simulations (e.g., Gillespie algorithm),
and compared to computer-based simulations.

Object clustering and assembling

Object clustering and assembling refer to behaviors in which the robots create
aggregates of objects. The difference between object clustering and assembling
is that in the former the aggregates are clusters of unconnected objects, whereas
in the latter the objects must be connected by some kind of physical link. These
two behaviors are at the basis of any swarm construction system. For the design
of object clustering and assembling, researchers often take inspiration from social
insects: brood clustering has been observed in ants, termites can build mounds
that are orders of magnitude larger than the single individuals.
Object clustering is usually obtained using a probabilistic finite state ma-
chine. The robots explore randomly the environment and react with appropriate
responses when they find an object or partially formed clusters. In the simplest
form of object clustering, a robot picks up an object and deposit it with a
probability that is proportional to the number of other objects perceived [119].
The final position of the clusters can be controlled by marking the ground with
colors or using other signals recognizable by the robots [120, 121]. Object clus-
tering can also be obtained via automatic design methods. Recently, a clustering
behavior for extremely simple robots was successfully developed through evo-
lutionary robotics [122]: the robots are not capable of arithmetic computation
and are only able to detect the presence of an object or another robot in their
direct line of sight. Despite these limitations, the swarm is able to successfully
create clusters of objects within a limited amount of time. Object clustering was
modeled in a seminal work on the use of rate equations in swarm robotics [75].
Concerning assembling, a recent work demonstrated a behavior that enables
the creation of arbitrary 3D structures [36]. This solution generates off-line
a set of traffic rules and assigns them to the robots, together with a static
representation of the target structure. Respecting the traffic rules, a group of
climbing robots builds the structure by placing a building block at a time. More
details on this work can be found in Section 5.

16
4.2 Navigation behaviors
Navigation behaviors are collective behaviors that aim at coordinating the move-
ments of a robot swarm.

Collective exploration
Collective exploration includes behaviors whose goal is to explore an environ-
ment, or interesting portions of it. Work on collective exploration takes fre-
quently inspiration from behaviors observed in natural systems. Control soft-
ware for collective exploration is typically implemented in the form of proba-
bilistic finite state machines. Often the swarm relies on static robots that act
as way points to guide the navigation of moving robots. To do that, the static
robots can form either physical or virtual structures.
Physical structures are usually the result of pattern formation and chain
formation (see Section 4.1). Once the physical structure is formed, the moving
robots can follow it, way point after way point, to navigate in the environment.
In virtual structures, the static robots are not necessarily close to each other,
but they are connected by a virtual medium. For example, pre-deployed robots
can create a virtual structure between two locations by exchanging messages.
Moving robots can exploit these messages for navigation [123, 124]. Similarly, a
network of pre-deployed sensors can be used by the robots to navigate towards
their goal location [125]. The navigation route is calculated by the robots via
a distributed variant of the Bellman-Ford algorithm. An hybrid solution was
developed in the Swarmanoid project [126] (see also Section 5). In this solution,
a set of aerial robots deploy sequentially to form a chain, using the position of the
previously deployed robots to determine their target position. Once deployed,
the robots establish also a virtual structure by acting as communication relays.
Lastly, a solution has been proposed in which the robots of a swarm both nav-
igate and guide the navigation of others, simultaneously [127]. While moving,
the robots share navigation information between them and hence cooperatively
guide each other towards a target location. The advantage of this solution is
that it does not bind any robots to a specific location. All the robots can thus
move and be involved in other tasks, possibly unrelated to navigation.

Coordinated motion
In coordinated motion, also known as flocking, the robots move in formation
through the environment, similarly to flocks of birds or schools of fish. In nature,
coordinated motion is used by many animals to reduce energy consumption and
increase the chance they survive attacks of predators. Flocking can be obtained
with either manual or automatic design methods. The most common design
method uses virtual physics. Virtual forces of attraction and repulsion maintain
a desired constant distance between the robots and a uniform alignment during
the motion [128]. The robots are capable of coordinated motion even in absence
of a common goal, thanks to the sole knowledge of heading and distance of their
neighbors [129]. Under this configuration, it is sufficient to insert few “informed”
robots to direct the movement of the other “uninformed” robots, and hence of the
whole swarm, toward a goal [130]. Further works showed that this behavior does
not require an explicit alignment rule, and thus robots do not need to perceive
the orientation of their neighbors. The swarm is still able to navigate with and

17
without the presence of informed robots [131]. Flocking of a swarm of aerial
robots was obtained through evolutionary robotics [132]. Without relying on
any external infrastructure, the aerial robots establish and maintain a wireless
communication network to connect a base station and a user station that are
located on the ground.
In the literature, flocking is typically modeled using differential equations.
An example is the application of a method based on a Fokker-Planck equa-
tion [79]. In another study, researchers performed preliminary steps towards
linking the models of flocking produced in statistical physics with the studies
produced in swarm robotics [133]. The authors focused on the alignment of
robots and verified the existence of a phase transition between order and disor-
der that depends on the level of noise and on the neighborhood size. The results
were validated using computer-based simulations.

Collective transport
Collective transport refers to a set of behaviors in which the goal of the swarm is
to cooperatively move objects from one location to another. The objects are too
heavy for a single robot, thus cooperation is necessary. Collective transport can
be observed in ant colonies. To achieve collective transport, ants use a trial-and-
error process in order to determine the right pulling/pushing direction [134].
Collective transport is usually designed via manual methods or artificial evo-
lution. Different strategies can be employed for transporting the object: robots
can connect directly to the object and move it, they can connect to each other
and then to the object, or they can surround the object and push it with their
movement [114, 135]. Consensus on the direction of movement and cooperation
are achieved either through direct or indirect communication. For example,
when direct communication is used, robots can agree on a common direction of
movement by averaging their individual desired directions [136]. When commu-
nication is indirect, robots can position themselves around the object depending
on the position already taken by other robots [135], or depending on an esti-
mation of the forces applied by other robots on the object or on their own
chassis [137].
As an alternative to reaching consensus on the direction on movement, some
robots can form a chain to connect the source and the destination of the objects
(see Section 4.1). The chain is then used as navigation aid by other robots that
transport the objects [138].

4.3 Collective decision-making

Collective decision making focuses on how a robot swarm can make decisions.
Two categories of situations can require a swarm to make a choice: the first
category comprises situations in which the robots have to reach a consensus on
a single choice among a set of possible alternatives. The behaviors that aim at
solving these problems are called consensus achievement. The second category is
composed of problems in which the robots have to distribute themselves among
a set of possible tasks and operate in parallel on those tasks in order to maximize
the performance of the system. This process is called task allocation.

18
Consensus achievement
Consensus achievement behaviors allow a robot swarm to converge on a single
choice among a set of alternatives. The choice is usually the one that maximizes
the performance of the system. Consensus achievement can be observed in many
insect species; for example, ants can determine the shortest of different paths
using pheromone [139] and bees collectively choose the best nest location among
several alternatives [140].
The robots of a swarm can reach consensus using either direct or indirect
communication. Several strategies have been proposed in the literature. By
mimicking a form of consensus achievement used by cockroaches, it is possible
to develop robot swarms that choose a single aggregation zone using indirect
communication [99]—that is, decisions are taken using indirect clues, such as
the density of neighbors. Consensus achievement can be achieved via quorum
sensing, an algorithm inspired by the choice of the best over N alternatives in
ants and bees [141]. The algorithm is based on direct communication: robots
evaluate alternatives and advertise them through recruiting messages that are
broadcast with a frequency proportional to their perceived quality. This al-
lows the swarm to eventually converge on the best alternative. The nest-site
selection in honeybee colonies inspired other work: a thorough study of its ana-
lytical model and the identification of the parameters that determine its working
regime enabled the definition of guidelines for the implementation of the indi-
vidual robot behavior. This approach was applied to the shortest path selection
problem [33]. Consensus achievement has been used also to let the robots agree
on a common reference orientation [142]. The algorithm, which uses only rela-
tive positioning and local communication, can be used as a preliminary step for
collective motion. Finally, two novel strategies were recently proposed: the first
is based on a weighted voter model and ensures a high decision accuracy and
robustness to noisy assessments of alternatives [83]; the second couples a mecha-
nism of time-modulation with individual robots’ decisions based on the majority
rule and has been shown to speed up the decision process considerably [143, 144,
145].
Consensus achievement is typically modeled using Markov chains [95, 96,
97, 96]. Another modeling technique used to model consensus achievement is
based on Bio-PEPA [70], a process algebra originally introduced for modeling
biochemical systems. From a formal specification of the system in Bio-PEPA,
the authors were able to analyze the behavior of a swarm that had to identify
the shortest path between two possible choices.

Task allocation
Task allocation behaviors deal with the distribution of robots over a set of
different tasks. The allocation is usually dynamic and aims at maximizing the
overall performance of the swarm. Ants and bees use task allocation. For
example, while some individuals are foraging, a number of other individuals are
in charge of looking after the larvae. Task allocation is usually obtained through
manual design methods. In the first works on task allocation the choice of the
robots was limited to whether to engage in a foraging task or remain in the nest,
resting. The robots would choose on the basis of the level of energy in the nest
(level raised by the preys collected and consumed by robots resting) [146], or on

19
the basis of individual observation of the environment and of other robots [147].
A similar approach is used in recent works to design task allocation for two
sequentially interdependent tasks [148, 149]. The process is based on individual
observations of the environment and of the current performance of the swarm,
and hence it does not require direct communication between robots. Other
works focus on whether, when, and how the robots should perform the overall
task or partition it and allocate themselves to one of the subtasks [150]. For
example, a robot could choose whether to carry a prey from the source to the
nest or store it in a cache [151]. In the latter case, the prey would then be
collected by robots waiting on the other side of the cache and carried to the
nest. The choice is made on the basis of the estimation of the costs involved.
Some attempts have been made to design task allocation via automatic meth-
ods. A recent study demonstrated how self-organized task allocation can be
obtained through evolutionary robotics [152]. Differently from the previous at-
tempts that used artificial neural networks [153, 154, 155], the authors used a
method of grammatical evolution to automatically design task allocation strate-
gies. This method was proven more effective, and was able to successfully design
task allocation both when provided few behavioral building blocks and when the
strategy had to be defined anew.
Task allocation is usually tested on foraging. However, there are examples
of application of task allocation to other practical problems. The allocation of
robots to different operations on a construction site [156] and a stick-pulling
problem [157] are two further examples of application. Often, task allocation is
used as a testbed for modeling techniques. A set of advection-diffusion-reaction
partial differential equations was used to model and design a behavior for task
allocation [23]. Delay differential equations were used to model task allocation
and verify the stability of the system obtained [87]. A population dynamics
model was used to determine some parameters of the individual behavior and
the optimal distribution of robots in two task-allocation scenarios [158]. Finally,
a Poisson process was used to describe the performance of a swarm perform-
ing a set of tasks and to design the individual switching probabilities for task
allocation [26].

4.4 Interaction with humans

Robot swarms are designed to work autonomously and to act in a distributed
way. These characteristics limit the degree of control that a human operator
can exercise on the system. However, there are several cases in which forms of
human control over the swarm are necessary. Human-swarm interaction studies
how a human operator can control a robot swarm and receive feedback from it.
Studies in this field can be categorized on the basis of the nature of interactions
that they propose.
The most common approach relies on an intermediate modeling layer be-
tween the operator and the swarm. Usually, the modeling layer produces an ab-
stract representation of the robots and their environment that is then displayed
to the operator through a graphical user interface. By acting on the GUI, the
operator can select robots and send them commands. The selection can contain
single robots [159], or a group of robots, which can be selected, for instance,
by drawing a rectangular zone that contains the robots in the GUI [160]. A
robot controlled by a human operator is perceived by the swarm as just another

20
robot, and thus the influence of the operator is very limited. This problem can
be solved partially by a hierarchical communication architecture in which the
operator sends orders to the selected robot, which is called ”the sergeant" [161].
Other studies focus on the use of augmented reality. Part of the studies that
use augmented reality propose solutions only for the visualization of feedback
from the robots to the operator. For instance, an optical see-through head-worn
device receives robots’ messages, analyses them and augments the environment
with their representation [162]. Similarly, firefighters are helped in their mis-
sion by a robot swarm, which gives them direction information displayed by
augmented helmets [163]. Other works provide bi-directional communication
solutions: through a device that display the augmented environment, the op-
erator can also give commands to the robots by acting on the real-time video
stream [164].
Finally, there are studies that aim at realizing a direct interaction,without
relying on intermediate modeling levels. In fact, creating and maintaining an
updated model of the robots and their environment is a demanding task. It
often requires ad-hoc infrastructures and it becomes intractable in dynamic
(real) environments or when the number of robots is greater than few units.
For these reasons techniques of direct interaction based on gestures recognition,
face engagement and speech recognition have been proposed. Combinations of
such techniques are possible [165, 166, 167]. Performing gesture recognition
directly on the robots might lead to mismatches, and thus require distributed
consensus algorithms in order to reach an agreement of all the robots on the
same gesture [168]. Other works proposed the use of external sensors, such
as the Microsoft Kinetic sensor to give commands through gestures to a robot
swarm [169].

4.5 Other behaviors

Some notable studies in swarm robotics do not belong in any of the previous
categories.
In collective fault detection, the swarm recognizes faulty robots and initiates
appropriate responses. Despite being robust to individual robot failures, there
might be situations in which robot swarms need to be aware of the presence
of faulty robots and react properly. In a pioneering work on collective fault
detection, the robots use an algorithm inspired by firefly synchronization [170].
The robots emit a periodic signal and eventually synchronize with each other
using a model of pulse-coupled oscillators [171]. If a robot does not synchronize
with its neighbors, it is assumed to be faulty and a response can be triggered.
In group size regulation, the robots have the ability to estimate and regulate
their number in a group. This ability is useful, for example, when an excessive
number of robots in a group lowers the performance of the swarm. One of the
first solutions proposed to estimate the number of robots in a group takes in-
spiration from the behavior of fireflies [172]: the robots emit a signal at random
times and count the number of signals perceived over a period. This number
is used to estimate and tune the size of the group. An improvement of this
algorithm based on a more strict signaling order can obtain more reliable group
size estimates [173]. In other studies a set of flying robots aids the aggregation
of ground robots [174]. The aerial robots estimate the size of the aggregate
and communicate to the ground robots the accordingly adjusted probabilities

21
Figure 2: Swarm-bot – Previously Figure 3: Swarmanoid – Still from
unreleased photo. Copyright: Marco the video: Swarmanoid, the movie.
Dorigo. Copyright: Mauro Birattari et al.
Reprinted by permission.

of joining or leaving the aggregate. Through this mechanism the ground robots
are able to form groups of different sizes. Finally, a recent study proposed an
algorithm inspired by cockroaches aggregation under shelters that is able to par-
tition the swarm in groups of different sizes [175]. The aggregation in different
groups develops in parallel, therefore the convergence time of the algorithm is
independent of the number of groups.

5 Notable systems
Among the vast production of fundamental research work in swarm robotics,
a few systems have been shown to be able to perform complex missions. In
the following we describe a selection of six notable systems. For each system,
we report an illustrative picture that includes a QR code in the bottom-right
corner. To watch a demonstrative video, either click on the QR code or scan it
through the camera of a mobile device.
Swarm-bot [12] is a robot swarm composed of relatively simple robots,
the s-bots, that are able to attach to each other—see Figure 2. The ability to
self-assemble, along with control algorithms inspired by self-organized behaviors
of social insects, allow the swarm-bot to effectively adapt to its environment.
For example, by self-assembling in different shapes, the swarm-bot can navigate
through rough terrains and drag objects that are too heavy for single s-bots.
The swarm-bot has been shown to be able to find a target, heavy object and
retrieve it. In the first phase, the s-bots form a chain between the object and
the nest (see chain formation in Section 4). In the second phase, a group of
s-bots surround the object, self-assemble, and drag the object to the nest along
the path described by the chain.
Swarmanoid [176] is a heterogeneous swarm composed of three types of
robots: eye-bots, hand-bots, and foot-bots—see Figure 3. Eye-bots are flying

22
Figure 4: TERMES Still from the Figure 5: Thousand-robot Swarm
video: Designing collective behavior – Still from the video: Programmable
in a termite-inspired robotic construc- self-assembly in a thousand-robot
tion team. Copyright: Justin Werfel swarm. Copyright: Michael Ruben-
et al. Reprinted by permission. stein et al. Reprinted by permission.

robots specialized in sensing the environment and providing an overview to foot-

bots and hand-bots. Hand-bots can climb walls or other vertical surfaces and
grab objects, but they cannot move on the ground without the help of other
robots. Foot-bots are specialized in moving on the ground and transporting
either objects or other robots. The Swarmanoid has been shown to be able to
explore an unknown indoor environment, locate a target object (a book), and
retrieve it. First, the eye-bots explore the environment, find the book on a shelf,
and highlight the path to it. Then, the foot-bots transport a hand-bot to the
shelf following the path indicated by the eye-bots. At this point, the hand-bot
climbs the shelf, grab the object, and returns to the ground where the foot-bots
transport it back to the initial location.
TERMES [36] is a robot swarm inspired by how termites construct
mounds—see Figure 4. TERMES allows a user to specify a high level rep-
resentation of the target 3D structure. From the specified structure, an off-line
software generates a set of traffic rules that direct the flow of robots over the
growing structure and regulate the building activity. Essentially, this set of
rules, namely structpath, is a 2D representation of the structure in which each
stack is annotated with its height and a travel direction between each adjacent
pair of stacks. Thanks to the structpath and to a static internal representation
of the target structure, a group of custom-designed climbing robots proceeds au-
tonomously to the construction process by depositing one brick at a time. The
effectiveness of the system has been shown both in a simulation environment
and with a real-world implementation.
Thousand-robot Swarm [35] is a robot swarm composed of 1 024 Kilobots—
see Figure 5. The Kilobot is a small, low-cost robot equipped only with vibration
motors for movement and an infrared transceiver for communication and dis-
tance sensing. A swarm of 1 024 Kilobots was demonstrated capable of forming
different user-specified patterns. The pattern is built gradually, starting from

23
Figure 6: CoCoRo – Still from the Figure 7: BioMachines Lab’s
video: TYOC#52/52: Final Demon- Aquatic Robot Swarm – Previ-
strator. Copyright: Thomas Schmickl ously unreleased photo. Copyright:
et al. Reprinted by permission. BioMachines Lab. Printed by permis-
sion.

few central robots that act as a seed. The other robots travel along the edge
of the pattern under formation and stop in a proper position, according to an
internal representation of the target pattern. The thousand-robot swarm is the
largest robot swarm demonstrated so far.
CoCoRo [177] is a heterogeneous swarm of underwater robots—see Figure 6.
The swarm is composed of a base station and two types of robots: Jeff robots
and Lily robots. Jeff robots are fast searching robots, Lily robots are slow
information carriers. A CoCoRo can monitor, search, maintain, explore and
harvest resources in underwater habitats. The CoCoRo was shown able to locate
an object and guide the base station to its position. The Jeff robots search the
seabed, until one of them finds the target object and starts recruiting more Jeff
robots. At this point the Lily robots start to build a relay chain between the
base station and the cluster of Jeff robots. The base station can then navigate
to the object location using the information carried by the chain.
BioMachines Lab’s Aquatic Robot Swarm [178, 179, 180] is a swarm
of 10 aquatic surface robots—see Figure 7. The robots are equipped with few
sensors and actuators and thus are relatively inexpensive. The control software
was designed automatically using evolutionary robotics (see Section 2). Four
collective behaviors were developed to perform four different tasks: flocking,
clustering, dispersion, and area coverage. By combining these collective behav-
iors, the swarm was shown able to perform a complex mission of environmental
monitoring: the robot swarm collectively navigates towards an area of interest,
optimizes the coverage of the area, monitors the water temperature in the area,
clusters, and finally heads back to the base.

24
6 Some prospective applications
There are a number of possible applications that appear to be appropriate for
robot swarms: search and rescue of survivors in disaster areas, humanitarian de-
mining, waste and pollutant removal (i.e., urban waste, oil spills), surveillance,
automation and management, health care and medical treatments, exploration
of hazardous environments or extraterrestrial planets. In the following, we high-
light some of these applications, for which initial promising results have been
achieved.
A swarm of aquatic surface robots is currently under development for per-
forming maritime tasks such as patrolling, intruder detection, aquaculture in-
spection, and environmental monitoring [179]. A swarm of 10 robots has been
shown capable of performing a complex environmental monitoring mission (see
Section 5). A fundamental characteristic of this swarm is the decentralized
control based on a combination of artificial evolution, manual design, and hi-
erarchical decomposition of behaviors. The swarm is composed of relatively
simple and cheap robots. Simulated experiments have shown that 1 000 drones
are able to patrol the 20 km long coastal strip of the Lampedusa Island [180].
Warehouse automation is a promising application area for robot swarms.
Researchers are working on the development of robot swarms to improve the
flow of materials in warehouses [181]. The robots will retrieve items, will trans-
port them to human operators, and will keep the warehouse organized. The
robots will use communication and ant-inspired algorithms to choose the best
paths while they perform these operations [182]. Autonomy and decentralization
would allow the addition of more robots at any time, if required by contingencies.
Failures of individual robots would not be problematic, as other robots could
take over the pending tasks. The parallelization introduced by robot swarms
could drastically reduce the average order processing time.
Agriculture appears to be another promising application area. Prototype
robot swarms have been developed to automate agricultural processes [183,
184]. Further, several academic projects investigate multi-robot coordination
for application in agriculture [185, 186]. These projects focus on the collab-
oration between one or more aerial robots and a fleet of ground robots. The
aerial robots scan the environment and provide information to the ground robots
(e.g., locations of weeds to be removed). This information, along with the data
collected on the field, allows the ground robots to make decisions about their
movements and operations.
Lastly, swarm robotics appears to be the appropriate approach to coordinate
large groups of nano-robots for medical applications. Swarms of nano-robots
are meant to be injected in a patient to perform tasks including diagnosis and
targeted drug-delivery. The main challenge to be faced is the design and re-
alization of the nano-robots. Two different approaches are arising. The first
approach focuses on minimalist nano-robots that do not have any sensing/ac-
tuating capabilities. Their behavior is fully determined at production time by
engineering their material, shape, charge, and coating [187, 188]. The second
approach focuses on relatively more complex nano-robots that mimic single-
cell organisms [189, 190]. These robots move autonomously, recognize a target,
chemically process molecules, and release them on demand. In both approaches,
nano-robots rely on swarm behaviors to navigate in blood vessels, protect them-
selves from macrophages, adhere to target tissues, and accomplish the intended

25
mission.

7 Conclusions
Swarm robotics is a promising approach for coordinating large groups of robots
that need to be fault tolerant, scalable, and flexible. So far, swarm robotics
research has focused on developing methods to design collective behaviors (see
Section 2) and to model and analyze them (see Section 3). A number of col-
lective behaviors have been produced (see Section 4) and a few notable robot
swarms have been demonstrated (see Section 5). The adoption of robot swarms
in the real-world is on the horizon, as there exist several prospective fields of
application. Among these, the most promising appear to be surveillance, envi-
ronmental monitoring, agriculture, warehouse automation, and health care (see
Section 6). Notwithstanding the significant results achieved so far, there are
issues that prevent the immediate uptake of swarm robotics.
The first issue concerns the hardware. Robots that are currently available
have limited functionalities and are not sufficiently reliable. Further research is
required to produce autonomous robots that can reliably operate in dynamic and
possibly hazardous environments. Another issue is the lack of effective ways to
interact and command a robot swarms. Although autonomy is a defining prop-
erty of a robot swarm, a human operator should always be able to control the
activity of a swarm. For example, a human operator should be able to stop a
robot swarm that is behaving in an unpredicted or dangerous way. Research in
human-swarm interaction aims at solving this issue by proposing forms of bidi-
rectional interaction between human operators and robot swarms. However,
due mainly to the autonomous and distributed nature of a robot swarm, this
issue has not yet been addressed in a fully satisfactory way. A last fundamental
issue is the lack of a reliable engineering methodology for specifying, design-
ing, analyzing, verifying, validating, and maintaining a robot swarm. Further
research is required to define or adapt formal languages to specify the require-
ments of robot swarms. In the design process, the definition of general methods
to derive the individual behavior from the requirement specification remains an
open challenge. The analysis of a robot swarms can be performed by means of
several methods, but the lack of well-defined common metrics prevents a proper
validation and verification of its properties. In particular, properties such as
reliability and safety are often claimed by definition, without the support of an
empirical or theoretical analysis. Only few preliminary works have performed
steps towards the definition of methods for analyzing, validating, and verifying
reliability and safety in robot swarms [191, 192]. The definition of a reliable en-
gineering methodology with all the described components is required to promote
the real-world uptake of swarm robotics.

References
[1] N. J. Nilsson. Shakey the robot. Tech. rep. 323. Menlo Park, CA, USA:
AI Center, SRI International, 1984.
[2] G. Dudek, M. R. M. Jenkin, E. Milios, and D. Wilkes. In: Auton. Robot.
3(4), 1996, pp. 375–397.

26
 [3] L. E. Parker. In: Distributed Autonomous Robotic Systems 4. Ed. by L.
E. Parker et al. Tokyo, Japan: Springer, 2000, pp. 3–12.
[4] L. Iocchi, D. Nardi, and M. Salerno. In: Balancing Reactivity and Social
Deliberation in Multi-agent Systems. Vol. 2103. LNCS. Berlin, Germany:
Springer, 2001, pp. 9–32.
[5] B. P. Gerkey and M. J. Matarić. In: Int. J. Robot. Res. 23(9), 2004,
pp. 939–954.
[6] S. Mitri, S. Wischmann, D. Floreano, and L. Keller. In: Biol. Rev. 88(1),
2013, pp. 31–39.
[7] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo. In: Swarm Intell.
7(1), 2013, pp. 1–41.
[8] E. Şahin. In: Swarm Robotics. Vol. 3342. LNCS. Berlin, Germany:
Springer, 2005, pp. 10–20.
[9] L. Bayindir and E. Şahin. In: Turk. J. Elec. Eng. & Comp. Sci. 15(2),
2007, pp. 115–147.
[10] V. Gazi and B. Fidan. In: Swarm Robotics. Vol. 4433. LNCS. Berlin,
Germany: Springer, 2007, pp. 71–102.
[11] M. O. Rabin. In: Inform. Control 6(3), 1963, pp. 230–245.
[12] S. Nouyan, R. Groß, M. Bonani, F. Mondada, and M. Dorigo. In: IEEE
T. Evolut. Comput. 13(4), 2009, pp. 695–711.
[13] W. Liu and A. F. Winfield. In: Int. J. Robot. Res. 29(14), 2010, pp. 1743–
1760.
[14] W. M. Spears, D. F. Spears, J. C. Hamann, and R. Heil. In: Auton.
Robot. 17(2–3), 2004, pp. 137–162.
[15] J. E. Jones. In: P. Roy. Soc. Lond. A Mat. 106(738), 1924, pp. 463–477.
[16] W. M. Spears and D. F. Spears. Physicomimetics: Physics-Based Swarm
Intelligence. Springer, 2012.
[17] C. Pinciroli, M. Birattari, E. Tuci, M. Dorigo, M. del Rey, T. Vinko,
and D. Izzo. In: Proceedings of the Sixth International Conference on
Ant Colony Optimization and Swarm Intelligence (ANTS-2008). LNCS
5217. Berlin, Germany: Springer, 2008, pp. 347–354.
[18] S. Kazadi. In: IJICC 2(4), 2009, pp. 672–694.
[19] V. Gazi. In: IEEE T. Robot. 21(6), 2005, pp. 1208–1214.
[20] V. Gazi and K. M. Passino. In: IEEE T. Syst. Man Cy. B 34(1), 2004,
pp. 539–557.
[21] P. Ögren, M. Egerstedt, and X. Hu. In: Proceedings of the 40th IEEE
Conference on Decision and Control 2001. Vol. 2. Piscataway, NJ, USA:
IEEE Press, 2001, pp. 1150–1155.
[22] M. Egerstedt and X. Hu. In: IEEE T. Robotic. Autom. 17(6), 2001,
pp. 947–951.
[23] S. Berman, Á. M. Halász, M. A. Hsieh, and V. Kumar. In: IEEE T.
Robot. 25(4), 2009, pp. 927–937.

27
[24] S. Berman, V. Kumar, and R. Nagpal. In: IEEE International Conference
on Robotics and Automation (ICRA 2011). Piscataway, NJ, USA: IEEE
Press, 2011, pp. 378–385.
[25] G. Mermoud, U. Upadhyay, W. C. Evans, and A. Martinoli. In: Experi-
mental Robotics. Ed. by O. Khatib et al. Vol. 79. STAR. Berlin, Germany:
Springer, 2014, pp. 615–629.
[26] Y. Khaluf, M. Birattari, and H. Hamann. In: From Animals to Animats
13. Ed. by A. P. del Pobil et al. Vol. 8575. LNCS. Berlin, Germany:
Springer, 2014, pp. 270–279.
[27] Y. K. Lopes, A. B. Leal, T. J. Dodd, and R. Groß. In: Swarm Intelligence,
ANTS 2014. Ed. by M. Dorigo et al. Vol. 8667. LNCS. Berlin, Germany:
Springer, 2014, pp. 62–73.
[28] F. Mondada, M. Bonani, X. Raemy, J. Pugh, C. Cianci, A. Klaptocz, S.
Magnenat, J.-C. Zufferey, D. Floreano, and A. Martinoli. In: 9th Confer-
ence on Autonomous Robot Systems and Competitions, Robótica 2009.
Castelo Branco, Portugal: IPCB-Instituto Politécnico de Castelo Branco,
2009, pp. 59–65.
[29] M. Rubenstein, C. Ahler, N. Hoff, A. Cabrera, and R. Nagpal. In: Robot.
Auton. Syst. 62(7), 2014, pp. 966–975.
[30] M. Brambilla, A. Brutschy, M. Dorigo, and M. Birattari. In: ACM Trans.
Auton. Adap. 9(4), 2014, pp. 17.1–28.
[31] O. Babaoglu, G. Canright, A. Deutsch, G. A. Di Caro, F. Ducatelle, L. M.
Gambardella, N. Ganguly, M. Jelasity, R. Montemanni, A. Montresor,
and T. Urnes. In: ACM Trans. Auton. Adap. 1(1), 2006, pp. 26–66.
[32] L. Gardelli, M. Viroli, and A. Omicini. In: Multi-Agent Systems and
Applications V. Ed. by H.-D. Burkhard et al. Vol. 4696. LNCS. Berlin,
Germany: Springer, 2007, pp. 123–132.
[33] A. Reina, R. Miletitch, M. Dorigo, and V. Trianni. In: Swarm Intell.
9(2-3), 2015, pp. 75–102.
[34] A. Reina, G. Valentini, C. Fernández-Oto, M. Dorigo, and V. Trianni.
In: PLoS ONE 10(10), 2015, e0140950.
[35] M. Rubenstein, A. Cornejo, and R. Nagpal. In: Science 345(6198), 2014,
pp. 795–799.
[36] J. Werfel, K. Petersen, and R. Nagpal. In: Science 343(6172), 2014,
pp. 754–758.
[37] R. Nagpal. In: Proceedings of the First International Joint Conference
on Autonomous Agents and Multiagent Systems: Part 1. AAMAS ’02.
New York, NY, USA: ACM, 2002, pp. 418–425.
[38] J. Bachrach, J. Beal, and J. McLurkin. In: Neural Comput. Appl. 19(6),
2010, pp. 825–847.
[39] C. Pinciroli, A. Lee-Brown, and G. Beltrame. Buzz: an extensible pro-
gramming language for self-organizing heterogeneous robot swarms. Avail-
able online at http://arxiv.org/abs/1507.05946. 2015.

28
[40] J. Beal. In: Proceedings of the International Workshop on Unconven-
tional Programming Paradigms (UPP). Vol. 3566. LNCS. Berlin, Ger-
many: Springer, 2004, pp. 97–97.
[41] M. Viroli, F. Damiani, and J. Beal. In: Advances in Service-Oriented
and Cloud Computing. Ed. by C. Canal and M. Villari. Vol. 393. CCIS.
Berlin, Germany: Springer, 2013, pp. 114–128.
[42] J. Beal and M. Viroli. In: Philos. T. Roy. Soc. A 373(2046), 2015.
[43] S. Nolfi and D. Floreano. Evolutionary Robotics. Intelligent Robots and
Autonomous Agents. Cambridge, MA, USA: MIT Press, 2000.
[44] V. Trianni and S. Nolfi. In: Artif. Life 17(3), 2011, pp. 183–202.
[45] V. Trianni and M. López-Ibáñez. In: PLoS ONE 10(8), 2015, e0136406–
27.
[46] J. Lehman and K. O. Stanley. In: Evol. Comput. 19(2), 2011, pp. 189–
223.
[47] J. Gomes, P. Urbano, and A. L. Christensen. In: Swarm Intell. 7(2-3),
2013, pp. 115–144.
[48] M. Duarte, S. M. Oliveira, and A. L. Christensen. In: Artificial Life
14: Proceedings of the International Conference on the Synthesis and
Simulation of Living Systems. Ed. by H. Sayama et al. Cambridge, MA,
USA: MIT Press, 2014, pp. 657–664.
[49] M. Duarte, S. M. Oliveira, and A. L. Christensen. In: J. Intell. Robot.
Syst. 78(3-4), 2014, pp. 463–484.
[50] V. Trianni. In: Front. Robot. AI 1(13), 2014, pp. 1–6.
[51] L. P. Kaelbling, M. L. Littman, and A. W. Moore. In: J. Artif. Intell.
Res. 4, 1996, pp. 237–285.
[52] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction.
Cambridge, MA, USA: MIT Press, 1998.
[53] L. Panait and S. Luke. In: Auton. Agent Multi-ag. 11(3), 2005, pp. 387–
434.
[54] L. Bušoniu, R. Babuška, and B. De Schutter. In: Innovations in Multi-
Agent Systems and Applications - 1. Ed. by D. Srinivasan and L. Jain.
Vol. 310. SCI. Berlin, Germany: Springer, 2010, pp. 183–221.
[55] M. J. Matarić. In: Auton. Robot. 4(1), 1997, pp. 73–83.
[56] M. J. Matarić. In: J. Exp. Theor. Artif. In. 10(3), 1998, pp. 357–369.
[57] R. A. Watson, S. G. Ficici, and J. B. Pollack. In: Robot. Auton. Syst.
39(1), 2002, pp. 1–18.
[58] N. Bredeche, J.-M. Montanier, W. Liu, and A. F. Winfield. In: Math.
Comp. Model. Dyn. 18(1), 2012, pp. 101–129.
[59] E. Haasdijk, N. Bredeche, and A. E. Eiben. In: PLoS ONE 9(6), 2014,
e98466.
[60] L. König and S. Mostaghim. In: IJICC 2(4), 2009, pp. 695–723.
[61] A. F. Winfield and M. D. Erbas. In: Memetic Computing 3(4), 2011,
pp. 261–270.

29
[62] J. Pugh and A. Martinoli. In: Proceedings of the IEEE Congress on Evolu-
tionary Computation. Piscataway, NJ, USA: IEEE press, 2007, pp. 3839–
3846.
[63] E. Di Mario and A. Martinoli. In: Robotica 32(2), 2014, pp. 193–208.
[64] J. P. Hecker, K. Letendre, K. Stolleis, D. Washington, and M. E. Moses.
In: Swarm Intelligence, ANTS 2012. Ed. by M. Dorigo et al. Vol. 7461.
LNCS. Berlin, Germany: Springer, 2012, pp. 252–259.
[65] M. Gauci, J. Chen, W. Li, T. J. Dodd, and R. Groß. In: Autonomous
Agents and Multiagent Systems (AAMAS 2014). Ed. by A. Lomuscio et
al. Richland, SC, USA: IFAAMAS, 2014, pp. 421–428.
[66] M. Gauci, J. Chen, W. Li, T. J. Dodd, and R. Groß. In: Int. Journal of
Robotics Research 33(8), 2014, pp. 1145–1161.
[67] S. D. Hettiarachchi. “Distributed evolution for swarm robotics”. PhD
thesis. Laramie, WY: University of Wyoming, 2007.
[68] G. Francesca, M. Brambilla, A. Brutschy, V. Trianni, and M. Birattari.
In: Swarm Intell. 8(2), 2014, pp. 89–112.
[69] G. Francesca, M. Brambilla, A. Brutschy, L. Garattoni, R. Miletitch, G.
Podevijn, A. Reina, T. Soleymani, M. Salvaro, C. Pinciroli, F. Mascia,
V. Trianni, and M. Birattari. In: Swarm Intell. 9(2-3), 2015, pp. 125–152.
[70] M. Massink, M. Brambilla, D. Latella, M. Dorigo, and M. Birattari. In:
Swarm Intell. 7(2–3), 2013, pp. 201–228.
[71] N. Napp, S. Burden, and E. Klavins. In: Auton. Robot. 30(1), 2010,
pp. 57–71.
[72] A. Martinoli, K. Easton, and W. Agassounon. In: Int. J. Robot. Res.
23(4-5), 2004, pp. 415–436.
[73] C. Pinciroli, V. Trianni, R. O’Grady, G. Pini, A. Brutschy, M. Brambilla,
N. Mathews, E. Ferrante, G. A. Di Caro, F. Ducatelle, M. Birattari, L.
M. Gambardella, and M. Dorigo. In: Swarm Intell. 6(4), 2012, pp. 271–
295.
[74] J. Kramer and M. Scheutz. In: Auton. Robot. 22(2), 2007, pp. 101–132.
[75] A. Martinoli, A. J. Ijspeert, and F. Mondada. In: Robot. Auton. Syst.
29(1), 1999, pp. 51–63.
[76] K. Lerman, A. Galstyan, A. Martinoli, and A. J. Ijspeert. In: Artif. Life
7(4), 2001, pp. 375–393.
[77] A. F. Winfield, W. Liu, J. Nembrini, and A. Martinoli. In: Swarm Intell.
2(2-4), 2008, pp. 241–266.
[78] R. O’Grady, C. Pinciroli, A. L. Christensen, and M. Dorigo. In: 9th
Conference on Autonomous Robot Systems and Competitions, Robótica
2009. Castelo Branco, Portugal: IPCB-Instituto Politécnico de Castelo
Branco, 2009, pp. 113–119.
[79] H. Hamann and H. Wörn. In: Swarm Intell. 2(2–4), 2008, pp. 209–239.
[80] T. Schmickl, H. Hamann, H. Wörn, and K. Crailsheim. In: Robot. Auton.
Syst. 57(9), 2009, pp. 913–921.

30
[81] A. Prorok, N. Correll, and A. Martinoli. In: Int. J. Robot. Res. 30(5),
2011, pp. 574–589.
[82] D. T. Gillespie. In: J. Phys. Chem. 81(25), 1977, pp. 2340–2361.
[83] G. Valentini, H. Hamann, and M. Dorigo. In: Proceedings of the 13th In-
ternational Conference on Autonomous Agents and Multiagent Systems.
Ed. by A. Lomuscio et al. AAMAS ’14. Richland, SC, USA: IFAAMAS,
2014, pp. 45–52.
[84] H. Hamann, G. Valentini, Y. Khaluf, and M. Dorigo. In: Parallel Prob-
lem Solving from Nature – PPSN XIII. Ed. by T. Bartz-Beielstein et al.
Vol. 8672. LNCS. Berlin, Germany: Springer, 2014, pp. 181–190.
[85] Y. Liu, K. M. Passino, and M. M. Polycarpou. In: IEEE T. Automat.
Contr. 48(1), 2003, pp. 76–95.
[86] Y. Liu and K. M. Passino. In: IEEE T. Automat. Contr. 49(1), 2004,
pp. 30–44.
[87] M. A. Hsieh, Á. Halász, S. Berman, and V. Kumar. In: Swarm Intell.
2(2–4), 2008, pp. 121–141.
[88] A. F. Winfield, J. Sa, M. C. Fernandez-Gago, C. Dixon, and M. Fisher.
In: Int. J. Adv. Robot. Syst. 2(4), 2005, pp. 363–370.
[89] P. Kouvaros and A. Lomuscio. In: Proceedings of the Twenty-Ninth AAAI
Conference on Artificial Intelligence. Palo Alto, CA, USA: AAAI Press,
2015, pp. 2081–2088.
[90] P. Kouvaros and A. Lomuscio. In: Proceedings of the Twenty-Fourth In-
ternational Joint Conference on Artificial Intelligence. Palo Alto, CA,
USA: AAAI Press / IJCAI, 2015, pp. 1083–1089.
[91] J. G. Kemeny and J. L. Snell. Finite Markov Chains. Vol. 356. Princeton,
NJ: Van Nostrand, 1960.
[92] O. Soysal, E. Bahçeci, and E. Şahin. In: Turk. J. Elec. Eng. & Comp.
Sci. 15(2), 2007, pp. 199–225.
[93] N. Correll and A. Martinoli. In: Int. J. Robot. Res. 30(5), 2011, pp. 615–
626.
[94] N. Mathews, G. Valentini, A. L. Christensen, R. O’Grady, A. Brutschy,
and M. Dorigo. In: Auton. Robot. 38(4), 2015, pp. 439–457.
[95] G. Valentini, M. Birattari, and M. Dorigo. In: Proceedings of the Eu-
ropean Conference on Complex Systems 2012. Ed. by T. Gilbert et al.
Springer Proceedings in Complexity. Berlin, Germany: Springer, 2013,
pp. 651–658.
[96] G. Valentini and H. Hamann. In: Swarm Intell. 8(2–3), 2015, pp. 153–
176.
[97] H. Hamann. In: Swarm Intell. 7(2-3), 2013, pp. 145–172.
[98] Y. Khaluf, M. Pace, F. Rammig, and M. Dorigo. Integrals of Markov
processes with application to swarm robotics modelling. Tech. rep.
TR/IRIDIA/2012-020. IRIDIA, Université Libre de Bruxelles, Belgium,
2012.

31
 [99] S. Garnier, C. Jost, R. Jeanson, J. Gautrais, M. Asadpour, G. Caprari,
and G. Theraulaz. In: Advances in Artificial Life. Vol. 3630. LNAI.
Berlin, Germany: Springer, 2005, pp. 169–178.
[100] O. Soysal and E. Şahin. In: Proceedings of the IEEE Swarm Intelligence
Symposium. Piscataway, NJ, USA: IEEE Press, 2005, pp. 325–332.
[101] V. Trianni, R. Groß, T. H. Labella, E. Şahin, and M. Dorigo. In: Advances
in Artificial Life: 7th European Conference – ECAL 2003. Vol. 2801.
LNAI. Berlin, Germany: Springer, 2003, pp. 865–874.
[102] V. Gazi and K. M. Passino. In: IEEE T. Automat. Contr. 48(4), 2003,
pp. 692–696.
[103] V. Gazi and K. M. Passino. In: Int. J. Control 77(18), 2004, pp. 1567–
1579.
[104] B. Shucker and J. K. Bennett. In: Distributed Autonomous Robotic Sys-
tems 6. Ed. by R. Alami et al. Tokyo, Japan: Springer, 2007, pp. 379–
388.
[105] B. Shucker, T. D. Murphey, and J. K. Bennett. In: IEEE T. Robot. 24(6),
2008, pp. 1405–1415.
[106] A. Howard, M. J. Matarić, and G. S. Sukhatme. In: Proceedings of the
2002 International Symposium on Distributed Autonomous Robotic Sys-
tems (DARS 2002). Piscataway, NJ, USA: IEEE Press, 2002, pp. 299–
308.
[107] J.-L. Deneubourg, S. Aron, S. Goss, and J. M. Pasteels. In: J. Insect
Behav. 3(2), 1990, pp. 159–168.
[108] S. Nouyan, A. Campo, and M. Dorigo. In: Swarm Intell. 2(1), 2008,
pp. 1–23.
[109] F. Ducatelle, G. A. Di Caro, C. Pinciroli, F. Mondada, and L. M. Gam-
bardella. In: Proceedings of the IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS 2011). Los Alamitos, CA, USA:
IEEE Computer Society Press, 2011, pp. 4981–4988.
[110] P. M. Maxim, W. M. Spears, and D. F. Spears. In: Proceedings of the
IFAC Workshop on Networked Robotics. Oxford, UK: Elsevier, 2009,
pp. 19–24.
[111] V. Sperati, V. Trianni, and S. Nolfi. In: Proceedings of the 7th Interna-
tional Conference on Swarm Intelligence (ANTS 2010). Vol. 6234. LNCS.
Berlin, Germany: Springer, 2010, pp. 155–166.
[112] R. O’Grady, R. Groß, A. L. Christensen, and M. Dorigo. In: Auton.
Robot. 28(4), 2010, pp. 439–455.
[113] F. Mondada, M. Bonani, A. Guignard, S. Magnenat, C. Studer, and D.
Floreano. In: Proceedings of the VIIIth European Conference on Artificial
Life. Vol. 3630. LNCS. Berlin, Germany: Springer, 2005, pp. 282–291.
[114] R. Groß and M. Dorigo. In: Int. J. Bio-Inspir. Com. 1(1–2), 2009, pp. 1–
13.
[115] P. Levi and S. Kernbach. Symbiotic Multi-Robot Organisms: Reliability,
Adaptability, Evolution. 1st. Springer, 2010.

32
[116] R. O’Grady, A. L. Christensen, and M. Dorigo. In: IEEE T. Robot. 25(3),
2009, pp. 738–743.
[117] N. Mathews, A. Stranieri, A. Scheidler, and M. Dorigo. In: Proceedings
of 11th International Conference on Autonomous Agents and Multiagent
Systems (AAMAS 2012). Ed. by V. Conitzer et al. Richland, SC, USA:
IFAAMAS, 2012, pp. 97–104.
[118] W. C. Evans, G. Mermoud, and A. Martinoli. In: IEEE International
Conference on Robotics and Automation (ICRA 2010). Piscataway, NJ,
USA: IEEE Press, 2010, pp. 1438–1445.
[119] R. Beckers, O. Holland, and J.-L. Deneubourg. In: Prerational Intelli-
gence: Adaptive Behavior and Intelligent Systems Without Symbols and
Logic, Volume 1, Volume 2 Prerational Intelligence: Interdisciplinary
Perspectives on the Behavior of Natural and Artificial Systems, Volume
3. Ed. by H. Cruse et al. Vol. 26. Studies in Cognitive Systems. Nether-
lands: Springer, 2000, pp. 1008–1022.
[120] C. Melhuish, J. Welsby, and C. Edwards. In: Proceedings of Towards
Intelligent Mobile Robots (TIMR’99). Manchester, UK: The University
of Manchester, 1999.
[121] R. L. Stewart and R. A. Russell. In: Adapt. Behav. 14(1), 2006, pp. 21–
51.
[122] M. Gauci, J. Chen, W. Li, T. J. Dodd, and R. Groß. In: Proceedings of the
2014 International Conference on Autonomous Agents and Multi-agent
Systems. AAMAS ’14. Richland, SC, USA: IFAAMAS, 2014, pp. 421–
428.
[123] D. Payton, M. Daily, R. Estowski, M. Howard, and C. Lee. In: Auton.
Robot. 11(3), 2001, pp. 319–324.
[124] G. A. Di Caro, F. Ducatelle, and L. M. Gambardella. In: Applications of
Evolutionary Computing. Vol. 5484. LNCS. Berlin, Germany: Springer,
2009, pp. 21–30.
[125] K. J. O’Hara and T. R. Balch. In: Distributed Autonomous Robotic Sys-
tems 6. Ed. by R. Alami et al. Tokyo, Japan: Springer, 2007, pp. 305–
314.
[126] T. Stirling and D. Floreano. In: Proceedings of the 7th International Con-
ference on Swarm Intelligence (ANTS 2010). LNCS. Berlin, Germany:
Springer, 2010, pp. 562–563.
[127] F. Ducatelle, G. A. Di Caro, A. Förster, M. Bonani, M. Dorigo, S. Mag-
nenat, F. Mondada, R. O’Grady, C. Pinciroli, P. Rétornaz, et al. In:
Swarm Intell. 8(1), 2014, pp. 1–33.
[128] C. W. Reynolds. In: Comp. Graph. 21(4), 1987, pp. 25–34.
[129] A. E. Turgut, H. Çelikkanat, F. Gökçe, and E. Şahin. In: Swarm Intell.
2(2–4), 2008, pp. 97–120.
[130] H. Çelikkanat and E. Şahin. In: Neural Comput. Appl. 19(6), 2010,
pp. 849–865.
[131] E. Ferrante, A. E. Turgut, C. Huepe, A. Stranieri, C. Pinciroli, and M.
Dorigo. In: Adapt. Behav. 20(6), 2012, pp. 460–477.

33
[132] S. Hauert, J.-C. Zufferey, and D. Floreano. In: Auton. Robot. 26(1), 2009,
pp. 21–32.
[133] A. E. Turgut, C. Huepe, H. Çelikkanat, F. Gökçe, and E. Şahin. In: Pro-
ceedings of the 6th International Conference on Ant Colony Optimization
and Swarm Intelligence, ANTS 2008. Vol. 5217. LNCS. Berlin, Germany:
Springer, 2008, pp. 108–119.
[134] C. R. Kube and E. Bonabeau. In: Robot. Auton. Syst. 30(1–2), 2000,
pp. 85–101.
[135] S. Wilson, T. P. Pavlic, G. P. Kumar, A. Buffin, S. C. Pratt, and S.
Berman. In: Swarm Intell. 8(4), 2014, pp. 303–327.
[136] A. Campo, S. Nouyan, M. Birattari, R. Groß, and M. Dorigo. In: Ant
Colony Optimization and Swarm Intelligence. Ed. by M. Dorigo et al.
Vol. 4150. LNCS. Berlin, Germany: Springer, 2006, pp. 191–202.
[137] G. Baldassarre, D. Parisi, and S. Nolfi. In: Artif. Life 12(3), 2006,
pp. 289–311.
[138] S. Nouyan, R. Groß, M. Dorigo, M. Bonani, and F. Mondada. In: Pro-
ceedings of the 9th International Conference on Intelligent Autonomous
Systems. Amsterdam, Netherlands: IOS Press, 2006, pp. 433–442.
[139] S. Camazine, J.-L. Deneubourg, N. R. Franks, J. Sneyd, G. Theraulaz,
and E. Bonabeau. Self-Organization in Biological Systems. Princeton
Studies in Complexity. Princeton, NJ: Princeton University Press, 2001.
[140] I. D. Couzin, J. Krause, N. R. Franks, and S. A. Levin. In: Nature
433(7025), 2005, pp. 513–516.
[141] C. A. C. Parker and Z. Hong. In: Int. J. Robot. Res. 30(5), 2011, pp. 524–
535.
[142] I. Navarro and F. Matía. In: Auton. Robot. 33(4), 2012, pp. 445–465.
[143] M. A. Montes de Oca, E. Ferrante, A. Scheidler, C. Pinciroli, M. Birattari,
and M. Dorigo. In: Swarm Intell. 5(3-4), 2011, pp. 305–327.
[144] G. Valentini, H. Hamann, and M. Dorigo. In: Proceedings of the 14th Int.
Conf. on Autonomous Agents and Multiagent Systems. Ed. by R. Bordini
et al. AAMAS ’15. Richland, SC, USA: IFAAMAS, 2015, pp. 1305–1314.
[145] G. Valentini, E. Ferrante, H. Hamann, and M. Dorigo. In: JAAMAS,
2015, pp. 1–28.
[146] M. J. B. Krieger and J.-B. Billeter. In: Robot. Auton. Syst. 30(1–2), 2000,
pp. 65–84.
[147] W. Agassounon and A. Martinoli. In: Proceedings of the First Interna-
tional Joint Conference on Autonomous Agents and Multi-Agent Sys-
tems. Richland, SC, USA: IFAAMAS, 2002, pp. 1090–1097.
[148] G. Pini, A. Brutschy, C. Pinciroli, M. Dorigo, and M. Birattari. In: Adapt.
Behav. 21(2), 2013, pp. 117–135.
[149] A. Brutschy, G. Pini, C. Pinciroli, M. Birattari, and M. Dorigo. In: Au-
ton. Agent Multi-ag. 28(1), 2014, pp. 101–125.
[150] G. Pini, A. Brutschy, M. Frison, A. Roli, M. Dorigo, and M. Birattari.
In: Swarm Intell. 5(3–4), 2011, pp. 283–304.

34
[151] A. Brutschy, L. Garattoni, M. Brambilla, G. Francesca, G. Pini, M.
Dorigo, and M. Birattari. In: Swarm Intell. 9(1), 2015, pp. 1–22.
[152] E. Ferrante, A. E. Turgut, E. Duéñez-Guzmán, M. Dorigo, and T. Wense-
leers. In: PLOS Comput. Biol. 11(8), 2015, e1004273.
[153] E. Tuci. In: ed. by M. Dorigo et al. Vol. 8667. LNCS. Berlin, Germany:
Springer, 2014, pp. 98–109.
[154] G. S. Nitschke, M. C. Schut, and A. E. Eiben. In: Swarm Evol. Comput.
2, 2012, pp. 25–38.
[155] G. S. Nitschke, A. E. Eiben, and M. C. Schut. In: Genet. Program. Evol.
M. 13(4), 2012, pp. 493–536.
[156] S.-k. Yun, M. Schwager, and D. Rus. In: Robotics Research. Ed. by C.
Pradalier et al. Vol. 70. STAR. Berlin, Germany: Springer, 2011, pp. 607–
623.
[157] Á. M. Halász, Y. Liang, M. A. Hsieh, and H.-J. Lai. In: Distributed
Autonomous Robotic Systems. Ed. by A. Martinoli et al. Vol. 83. STAR.
Berlin, Germany: Springer, 2013, pp. 403–416.
[158] N. Correll. In: IEEE International Conference on Robotics and Automa-
tion (ICRA 2008). Piscataway, NJ, USA: IEEE Press, 2008, pp. 3302–
3307.
[159] S. Bashyal and G. K. Venayagamoorthy. In: Swarm Intelligence Sympo-
sium, 2008. SIS 2008. IEEE. Piscataway, NJ, USA: IEEE Press, 2008,
pp. 1–8.
[160] A. Kolling, K. Sycara, S. Nunnally, and M. Lewis. In: JHRI 2(2), 2013,
pp. 103–128.
[161] D. J. Bruemmer, D. D. Dudenhoeffer, and J. L. Marble. Mixed-initiative
remote characterization using a distributed team of small robots. Tech.
rep. WS-01-01/WS01-01-005. AAAI, 2001.
[162] M. Daily, Y. Cho, K. Martin, and D. Payton. In: Proceedings of the
36th Annual Hawaii International Conference on System Sciences, 2003.
Piscataway, NJ, USA: IEEE Press, 2003, pp. 125–130.
[163] A. M. Naghsh, J. Gancet, A. Tanoto, and C. Roast. In: Robot and Hu-
man Interactive Communication, 2008. RO-MAN 2008. The 17th IEEE
International Symposium on. Piscataway, NJ, USA: IEEE press, 2008,
pp. 255–260.
[164] F. Ghiringhelli, J. Guzzi, G. A. Di Caro, V. Caglioti, L. M. Gambardella,
and A. Giusti. In: 2014 IEEE/RSJ International Conference on Intel-
ligent Robots and Systems (IROS 2014). Piscataway, NJ, USA: IEEE
Press, 2014, pp. 1195–1201.
[165] A. Couture-Beil, R. T. Vaughan, and G. Mori. In: Computer and Robot
Vision (CRV), 2010 Canadian Conference on Computer and Robot Vi-
sion (CRV). Piscataway, NJ, USA: IEEE Press, 2010, pp. 159–166.
[166] S. Pourmehr, V. M. Monajjemi, R. T. Vaughan, and G. Mori. In: 2013
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS 2013). Piscataway, NJ, USA: IEEE Press, 2013, pp. 137–142.

35
[167] J. Nagi, A. Giusti, L. M. Gambardella, and G. A. Di Caro. In: Proceedings
of the 27th IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS). Piscataway, NJ, USA: IEEE Press, 2014, pp. 3834–
3841.
[168] A. Giusti, J. Nagi, L. M. Gambardella, and G. A. Di Caro. In: 2012
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS 2012). Piscataway, NJ, USA: IEEE Press, 2012, pp. 551–558.
[169] G. Podevijn, R. O’Grady, Y. S. G. Nashed, and M. Dorigo. In: Towards
Autonomous Robotic Systems - 14th Annual Conference, TAROS 2013.
Ed. by A. Natraj et al. Vol. 8069. LNCS. Berlin, Germany: Springer,
2013, pp. 390–403.
[170] A. L. Christensen, R. O’Grady, and M. Dorigo. In: IEEE T. Evolut.
Comput. 13(4), 2009, pp. 754–766.
[171] E. M. Izhikevich. In: IEEE T. Neural Networ. 10(3), 1999, pp. 508–526.
[172] C. Melhuish, O. Holland, and S. Hoddell. In: Robot. Auton. Syst. 28(2–3),
1999, pp. 207–216.
[173] M. Brambilla, C. Pinciroli, M. Birattari, and M. Dorigo. In: Fourteenth
International Conference on Advanced Robotics – ICAR 2009. Proceed-
ings on CD-ROM, paper ID 137. 2009, p. 6.
[174] C. Pinciroli, R. O’Grady, A. L. Christensen, and M. Dorigo. In: Pro-
ceedings of the Seventh International Conference on Ant Colony Opti-
mization and Swarm Intelligence (ANTS-2010). Vol. 6234. LNCS. Berlin,
Germany: Springer, 2010, pp. 558–559.
[175] C. Pinciroli, R. O’Grady, A. L. Christensen, M. Birattari, and M. Dorigo.
In: Journal of SICE 52(3), 2013, pp. 213–226.
[176] M. Dorigo, D. Floreano, L. M. Gambardella, F. Mondada, S. Nolfi,
T. Baaboura, M. Birattari, M. Bonani, M. Brambilla, A. Brutschy, D.
Burnier, A. Campo, A. L. Christensen, A. Decugnière, G. A. Di Caro,
F. Ducatelle, E. Ferrante, A. Förster, J. Guzzi, V. Longchamp, S. Mag-
nenat, J. Martinez Gonzales, N. Mathews, M. A. Montes de Oca, R.
O’Grady, C. Pinciroli, G. Pini, P. Rétornaz, J. Roberts, V. Sperati, T.
Stirling, A. Stranieri, T. Stützle, V. Trianni, E. Tuci, A. E. Turgut, and
F. Vaussard. In: IEEE Robot. Autom. Mag. 20(4), 2013, pp. 60–71.
[177] T. Schmickl, R. Thenius, C. Moslinger, J. Timmis, A. Tyrrell, M. Read,
J. Hilder, J. Halloy, A. Campo, C. Stefanini, L. Manfredi, S. Orofino,
S. Kernbach, T. Dipper, and D. Sutantyo. In: Fifth IEEE Conference on
Self-Adaptive and Self-Organizing Systems Workshops (SASOW 2011).
Piscataway, NJ, USA: IEEE Press, 2011, pp. 120–126.
[178] M. Duarte, V. Costa, J. C. Gomes, T. Rodrigues, F. Silva, S. M. Oliveira,
and A. L. Christensen. In: arXiv–CoRR abs/1511.03154, 2015.
[179] A. L. Christensen, S. M. Oliveira, O. Postolache, M. J. de Oliveira, S.
Sargento, P. Santana, L. Nunes, F. Velez, P. Sebastiao, V. Costa, M.
Duarte, J. Gomes, T. Rodrigues, and F. Silva. In: Proceedings of the In-
ternational Conference on Agents and Artificial Intelligence (ICAART).
Setúbal, Portugal: SCITEPRESS, 2015, pp. 548–555.

36
[180] M. Duarte, S. M. Oliveira, and A. L. Christensen. In: Proceedings of
the International Conference on the Synthesis and Simulation of Living
Systems (ALIFE). Cambridge, MA, USA: MIT Press, 2014, pp. 785–792.
[181] Warehouse robots get smarter with ant intelligence. Available at http:
//spectrum.ieee.org/automaton/robotics/industrial- robots/
warehouse-robots-get-smarter-with-ant-intelligence. 2012.
[182] Swarming and transporting. Available at http://www.fraunhofer.de/
en/press/research-news/2012/march/swarming-and-transporting.
html. 2012.
[183] Prospero. Dorhout R&D LLC. Available at http://www.businessinsider.
com/presenting-the-robot-farmers-of-the-future-2011-12. Ac-
cessed January, 2016.
[184] Harvest Automation, Inc. Available at https://www.harvestai.com.
Accessed January, 2016.
[185] RHEA project. Available at http://www.rhea- project.eu. Accessed
January, 2016.
[186] ASETA project. Available at http://plen.ku.dk/english/research/
crop_sciences/plant_protection/aseta/. Accessed January, 2016.
[187] S. Hauert and S. N. Bhatia. In: Trends Biotechnol. 32(9), 2014. Special
Issue: Next Generation Therapeutics, pp. 448–455.
[188] S. Hauert, S. Berman, R. Nagpal, and S. N. Bhatia. In: Nano Today 8(6),
2013, pp. 566–576.
[189] Chemical robots. Available at http://www.chobotix.cz.
[190] N. Sarvašová, P. Ulbrich, V. Tokárová, A. Zadražil, and F. Štěpánek. In:
Powder Technol. 278, 2015, pp. 17 –25.
[191] A. F. Winfield and J. Nembrini. In: IJMIC 1(1), 2006, pp. 30–37.
[192] J. D. Bjerknes and A. F. Winfield. In: Distributed Autonomous Robotic
Systems. Ed. by A. Martinoli, F. Mondada, N. Correll, G. Mermoud,
M. Egerstedt, M. A. Hsieh, L. E. Parker, and K. Støy. Vol. 83. STAR.
Berlin, Germany: Springer, 2013, pp. 431–444.

37