Swarm robotics is a promising approach to coordinating
large groups of autonomous robots. Unfortunately, the lack
of a general methodology for designing collective
behaviors for robot swarms hinders its real-world
application.
An important share of the research in swarm robotics
has been dedicated to optimization-based design. In
optimization-based design, the mission to be accomplished
by the swarm is specified by defining a performance
measure, a function that evaluates the extent to which the
swarm attains the goals and/or satisfies the constraints
of the mission. The design problem is formulated as an
optimization problem: the possible individual behaviours
are the search space explored by an optimization algorithm
that maximises the aforementioned performance measure
It is our contention that two approaches to
optimization-based design should be disentangled:
semi-automatic design and automatic (or, more explicitly,
fully automatic) design.
In semi-automatic design, a human designer operates an
optimization algorithm as its main design
tool. Semi-automatic design is an iterative process in
which the designer, guided by their intuition and
experience, instantiates a first optimization process,
evaluates the behavior generated, modifies the
optimization process on the basis of the results observed…
these steps are repeated until the designer is satisfied
with the behavior obtained and/or feels that it cannot be
improved any further. All in all, semi-automatic design
relies on an optimization algorithm, but features a human
designer in the loop. In automatic design, the
optimization process is performed in a fully automatic way
and does not provide for any per-mission intervention of a
human designer.
In our vision, both semi-automatic and automatic design
are relevant to the development of swarm robotics; they
will occupy two different niches and will address
different contexts of application. Semi-automatic design
is appealing to handle an individual, complex design
problem for a specific mission that would be too difficult
or time-consuming to be solved manually. On the other
hand, automatic design is appealing when a design process
is to be executed repeatedly on different missions
belonging to a given class and it is impossible,
impractical, or economically unfeasible that a human
designer performs, supervises, or checks the design
process itself.
Disentangling semi-automatic and automatic design is
crucial to properly frame the future research. The clear
understanding of the specificities of semi-automatic and
automatic design will allow the community to properly
state the relevant research questions and to define
appropriate experimental protocols to address them. It
will also contribute to set correct and realistic
expectations on what each of the two approaches could and
should produce.
This work is a core conceptual contribution of the
DEMIURGE project,
funded by the European Research Council via an ERC
Consolidator Grant awarded to
Mauro Birattari
of IRIDIA, the artificial intelligence laboratory of the
ULB. The scope of the DEMIURGE project is indeed the
automatic design of robot swarms. This article frames
the central ideas of the field, disentangles them from
other ideas that are related but different, presents the
existing results, isolates the fundamental problems,
sets out the research questions that characterize it and
outlines the path of its future development.
