Comparison of Different Exploration Schemes
        in the Automatic Modular Design
                of Robot Swarms?

 Gaëtan Spaey, Miquel Kegeleirs, David Garzón Ramos, and Mauro Birattari

                    IRIDIA, Université Libre de Bruxelles, Belgium
                                 mbiro@ulb.ac.be

The main challenge on the design of robot swarms is that no general methodology
exist to anticipate the global behavior of a robot swarm based on the behavior
of a single individual [2]. A promising approach to overcome this difficulty is the
design of collective behaviors by automatic design methods [1].
    In this work, we study different exploration schemes in the context of au-
tomatic modular design methods—such as AutoMoDe-Chocolate, proposed by
Francesca et al. [4]. Exploration schemes are an essential part of many collective
behaviors, and yet, they have never been thoroughly evaluated in the context of
automatic modular design. Our hypothesis is that the exploration schemes—such
as random walks [3, 5] and the ballistic motion used in Chocolate—have a no-
ticeable influence on the exploration capabilities of automatically designed robot
swarms. To test our hypothesis, we conceived AutoMoDe-Coconut—a variant of
Chocolate with multiple configurable exploration schemes embedded within its
modules. We assess Coconut in the design of collective behaviors for missions that
require robot swarms to explore in different manners: aggregation—searching for
a specific location to aggregate; foraging—traveling back and forth from two lo-
cations; and grid exploration: uniformly covering the workspace. We conduct
realistic simulations in workspaces that are either bounded or unbounded (i.e.
robots might leave the workpsace) and we compare the performance of Coconut
and Chocolate. We expect Coconut to appropriately select and configure its
exploration schemes for the missions in hand. We also expect Coconut to out-
perform Chocolate—we conjecture that Chocolate produces swarms with lesser
exploration capabilities due to its single exploration scheme, the ballistic motion.
    We observed that Coconut is prone to select exploration schemes that fit the
requirements of the workspace, as shown by Figure 1a. In bounded workspaces,
Coconut uses mainly the ballistic motion—this allows robots to cover large dis-
tances and lead the swarm to explore widely. In unbounded workspaces, Coconut
uses mainly the random walk—this maintains robots to explore in small areas
and lead to behaviors that keep the robots within the workspace. However, con-
trary to our expectations, Coconut performs similarly to Chocolate, as shown by
Figure 1b. Originally, we expected that ballistic motion would perform poorly in
unbounded workspaces—the steady straight motion could easily lead robots to
?
    GS and MK contributed equally to this work and should be considered as co-first au-
    thors. Copyright c 2019 for this paper by its authors. Use permitted under Creative
    Commons License Attribution 4.0 International (CC BY 4.0).
2        G. Spaey et al.




    (a) Exploration schemes distribution       (b) Comparative performance results

Fig. 1: Aggregated experimental results in bounded and unbounded workspaces:
(a) Distribution of exploration schemes in Coconut when the robots explore; (b)
Comparative performance results between Coconut and Chocolate.

leave the workspace. However, Chocolate is able to design collective behaviors
that prevent the robots to leave the workspace by combining its different mod-
ules. This result is interesting as it allows us to make the following observation:
the exploration capabilities of modular control software come from the high level
interaction between modules and not only from the exploration schemes embed-
ded within them. In this sense, automatic modular methods such as Chocolate
can produce complex exploration strategies even if they are endowed with sim-
ple exploration schemes—in part confirmed by the effective design of collective
behaviors in previous studies [4]. We hence conclude that ballistic motion is a
sufficiently appropriate exploration scheme for classes of missions with bounded
workspaces. Still, whether an exploration strategy based on random walk could
be a suitable solution in other contexts needs to be further explored.
Funding: ERC (Demiurge: grant No 681872); FRS-FNRS (MB); COLCIEN-
CIAS (DGR).

References
1. Birattari, M., Ligot, A., Bozhinoski, D., Brambilla, M., Francesca, G., Garattoni, L.,
   Garzón Ramos, D., Hasselman, K., Kegeleirs, M., Kuckling, J., Pagnozzi, F., Roli,
   A., Salman, M., Stützle, T.: Automatic off-line design of robot swarms: a manifesto.
   Frontiers in Robotics and AI 1(1), 1 (2019)
2. Brambilla, M., Ferrante, E., Birattari, M., Dorigo, M.: Swarm robotics: a review
   from the swarm engineering perspective. Swarm Intelligence 7(1), 1–41 (2013)
3. Dimidov, C., Oriolo, G., Trianni, V.: Random Walks in Swarm Robotics: An Ex-
   periment with Kilobots. In: Swarm Intelligence, vol. 9882, pp. 185–196. Springer,
   Cham, Switzerland (2016)
4. Francesca, G., Brambilla, M., Brutschy, A., Garattoni, L., Miletitch, R., Podevijn,
   G., Reina, A., Soleymani, T., Salvaro, M., Pinciroli, C., Birattari, M.: AutoMoDe-
   Chocolate: automatic design of control software for robot swarms. Swarm Intelli-
   gence 9(2/3), 125–152 (2015)
5. Kegeleirs, M., Garzón Ramos, D., Birattari, M.: Random walk exploration for swarm
   mapping. In: Towards Autonomous Robotic Systems. TAROS 2019. LNCS, vol.
   11650, pp. 211–222. Springer, Cham, Switzerland (2019)