Neuro-evolution of robot swarms

About

Neuro-evolutionary robotics is an appealing approach to realizing collective behaviors for robot swarms. In its typical application, known as off-line automatic design, the neural networks controlling the robots are optimized in simulation. Notwithstanding the large number of studies that have been devoted to the application of neuro-evolution to swarm robotics, and although many methods and ideas have been proposed, empirical assessments and comparative analyses are rare and a proper state of the art has not been established, yet.

In this paper, we present an empirical study in which we compared some of the most popular and advanced neuro-evolutionary methods for the off-line design of robot swarms.

The results show that the control software produced by most of the methods under analysis performed well in simulation. Unfortunately, in the real-robot experiments, all differences were practically wiped out and all the control software produced by the different methods performed unsatisfactorily. We find this to be compelling evidence that real-robot experiments are needed to reliably assess the performance of neuro-evolutionary methods, and that the robustness to the reality gap is the main issue to be addressed to advance the application of neuro-evolution to the design of robot swarms.

This work is a core 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 Université libre de Bruxelles. The scope of the DEMIURGE project is the automatic design of robot swarms.

Publication

Read the original article:

Ken Hasselmann^‡, Antoine Ligot^‡, Julian Ruddick, and Mauro Birattari^* (2021)
Empirical assessment and comparison of neuro-evolutionary methods for the automatic off-line design of robot swarms.

Nature Communications, 12:4345

Read it online: 10.1038/s41467-021-24642-3

^‡These authors contributed equally: Ken Hasselmann and Antoine Ligot
^*Corresponding author: Mauro Birattari <mbiro@ulb.ac.be>

Authors

Ken Hasselmann

A PhD candidate at IRIDIA and teaching assistant at École Polytechnique de Bruxelles, Université libre de Bruxelles. He received a Master's degree in Electronics and Embedded Systems Engineering from INP-ENSEEIHT in Toulouse and a Master's degree in Innovation from Toulouse School of Management in 2014. His research interests include swarm robotics, machine learning, and the automatic design of collective behaviors and communication protocols for robot swarms.

Antoine Ligot

A PhD candidate at IRIDIA, Université libre de Bruxelles. He received a Master's degree in Computational Intelligence from Université libre de Bruxelles in 2016. His research interests include swarm robotics, automatic design of collective behaviors, design methodologies and methods to handle the reality gap.

Julian Ruddick

A master student at IRIDIA, Université libre de Bruxelles at the moment of contributing to this work. He is now a PhD candidate at MOBI, Vrije Universiteit Brussel. His research interests include reinforcement learning, energy communities and machine learning based energy management systems.

Mauro Birattari

A Research Director of the Belgian Fonds de la Recherche Scientifique—FNRS at IRIDIA, Université libre de Bruxelles. He received a Master's degree in Electrical and Electronic Engineering from Politecnico di Milano in 1997 and a doctoral degree in Information Technologies from the Faculty of Engineering of the Université libre de Bruxelles in 2004. His research focuses on swarm intelligence, collective robotics, machine learning, and on the application of artificial intelligence techniques to the automatic design of algorithms.

Dr. Birattari is the principal investigator of the project "DEMIURGE: automatic design of robot swarms," funded by the European Research Council through an ERC Consolidator Grant.

Comparison of neuro-evolutionary methods for swarm robotics