Evolving an Integrated Phototaxis and Hole-avoidance Behavior for a Swarm-bot

Abstract

This article is on the subject of evolving neural network controllers for cooperative, mobile robots. We evolve controllers for combined hole-avoidance and phototaxis in a group of physically connected, autonomous robots called s-bots, each with limited sensing capabilities. We take a systematic approach to finding a suitable fitness function, an appropriate neural network structure, and we explore and compare three evolutionary algorithms commonly used in evolutionary robotics: genetic algorithms, (µ,λ) evolutionary strategies, and cooperative coevolutionary genetic algorithms for optimizing weights in neural robot controllers. Finally, we show that solutions evolved in our software simulator can be transferred successfully to real robots.