Autonomous Swarm Navigation Using Multi-Agent Reinforcement Learning and Neuro-Evolution

In recent years, swarm robotics has proven to be a promising approach to develop collective behaviors that permit to extend the capabilities of a single robot in solving complex tasks. Here, the major challenge consists of designing individual behaviors that lead to the emergence of the desired collective behavior. In this regard, several automatic design methods have been proposed over the last decades. In this work, we present a hybrid approach that combines two of the main techniques of such design methods: multi-agent reinforcement learning and neuro-evolution. The idea is to evolve a pre-trained individual policy, in such a way that we can learn an effective collective behavior for vision-based navigation tasks. Specifically, we want a homogeneous swarm of low-cost robots to autonomously explore a simulated indoor corridor, while avoiding static and dynamic obstacles in a flexible and intelligent way. Through several experiments, we show how the learned collective policy outperforms the individual one in solving the previous task. Furthermore, our approach becomes highly memory-efficient, thanks primarily to a particular autoencoder able to compress input images into very compact visual representations.