Multi-agent Deep Reinforcement Learning for Self-organized Aggregation of a Swarm of Robots

Self-organized aggregation is a fundamental behavior for a swarm of robots wherein the robots are required to form one aggregate. It allows interaction between the swarm members and it is considered as a prerequisite for other complex behaviors. Designing automatically such behavior is attractive and promising, yet challenging, particularly in a homogeneous environment and without communication. In this paper, we tackle this problem using the vanilla policy gradient (VPG) algorithm following the centralized training decentralized execution scheme (CTDE). We assess its performance in our implemented physics-based simulator following the OpenAI Gym framework. Our results demonstrate the viability of this approach for designing automatically a scalable self-organized aggregation.