Self-Organised Swarm Flocking with Deep Reinforcement Learning

Optimising a set of parameters for swarm flocking is a tedious task as it requires hand-tuning of the parameters. In this paper, we developed a self-organised flocking mechanism with a swarm of homogeneous robots. The proposed mechanism used deep reinforcement learning to teach the swarm to perform the flocking in a continuous state and action space. Collective motion was represented by a self-organising dynamic model that is based on linear spring-like forces between self-propelled particles in an active crystal. We tuned the inverse rotational and translational damping coefficients of the dynamic model for swarm populations of $N\in \{25,\ 100\}$ E {25, 100} robots. We study the application of reinforcement learning in a centralised multi-agent approach, where we have a global state space matrix that is accessible by actor and critic networks. Furthermore, we showed that our method could train the system to flock regardless of the sparsity of the swarm population, which is a significant result.