A Dual-Task Deep Reinforcement Learning and Domain Transfer Architecture for Bootstrapping Swarming Collective Motion Skills

Recent research has shown it is possible for groups of robots to automatically “bootstrap” their own collective motion behaviors, particularly movement in a group. However, existing work has primarily provided proof of concept in regular, open arenas without obstacles. For practical applications on real robots, multiple collective motion skills are required. This article proposes a novel, multitask deep reinforcement learning algorithm and domain transfer architecture permitting multiple collective motion skills to be bootstrapped automatically and applied to real robots. The proposed approach is tested for tuning two collective motion skills for grouped movement and obstacle avoidance, without requiring a map of the environment. We show that our approach can tune obstacle avoidance parameters while maintaining high-quality swarming collective behavior when an obstacle is detected. Furthermore, learned collective motion skills can be transferred from a point mass simulation onto real mobile robots using our domain transfer architecture, without loss of quality. Transferability is comparable to that of an evolutionary algorithm run in a high-fidelity simulator.