Decentralized Multi-Robot Collision Avoidance in Complex Scenarios With Selective Communication

Abstract

Deep reinforcement learning has been demonstrated to be an effective solution to the multi-robot collision avoidance problem. However, with existing methods, robots typically generate actions only based on local observations, sometimes augmented with global communication. Their performance deteriorates in limited bandwidth environments and complex scenarios with various obstacles and high robot density. We propose SelComm, a selective communication framework to generate cooperative and collision-free actions for robots in multi-robot navigation tasks. Specifically, we develop a decentralized message selector, enabling each robot to calculate relations with other robots using both agent-level information and sensor-level information, and select the most valuable messages to meet the bandwidth limitation. Then we introduce the attentional communication channel for efficient communication. Our experimental evaluations based on various scenarios demonstrate that SelComm learns more cooperative behaviors and outperforms state-of-the-art methods in limited bandwidth environments and complex scenarios.