Self-Attention機構を用いた深層強化学習による自律分散群ロボットの協調障害物除去行動の獲得

―ロボットの観測情報による性能差の検証―

抄録

At disaster sites, swarm robot systems are required to cooperatively remove obstacles of different number and mass, but an autonomous distributed swarm robot system with scalability, flexibility, and robustness has not been established. In this paper, we examine whether deep reinforcement learning based on Self-Attention can be applied to autonomous distributed swarms of robots to acquire cooperative obstacle removal behavior. Specifically, we verified the robustness, flexibility, and scalability of learning behaviors using the relative coordinates of obstacles and points on the robot’s surface as the robot’s observation information. As a result, it was confirmed that scalability and flexibility can be obtained by the proposed learning method.