Impact of Environment on Navigation Performance for Autonomous Mobile Robots in Crowds

摘要

We have been studying methods to improve both the safety and efficiency of navigation for autonomous mobile robots in crowds by switching between multiple control policies, including deep reinforcement learning. Compared to existing methods, we have shown that the method improves the collision rate (safety metrics) even in narrow corridors, but more detailed analysis is needed to address deviations in results due to environments. In this paper, we clarify the impact of different environments on navigation performance by conducting experiments under different policy-switching conditions in multiple environments. Moreover, we examine the ratio of policies used in each environment and the specific switching procedures. The results indicate what policies should be used in the target environment.