Motion Planning with Deep Reinforcement Learning without Using the Pre-Environment Map for Autonomous Mobile Robot

Abstract

In this paper, we propose a learning-based motion planner for autonomous mobile robot which output continuous motion commands by referring from the scan data of 2D-LiDAR and the target point without using environmental map. We aim to use this method for a safe and efficient navigation system of an autonomous mobile robot with high generalization performance in urban environment. We show that a mapless motion planner can be trained through a deep reinforcement learning method in only simulator. In order to verify the effectiveness of the proposed method, collision avoidance and navigation performance are evaluated by directly applying the learned planner to the actual machine in the real world. As a result, we showed that we can obtain navigation performance equivalent to conventional method without using pre-environment map.