An autonomous mobile robot that follows a specific target by deep reinforcement learning considering the existence of the target human

Abstract

This study introduces an action model for mobile robots that follows the specific walking human ahead of the robot and reaches destinations with the target human. The action model is trained by deep reinforcement learning in simulation environments including the target human. A conventional study of the authors shows that the training using 2D-LiDAR scan can provide action models for navigation to the destinations and following human flows. In this study, the robot distinguishes humans and the other static obstacles and distances to the human observed by the robot in addition to scan data are adopted as input state of the training. As the results, following of the specific human was achieved in the simple simulation environments.