Room for Me?: Mobile Navigation for Entering a Confined Space Using Deep Reinforcement Learning

Abstract

We consider the navigation problem where a robot tries to enter a confined space where obstacles are populated randomly. The objective is to move into and stay inside the confined space while not colliding with obstacles. Since the robot does not know the exact locations of the obstacles beforehand, the goal location inside the space and the path to the goal cannot precomputed. Instead of determining the goal and planning path once the configuration space is exactly known, we propose a method based on deep reinforcement learning (DRL). The method directly generates the control input of the robot from the state and sensing information. We train the DRL agent in a simulated environment with a mobile robot in a confined space with randomly located obstacles (up to three). The robot achieves at most 95% of success rate where the robot enters the space without collisions and stays inside for a certain period of time.