Trajectory Changes Based on Rewards in Double Deep Q-Network for Autonomous Mobile Robot Navigation
Takumi Furuya, Yuki Kato, Kazuyuki Morioka · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2019
We developed an autonomous mobile robot system based on behaviors acquired by deep reinforcement learning. Navigation performances including traveling trajectories are affected by the design of state, action and reward in deep reinforcement learning. This paper focuses on rewards given in training of action policies on the simulator. For example, negative rewards are given to the situations that the robot approaches to obstacles closely. Then, the robot has a tendency to run far from obstacles. In the paper, robot navigation experiments in a real world were performed. Differences of the trajectories according to several rewards are discussed.