Reinforcement Learning-based Motion Generation for a Tracked Robot to Go Over a Sphere-shaped Non-fixed Obstacle

Hidenori Takamiya, Ryosuke Yajima, Jun Younes Louhi Kasahara, Ren Komatsu, Keiji Nagatani, Atsushi Yamashita, Hajime Asama

Abstract: Tracked robots have high traversability over rough terrain. However, even for such robots, it is still challenging to traverse terrain with non-fixed obstacles which may move when the robots go over them. Therefore, we propose a reinforcement learning-based method to generate the motion of the tracked robot to go over the obstacle. We set a task where the robot attempts to go over a sphere-shaped non-fixed obstacle and reach the goal. To succeed in the task, we designed a reward function so that the robot can reach the goal as straight as possible. As a training algorithm, Deep Q-Network was used and the robot was trained in a dynamics simulator. It was confirmed that the robot succeeded in the task using the trained network, which generated motion for going over a sphere-shaped non-fixed obstacle.