Safe Locomotion Within Confined Workspace using Deep Reinforcement Learning

Abstract

Legged Robots have emerged as new and efficient alternatives to conventional wheeled robots for versatile locomotion techniques and stable walking gaits over varying rough terrains. Unfortunately, these promises of the legged robots arrive with an issue of handling high complexity in footstep planning and synchronous joints control. On the other side, deep reinforcement learning (DRL) has revolutionized the method for learning complex policies without any prior knowledge of the environment and robotic entity itself. This paper tried to bind these two advantages so that stable and safe locomotion can be learned for a complex six-legged hexapod robot through DRL. Based on tripod gait planning by spiders, a safe controller has also been used as a back-end action sequence governor in this work to reduce number of falls and avoid self-collision when robot is moving in high speed. As well, to keep the robot within a limited work-space, the agent is required to develop autonomously the capabilities of walking along at least two directions and learning variable speed-control while walking.