Fault-Tolerant Gait of a Quadruped Robot at One Joint Using Deep Reinforcement Learning

Abstract

In recent years, legged rescue robots have been attracting attention. In disaster situations, the legs of the robot may fail, but the robot must still be able to perform its mission. However, few studies have been conducted on fault-tolerant gaits of quadruped robots using deep reinforcement learning. The objective of this study was to create a gait that can cope with failures during walking. The two cases of failure are considered, a fixed motor and a free motor. We used deep reinforcement learning to create a gait that allows the robot to continue walking without tipping over, even if it suddenly breaks down while walking. The results of the validation in the simulation showed the effectiveness of the proposed method.