Trot Gait Acquisition and Improvement in Movement of Four-Legged Running Robot Based on Reinforcement Learning

Abstract

When we develop robots, we cannot completely deal with unexpected situations and unknown environments in advance. One way to solve this problem is reinforcement learning, which is a technique for finding an optimal behavior pattern with respect to the target depending on the circumstances and environment through trial and error. This learning enables an action suitable for the achievement of the aim tailored to the environment without changing the control program. However, it takes a lot of time to complete learning. In this study, we applied reinforcement learning to our four-legged robot modeled on four-legged animals and examined whether the robot could acquire an improved trotting gait. As a result, the robot acquired the trotting gait. Furthermore, in repeated learning, running speed became 31% faster, and energy consumed fell by 23% because running movement was improved.