Gait Generation of Four-legged Running Robot Based on Reinforcement Learning to Reach a Goal

Abstract

A robot cannot completely deal with unexpected situations and unknown environments by simple control. One approach to solve this problem is Reinforcement Learning. By incorporating this learning into the control, a robot can learn and select actions toward achieving a target in the environment. In this study, we applied Reinforcement Learning to the task of reaching a goal using our robot modeled on four-legged animals. We examined the influence of differences in states and rewards on the learning effect. As a result, when we used information on the posture of the robot as well as the goal and gave both short-term and long-term rewards, the robot recorded the fastest time to the goal. Focusing on goal-reaching actions, we confirmed that the robot ran by gradually adjusting its direction in the initial stage of learning and ran straight after first turning in repeated learning.