Learning Impulse-Reduced Gait for Quadruped Robot using CMA-ES

This study proposes reinforcement learning (RL) for a policy that can create impulse-reduced quadrupedal walking. Reducing the impulse at the foot can be helpful since a large impulse generates loud noise and it adversely affects the durability of the robot. In this study, we constructed a neural network for the gait scheduling policy and trained it to minimize impulse. The RL and covariance matrix adaptation evolution strategy are adopted and learning is conducted by simulations. The performance of the developed learned policy is verified through experiments with the quadruped robot A1. As a result, the impulse is successfully reduced in addition to the capability of robust walking.