Skill Latent Space Based Multigait Learning for a Legged Robot

Using proprioception to reproduce multiple gaits is nontrivial for legged robots, especially when encountering different terrains and velocity commands. Recently, reinforcement learning (RL) has been utilized to design powerful blind locomotion controllers. However, many RL-based blind locomotion methods are studied based on single gait. In this work, we propose an end-to-end training framework capable of learning multiple gaits for a quadruped robot. A latent space is constructed concurrently by a gait encoder and a gait generator, which helps the robot to reuse multiple gait skills to achieve adaptive gait (AG) behaviors. The trained controller enables smooth transitions between gaits and generates an AG. This means that the robot can effectively apply different gaits based on the current state and velocity command. To learn natural behaviors for multiple gaits, the reward explicitly constructed from gait parameters and the reward implicitly constructed from conditional adversarial motion priors together form the gait-dependent reward, which is subsequently added to the total reward. In the experiment, we demonstrate good performance of the multiple gaits control on a quadruped robot with only proprioceptive sensors.