High-speed quadrupedal locomotion by imitation-relaxation reinforcement learning

Fast and stable locomotion of legged robots involves demanding and contradictory requirements, in particular rapid control frequency as well as an accurate dynamics model. Benefiting from universal approximation ability and offline optimization of neural networks, reinforcement learning has been used to solve various challenging problems in legged robot locomotion; however, the optimal control of quadruped robot requires optimizing multiple objectives such as keeping balance, improving efficiency, realizing periodic gait and following commands. These objectives cannot always be achieved simultaneously, especially at high speed. Here, we introduce an imitation-relaxation reinforcement learning (IRRL) method to optimize the objectives in stages. To bridge the gap between simulation and reality, we further introduce the concept of stochastic stability into system robustness analysis. The state space entropy decreasing rate is a quantitative metric and can sharply capture the occurrence of period-doubling bifurcation and possible chaos. By employing IRRL in training and the stochastic stability analysis, we are able to demonstrate a stable running speed of 5.0 m s(-1) for a MIT-MiniCheetah-like robot.