Abstract
The mechanisms of locomotion in mammals have been extensively studied and inspire the related researches on designing the control architectures for the legged robots. Reinforcement learning (RL) is a promising approach allowing robots to automatically learn locomotion policies. However, careful reward-function adjustments are often required via trial-and-error until achieving a desired behavior, as RL policy behaviors are sensitive to the rewards. In this paper, we draw inspiration from the rhythmic locomotion behaviors of animals and propose a new control architecture by incorporating a rhythm generator to naturally stimulate periodic motor patterns, which actively participates in the timing of phase transitions in the robot step cycle. To speed up training, we use the joint position increments rather than the conventional joint positions as the outputs of the RL policy. During deployment, the rhythm generator can be reused for the state estimation of quadruped robots. We validate our method by realizing the full spectrum of quadruped locomotion in both simulated and real-world scenarios.
Funding Information
  • National Key Research and Development Plan of China (2018AAA0102504)
  • National Natural Science Foundation of China (U1913204, 61991411)
  • Natural Science Foundation of Shandong Province for Distinguished Young Scholars (ZR2020JQ29)
  • Project for Self-Developed Innovation Team of Jinan City (2021GXRC038)

This publication has 32 references indexed in Scilit: