Learning Robust and Agile Legged Locomotion Using Adversarial Motion Priors
Jinze Wu, Guiyang Xin, Chenkun Qi, Yufei Xue
Abstract
Developing both robust and agile locomotion skills for legged robots is non-trivial. In this work, we present the first blind locomotion system capable of traversing challenging terrains robustly while moving rapidly over natural terrains. Our approach incorporates the Adversarial Motion Priors (AMP) in locomotion policy training and demonstrates zero-shot generalization from the motion dataset on flat terrains to challenging terrains in the real world. We show this result on a quadruped robot Go1 using only proprioceptive sensors consisting of the IMU and joint encoders. Experiments on the Go1 demonstrate the robust and natural motion generated by the proposed method for traversing challenging terrains while moving rapidly over natural terrains.