Learning Gait-conditioned Bipedal Locomotion with Motor Adaptation

Whole body locomotion in humanoid robots remains a significant challenge due to the requirement of whole body coordination, natural bipedal walking gait, and accurate state estimation to enable them to traverse plain and uneven terrain. In this paper, we propose a learning-based humanoid locomotion controller that can adapt to disturbance and uneven terrain. We leverage the advances in rapid adaptation for quadruped locomotion control, and expend them to the humanoid robots, resulting in a well-performed whole body locomotion policy, gait-conditioned RMA without any reference trajectory during training. In simulation test, our trained policy demonstrates the ability to generalize in unseen turning tasks and showcases its robustness in more complex environment include hill and steps.