Robust Walking and Sim-to-Real Optimization for Quadruped Robots via Reinforcement Learning

Abstract

Achieving robust walking for different stairs is one of the most challenging tasks for quadruped robots in real world. Traditional model-based methods heavily rely on environmental factors, are burdened by intricate modelling complexities, and lack generalizability. The potential for advancements in adaptive locomotion control, often impeded by complex modelling processes, can be substantially enhanced through the application of Reinforcement Learning (RL). In this paper, a learning-based method is proposed to directionally enhance the stair-climbing skill of quadruped robots under different stair conditions. First, the general policy model based on proprioceptive perception is trained as a pre-training model. Then, the pre-training model was initialized, and different terrain information from the stairs was introduced for customized training to enhance the stair-climbing skill without affecting the existing locomotion performance. Finally, the customized control policy is deployed to the real robot to realize motion control in real environments. The experimental results demonstrate that the customized control policy can significantly improve the motion performance of quadruped robots when facing complex stair terrains and has certain generalizability in other complex terrains. The proposed algorithm can be extended to various terrestrial environments.