Learning quadrupedal locomotion on tough terrain using an asymmetric terrain feature mining network

The development of robust and agile locomotion skills for legged robots using reinforcement learning is challenging, particularly in demanding environments. In this study, we propose a blind locomotion control learning framework that enables fast and stable walking on challenging terrains. First, we construct an asymmetric terrain feature extraction network that uses a multilayer perceptron to effectively infer terrain features from the history of proprioceptive states, consisting only of inertial measurement unit and joint encoder data. Additionally, our asymmetric actor-critic framework implicitly infers terrain features, thereby enhancing the accuracy of terrain representation. Second, we introduce a foot trajectory generator based on prior gait behaviors, which improves the gait periodicity and provides accurate state information for terrain feature inference. Compared to state-of-the-art methods, our approach significantly increases the learning efficiency by 26.0% and enhances terrain adaptation by 5.0%. It also achieved a more periodic gait, with the state-command tracking error reduced by 38.5% compared with advanced methods. The success rate for traversing complex terrains was similar to that of the baseline methods, with a 31.3% increase in the step height on stair-like terrains. The experimental results demonstrate that the proposed method enables fast and stable walking on challenging terrains.