Offline-Learned Visual Encoder for Privileged-to-Real Bipedal Locomotion

This paper proposes a reinforcement learning framework for vision-based locomotion control of bipedal robots on complex, unstructured terrains, integrating privileged perception transfer, offline Transformer-based depth reconstruction, and historical observation encoding. During training, the critic utilizes dense privileged height samples to improve value estimation accuracy, while the actor operates on sparse observations consistent with the deployment phase, ensuring stable policy transfer from training to real-world application. The offline visual module converts depth images into structured height representations, significantly alleviating the dependence of online training on high simulation parallelism and computational resources. The historical encoder further enhances temporal modeling capabilities for dynamic events such as terrain transitions and contact delays. Validation results on the MuJoCo simulation platform and the TRON1 hardware platform demonstrate that the proposed method consistently outperforms existing end-to-end visual policies and proprioception-based strategies across multiple performance metrics, exhibiting superior stability and adaptability. This framework offers a scalable new paradigm for achieving efficient and robust vision-based locomotion control without requiring high-precision localization or calibration.