BIE: Bipedal Robots Using Implicit-Explicit Learning Framework

Bipedal robots possess the unique capability to adapt human-centric environments, yet achieving agile and robust locomotion on complex terrains remains a formidable challenge due to their inherent instability and the noise associated with visual perception. Current approaches often struggle to effectively integrate high-dimensional visual data with precise motor control, leading to conservative behaviors or failure in harsh environments. In this paper, we propose a novel learning-based framework that synergizes a Teacher-Student architecture with a dual-level Implicit-Explicit learning mechanism to address these limitations. Our method leverages a teacher policy trained with privileged information to guide a student policy that operates solely on proprioceptive and noisy visual inputs. Crucially, we incorporate an implicit-explicit estimation module within the student network: the explicit component reconstructs terrain features to provide a direct understanding of the environment, while the implicit component captures latent dynamic parameters to compensate for actuation and sensing uncertainties. This hybrid approach enhances the student's ability to generalize across diverse terrains by aligning both behavioral policies and state representations. We validate our framework through extensive simulation and real-world experiments on a bipedal robot. Results demonstrate that our method achieves great performance in traversing challenging terrains with high agility and robustness, successfully bridging the sim-to-real gap without requiring complex pre-mapping.