ShoesMimic: One-Policy Whole-Body Control for Multi-Shoe Humanoid Locomotion

Humanoid robots are commonly designed with rigid and flat feet. Although this structure simplifies control, it often leads to large landing impacts and contact noise during walking, which limits the range of potential applications and shortens mechanical lifespan. Although wearing shoes can help alleviate the above problems by providing cushioning and compliance, it also introduces new challenges such as inconsistency among shoe types and reduced walking stability. To address this issue, a unified framework named ShoesMimic is proposed for humanoid walking control across multiple shoe types. In this framework, variations in sole geometry and physical properties are systematically introduced during both training and deployment, allowing stable whole-body gait control to be maintained under diverse contact conditions. A humanoid shoe asset library is constructed, containing various shoe types. During physical experiments, multiple shoes are tested, and stable gait and posture control are maintained by ShoesMimic using a single policy. These results demonstrate robustness and noise resistance under changes in contact stiffness and geometry. ... mehrThe study provides the first systematic solution for humanoid walking control across different shoe types and establishes a foundation for future research on physics-consistent imitation learning and contact-adaptive control.