Sim-to-Real Latent Dynamics Adaptation in High-Gear-Ratio Humanoid Robots for Robust and Expressive Locomotion *

A key challenge in humanoid robot control arises from the significant mismatch between simulation and reality, especially when high gear ratio actuators introduce backlash and nonlinear friction. Existing methods such as system identification and domain randomization, typically rely on intricate parameter tuning and tend to produce overly conservative policies under intense motion. To address these limitations, this paper presents a novel framework for humanoid robots with high gear reduction, integrating a Latent Dynamics Adaptation module and a high-dimensional gait library. The proposed encoder-decoder architecture extracts key dynamic parameters, including joint-level control PD gains, from limited proprioceptive data, enabling real-time adaptation to varying gains, uncertain payloads, and external disturbances. A parametric gait library is constructed using velocity commands, the von Mises distribution, and inverse kinematics, while retargeted human motion-capture data provides expressive upper-body reference trajectories. The policy by reinforcement learning in simulation could be deployed directly on physical hardware without further fine-tuning. Experimental results demonstrate superior performance over conventional methods in disturbance rejection, extreme-motion capability, and overall flexibility. Under external pushes, the robot maintains higher stability, achieves faster omnidirectional locomotion, and performs continuous heel-to-toe-to-heel foot placement with expressive upper-body motions. Overall, this simulation-to-reality transfer framework substantially improves adaptation to high-gear-ratio actuators, offering a viable pathway for humanoid robots to achieve robust yet expressive whole-body motion in real-world humanoid applications.