Style-Aware and Robust Sim-to-Real Humanoid Locomotion with Contact-Rich Motion Priors and Online Dynamics Adaptation^*

Achieving natural and expressive full-body motion remains a significant challenge for humanoid robots. This work proposes a sim-to-real control framework integrating an online dynamics inference and adaptation module with a Contact-Rich Adversarial Motion Prior (CRAMP) to enable robust and style-aware motion control for humanoid robots with high gear ratio actuation. The dynamics adaptation module relies solely on proprioceptive feedback and multi-point foot contact modeling to online estimate and adapt critical dynamic parameters, effectively narrowing the simulation-to-reality gap. Meanwhile, CRAMP extends conventional adversarial motion priors by explicitly incorporating the robot-environment contact structure into the discriminator input, enabling the policy to learn end-to-end expressive behaviors with consistent styles and coordinated contact transitions, without relying on sep-arate upper-lower body models or multi-stage training pro-cesses. The control policy is trained via reinforcement learning in simulation and successfully transferred to real hardware without additional tuning. Experimental results from G TX-III demonstrate that the proposed method stably generates natural and stylistically diverse full-body motions exhibiting clear capabilities in gait rhythm modulation, posture tension control, asymmetric adaptation, contact strategy adjustment, and delivering multidimensional stylistic responses in context-rich human-like scenarios.