PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization
PhysMoDPO introduces a method to improve human motion generation for robotics by integrating Whole-Body Control into a Direct Preference Optimization (DPO) framework. The system trains motion generators using physics-grounded rewards derived from tracking, sliding, and task adherence metrics. This process refines the generator to produce motions that are both physically plausible and accurately follow original text instructions.