PhysDrift: Bridging the Embodiment Gap in Humanoid Co-Speech Motion Generation

PhysDrift is a new embodiment-aware co-speech motion generation framework designed for humanoid robots, which directly predicts executable joint trajectories from speech, bypassing the traditional human-centric retargeting methods. The framework introduces IK-EER, optimizing kinematic feasibility and speech-motion alignment, and shows improvements in speech-motion synchronization, physical plausibility, and real-time interaction capabilities. This advancement addresses the embodiment gap in humanoid motion generation, enhancing expressive behaviors and motion diversity, which is crucial for practitioners developing more capable humanoid systems.