PhysDrift Improves Humanoid Co-Speech Motion Generation

Current methods for generating co-speech motions for humanoid robots typically involve first creating motions for human body models and then adapting them to robots. This paper identifies a significant "embodiment gap" in this approach, where the inherent differences between human and robot body mechanics lead to inconsistencies and a reduction in motion diversity during the transfer process. This ultimately limits the expressiveness of humanoid behaviors. To overcome this, the researchers propose IK-EER, a framework for curating robot-native motion data that ensures both kinematic feasibility and precise speech-motion synchronization. Building on this, they introduce PhysDrift, a novel embodiment-aware generation framework. PhysDrift directly predicts executable humanoid joint movements from speech input, eliminating the need for intermediate human-body representations. By maintaining embodiment consistency throughout training and inference, and incorporating physical regularization, PhysDrift significantly enhances speech-motion alignment, physical plausibility, and smoothness. Extensive experiments and real-world robot deployments confirm its effectiveness, demonstrating improved efficiency and real-time interaction capabilities compared to traditional human-centric pipelines.