DeVI: Physics-based Dexterous Human-Object Interaction via Synthetic Video Imitation
DeVI learns physics-based control for dexterous human-object interaction by imitating synthetic videos generated from text prompts. It solves issues with unreliable 3D motion reconstruction by using a hybrid imitation target combining 3D human tracking and 2D object trajectories. This framework enables physically plausible manipulation without needing real motion capture data.