DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models
DIDO distills complex, multi-step video generation dynamics into a single denoising step to reduce latency in robotic manipulation. It addresses the problem where critical interaction dynamics between a gripper and an object are lost when models are truncated early. DIDO uses interaction-centric reasoning tokens and alignment with pre-trained encoders to recover these crucial details, leading to faster inference without losing manipulation accuracy.