换个地面/身体，机器人策略就失灵？让扩散策略读懂「动力学」| ICML'26

TL;DR - DADP is a domain-adaptive diffusion policy for zero-shot robot control across changing physical dynamics. It improves generalization by learning stable domain representations and embedding them directly into the diffusion process.

• Lagged contexts from separate episodes suppress transient state cues and isolate static factors such as friction, mass, damping, and morphology.
• Domain representations shift the diffusion prior toward domain-specific action manifolds instead of serving only as concatenated inputs.
• DADP performed strongly across Seen, Unseen, and OOD MuJoCo and Adroit settings, with especially large OOD gains.
• The method retained more performance with one-step DDIM sampling, but primarily targets static or piecewise-stable dynamics.