DADP makes one diffusion policy work across changing robot dynamics
新智元 · wechat · 2026-07-27
A joint team from UC Berkeley, Peking University, CMU, and Tsinghua proposes DADP, a Domain-Adaptive Diffusion Policy that learns one policy for multiple robot dynamics settings instead of training separate controllers for each domain.
- The method tackles a common failure mode in robotics: policies overfit to a specific mix of friction, mass, damping, and body shape, then break when deployed in a new environment.
- DADP uses a lagged-context prediction scheme to extract more stable domain information, then injects that domain embedding into the diffusion process so generation starts from a domain-aware prior instead of the same generic noise.
- In MuJoCo and Adroit zero-shot cross-domain tasks, the paper reports strong results, especially under OOD shifts; Walker2d is highlighted as a standout case.
- Ablations show that pushing context further away from the current timestep improves representation quality, and that combining domain-aware priors with a modified diffusion objective matters more than simply concatenating embeddings to the input.
More from Embodied
- Astribot’s Lumo-2 uses latent world dynamics to improve long-horizon robot tasks — jiqizhixin · 2026-07-27
- SynapseSemi puts a neural network inside the camera pixel — ycombinator · 2026-07-27
- Enigma raises $70 million to simplify robot control — luisdans · 2026-07-27
- IROS 2026 to host a Pittsburgh workshop on scalable tactile sensing for dexterous robots — mangahomanga · 2026-07-27
- AGIBOT and JD Logistics unveil a 50 kg humanoid for warehouse operations — CyberRobooo · 2026-07-27
- U-BOT enters early sketching as a small outdoor utility robot concept — _Stocko_ · 2026-07-27