Shanghai AI Lab's InternW0 Is a Physical World Model Unifying Video Prediction and Robot Control
Jisong Cai · hf · 2026-09-24
Shanghai AI Laboratory introduces InternW0, the first instantiation of its InternW physical world model series, built on omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling. It jointly learns future visual dynamics and continuous robot control via an asymmetric video-action architecture with flow matching: a large video expert provides longer-horizon context while a lightweight action expert runs at a faster timescale.
- Reuses layerwise K/V with observation-conditioned context routing instead of regenerating the future per action
- Contact-aware post-training incorporates force and tactile signals; soft prompts support heterogeneous embodiments
- Trained on 7,200 hours of heterogeneous robot and egocentric data, including the 275-hour real-lab EgoLab dataset
- Evaluated on simulation benchmarks and real science tasks like 15-stage MOF synthesis and force-aware dexterous pipetting
More from Embodied
- Programmable quantum photonic processor runs in orbit for the first time on a nanosatellite — jwt0625 · 2026-09-24
- Humane AI Pin gets rooted and runs Rabbit OS3 in community hack — jesselyu · 2026-09-24
- No-camera Meta glasses spark debate: killer use case is blink-controlled AI prompts — tekbog · 2026-09-24
- Tesla Robotaxi in Austin rides smoother than any Uber or Lyft, user reports — EricETesla · 2026-09-24
- GPT-6 with a robot arm executed most of 5 dangerous tasks: stabbing a dummy, tossing gas canisters into a furnace — FuSheng_0306 · 2026-09-24
- "Turns out robots are just computers. Very anticlimactic." — danfei_xu · 2026-09-24