InternW0: A Foundational Physical World Model for Efficient Real-World Interactions

InternW0. A Foundational Physical World Model for Efficient Real-World Interactions. By Jisong Cai and twenty-five others, from the Shanghai AI Laboratory. Published on the twenty-third of September. No code repository is linked to it.

Physical intelligence requires more than predicting how the world may evolve. The predictions have to remain actionable while the world keeps changing. A forecast that arrives after the scene has moved on is no help to a robot that has to act now. The design problem is to keep prediction useful at the speed that control requires.

InternW0 is the first instantiation of the InternW physical world model series from Shanghai AI Laboratory. It is built around omnimodal interfaces, asynchronous processing at multiple frequencies, and local physical modeling under partial observations and external influences. It learns future visual dynamics and continuous robot control jointly, through an asymmetric video and action architecture with flow matching. A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. The split mirrors the problem. Looking ahead is expensive and can run slower. Acting has to be quick, and is kept light.

Rather than regenerating the future for every action update, InternW0 reuses its layerwise K and V, and adapts them to newly observed states through observation- conditioned context routing. Domain-specific interfaces and soft prompts support heterogeneous embodiments. Contact-aware post-training brings in force and tactile signals, for contact-rich manipulation. The reuse is what lets the fast expert keep pace. The expensive predictive context is carried forward and adjusted to the latest observation, instead of being produced again from scratch each time the action changes.

Training uses about seven thousand two hundred hours of heterogeneous robot and egocentric data. That includes EgoLab, a real-laboratory egocentric dataset of two hundred seventy-five hours. Evaluation spans simulation benchmarks and real- world scientific tasks. One is a fifteen-stage workflow for synthesizing a metal-organic framework. Another is five-stage dexterous manipulation for general-purpose quantitative pipetting, aware of contact and force. These are long, multi-step procedures from a real laboratory, the kind of setting where a robot's model of the world has to stay current from one stage to the next.

The emphasis throughout is on keeping prediction and action in step. A slow expert looks further ahead, a fast one acts, and earlier context is adapted rather than rebuilt. The authors present this as advancing scalable, asynchronous physical world models that are native to science.