Shanghai AI Laboratory has introduced InternW0, described as the first instantiation of its InternW physical world model series. The work targets a central challenge in physical intelligence: predictions about the world must remain actionable even as conditions continue to change.
The model is built around three design pillars: omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and external influences. InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video-action architecture with flow matching.
In this setup, a high-capacity video expert supplies longer-horizon predictive context, while a lightweight action expert runs at a faster timescale. Rather than regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states via observation-conditioned context routing.
To handle varied hardware, the model uses domain-specific interfaces and soft prompts that support heterogeneous embodiments. A contact-aware post-training stage incorporates force and tactile signals for contact-rich manipulation.
Training drew on roughly 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. Evaluation covered simulation benchmarks and real-world scientific tasks, such as a 15-stage metal-organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting.
The authors position these results as a step toward scalable, asynchronous, and science-native physical world models for efficient real-world interaction.