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

The paper introduces InternW0, a physical world model featuring omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling. InternW0 jointly learns future visual dynamics and robot control through an asymmetric video-action architecture with flow matching. A high-capacity video expert handles long-horizon context, while a lightweight action expert operates at a faster timescale, reusing layerwise K/V via observation-conditioned routing. Trained on approximately 7,200 hours of heterogeneous data, InternW0 excels in simulation benchmarks and real-world scientific manipulation workflows.

Trained on approximately 7,200 hours of heterogeneous robot and egocentric data.

Includes EgoLab, a 275-hour real-laboratory egocentric dataset.

Demonstrates a 15-stage metal-organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation.