Shanghai AI Lab trains a robot world model on 7,200 hours of data

single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC

What this means for you

Nothing to act on yet. The abstract gives no weights, code or release date, so there is nothing to run. Watch for the evaluation numbers: until they appear, the training scale and the pipetting demonstration are the authors' claims, not measurements.

Shanghai AI Laboratory has published InternW0, the first of a planned series of physical world models — systems that predict how a scene will evolve so a robot can act on that prediction. A large video model runs slowly and supplies long-horizon context; a small action model runs faster, reusing stored intermediate state rather than regenerating video for every control update. It was trained on roughly 7,200 hours of robot and first-person video, including a 275-hour laboratory dataset. Evaluation covers simulation and two real workflows: 15-stage metal–organic framework synthesis and five-stage pipetting. The preprint gives no scores, weights, licence or release date.

Key facts

  • ·InternW0 is the first model in Shanghai AI Laboratory's InternW series of physical world models, published as an arXiv preprint dated 22 September 2026. source
  • ·It was trained on approximately 7,200 hours of heterogeneous robot and egocentric (first-person) data. source
  • ·The training set includes EgoLab, a 275-hour first-person dataset collected in a real laboratory. source
  • ·The architecture pairs a high-capacity video expert with a lightweight action expert that operates at a faster timescale. source
  • ·Reported evaluations include a 15-stage metal–organic framework synthesis workflow and 5-stage contact- and force-aware manipulation for quantitative pipetting. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire