Shanghai AI Laboratory's Physical Intelligence Team submitted InternW0 on Sept. 23, 2026, as the first instantiation of its InternW physical world-model series. The work couples future visual prediction with continuous robot control so predictions stay usable while the world keeps changing under partial observations and contact.
In plain terms, InternW0 is not a redeploy of already-live OpenWAM or LingBot-World-2 product landings. It is an asynchronous duplex world-action model: a high-capacity video expert plans on a slower clock, a lighter action expert updates faster, and observation-conditioned layerwise K/V routing adapts cached video context to newly seen states instead of regenerating a full video plan on every control step.
Authors report the clearest listed out-of-domain win on RoboTwin 2.0 Clean2Random, where clean-only post-training yields 83.20% clean / 68.0% randomized / 75.60% average, +8.30 percentage points over GigaBrain-0.7's 67.3% average. On RoboTwin Full, InternW0 averages 93.12%, which trails ABot-M0.5 at 94.10%. On LIBERO Full post-training the average is 98.6% (Spatial / Object / Goal / Long 99.4 / 99.4 / 98.6 / 97.0), only 0.1 points above LingBot-VA's 98.5 near the listed ceiling.
On five real-world tasks over 15 trials each, authors report MoF Experiments ordered-subtask progress of 68.4% and Quantitative Pipetting progress of 65.3%, versus 50.2% / 46.7% for π0.5 and 10.7% / 18.7% for Fast-WAM under a shared VLM subtask generator. Those figures are ordered-subtask progress rates, not all-or-nothing episode success. Make Sandwich episode success ties π0.5 at 73.3%; Pick Industrial Parts object-level success reaches 82.7%. Critical-path action generation on an RTX 5090D is reported at 60.73 ms, or 16.47 Hz model-side updates and a 3.13× speedup over Fast-WAM; the number covers observation encoding, context routing, and action denoising, and excludes asynchronous Video-DiT plan generation, so it is not full end-to-end system latency.
Pretraining spans about 7,233.5 hours and 811,969 episodes across 25 domains, including EgoLab's 275.4 hours of real wet-lab egocentric video on a video-only pathway. Contact-aware post-training can add force and tactile history and jointly predict wrenches with actions for contact-rich tool use such as force-aware pipetting.
Hugging Face Papers listed the preprint on its Sept. 24, 2026 Daily Papers feed, alphaXiv indexes the same arXiv record, and the InternRobotics project page hosts architecture figures plus real-world demo videos for sandwich, parts, tubes, MoF, and pipetting. What remains open is independent replication of the Clean2Random, wet-lab progress, and critical-path latency claims outside the authors' protocol, and public Hugging Face or ModelScope checkpoints, which the project page still marks as pending. Lab claims, community indexes, and demos should be weighed separately until those points are settled.