InternW0: A Foundational Physical World Model for Efficient Real-World Interactions
Jisong Cai, Yao Mu, Ganlin Yang, Zhe Cao, Zhangzheng Tu, Xing Gao, Kailin Li, Xinyu Zhan, Lixin Yang, Yangkun Zhu, Haoxiang Ma, Ming Zhou, Qiaojun Yu, Yufei Xue, Liqun He, Yifei Yao, Yifan Zhu, Long Ling, Bingqi Jiang, Haoyu Guo, Xueyue Zhu, Bowen Zhou, Bin Zhao, Tianfan Xue, Chunhua Shen, Weinan Zhang
cs.RO, cs.AI
2026-09-23
Shanghai AI Lab’s InternW0 jointly trains video and action experts on ~7,200 hours, hits 98.6% on LIBERO, and reaches 68.4% progress on a 15-stage real MOF workflow.
A world action model has to bind two jobs: how the scene will look next, and which motor command to send. Most current WAMs run expensive video prediction in lockstep with fast action generation, and they struggle to absorb heterogeneous robots plus force or touch. The world keeps moving. A forecast that arrives too late cannot change the next command.
InternW0 is the first model in Shanghai AI Laboratory’s InternW series. Three constraints sit on top: omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and outside disturbances.
The backbone is a mixture of transformers. A high-capacity video expert predicts future frames; a lighter action expert emits continuous control on a faster clock. Both train with flow matching. Vision uses a frozen Wan VAE and DINOv3; language uses precomputed text embeddings. The video expert never sees robot action tokens, so the predictive representation is shared across embodiments. The action expert reads video context, the latest observation, proprioception, and domain interfaces.
The efficiency trick is reuse. The video expert writes a slower-timescale plan and caches layerwise K/V. For each action chunk, a chunk K/V editor uses the new observation as query: retrieve from the cached plan, route the retrieved content back onto video-token positions, predict residual ΔK/ΔV, and add them through a learned gate. The underlying video context stays put; each chunk sees an edited view conditioned on what just happened. Residual projections start at zero, so the editor begins as identity.
Cross-embodiment training uses domain soft prompts and a 37-dimensional unified action/state interface, with validity masks for missing channels. Pretraining combines video flow matching with a joint action loss when labels exist. A second independent-denoising video stream supplies condition K/V that is not detached, so the action loss shapes the video expert. Contact-rich post-training appends force and tactile history on the action side and jointly predicts future 6-D wrenches.
The corpus is about 7,234 hours, 811,969 episodes, 25 domains: InternData-A1 (3,494 h sim), AgibotWorld (2,620 h), plus RoboCOIN, Galaxea, MolmoAct, RoboDojo, and EgoLab, a 275-hour in-house egocentric wet-lab set. Domains are sampled with square-root weights on effective anchors.
LIBERO averages 98.6% (Spatial 99.4, Long 97.0), a hair above LingBot-VA at 98.5. RoboTwin 2.0 Full averages 93.12 (clean 93.20 / randomized 93.04), below ABot-M0.5 at 94.10 and above AHA-WAM at 92.80. Clean2Random is the sharper test: fine-tune only on clean demos, then hit 68.0% on randomized scenes, 75.60 average, 8.3 points above GigaBrain-0.7. In a small-data ablation, adding 50 h of EgoLab on top of 50 h of robot data lifts randomized success from 13.73% to 21.97%.
Real-robot, 15 trials each:
| Task | π0.5 | Fast-WAM | InternW0 |
| Make sandwich | 73.3 | 40.0 | 73.3 |
| Industrial parts | 53.1 | 50.6 | 82.7 |
| Sort tubes | 86.7 | 66.7 | 88.9 |
| 15-stage MOF progress | 50.2 | 10.7 | 68.4 |
| 5-stage pipetting progress | 46.7 | 18.7 | 65.3 |
On MOF, π0.5 falls to 46.7% at step 8 and to 6.7% from step 11; InternW0 holds 66.7% through step 12 and dies on the stirrer buttons (26.7%). Contact-heavy pipetting subtasks beat both baselines; pickup-and-reorient is 86.7% against π0.5’s 93.3%. Critical-path action latency is 60.73 ms, about 16.47 Hz, 3.13× faster than Fast-WAM.
This is a WAM pushed from simulation leaderboards toward multi-step lab work. Cached K/V plus the chunk editor take video prediction off the control critical path; force post-training is aimed at contact. Teams already running π0 or a WAM should steal the editor and the 37-D interface before trying to clone 7,200 hours of data. LIBERO and RoboTwin Full are already packed near 98%/93%. The gap that still moves is Clean2Random and wet-lab progress.
The roadmap asks for touch, geometry, hierarchical time, and a 4B then sub-1B model, but the paper never states InternW0’s size. RoboTwin Full still trails ABot-M0.5. Clean-to-random still drops 15.2 points. Real-robot N is 15; the MOF button stage is 26.7%, so millimeter precision is unsolved. Force and tactile enter only in post-training. EgoLab has video labels, no actions. The InkStone plus Intern-S2 science loop is described as a direction; there is no end-to-end success rate for “reasoner writes a protocol, InternW0 runs it.”