InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
Xingyu Miao, Zizun Li, Baole Fang, Kaiwen Song, Tenghui Wang, Hanxue Zhang, Yating Wang, Xudong Li, Yuping He, Xueyuan Wei, Chao Gao, Xijie Yang, Yingxiang Xu, Kerui Ren, Wenqi Guo, Jianjun Zhou, Xinzhe Wang, Weiguang Zhao, Ni Yang, Zetao Cai, Yufei Xue, Hengjie Li, Zeyu He, Yuanzhen Zhou, Rong Fu, Jianyang Zhang, Siwei Cui, Fuxian Huang, Yunsong Zhou, Xing Gao, Yifei Yao, Qiaojun Yu, Kailin Li, Ming Zhou, Mu Huang, Xinyue Li, Wenze Cui, Bingqi Jiang, Xueyue Zhu, Junting Dong, Haoyu Guo, Tao Lu, Mulin Yu, Bowen Zhou, Bin Zhao, Tianfan Xue, Weinan Zhang, Chunhua Shen
cs.RO, cs.CV
2026-09-25
InternW0-Δ, pretrained on a large open heterogeneous corpus, reaches 92.8% on LIBERO-Plus under shift and predicts actions without future-video rollout.
World Action Models jointly learn visual dynamics and robot actions. Video pretraining knows how scenes evolve, but predicting the next frame does not yield a control command. Many WAMs also roll out future video at inference, which is too slow for a controller. Open embodied datasets further disagree on embodiment, action space, cameras, and frame rate, so naively mixing them assigns different physics to the same action index.
InternW0-Δ, from Shanghai AI Lab, turns predictive dynamics into features the action head can read, samples no future video at inference, and trains on a canonical state-action layout over a claimed 20K+ hours of open heterogeneous data.
The backbone is a 30-block directed Mixture-of-Transformers: a Wan2.2-TI2V-5B video expert and an ActionDiT action expert. A frozen VLM feeds scene semantics to the action head; T5 keeps the video expert's original language conditioning. Visual context is sparse memory: an episode anchor, a recent frame from before the last action chunk, and the current frame.
Causal Imprint is the interface. Learnable tokens see only recent and current observations. Training supervises them with adjacent differences of clean future video latents, plus cosine alignment to deeper future features of the video expert. The attention mask blocks these tokens and the action tokens from reading realized futures. At inference the video expert prefills once, caches K/V for A/R/C/Δ, and only the action stream is denoised with flow matching. Action chunks are length 32 in an 80-D canonical action space.
A frozen Track4World teacher distills a 1430-D clip descriptor during training (cached offline). Teacher and student branches are dropped at inference. Robot demos, UMI, egocentric human video, and Ego2Robot are mapped into fixed semantic slots: joint actions as absolute targets, end-effector actions as relative increments. After filtering, Table 2 lists about 11.3K hours of robot data, 2,075 of UMI, and 4,061 of egocentric video. Joint pretraining runs about 245K steps over 14 days, then embodiment-specific post-training. Training-time RTC keeps the chunk-boundary action jump ratio at 1.12, versus 8.94 for VJP and 7.13 for a hard prefix. Controller-observed round trip on an RTX 5090 is 152.8 ms, 5.11× the unoptimized runtime.
Post-training uses unperturbed demos only; evaluation follows official shift protocols.
| Benchmark | InternW0-Δ | Strongest listed baseline |
| LIBERO-Plus success | 92.8% | Qwen-RobotManip-Context 91.4%; best prior WAM Being-H0.7 84.8% |
| RoboTwin Clean2Random | 71.9% (Clean2Clean 90.0%) | Qwen 69.4%; OpenWAM-α 48.7% |
| EBench score / success | 66.0 / 49.2% | OpenWAM-α 64.7 / 49.4% |
| RoboDojo mean success | 23.91% | OpenWAM-α 11.92% |
On LIBERO-Plus, robot perturbation is 91.1% versus a previous best 87.4%. On EBench long-horizon tasks, success 49.4 and score 76.5 lead the table; overall success is a hair below OpenWAM-α at 49.4%, while the score leads. RoboDojo stays hard: 23.91% nearly doubles the best prior WAM and is still a low absolute number.
From-scratch LIBERO-Plus ablations: 49.59% baseline → 53.47% with sparse memory → 69.08% with RynnBrain1.1-2B → 70.80% with Causal Imprint → 76.45% with the alignment loss → 78.37% with 4D distillation. The headline 92.8% is pretrained then post-trained; do not mix the two ladders.
Four real platforms (gripper AC-One / Arx5, dexterous Franka+XHand / TianJi+Wuji) share one pretrained checkpoint. On AC-One, toast bread goes 4/20 → 19/20 and Luminol 0/20 → 19/20 with pretraining. The paper does not publish a full success table for all eleven real tasks.
The practical WAM problem is expensive video priors plus even more expensive future rollouts. Causal Imprint stamps "what will change" into currently observed features so the action head does not imagine video. The 80-D canonical action layout and filtered indices are the parts other labs can actually reuse. The LIBERO-Plus gap over the best VLA is 1.4 points, a small step. The pull-away versus prior WAMs on RoboTwin randomization and RoboDojo is larger.
Egocentric human data is a big slice of pretraining, and source, mix ratio, and supervision form are not swept. Agent-in-the-loop control is a small correction study, not a planner. EBench overall success does not beat OpenWAM-α. RoboDojo at 23.91% still fails memory and open-instruction tasks often. Real-robot reporting is incomplete; the pretraining ablation covers two AC-One tasks. The text claims 20K+ processed hours; summing Table 2's filtered buckets is about 17.4K. Use the table.