Hugging Face Daily Papers · September 2, 2026 · 4 min read

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

We present ZimaBlue, framing video scaling as a practical route toward generalizable World Action Models, and provide empirical evidence that scaling video pre-training significantly improves zero-shot generalization and strengthens performance on challenging manipulation benchmarks.</p>\n","updatedAt":"2026-09-02T02:28:41.335Z","author":{"_id":"63660fc964bcbbd03e2d227e","avatarUrl":"/avatars/4d7ab3a8cf27a53093e2fcc5f9ee84fb.svg","fullname":"Vincent Lee","name":"fenglinglwb","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8538641929626465},"editors":["fenglinglwb"],"editorAvatarUrls":["/avatars/4d7ab3a8cf27a53093e2fcc5f9ee84fb.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.00188","authors":[{"_id":"6a9789d7fe3c2f89286c3950","name":"Xionghao Wu","hidden":false},{"_id":"6a9789d7fe3c2f89286c3951","name":"Yijun Yang","hidden":false},{"_id":"6a9789d7fe3c2f89286c3952","name":"Shiyang Zhou","hidden":false},{"_id":"6a9789d7fe3c2f89286c3953","name":"Haoze Sun","hidden":false},{"_id":"6a9789d7fe3c2f89286c3954","name":"Jianhui Liu","hidden":false},{"_id":"6a9789d7fe3c2f89286c3955","name":"Songsong Yu","hidden":false},{"_id":"6a9789d7fe3c2f89286c3956","name":"Jiyao Zhang","hidden":false},{"_id":"6a9789d7fe3c2f89286c3957","name":"Wenbo Li","hidden":false},{"_id":"6a9789d7fe3c2f89286c3958","name":"Bo Wang","hidden":false},{"_id":"6a9789d7fe3c2f89286c3959","name":"Guoqing Ma","hidden":false},{"_id":"6a9789d7fe3c2f89286c395a","name":"Lin Song","hidden":false},{"_id":"6a9789d7fe3c2f89286c395b","name":"Renjie Liao","hidden":false},{"_id":"6a9789d7fe3c2f89286c395c","name":"Shenghe Zheng","hidden":false},{"_id":"6a9789d7fe3c2f89286c395d","name":"Wei Tang","hidden":false},{"_id":"6a9789d7fe3c2f89286c395e","name":"Xiaojuan Qi","hidden":false},{"_id":"6a9789d7fe3c2f89286c395f","name":"Yanwei Li","hidden":false},{"_id":"6a9789d7fe3c2f89286c3960","name":"Yuan Zhang","hidden":false},{"_id":"6a9789d7fe3c2f89286c3961","name":"Zhuotao Tian","hidden":false},{"_id":"6a9789d7fe3c2f89286c3962","name":"Haoyang Huang","hidden":false},{"_id":"6a9789d7fe3c2f89286c3963","name":"Nan Duan","hidden":false}],"publishedAt":"2026-08-31T00:00:00.000Z","submittedOnDailyAt":"2026-09-02T00:00:00.000Z","title":"ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training","submittedOnDailyBy":{"_id":"63660fc964bcbbd03e2d227e","avatarUrl":"/avatars/4d7ab3a8cf27a53093e2fcc5f9ee84fb.svg","isPro":false,"fullname":"Vincent Lee","user":"fenglinglwb","type":"user","name":"fenglinglwb"},"summary":"Robotic manipulation faces a fundamental scaling challenge: robust generalization demands broad physical experience, yet action-labeled robot trajectories are expensive to collect and inherently limited in diversity. Egocentric videos offer a far more scalable source of embodied experience, capturing object interactions, contact dynamics, tool use, and long-horizon behaviors across diverse environments. The central challenge is how to convert this abundant but action-free experience into effective robot control. We introduce ZimaBlue, a scalable framework for learning generalizable World Action Models (WAMs) from large-scale video. ZimaBlue follows a three-stage training curriculum: it first performs causal embodied video pre-training on large-scale human and robot egocentric videos, then grounds the learned visual dynamics in heterogeneous robot trajectories through video-action mid-training with a unified action representation, and finally specializes the model to a target robot for deployment. To make generative WAMs practical for real-time control, ZimaBluefurther adopts an asynchronous Slow-Fast dual-system architecture, where a high-capacity Slow world model provides generalizable spatiotemporal representations and a lightweight Fast branch enables 30 Hz action prediction on NVIDIA RTX 4090. On real-robot zero-shot evaluations, scaling from target-robot data alone to over 120,000 hours of embodied video improves success from 36.1% to 77.8%. ZimaBlue further delivers strong performance across multiple benchmarks, with particularly pronounced gains on unseen tasks.","upvotes":35,"discussionId":"6a9789d8fe3c2f89286c3964","projectPage":"https://zimablue-wam.github.io/","githubRepo":"https://github.com/ZimaBlue-WAM/ZimaBlue","githubRepoAddedBy":"user","ai_summary":"ZimaBlue learns generalizable world action models from large-scale egocentric video via a three-stage curriculum and a slow-fast architecture, substantially improving zero-shot robotic manipulation.","ai_keywords":["World Action Models","causal embodied video pre-training","video-action mid-training","unified action representation","Slow-Fast dual-system architecture","asynchronous inference","zero-shot evaluation"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":3,"organization":{"_id":"69d76efa4c66497b54730199","name":"JoyFutureAcademy","fullname":"Joy Future Academy","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/63660fc964bcbbd03e2d227e/VH1Qey0Jb40CjBVr4M128.webp"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"63660fc964bcbbd03e2d227e","avatarUrl":"/avatars/4d7ab3a8cf27a53093e2fcc5f9ee84fb.svg","isPro":false,"fullname":"Vincent Lee","user":"fenglinglwb","type":"user"},{"_id":"66e1557c75d9226ba13b38d2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/vAwxAOynnsF2iqYeO3UBl.png","isPro":false,"fullname":"shenghe zheng","user":"desimfj","type":"user"},{"_id":"6784a033eb390961206f542e","avatarUrl":"/avatars/d2b5bd89535618a8ca383a3dd2537055.svg","isPro":false,"fullname":"Jiyao Zhang","user":"JiyaoZhang","type":"user"},{"_id":"6527b7280ae663e384eb8499","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6527b7280ae663e384eb8499/73yF3eu2cUx7jVZrhXnXx.jpeg","isPro":false,"fullname":"Senqiao Yang","user":"Senqiao","type":"user"},{"_id":"6447756ae6161a1f32e1c734","avatarUrl":"/avatars/27fabb6e85b7405c1668201ce7cd51aa.svg","isPro":false,"fullname":"bowang","user":"bwang3579","type":"user"},{"_id":"64b8e6c28c12f0008b956257","avatarUrl":"/avatars/8e8a3925b4a41837392fa7bae7671163.svg","isPro":false,"fullname":"Wei Tang","user":"weitang","type":"user"},{"_id":"64b6523a16052bf82c697a3f","avatarUrl":"/avatars/a6b48f55ae6574b17495b673a48f0e25.svg","isPro":false,"fullname":"Clausy","user":"umu","type":"user"},{"_id":"673c3d68a90d746ff6b8ba64","avatarUrl":"/avatars/9fcd3f915e6cbacc304bd1f54d0223a2.svg","isPro":false,"fullname":"Yijun Yang","user":"scott-yjyang","type":"user"},{"_id":"64b3ef335e1230a79fad01db","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/mO95xQ-qQaQbRw5e8sy5L.jpeg","isPro":false,"fullname":"zhangyuan","user":"Zianer","type":"user"},{"_id":"686f974ccc59651f3ae60861","avatarUrl":"/avatars/9dfc06d52e124cda90ba47f280476a94.svg","isPro":true,"fullname":"Zhicheng Zhang","user":"nku-zhichengzhang","type":"user"},{"_id":"678dfece148f7a067dad9e80","avatarUrl":"/avatars/3a133a32c66ed3a57a7b3397e9402689.svg","isPro":false,"fullname":"金","user":"bigqingqingqingqing","type":"user"},{"_id":"6496f5754a3c31df8e3139f6","avatarUrl":"/avatars/cf789d1986f976373c82b2976df4542a.svg","isPro":false,"fullname":"Zhongwei Zhang","user":"zzwustc","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"69d76efa4c66497b54730199","name":"JoyFutureAcademy","fullname":"Joy Future Academy","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/63660fc964bcbbd03e2d227e/VH1Qey0Jb40CjBVr4M128.webp"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.00188.md","query":{}}">

Papers

arxiv:2609.00188

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

Published on Aug 31

· Submitted by

Vincent Lee on Sep 2

Joy Future Academy

Upvote

Authors:

Abstract

ZimaBlue learns generalizable world action models from large-scale egocentric video via a three-stage curriculum and a slow-fast architecture, substantially improving zero-shot robotic manipulation.

Generated by thinkingmachines/Inkling-Small

Robotic manipulation faces a fundamental scaling challenge: robust generalization demands broad physical experience, yet action-labeled robot trajectories are expensive to collect and inherently limited in diversity. Egocentric videos offer a far more scalable source of embodied experience, capturing object interactions, contact dynamics, tool use, and long-horizon behaviors across diverse environments. The central challenge is how to convert this abundant but action-free experience into effective robot control. We introduce ZimaBlue, a scalable framework for learning generalizable World Action Models (WAMs) from large-scale video. ZimaBlue follows a three-stage training curriculum: it first performs causal embodied video pre-training on large-scale human and robot egocentric videos, then grounds the learned visual dynamics in heterogeneous robot trajectories through video-action mid-training with a unified action representation, and finally specializes the model to a target robot for deployment. To make generative WAMs practical for real-time control, ZimaBluefurther adopts an asynchronous Slow-Fast dual-system architecture, where a high-capacity Slow world model provides generalizable spatiotemporal representations and a lightweight Fast branch enables 30 Hz action prediction on NVIDIA RTX 4090. On real-robot zero-shot evaluations, scaling from target-robot data alone to over 120,000 hours of embodied video improves success from 36.1% to 77.8%. ZimaBlue further delivers strong performance across multiple benchmarks, with particularly pronounced gains on unseen tasks.