Hugging Face Daily Papers · · 3 min read

Light-WAM: Efficient World Action Models with State-Fusion Action Decoding

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

\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/66a9cb830751ea3455c0618c/6PBPnlUKkvPCpyuGO4oBP.png\" alt=\"Light-WAM architecture\" width=\"515\">\n</p>\nLight-WAM: Efficient World Action Models with State-Fusion Action Decoding","html":"<p align=\"left\">\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/66a9cb830751ea3455c0618c/6PBPnlUKkvPCpyuGO4oBP.png\" alt=\"Light-WAM architecture\" width=\"515\">\n</p>\nLight-WAM: Efficient World Action Models with State-Fusion Action Decoding","updatedAt":"2026-06-09T14:28:36.320Z","author":{"_id":"66a9cb830751ea3455c0618c","avatarUrl":"/avatars/4e09d66c2ce919ec948e7b1cf1138f80.svg","fullname":"SII-L1ziang","name":"l1ziang","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.49749231338500977},"editors":["l1ziang"],"editorAvatarUrls":["/avatars/4e09d66c2ce919ec948e7b1cf1138f80.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.08242","authors":[{"_id":"6a27ab606dde1c5ef75bd16e","user":{"_id":"66a9cb830751ea3455c0618c","avatarUrl":"/avatars/4e09d66c2ce919ec948e7b1cf1138f80.svg","isPro":false,"fullname":"SII-L1ziang","user":"l1ziang","type":"user","name":"l1ziang"},"name":"Ziang Li","status":"claimed_verified","statusLastChangedAt":"2026-06-09T12:41:13.106Z","hidden":false},{"_id":"6a27ab606dde1c5ef75bd16f","name":"Dongzhou Cheng","hidden":false},{"_id":"6a27ab606dde1c5ef75bd170","name":"Yibin Wang","hidden":false},{"_id":"6a27ab606dde1c5ef75bd171","name":"Shiyue Wang","hidden":false},{"_id":"6a27ab606dde1c5ef75bd172","name":"Xiaoyang Xu","hidden":false},{"_id":"6a27ab606dde1c5ef75bd173","name":"Lingxuan Weng","hidden":false},{"_id":"6a27ab606dde1c5ef75bd174","name":"Juan Wang","hidden":false},{"_id":"6a27ab606dde1c5ef75bd175","name":"Jiaqi Wang","hidden":false}],"publishedAt":"2026-06-06T00:00:00.000Z","submittedOnDailyAt":"2026-06-09T00:00:00.000Z","title":"Light-WAM: Efficient World Action Models with State-Fusion Action Decoding","submittedOnDailyBy":{"_id":"66a9cb830751ea3455c0618c","avatarUrl":"/avatars/4e09d66c2ce919ec948e7b1cf1138f80.svg","isPro":false,"fullname":"SII-L1ziang","user":"l1ziang","type":"user","name":"l1ziang"},"summary":"World Action Models (WAMs) extend robot policy learning by incorporating future prediction as an additional training objective, encouraging the policy to encode task-relevant temporal structure in its representations. Current WAMs often rely on large-scale generative architectures that incur high training costs and inference latency, making them difficult to deploy as efficient closed-loop policies. We propose Light-WAM, a lightweight World Action Model for efficient robot manipulation. Specifically, it is built with a compact video backbone and performs future-video supervision in a downsampled latent space, reducing the cost of video co-training while retaining its benefits for representation learning. For action prediction, Light-WAM introduces the StateFusionActionExpert, which reads adapted states from multiple backbone layers, fuses them through learned-query pooling, and directly predicts action chunks in a single forward pass. This design provides an efficient interface between video backbone representations and robot actions, avoiding the need for heavy generative action experts. Experiments demonstrate that Light-WAM maintains strong performance on LIBERO and achieves usable multi-task performance on RoboTwin 2.0, while using only 0.44B trainable parameters. It also achieves 72.03ms inference latency with 4.1GiB peak GPU memory and improved training throughput.","upvotes":7,"discussionId":"6a27ab606dde1c5ef75bd176","githubRepo":"https://github.com/L1ziang/Light-WAM","githubRepoAddedBy":"user","ai_summary":"Light-WAM is a lightweight world action model for robot manipulation that uses a compact video backbone and downsampled latent space for efficient future-video supervision, combined with a StateFusionActionExpert for direct action prediction.","ai_keywords":["World Action Models","robot policy learning","future prediction","generative architectures","video backbone","downsampled latent space","StateFusionActionExpert","learned-query pooling","action chunks","inference latency","training throughput"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":25,"organization":{"_id":"6350bdf559bfa9a85d42fea4","name":"WuhanUniversity","fullname":"Wuhan Univeristy","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6350bd20aaee2ec378dfe506/Bu1Fwz4dAwjwzWv-vZ2FN.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66a9cb830751ea3455c0618c","avatarUrl":"/avatars/4e09d66c2ce919ec948e7b1cf1138f80.svg","isPro":false,"fullname":"SII-L1ziang","user":"l1ziang","type":"user"},{"_id":"654c6845bac6e6e49895a5b5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/KXQaAxulqr8jNBSpEaYM4.png","isPro":false,"fullname":"SII-Yibin Wang","user":"CodeGoat24","type":"user"},{"_id":"6842f3fa0ffbfcaee7d5f9f4","avatarUrl":"/avatars/c81cb1be2683e75e8da5354619be8812.svg","isPro":false,"fullname":"Willms Mihayo","user":"iMihayo","type":"user"},{"_id":"68f850d3fdcd856c15b23f66","avatarUrl":"/avatars/7caf934d09cddb7009e4ab87a7b9daaf.svg","isPro":false,"fullname":"Shiyue Wang","user":"wangsh1yue","type":"user"},{"_id":"658c19a2539b68adc77db11a","avatarUrl":"/avatars/af4c7ecf96253f696dec6c50f0326935.svg","isPro":false,"fullname":"Jack Smith","user":"x1a0yue","type":"user"},{"_id":"6a282c4f4b8574890c1eb1e6","avatarUrl":"/avatars/d33b8bdac16053c00683441a26503729.svg","isPro":false,"fullname":"ncca402","user":"ncca402","type":"user"},{"_id":"63849307b4d5a5b7f43f59fc","avatarUrl":"/avatars/7171d2652a916a2c13674cb717e870c1.svg","isPro":false,"fullname":"anthony Saint","user":"Anthony52233","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6350bdf559bfa9a85d42fea4","name":"WuhanUniversity","fullname":"Wuhan Univeristy","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6350bd20aaee2ec378dfe506/Bu1Fwz4dAwjwzWv-vZ2FN.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.08242.md"}">
Papers
arxiv:2606.08242

Light-WAM: Efficient World Action Models with State-Fusion Action Decoding

Published on Jun 6
· Submitted by
SII-L1ziang
on Jun 9
Authors:
,
,
,
,
,
,

Abstract

Light-WAM is a lightweight world action model for robot manipulation that uses a compact video backbone and downsampled latent space for efficient future-video supervision, combined with a StateFusionActionExpert for direct action prediction.

World Action Models (WAMs) extend robot policy learning by incorporating future prediction as an additional training objective, encouraging the policy to encode task-relevant temporal structure in its representations. Current WAMs often rely on large-scale generative architectures that incur high training costs and inference latency, making them difficult to deploy as efficient closed-loop policies. We propose Light-WAM, a lightweight World Action Model for efficient robot manipulation. Specifically, it is built with a compact video backbone and performs future-video supervision in a downsampled latent space, reducing the cost of video co-training while retaining its benefits for representation learning. For action prediction, Light-WAM introduces the StateFusionActionExpert, which reads adapted states from multiple backbone layers, fuses them through learned-query pooling, and directly predicts action chunks in a single forward pass. This design provides an efficient interface between video backbone representations and robot actions, avoiding the need for heavy generative action experts. Experiments demonstrate that Light-WAM maintains strong performance on LIBERO and achieves usable multi-task performance on RoboTwin 2.0, while using only 0.44B trainable parameters. It also achieves 72.03ms inference latency with 4.1GiB peak GPU memory and improved training throughput.

Community

Paper author Paper submitter about 5 hours ago

Light-WAM architecture

Light-WAM: Efficient World Action Models with State-Fusion Action Decoding
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images

· Sign up or log in to comment

Get this paper in your agent:

hf papers read 2606.08242
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 1

Datasets citing this paper 1

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2606.08242 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.

Discussion (0)

Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.

Sign in →

No comments yet. Sign in and be the first to say something.

More from Hugging Face Daily Papers