TL;DR. Behavior cloning (BC) finetuning slowly overwrites the pretrained VLM representations a VLA relies on for visual and semantic generalization. Anchor-Align adds two objectives to standard BC: Vision-Language Anchoring (layer-wise distillation from a frozen VLM copy, which keeps the representations from drifting) and Language-Action Alignment (each action target becomes a discrete motion-direction label, and language and action are trained jointly on the same robot observation).</p>\n<p>Anchor-Align achieves stronger OOD generalization. In simulation, it improves robustness to unseen spatial rearrangements, camera perturbations, and other shifts absent from training on LIBERO-PRO and LIBERO-Plus, plus long-horizon control on CALVIN. On a physical xArm7, real-robot success under unseen spatial rearrangements, semantic perturbations, and cluttered scenes rises from 28% to 54% and from 37% to 60% across two VLA architectures.</p>\n","updatedAt":"2026-07-23T02:58:51.720Z","author":{"_id":"63c949af51b03e8e01b23a95","avatarUrl":"/avatars/280b47d07977db9c4b3f91e9ab845af7.svg","fullname":"Dwip Dalal","name":"Dwipz","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.829502284526825},"editors":["Dwipz"],"editorAvatarUrls":["/avatars/280b47d07977db9c4b3f91e9ab845af7.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.13429","authors":[{"_id":"6a59713c6c2e371e6ca38006","user":{"_id":"63c949af51b03e8e01b23a95","avatarUrl":"/avatars/280b47d07977db9c4b3f91e9ab845af7.svg","isPro":false,"fullname":"Dwip Dalal","user":"Dwipz","type":"user","name":"Dwipz"},"name":"Dwip Dalal","status":"claimed_verified","statusLastChangedAt":"2026-07-21T09:28:57.087Z","hidden":false},{"_id":"6a59713c6c2e371e6ca38007","name":"Shivansh Patel","hidden":false},{"_id":"6a59713c6c2e371e6ca38008","name":"Chahit Jain","hidden":false},{"_id":"6a59713c6c2e371e6ca38009","name":"Jeonghwan Kim","hidden":false},{"_id":"6a59713c6c2e371e6ca3800a","name":"Utkarsh Mishra","hidden":false},{"_id":"6a59713c6c2e371e6ca3800b","name":"Alex Baratian","hidden":false},{"_id":"6a59713c6c2e371e6ca3800c","name":"Hyeonjeong Ha","hidden":false},{"_id":"6a59713c6c2e371e6ca3800d","name":"Heng Ji","hidden":false},{"_id":"6a59713c6c2e371e6ca3800e","name":"Svetlana Lazebnik","hidden":false},{"_id":"6a59713c6c2e371e6ca3800f","name":"Unnat Jain","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/63c949af51b03e8e01b23a95/PJ0xxVWKIgmWDaLjjP3LS.mp4"],"publishedAt":"2026-07-15T00:00:00.000Z","submittedOnDailyAt":"2026-07-23T00:00:00.000Z","title":"Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment","submittedOnDailyBy":{"_id":"63c949af51b03e8e01b23a95","avatarUrl":"/avatars/280b47d07977db9c4b3f91e9ab845af7.svg","isPro":false,"fullname":"Dwip Dalal","user":"Dwipz","type":"user","name":"Dwipz"},"summary":"Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, does not prevent this; it applies language and action losses to separate observations, leaving VLAs with language-action misalignment that standard manipulation benchmarks do not expose. We propose Anchor-Align, which augments BC with two objectives: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy to prevent this drift, while Language-Action Alignment converts each action target into a discrete motion-direction label and jointly trains language and action prediction on the same robot observation. On a physical xArm7 robot, across two widely used VLA architectures, Anchor-Align improves real-robot success on both (28% to 54% and 37% to 60%). At scale in simulation, we demonstrate consistent improvements on OOD perturbations, perceptual robustness, and long-horizon control across LIBERO-PRO, LIBERO-Plus, and CALVIN, respectively, suggesting that preserving pretrained representations and effective action learning are not fundamentally at odds. Project page: anchoralignvla.github.io","upvotes":3,"discussionId":"6a59713c6c2e371e6ca38010","projectPage":"https://anchoralignvla.github.io/","githubRepo":"https://github.com/dwipddalal/Anchor-Align","githubRepoAddedBy":"user","githubStars":11},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"63c949af51b03e8e01b23a95","avatarUrl":"/avatars/280b47d07977db9c4b3f91e9ab845af7.svg","isPro":false,"fullname":"Dwip Dalal","user":"Dwipz","type":"user"},{"_id":"6998202e3fd7ac0571aa5c28","avatarUrl":"/avatars/711ab074fb35a817d7ed301f51fcdabc.svg","isPro":false,"fullname":"Mishra","user":"Utkarsh-Mishra","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.13429.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2607.13429 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.13429 in a Space README.md to link it from this page.

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