We introduce VLA-Corrector, a lightweight detect-and-correct inference framework for action-chunked Vision-Language-Action policies.</p>\n<p><video src=\"https://cdn-uploads.huggingface.co/production/uploads/6485bd278d14bcd5cdbb7c8d/AoHHQtSGsY28pGj6Oqfi4.mp4\" controls=\"\" class=\"max-w-full!\"></video></p>\n\n\n<p>Modern VLA policies often predict and execute action chunks to reduce policy-call frequency and improve temporal smoothness. However, this fixed-horizon execution creates an open-loop blind spot: when an object slips, the robot pose drifts, or the scene changes during execution, the policy may continue executing stale actions before querying the model again.</p>\n<p>VLA-Corrector addresses this issue without retraining or modifying the VLA backbone. It adds a lightweight external correction pathway that monitors latent visual dynamics during execution. When the observed visual evolution persistently deviates from the expected one, VLA-Corrector interrupts the current action chunk, discards stale actions, and triggers corrective replanning with Online Gradient Guidance.</p>\n<p>This turns a fixed action horizon into an adaptive one: long-horizon execution is preserved when the chunk remains reliable, while short-horizon corrective behavior is activated when execution starts to drift.</p>\n<p>Across MetaWorld, LIBERO, and real-world AgileX PiPER experiments, VLA-Corrector improves robustness and success-per-call efficiency across multiple VLA backbones, showing that small inference-time modules can provide meaningful gains for reliable robot manipulation.</p>\n<p>Project Page: <a href=\"https://zju-omniai.github.io/vla-corrector/\" rel=\"nofollow\">https://zju-omniai.github.io/vla-corrector/</a></p>\n<p>GitHub: <a href=\"https://github.com/ZJU-OmniAI/vla-corrector\" rel=\"nofollow\">https://github.com/ZJU-OmniAI/vla-corrector</a></p>\n<p>Paper: <a href=\"https://arxiv.org/abs/2607.01804\" rel=\"nofollow\">https://arxiv.org/abs/2607.01804</a></p>\n<p>Contact:<br>Yi Pan: <a href=\"mailto:[email protected]\" rel=\"nofollow\">[email protected]</a><br>Wenqi Zhang: <a href=\"mailto:[email protected]\" rel=\"nofollow\">[email protected]</a></p>\n","updatedAt":"2026-07-06T15:29:36.758Z","author":{"_id":"6485bd278d14bcd5cdbb7c8d","avatarUrl":"/avatars/1427cf1a72b5db0cb263ad45885cf925.svg","fullname":"Wenqi Zhang","name":"zwq2018","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":14,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.7556464672088623},"editors":["zwq2018"],"editorAvatarUrls":["/avatars/1427cf1a72b5db0cb263ad45885cf925.svg"],"reactions":[],"isReport":false}},{"id":"6a4b98de0ef1d556bb804a49","author":{"_id":"658412f93a84a40185adaf37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg","fullname":"Aamer Mihaysi","name":"O96a","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false},"createdAt":"2026-07-06T12:00:30.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"The focus on \"predict-then-blindly-execute\" in VLA models is a critical pain point for anyone actually deploying these in production. Fixed action horizons are a convenient abstraction for researchers, but they fail the moment a robot hits a physical snag or a surface slips. The VLA-Corrector's approach to adaptive horizons is the right move—closed-loop reactivity shouldn't be an afterthought. I'm interested to see how the \"detect-and-correct\" overhead impacts real-time latency, as that's usually where these lightweight wrappers struggle. If this can maintain temporal coherence without spiking the inference budget, it's a genuine step toward deployable embodied AI.","html":"<p>The focus on \"predict-then-blindly-execute\" in VLA models is a critical pain point for anyone actually deploying these in production. Fixed action horizons are a convenient abstraction for researchers, but they fail the moment a robot hits a physical snag or a surface slips. The VLA-Corrector's approach to adaptive horizons is the right move—closed-loop reactivity shouldn't be an afterthought. I'm interested to see how the \"detect-and-correct\" overhead impacts real-time latency, as that's usually where these lightweight wrappers struggle. If this can maintain temporal coherence without spiking the inference budget, it's a genuine step toward deployable embodied AI.</p>\n","updatedAt":"2026-07-06T12:00:30.847Z","author":{"_id":"658412f93a84a40185adaf37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg","fullname":"Aamer Mihaysi","name":"O96a","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9103544354438782},"editors":["O96a"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg"],"reactions":[],"isReport":false},"replies":[{"id":"6a4bc902b57af46876fd5b9d","author":{"_id":"69367c5e39abc914b3bbb2e3","avatarUrl":"/avatars/8ed0e0147bcc9c77040ad490ae8a107c.svg","fullname":"Yi Pan","name":"Yi0304","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false},"createdAt":"2026-07-06T15:25:54.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Thanks — I completely agree that wall-clock latency is the key question. In our profiling, the full LVM+OGG version increases amortized per-step inference from 12.32 ms to 20.25 ms, so the extra cost is about +7.93 ms/step. Most of this overhead comes from OGG, since it involves gradient-based recovery and is only triggered after an interrupt.\n\nA useful takeaway is that LVM-based truncation alone already gives a strong boost, e.g., 48.70% → 60.35% on MetaWorld, with almost negligible runtime overhead compared with VLA inference, since fresh visual observations are already available during execution. We are also exploring faster recovery variants, such as fewer OGG steps, selective guidance, and distilling the recovery guidance to further reduce latency.","html":"<p>Thanks — I completely agree that wall-clock latency is the key question. In our profiling, the full LVM+OGG version increases amortized per-step inference from 12.32 ms to 20.25 ms, so the extra cost is about +7.93 ms/step. Most of this overhead comes from OGG, since it involves gradient-based recovery and is only triggered after an interrupt.</p>\n<p>A useful takeaway is that LVM-based truncation alone already gives a strong boost, e.g., 48.70% → 60.35% on MetaWorld, with almost negligible runtime overhead compared with VLA inference, since fresh visual observations are already available during execution. We are also exploring faster recovery variants, such as fewer OGG steps, selective guidance, and distilling the recovery guidance to further reduce latency.</p>\n","updatedAt":"2026-07-06T15:25:54.397Z","author":{"_id":"69367c5e39abc914b3bbb2e3","avatarUrl":"/avatars/8ed0e0147bcc9c77040ad490ae8a107c.svg","fullname":"Yi Pan","name":"Yi0304","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.942569375038147},"editors":["Yi0304"],"editorAvatarUrls":["/avatars/8ed0e0147bcc9c77040ad490ae8a107c.svg"],"reactions":[],"isReport":false,"parentCommentId":"6a4b98de0ef1d556bb804a49"}}]}],"primaryEmailConfirmed":false,"paper":{"id":"2607.01804","authors":[{"_id":"6a47f8c13daa34221c7c1e1c","name":"Yi Pan","hidden":false},{"_id":"6a47f8c13daa34221c7c1e1d","name":"Miao Pan","hidden":false},{"_id":"6a47f8c13daa34221c7c1e1e","name":"Qi Lu","hidden":false},{"_id":"6a47f8c13daa34221c7c1e1f","name":"Jiaming Huang","hidden":false},{"_id":"6a47f8c13daa34221c7c1e20","name":"Man Zhang","hidden":false},{"_id":"6a47f8c13daa34221c7c1e21","name":"Siteng Huang","hidden":false},{"_id":"6a47f8c13daa34221c7c1e22","name":"Xin Li","hidden":false},{"_id":"6a47f8c13daa34221c7c1e23","name":"Jie Zhang","hidden":false},{"_id":"6a47f8c13daa34221c7c1e24","name":"Yongliang Shen","hidden":false},{"_id":"6a47f8c13daa34221c7c1e25","name":"Xuhong Zhang","hidden":false},{"_id":"6a47f8c13daa34221c7c1e26","user":{"_id":"6485bd278d14bcd5cdbb7c8d","avatarUrl":"/avatars/1427cf1a72b5db0cb263ad45885cf925.svg","isPro":false,"fullname":"Wenqi Zhang","user":"zwq2018","type":"user","name":"zwq2018"},"name":"Wenqi Zhang","status":"claimed_verified","statusLastChangedAt":"2026-07-06T20:09:32.425Z","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/6485bd278d14bcd5cdbb7c8d/pWfSFGZO9qE55za_P7i87.mp4"],"publishedAt":"2026-07-02T00:00:00.000Z","submittedOnDailyAt":"2026-07-06T00:00:00.000Z","title":"VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon","submittedOnDailyBy":{"_id":"6485bd278d14bcd5cdbb7c8d","avatarUrl":"/avatars/1427cf1a72b5db0cb263ad45885cf925.svg","isPro":false,"fullname":"Wenqi Zhang","user":"zwq2018","type":"user","name":"zwq2018"},"summary":"Vision-Language-Action (VLA) foundation models have recently achieved strong progress in embodied intelligence. To reduce policy-call frequency while preserving temporal coherence, most generative policies adopt an action chunk mechanism, executing multiple future actions in an open-loop manner under a fixed action horizon. However, this \"predict-then-blindly-execute\" paradigm sacrifices closed-loop reactivity: in contact-rich physical interactions, even small local perturbations can rapidly amplify within the open-loop blind spot, leading to compounding errors and ultimately task failure. To address this limitation, we propose VLA-Corrector, a lightweight corrective inference framework for action-chunked VLA policies. Without modifying the backbone policy weights, VLA-Corrector introduces a lightweight Latent-space Vision Monitor (LVM) that continuously compares predicted and actual visual feature evolution, enabling online detection of visual dynamics deviations. Once persistent deviation is detected, the system triggers a truncation event, discards the remaining stale actions, and invokes corrective replanning via Online Gradient Guidance (OGG). The detect-and-correct mechanism of VLA-Corrector naturally induces an event-triggered adaptive action horizon: it preserves long-horizon execution when the current chunk remains reliable, and invokes short-horizon corrective replanning when execution begins to drift. In doing so, VLA-Corrector mitigates the trade-off imposed by static horizons between execution robustness and policy-call frequency. It can be integrated into different VLA models without further retraining the VLA backbone, interrupting compounding errors while preserving much of the efficiency benefit of action chunking and substantially improving robustness in long-horizon, contact-rich robotic manipulation tasks.","upvotes":23,"discussionId":"6a47f8c13daa34221c7c1e27","projectPage":"https://zju-omniai.github.io/vla-corrector/","githubRepo":"https://github.com/ZJU-OmniAI/vla-corrector","githubRepoAddedBy":"user","ai_summary":"VLA-Corrector addresses limitations of action chunking in vision-language-action models by introducing a lightweight latent-space vision monitor that enables adaptive corrective replanning, improving robustness in contact-rich manipulation tasks.","ai_keywords":["Vision-Language-Action","action chunk mechanism","open-loop execution","closed-loop reactivity","latent-space Vision Monitor","visual feature evolution","truncation event","Online Gradient Guidance","event-triggered adaptive action horizon","corrective replanning"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":40,"organization":{"_id":"696461ab2d94e9a07cdb8efd","name":"OmniAI-ZJU","fullname":"ZJU-OmniAI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f2595830354d0ee043b25a/eEeRdHlGyJ148JQ6BAJ4O.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6485e686c7f19728a4a49b68","avatarUrl":"/avatars/8172c56c0e01163e190af3ea614e7449.svg","isPro":false,"fullname":"mayanna","user":"stena303","type":"user"},{"_id":"67543820c3af453d7b3e1d5e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67543820c3af453d7b3e1d5e/RbAZ9AQlxpy5E5Is-QN8b.jpeg","isPro":false,"fullname":"Dingming Li","user":"lidingm","type":"user"},{"_id":"6358b570aff68f72ac06113b","avatarUrl":"/avatars/c1df6bab08bc1a940b0af88a3f1e58ad.svg","isPro":false,"fullname":"DawnChase","user":"DawnChase","type":"user"},{"_id":"67e3de6b5269fa5ef3f2bd0d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/2QfglNGb0keGsllfHaS7Q.png","isPro":false,"fullname":"sfywtdiy","user":"xnxjiu","type":"user"},{"_id":"690f009eb9a679e969ece716","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/690f009eb9a679e969ece716/tDU-tAo1JyHKo__e7PLgv.jpeg","isPro":false,"fullname":"xuwang","user":"wx91726","type":"user"},{"_id":"6772300bf4c86f3302cb3e1d","avatarUrl":"/avatars/58c6210321d6304eba940686336b37bc.svg","isPro":false,"fullname":"HHH","user":"HHH10086","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"6485bd278d14bcd5cdbb7c8d","avatarUrl":"/avatars/1427cf1a72b5db0cb263ad45885cf925.svg","isPro":false,"fullname":"Wenqi Zhang","user":"zwq2018","type":"user"},{"_id":"69367c5e39abc914b3bbb2e3","avatarUrl":"/avatars/8ed0e0147bcc9c77040ad490ae8a107c.svg","isPro":false,"fullname":"Yi Pan","user":"Yi0304","type":"user"},{"_id":"67fc96221fd0efddecb9a009","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/q07B90UnvXQ8bCvnHqjyf.png","isPro":false,"fullname":"lee","user":"lee11111111","type":"user"},{"_id":"677b7d536b1653e39521d1db","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/DRdExsJdKI470ylktFOu0.png","isPro":false,"fullname":"Xuhong Zhang","user":"xuhongnever","type":"user"},{"_id":"69eb19da92b63efd5563805b","avatarUrl":"/avatars/96dd92e5c85402d6d978851c50dea4f8.svg","isPro":false,"fullname":"chenbao123ww","user":"Chenbao77","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"696461ab2d94e9a07cdb8efd","name":"OmniAI-ZJU","fullname":"ZJU-OmniAI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f2595830354d0ee043b25a/eEeRdHlGyJ148JQ6BAJ4O.png"},"query":{}}">

Cite arxiv.org/abs/2607.01804 in a model README.md to link it from this page.

Cite arxiv.org/abs/2607.01804 in a Space README.md to link it from this page.