Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.</p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/6385f7b969634850f8ddd541/gNYEQyLS0ALRY0GNyrwBb.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6385f7b969634850f8ddd541/gNYEQyLS0ALRY0GNyrwBb.png\" alt=\"teaser\"></a></p>\n","updatedAt":"2026-08-13T08:03:19.018Z","author":{"_id":"6385f7b969634850f8ddd541","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1669723465271-noauth.png","fullname":"Peidong Wang","name":"WDong","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.9347298741340637},"editors":["WDong"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1669723465271-noauth.png"],"reactions":[{"reaction":"👍","users":["moomight"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.11350","authors":[{"_id":"6a7d762e0ac8bee77474efd5","user":{"_id":"6385f7b969634850f8ddd541","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1669723465271-noauth.png","isPro":false,"fullname":"Peidong Wang","user":"WDong","type":"user","name":"WDong"},"name":"Peidong Wang","status":"claimed_verified","statusLastChangedAt":"2026-08-13T08:45:05.050Z","hidden":false},{"_id":"6a7d762e0ac8bee77474efd6","name":"Zhiming Ma","hidden":false},{"_id":"6a7d762e0ac8bee77474efd7","name":"Ying Chang","hidden":false},{"_id":"6a7d762e0ac8bee77474efd8","name":"Xufang Luo","hidden":false},{"_id":"6a7d762e0ac8bee77474efd9","name":"Xiaocui Yang","hidden":false},{"_id":"6a7d762e0ac8bee77474efda","name":"Shi Feng","hidden":false},{"_id":"6a7d762e0ac8bee77474efdb","name":"Yuqing Yang","hidden":false},{"_id":"6a7d762e0ac8bee77474efdc","name":"Dongsheng Li","hidden":false}],"publishedAt":"2026-08-11T00:00:00.000Z","submittedOnDailyAt":"2026-08-13T00:00:00.000Z","title":"Self-Evolving Embodied Agents via Skill-Harness Evolution","submittedOnDailyBy":{"_id":"6385f7b969634850f8ddd541","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1669723465271-noauth.png","isPro":false,"fullname":"Peidong Wang","user":"WDong","type":"user","name":"WDong"},"summary":"Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.","upvotes":4,"discussionId":"6a7d762e0ac8bee77474efdd","ai_summary":"SHAPER is a train-free framework that improves embodied agents by evolving reusable skills and a context-code harness around a frozen foundation model through environment rollouts.","ai_keywords":["embodied agents","foundation models","supervised fine-tuning","reinforcement learning","train-free adaptation","SHAPER","self-evolving framework","reusable skills","context-code harness","target-environment rollouts","planner and optimizer","VLABench","ESI-Bench"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6385f7b969634850f8ddd541","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1669723465271-noauth.png","isPro":false,"fullname":"Peidong Wang","user":"WDong","type":"user"},{"_id":"666af9e68b12eb564be39bf2","avatarUrl":"/avatars/4a73fe5d54621cbafb1efa2dc4c70fec.svg","isPro":false,"fullname":"Ying Chang","user":"moomight","type":"user"},{"_id":"64b8a72952b7353d8c669086","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b8a72952b7353d8c669086/3PUTNmx9kd17gtZ9-Yviw.jpeg","isPro":false,"fullname":"Qi Fan","user":"fanqiNO1","type":"user"},{"_id":"64a0ed5ed5374ca472cfb0ac","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64a0ed5ed5374ca472cfb0ac/n_wXamXfR_PPn0hRbnR1X.jpeg","isPro":false,"fullname":"ZhimingMa","user":"JimmyMa99","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.11350.md","query":{}}">
Self-Evolving Embodied Agents via Skill-Harness Evolution
Abstract
SHAPER is a train-free framework that improves embodied agents by evolving reusable skills and a context-code harness around a frozen foundation model through environment rollouts.
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.
Community
Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings. We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts. In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates. We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting. Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.11350 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.11350 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.11350 in a Space README.md 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.