Graph-as-Policy (GaP) builds interpretable computation graphs (like ROS) instead of a single end-to-end policy, targeting Variational Automation — repetitive industrial tasks where objects vary in shape and pose but the workcell is fixed (packing groceries, making coffee, sorting packages). A multi-agent harness built over coding agents decomposes the task and assembles a graph from an open library of 50+ modular robot skills; the graph rehearses in simulation, uses contact feedback to diagnose its own failures, and rewrites its own structure until performance plateaus, then transfers sim-to-real. Where model-free VLA policies drop to ~20% success under object pose variation, GaP reaches 93–99%, and on a bimanual crate-washing task it matches hand-engineered code via automated fine-tuning. Paper, code, an 8-task benchmark suite, and the full skill library are all open.</p>\n","updatedAt":"2026-07-07T19:22:54.978Z","author":{"_id":"66d2af23f040611f7cea1b1b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66d2af23f040611f7cea1b1b/vCToO_XgrLD0nfR8rzjZd.jpeg","fullname":"Kaiyuan Eric Chen","name":"keplerccc","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.850517213344574},"editors":["keplerccc"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/66d2af23f040611f7cea1b1b/vCToO_XgrLD0nfR8rzjZd.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.05369","authors":[{"_id":"6a4d51d825849b193a834955","name":"Kaiyuan Chen","hidden":false},{"_id":"6a4d51d825849b193a834956","name":"Shuangyu Xie","hidden":false},{"_id":"6a4d51d825849b193a834957","name":"Letian Fu","hidden":false},{"_id":"6a4d51d825849b193a834958","name":"Justin Yu","hidden":false},{"_id":"6a4d51d825849b193a834959","name":"William Pacini","hidden":false},{"_id":"6a4d51d825849b193a83495a","name":"Sandeep Bajamahal","hidden":false},{"_id":"6a4d51d825849b193a83495b","name":"Hudson Kim","hidden":false},{"_id":"6a4d51d825849b193a83495c","name":"Jaimyn Drake","hidden":false},{"_id":"6a4d51d825849b193a83495d","name":"Daehwa Kim","hidden":false},{"_id":"6a4d51d825849b193a83495e","name":"Haoru Xue","hidden":false},{"_id":"6a4d51d825849b193a83495f","name":"Jonathan Francis","hidden":false},{"_id":"6a4d51d825849b193a834960","name":"Christian Juette","hidden":false},{"_id":"6a4d51d825849b193a834961","name":"Peter Schaldenbrand","hidden":false},{"_id":"6a4d51d825849b193a834962","name":"Muhammet Yunus Seker","hidden":false},{"_id":"6a4d51d825849b193a834963","name":"Ruwan Wickramarachchi","hidden":false},{"_id":"6a4d51d825849b193a834964","name":"Uksang Yoo","hidden":false},{"_id":"6a4d51d825849b193a834965","name":"Guanzhi Wang","hidden":false},{"_id":"6a4d51d825849b193a834966","name":"Adithyavairavan Murali","hidden":false},{"_id":"6a4d51d825849b193a834967","name":"Balakumar Sundaralingam","hidden":false},{"_id":"6a4d51d825849b193a834968","name":"S. Shankar Sastry","hidden":false},{"_id":"6a4d51d825849b193a834969","name":"Spencer Huang","hidden":false},{"_id":"6a4d51d825849b193a83496a","name":"Yuke Zhu","hidden":false},{"_id":"6a4d51d825849b193a83496b","name":"Linxi \"Jim\" Fan","hidden":false},{"_id":"6a4d51d825849b193a83496c","name":"Ken Goldberg","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/66d2af23f040611f7cea1b1b/Bx43oDEKrIH10nuHOWWD8.mp4"],"publishedAt":"2026-07-06T00:00:00.000Z","submittedOnDailyAt":"2026-07-07T00:00:00.000Z","title":"GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks","submittedOnDailyBy":{"_id":"66d2af23f040611f7cea1b1b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66d2af23f040611f7cea1b1b/vCToO_XgrLD0nfR8rzjZd.jpeg","isPro":false,"fullname":"Kaiyuan Eric Chen","user":"keplerccc","type":"user","name":"keplerccc"},"summary":"For robots to work reliably in commercial and industrial applications, can recent advances in agentic coding systems combine interpretable robot programming with the open-world adaptability of model-free policies? We focus on \"Variational Automation\" (VA), a class of tasks that have larger variations in object geometry and pose than fixed automation. Model-free policies often struggle to close the reliability gap for VA tasks, which must be executed persistently and reliably in commercial and industrial applications. Motivated by prior work on Task and Motion Planning (TAMP) and the Robot Operating System (ROS), we introduce Graph-as-Policy (GaP), a multi-agent coding harness that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). GaP then generates an internal simulation environment to rehearse task instances with different graphs in parallel to iteratively refine the graph structure and parameters to improve success rates and throughput. Evaluation with 8 new open VA task benchmarks, 4 in-simulation and 4 in real-world, suggests that GaP can achieve success rates that significantly outperform baselines. Details, code, and data can be found online: https://graph-robots.github.io/gap","upvotes":1,"discussionId":"6a4d51d925849b193a83496d","projectPage":"https://graph-robots.github.io/gap/","githubRepo":"https://github.com/graph-robots/graph-as-policy","githubRepoAddedBy":"user","ai_summary":"Graph-as-Policy system combines modular robot skills with multi-agent coding to improve reliability in variable automation tasks through parallel simulation refinement.","ai_keywords":["agentic coding systems","model-free policies","variational automation","Task and Motion Planning","Robot Operating System","directed computation graphs","Modular Open Robot Skill Library","internal simulation environment","parallel refinement"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":13,"organization":{"_id":"61f20a9ce108f2cba2dc0730","name":"Berkeley","fullname":"UC Berkeley","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61ac8f8a00d01045fca0ad2f/0FjsTg2txEZZ4dEgmMnQL.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66d2af23f040611f7cea1b1b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66d2af23f040611f7cea1b1b/vCToO_XgrLD0nfR8rzjZd.jpeg","isPro":false,"fullname":"Kaiyuan Eric Chen","user":"keplerccc","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"61f20a9ce108f2cba2dc0730","name":"Berkeley","fullname":"UC Berkeley","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61ac8f8a00d01045fca0ad2f/0FjsTg2txEZZ4dEgmMnQL.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.05369.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2607.05369 in a model README.md to link it from this page.

Cite arxiv.org/abs/2607.05369 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.05369 in a Space README.md to link it from this page.

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