<a href=\"https://cdn-uploads.huggingface.co/production/uploads/687ecefef2f71aa022d04c44/sNl2bmKnJ0KEuaPSayzPD.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/687ecefef2f71aa022d04c44/sNl2bmKnJ0KEuaPSayzPD.png\" alt=\"terminal_states_r2\"></a></p>\n<p><em>Dexterous Insertion at the Maximum Demonstrated Speed. Representative terminal states of the parcel acquisition, reorientation, and insertion task in ParcelStow, with magnified views of the receptacle interiors. Captions show the final parcel orientation error (success requires a value of 10 deg.).</em></p>\n<p>Does imitation learning preserve an expert’s performance when task timing is compressed to increase execution speed, even within the demonstrated speed range?</p>\n<p>We study this question using ParcelStow, a dexterous manipulation benchmark with matched expert and learner initial conditions. In the parcel-insertion study reported in the paper, the scripted expert and ACT both achieve 100% task success at nominal speed, but at the maximum demonstrated speed, success falls to 84% for the expert and 53% for ACT. Stage-level and relative-motion analyses localize much of this difference to insertion rather than acquisition or free-space transport. The accompanying open-source benchmark now includes three contact-rich tasks---parcel insertion, upright placement, and keyed peg insertion---with scripted experts, ACT checkpoints, matched initial conditions, physical success predicates, episode records, and CPU-only result reproduction. Videos of each task are provided as media attachments alongside this comment.</p>\n<p>Hugging Face Page: <a href=\"https://huggingface.co/datasets/cenwerem/parcelstow\">https://huggingface.co/datasets/cenwerem/parcelstow</a></p>\n","updatedAt":"2026-09-02T23:12:46.005Z","author":{"_id":"687ecefef2f71aa022d04c44","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/687ecefef2f71aa022d04c44/QKPk0YEuDTv_Fq-kgzeV0.png","fullname":"Clinton Enwerem","name":"cenwerem","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8602074980735779},"editors":["cenwerem"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/687ecefef2f71aa022d04c44/QKPk0YEuDTv_Fq-kgzeV0.png"],"reactions":[],"isReport":false}},{"id":"6a98cdabe86d40b3ac31ecf2","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":379,"isUserFollowing":false},"createdAt":"2026-09-03T01:30:19.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [AutoIntervene: Calibrated Intervention for Action-Chunking Imitation Learning Policies](https://huggingface.co/papers/2608.07065) (2026)\n* [NestDex: Nested Policy Learning with Copilot Assisted Teleoperation for Dexterous Manipulation](https://huggingface.co/papers/2608.13362) (2026)\n* [ReForce: Learning Force-aware Retargeting for Dexterous Manipulation](https://huggingface.co/papers/2608.15560) (2026)\n* [FetchMan: Learning Visual Humanoid Loco-Manipulation Policies from Simulated Experiences](https://huggingface.co/papers/2608.17027) (2026)\n* [WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning](https://huggingface.co/papers/2608.22591) (2026)\n* [Policy-Induced Hand Priors in Humanoid Dual-Arm Manipulation: Diagnosing and Mitigating Initial-Pose Dependence](https://huggingface.co/papers/2608.11769) (2026)\n* [EvoHIL: Self-Evolving Reward and Flow-Matched Policy Optimization for Robust Human-in-the-Loop Reinforcement Learning](https://huggingface.co/papers/2608.03872) (2026)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"<p>This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. </p>\n<p>The following papers were recommended by the Semantic Scholar API </p>\n<ul>\n<li><a href=\"https://huggingface.co/papers/2608.07065\">AutoIntervene: Calibrated Intervention for Action-Chunking Imitation Learning Policies</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.13362\">NestDex: Nested Policy Learning with Copilot Assisted Teleoperation for Dexterous Manipulation</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.15560\">ReForce: Learning Force-aware Retargeting for Dexterous Manipulation</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.17027\">FetchMan: Learning Visual Humanoid Loco-Manipulation Policies from Simulated Experiences</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.22591\">WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.11769\">Policy-Induced Hand Priors in Humanoid Dual-Arm Manipulation: Diagnosing and Mitigating Initial-Pose Dependence</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.03872\">EvoHIL: Self-Evolving Reward and Flow-Matched Policy Optimization for Robust Human-in-the-Loop Reinforcement Learning</a> (2026)</li>\n</ul>\n<p> Please give a thumbs up to this comment if you found it helpful!</p>\n<p> If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space</p>\n<p> You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code>@librarian-bot recommend</code></p>\n","updatedAt":"2026-09-03T01:30:19.435Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":379,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7494528889656067},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.01453","authors":[{"_id":"6a978ab3fe3c2f89286c3989","user":{"_id":"687ecefef2f71aa022d04c44","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/687ecefef2f71aa022d04c44/QKPk0YEuDTv_Fq-kgzeV0.png","isPro":false,"fullname":"Clinton Enwerem","user":"cenwerem","type":"user","name":"cenwerem"},"name":"Clinton Enwerem","status":"claimed_verified","statusLastChangedAt":"2026-09-02T16:45:04.817Z","hidden":false},{"_id":"6a978ab3fe3c2f89286c398a","name":"John S. Baras","hidden":false},{"_id":"6a978ab3fe3c2f89286c398b","name":"Calin Belta","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/687ecefef2f71aa022d04c44/Qqqwu8mzye3UT1qv8i2Ru.mp4","https://cdn-uploads.huggingface.co/production/uploads/687ecefef2f71aa022d04c44/l-dF-xcyE4Px_oZnchjCV.mp4","https://cdn-uploads.huggingface.co/production/uploads/687ecefef2f71aa022d04c44/x1lYUbwoJwkNmUtP1SHtT.mp4"],"publishedAt":"2026-09-01T00:00:00.000Z","submittedOnDailyAt":"2026-09-02T00:00:00.000Z","title":"Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds","submittedOnDailyBy":{"_id":"687ecefef2f71aa022d04c44","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/687ecefef2f71aa022d04c44/QKPk0YEuDTv_Fq-kgzeV0.png","isPro":false,"fullname":"Clinton Enwerem","user":"cenwerem","type":"user","name":"cenwerem"},"summary":"Dexterous manipulation policies learned by imitation are typically evaluated for robustness to variation in scenes, objects, or instructions, but their performance across task execution speeds is less often examined. This leaves open how much temporal robustness a learner retains relative to the expert it imitates. We compare an expert and learner under the same task conditions, initial-condition draws, and speedup factors. We instantiate the evaluation in ParcelStow, a contact-rich task in which the robot acquires, reorients, and inserts a parcel. The demonstrations span the speedup range for the manipulation phases after parcel acquisition. A scripted expert and an Action Chunking with Transformers (ACT) policy trained from the expert's demonstrations both achieve 100 percent task success at nominal speed. Their success rates diverge within the demonstrated range: at its maximum, expert success is 84 percent and ACT success is 53 percent. Two ACT policies with different parameter initializations show similar degradation, decreasing by 34 and 48 percentage points from nominal speed to the maximum demonstrated speed, compared with 16 points for the expert. Stage-level analysis shows that 35 of ACT's 47 failures at the maximum demonstrated speed are insertion misalignments. Under the relative-motion handoff, every ACT acquisition retains the parcel through reorientation and transfer in free space, but only 64 percent complete the overall task, compared with 95 percent after expert acquisition. Across all evaluated policies and speeds, none of the 414 acquisitions without force closure completes the task. Equal nominal task success therefore does not imply preservation of expert performance across execution speeds. Code, data, and evaluation scripts are available at https://github.com/coenwerem/parcelstow.","upvotes":2,"discussionId":"6a978ab4fe3c2f89286c398c","ai_summary":"Imitation-learned dexterous manipulation policies degrade more sharply than expert policies when execution speed increases, with insertion misalignment being the primary failure mode.","ai_keywords":["Action Chunking with Transformers (ACT)"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"68b3c3bbc375e05b059370b2","name":"UMCP","fullname":"University of Maryland College Park","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68b3c2c3a4ea236d1a97871a/bji3nI5ZWm2r4JX_-HLo0.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"687ecefef2f71aa022d04c44","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/687ecefef2f71aa022d04c44/QKPk0YEuDTv_Fq-kgzeV0.png","isPro":false,"fullname":"Clinton Enwerem","user":"cenwerem","type":"user"},{"_id":"699eaf88ef8843d6e00c987e","avatarUrl":"/avatars/b472c58a6fc139689875127ef401567e.svg","isPro":false,"fullname":"Scarlett Harris","user":"scarlettharris","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"68b3c3bbc375e05b059370b2","name":"UMCP","fullname":"University of Maryland College Park","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68b3c2c3a4ea236d1a97871a/bji3nI5ZWm2r4JX_-HLo0.png"},"query":{}}">
Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds
Abstract
Imitation-learned dexterous manipulation policies degrade more sharply than expert policies when execution speed increases, with insertion misalignment being the primary failure mode.
Dexterous manipulation policies learned by imitation are typically evaluated for robustness to variation in scenes, objects, or instructions, but their performance across task execution speeds is less often examined. This leaves open how much temporal robustness a learner retains relative to the expert it imitates. We compare an expert and learner under the same task conditions, initial-condition draws, and speedup factors. We instantiate the evaluation in ParcelStow, a contact-rich task in which the robot acquires, reorients, and inserts a parcel. The demonstrations span the speedup range for the manipulation phases after parcel acquisition. A scripted expert and an Action Chunking with Transformers (ACT) policy trained from the expert's demonstrations both achieve 100 percent task success at nominal speed. Their success rates diverge within the demonstrated range: at its maximum, expert success is 84 percent and ACT success is 53 percent. Two ACT policies with different parameter initializations show similar degradation, decreasing by 34 and 48 percentage points from nominal speed to the maximum demonstrated speed, compared with 16 points for the expert. Stage-level analysis shows that 35 of ACT's 47 failures at the maximum demonstrated speed are insertion misalignments. Under the relative-motion handoff, every ACT acquisition retains the parcel through reorientation and transfer in free space, but only 64 percent complete the overall task, compared with 95 percent after expert acquisition. Across all evaluated policies and speeds, none of the 414 acquisitions without force closure completes the task. Equal nominal task success therefore does not imply preservation of expert performance across execution speeds. Code, data, and evaluation scripts are available at https://github.com/coenwerem/parcelstow.
Community
Dexterous Insertion at the Maximum Demonstrated Speed. Representative terminal states of the parcel acquisition, reorientation, and insertion task in ParcelStow, with magnified views of the receptacle interiors. Captions show the final parcel orientation error (success requires a value of 10 deg.).
Does imitation learning preserve an expert’s performance when task timing is compressed to increase execution speed, even within the demonstrated speed range?
We study this question using ParcelStow, a dexterous manipulation benchmark with matched expert and learner initial conditions. In the parcel-insertion study reported in the paper, the scripted expert and ACT both achieve 100% task success at nominal speed, but at the maximum demonstrated speed, success falls to 84% for the expert and 53% for ACT. Stage-level and relative-motion analyses localize much of this difference to insertion rather than acquisition or free-space transport. The accompanying open-source benchmark now includes three contact-rich tasks---parcel insertion, upright placement, and keyed peg insertion---with scripted experts, ACT checkpoints, matched initial conditions, physical success predicates, episode records, and CPU-only result reproduction. Videos of each task are provided as media attachments alongside this comment.
Hugging Face Page: https://huggingface.co/datasets/cenwerem/parcelstow
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
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/2609.01453 in a model README.md to link it from this page.
Cite arxiv.org/abs/2609.01453 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.