# A paper proposes using sparse human corrections to guide robot-policy refinement [Canonical reading](https://escapemeridian.com/readings/a-paper-proposes-using-sparse-human-corrections-to-guide-robot-bcd140d9) Anukriti Singh, Maxime Caniot, Pratap Tokekar and Firas Abi-Farraj argue that a robot policy can work through most of a manipulation trajectory yet fail where precision matters. They say updating the full policy or exploring residual actions everywhere treats a localized failure as a global learning problem. Their Adaptive Residual Correction (ARC) method keeps the policy frozen and uses reinforcement learning (RL) to learn residual corrections. Sparse human corrections guide when residual control activates and where autonomous refinement explores; task reward determines what residual is executed. The corrections are not treated as imitation targets. Across three precision-manipulation tasks on a real robot, the authors report that ARC raised average task success from 33.3% to 75.0%. They also report that human corrections define task-dependent 2–3 dimensional subspaces of the seven-joint residual space, retaining 96.0–97.7% of squared correction magnitude. The results suggest, in the authors’ account, that human input can constrain post-training without making the observed correction the final action target. How well does ARC’s reported improvement carry over to precision-manipulation tasks beyond the three tested? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-14T00:02:25.137Z. First observed: 2026-10-28T08:58:38.760Z. Retrieved: 2026-10-30T20:32:57.615Z. - [When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies](https://arxiv.org/abs/2610.37547) ## Limits The reported evaluation covers three precision-manipulation tasks on a real robot; the abstract does not establish how results transfer to other tasks or settings. The supplied evidence is the abstract and does not provide task-level results or further evaluation details. ## Open question How well does ARC’s reported improvement carry over to precision-manipulation tasks beyond the three tested? ## Source-stated The authors introduce ARC as a post-training method that learns residual corrections on top of a frozen policy, using sparse human corrections to guide autonomous exploration. > uses reinforcement learning (RL) to learn a residual correction on top of a frozen policy while using sparse human corrections to structure autonomous exploration - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated In ARC, human corrections guide when residual control activates and where autonomous refinement explores; task reward determines the residual executed. > pre-intervention states to determine 'when' residual control should become active and correction geometry to determine 'where' in residual action space autonomous refinement should explore; task reward determines 'what' residual is ultimately executed. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated Across three precision-manipulation tasks on a real robot, the authors report average task success rising from 33.3% to 75.0%. > Across three precision-manipulation tasks on a real robot, ARC improves average task success from 33.3% to 75.0%. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated The authors report that task-dependent subspaces derived from human corrections retain 96.0–97.7% of squared correction magnitude. > human corrections define task-dependent 2-3 dimensional subspaces of the seven-joint residual space that retain 96.0-97.7% of squared correction magnitude. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source and contributor context Singh, Anukriti, Caniot, Maxime, Tokekar, Pratap, Abi-Farraj, Firas · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Reinforcement learning Training by trial and error, in which a system is rewarded for the outcomes its designers want. - [Definition reference](https://developers.google.com/machine-learning/glossary#reinforcement-learning-rl) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 52cc78e5-feff-47b1-8280-76e12944dd02. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T00:02:25.137Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3fa32041794d534f2cbb413968c6dfca654fab33f6a756fa3c4e6613d3ef1c23. Source text hash: a4f5a82b8e0edfd168bdfc959ab280f9efa6980a333d52d501d76488167df681. ## Revision and source-check history ### Version 1 — 2026-11-14T00:02:25.137Z Anukriti Singh, Maxime Caniot, Pratap Tokekar and Firas Abi-Farraj argue that a robot policy can work through most of a manipulation trajectory yet fail where precision matters. They say updating the full policy or exploring residual actions everywhere treats a localized failure as a global learning problem. Their Adaptive Residual Correction (ARC) method keeps the policy frozen and uses reinforcement learning (RL) to learn residual corrections. Sparse human corrections guide when residual control activates and where autonomous refinement explores; task reward determines what residual is executed. The corrections are not treated as imitation targets. Across three precision-manipulation tasks on a real robot, the authors report that ARC raised average task success from 33.3% to 75.0%. They also report that human corrections define task-dependent 2–3 dimensional subspaces of the seven-joint residual space, retaining 96.0–97.7% of squared correction magnitude. The results suggest, in the authors’ account, that human input can constrain post-training without making the observed correction the final action target. How well does ARC’s reported improvement carry over to precision-manipulation tasks beyond the three tested? ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-14T00:02:25.137Z. First observed: 2026-10-28T08:58:38.760Z. Retrieved: 2026-10-30T20:32:57.615Z. - [When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies](https://arxiv.org/abs/2610.37547) ## Limits The reported evaluation covers three precision-manipulation tasks on a real robot; the abstract does not establish how results transfer to other tasks or settings. The supplied evidence is the abstract and does not provide task-level results or further evaluation details. ## Open question How well does ARC’s reported improvement carry over to precision-manipulation tasks beyond the three tested? ## Source-stated The authors introduce ARC as a post-training method that learns residual corrections on top of a frozen policy, using sparse human corrections to guide autonomous exploration. > uses reinforcement learning (RL) to learn a residual correction on top of a frozen policy while using sparse human corrections to structure autonomous exploration - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated In ARC, human corrections guide when residual control activates and where autonomous refinement explores; task reward determines the residual executed. > pre-intervention states to determine 'when' residual control should become active and correction geometry to determine 'where' in residual action space autonomous refinement should explore; task reward determines 'what' residual is ultimately executed. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated Across three precision-manipulation tasks on a real robot, the authors report average task success rising from 33.3% to 75.0%. > Across three precision-manipulation tasks on a real robot, ARC improves average task success from 33.3% to 75.0%. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source-stated The authors report that task-dependent subspaces derived from human corrections retain 96.0–97.7% of squared correction magnitude. > human corrections define task-dependent 2-3 dimensional subspaces of the seven-joint residual space that retain 96.0-97.7% of squared correction magnitude. - [Supporting source](https://arxiv.org/abs/2610.37547) ## Source and contributor context Singh, Anukriti, Caniot, Maxime, Tokekar, Pratap, Abi-Farraj, Firas · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Reinforcement learning Training by trial and error, in which a system is rewarded for the outcomes its designers want. - [Definition reference](https://developers.google.com/machine-learning/glossary#reinforcement-learning-rl) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 52cc78e5-feff-47b1-8280-76e12944dd02. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T00:02:25.137Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3fa32041794d534f2cbb413968c6dfca654fab33f6a756fa3c4e6613d3ef1c23. Source text hash: a4f5a82b8e0edfd168bdfc959ab280f9efa6980a333d52d501d76488167df681. ### uncertain — 2026-11-14T00:02:25.137Z — anthropic/claude-opus-5.5 Publication: 52cc78e5-feff-47b1-8280-76e12944dd02 attribution — pass: The paper is correctly attributed to Anukriti Singh, Maxime Caniot, Pratap Tokekar and Firas Abi-Farraj. Results are framed as the authors' reports rather than established fact. > Authors:Anukriti Singh, Maxime Caniot, Pratap Tokekar, Firas Abi-Farraj summary — uncertain: The reading is faithful but hard for non-specialists. Central terms are never explained in plain language: residual correction (a small learned adjustment added to the existing policy's actions), frozen policy (left unchanged), manipulation trajectory, and post-training. The subspace statistic (2-3 dimensions, squared correction magnitude) is opaque without explanation. The draft also says "a robot policy" where the source says imitation policies, which drops scope. Smallest repair: one plain-language clause per term, then a sentence saying why fixing only the failing region matters. > Imitation policies that reliably execute most of a manipulation trajectory can still fail in localized regions requiring last-mile precision. date — pass: The originalPublished value of 2026-10-27 matches the arXiv v1 submission date. The later feed date is correctly not used. > \[Submitted on 27 Oct 2026\] support — pass: All claims match the abstract: the frozen policy with an RL residual, the when/where/what roles, success rising from 33.3% to 75.0% on three real-robot tasks, and the 96.0-97.7% retention figure. The question about transfer beyond the three tasks follows from the stated scope. The limitations are appropriate, and no unsupported benefits are implied. > sparse human corrections can localize and constrain autonomous post-training without requiring the observed correction itself to serve as the final action target media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/bcd140d921db086bc6ff63e3d39ffde08f11b05d81cfa459e3dcebc32ae1fa1c/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/a-paper-proposes-using-sparse-human-corrections-to-guide-robot-bcd140d9.json)