{"id":"dd30c24e-7466-4a33-8834-37b042ee16d6","arxiv_id":"2509.07953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Robot policies trained on human interventions that rewind to a familiar state and then correct the mistake achieve higher long-horizon success and better data efficiency than imitation on full demonstrations alone.","lead":"This paper introduces a robot training recipe that mixes human recovery and correction segments into imitation data, and reports better success rates on long-horizon bimanual tasks with far less collected data than prior work. It also reports that success rises with the number of recovery maneuvers a policy performs, which the authors call test-time scaling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Figure 10's 'test-time scaling' confounds training round with recovery count; the central 'o1-style' scaling claim needs a fixed-checkpoint or round-controlled test.","rationale":"The reader's conditional verdict already identifies the Figure 10 correlation as a key weakness, and this stress-test sharpens it: the claim is not merely statistically thin, it is confounded by round index. This is load-bearing because test-time scaling is presented as a headline contribution ('o1-style' robot policies) and is part of the paper's strongest claim that success scales with recovery count. If the within-round analysis fails, the scaling claim should be removed or reframed, but the data-efficiency and recovery-composition results from the same-policy comparisons and ablations would remain plausible. Thus the appropriate verdict remains conditional, pending this check and the other verification issues the reader noted. I chose UNCHANGED because the reader's conditional assessment already covers this concern; the concrete test would tighten the condition but does not by itself overturn the overall verdict.","tokens_in":22757,"tokens_out":7339,"duration_ms":91528,"concrete_test":"Use the raw evaluation rollouts behind Figure 10. Fit a logistic regression of per-trial success on the number of recovery maneuvers, including round/checkpoint fixed effects (or equivalently recompute the Figure 10 correlation separately within each training round and pool the within-round estimates). If the within-round recovery coefficient is not positive and significant, or the pooled within-round slope is near zero, the claimed linear test-time scaling is confounded by training round and should be restated as a training-progress correlation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3, Result 2, and Figure 10 claim that RaC exhibits test-time scaling: success rate increases linearly with the number of recovery maneuvers. The evidence is a correlation computed across evaluation rollouts from different training rounds, with each point an average over one round. This conflates two variables: training progress (more rounds -> more aggregated data -> higher success) and recovery count (later checkpoints also happen to emit more recoveries). The fitted slope may therefore be entirely driven by round index, not by recovery count. Nothing in the paper demonstrates that, for a fixed policy checkpoint, rollouts with more recoveries are more likely to succeed, nor that an operator can increase recovery count at deployment to improve success. The abstract and Section 4.1 present this as a central contribution analogous to LLM test-time scaling, and the paper's own Section 6 weakens it to a correlation. Since the protocol in Algorithm 1 does not control recovery count at deployment—recovery is an emergent policy behavior—the causal scaling claim is currently unsupported. This concern can be settled from the authors' existing raw evaluation data; it does not require re-running the real-robot experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RaC, a human-in-the-loop data collection protocol for imitation learning. After an initial round of full expert demonstrations, human operators intervene during policy rollouts, first rewinding the robot to a state judged to be in-distribution and then performing a corrective segment that completes the current subtask; episodes terminate after the intervention. Policies are trained with a flow-matching transformer on the aggregated data. Experiments on three real bimanual tasks (shirt hanging, airtight lid sealing, takeout box packing) and one simulated assembly task report that RaC outperforms batched full demonstrations and HG-DAgger-style intervention collection in success rate and data efficiency, and the paper claims an order-of-magnitude data saving over prior work on shirt hanging. The paper also claims a form of \"test-time scaling\": success rate grows linearly with the number of recovery maneuvers exhibited by the policy, analogous to long chain-of-thought scaling in LLMs.","tokens_in":22971,"tokens_out":4710,"duration_ms":60301,"significance":"If the results hold, the paper makes a practically valuable contribution: a relatively simple, standardized human-in-the-loop protocol that improves the data efficiency and robustness of imitation-learned policies on long-horizon, contact-rich bimanual tasks. The real-robot evaluation is a genuine strength: three tasks, 60 trials per condition, comparisons against batched full demonstrations and HG-DAgger, plus ablations of the two protocol rules. The paper also honestly notes that some of the evidence is correlational. However, the two headline claims are currently stronger than the evidence: the 10x data-efficiency statement relies on a cross-paper estimate of prior data length, and the test-time scaling result conflates training round with recovery count. Both can be addressed with additional analysis of the authors' existing data, so the central idea is defensible and worth publishing after revision.","major_comments":[{"comment":"The \"test-time scaling\" claim conflates two variables. Each point in Figure 10 is a training round: the x-value is the average number of recoveries in that round's successful rollouts, and the y-value is the same round's success rate. Both quantities increase with round index, so the fitted linear relationship may be entirely driven by training progress (more rounds → more data → better policy → more recoveries) rather than by recovery count causing success. The paper's own Section 6 weakens this to \"correlate,\" but the abstract and Section 4.1 present it as a central contribution. To support the claim, the authors should analyze rollouts from a fixed policy checkpoint, grouping by recovery count, or otherwise control recovery count at deployment. This is answerable from the existing raw evaluation logs.","section":"Section 5.3, Result 2 / Figure 10"},{"comment":"The order-of-magnitude data-efficiency claim against ALOHA Unleashed is not a controlled comparison. The 89-hour figure is an estimate obtained by assuming 1 minute per trajectory (Appendix D), not a reported data length. The two systems differ in robot hardware, model size (368M vs. 217M), and evaluation protocol. The claim should be reframed: the controlled comparison in Figure 8 shows roughly a 2x efficiency improvement over the authors' own batched full-demonstration baseline on shirt hanging, while the comparison to prior work should be reported as indicative rather than as a measured 10x result. The abstract and Section 5.2 currently overstate the strength of this comparison.","section":"Table 1 / Appendix D"},{"comment":"The load-bearing premise of the method is that recovery is systematically easier to learn than correction, and that the human operator's rewind reliably returns the robot to the in-distribution region of the full demonstrations. Section 4.1 asserts the \"verification-generation gap\" without measuring it, and Section 4.3's heatmap is a heuristic visual aid with no quantitative verification. The paper should provide at least a post-hoc analysis: e.g., measure the distance between recovery endpoints and the full-demo state distribution, show that similar success can be obtained when recovery is performed from different initial failure states, or compare learning curves for recovery-only vs. correction-only segments. Without such evidence, the mechanism behind the data-efficiency gain remains a plausible hypothesis rather than an established result.","section":"Section 4.1 / 4.2 / 4.3"},{"comment":"The data accounting in Algorithm 1 is inconsistent with the reported efficiency numbers. Line 14 adds only human intervention transitions to the dataset, but line 15 charges the full episode length to the budget, including policy-driven frames before the intervention. The abstract's phrase \"10x less data collection time and samples\" is ambiguous: Table 1 reports 5 hours of RaC data, while Figure 8 reports human-collected frames. The authors should clarify whether the reported budget is wall-clock time, human-collected frames, or total episode frames, and ensure that the same accounting is used for all methods compared.","section":"Algorithm 1 / Section 4.4"}],"minor_comments":[{"comment":"The text says \"In Figure 10, we visualize the composition of intervention data,\" but the composition is shown in Figure 12 (left). Please correct the cross-reference.","section":"Section 5.4, first paragraph"},{"comment":"The x-axis is labeled \"Average Recoveries per Successful Trajectory,\" but the text and caption refer to \"recovery segments upon deployment\" and \"recovery frequency.\" Since only successful trajectories are used for the x-axis, this selection should be stated and its possible bias discussed.","section":"Figure 10"},{"comment":"The claim that \"longer length is correlated with better average performance\" is based only on successful rollouts. Reporting the length distribution of failed rollouts as well would make the point more compelling and would clarify whether length itself is predictive or merely a byproduct of success.","section":"Section 5.3, Result 3 / Figure 11"},{"comment":"Line 8 has a typo: \"Pair each recovery a correction\" should read \"Pair each recovery with a correction.\"","section":"Algorithm 1"},{"comment":"The statement that only 3.68% of DROID episodes contain recovery is based on a 1% sub-sample. Please report the sub-sample size and the counting definition (e.g., whether a recovery was identified by human annotation or by a heuristic).","section":"Section 4.2, DROID analysis"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for a robotics venue and the real-robot experiments are substantial. My main concern is that the two headline claims—10x data efficiency and test-time scaling—are currently supported by weaker evidence than the abstract suggests. The test-time scaling issue is the most serious because it is presented as an 'o1-style' contribution but is a round-confounded correlation. The good news is that the authors likely already have the raw evaluation logs to perform the fixed-checkpoint analysis, and the cross-paper comparison can be softened without changing the core contribution. I would encourage a major revision rather than rejection, because the underlying protocol and controlled comparisons (RaC vs. HG-DAgger vs. full demos) are valuable and the load-bearing gaps appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is simple and worth taking seriously: during human interventions, pair a recovery segment (rewind to a familiar state) with a correction segment, then terminate the episode. Train on that mixture. The real-robot results across three bimanual tasks with 60 trials each are genuine evidence that this protocol helps, and the simulation ablations support the importance of both rules. The authors also deserve credit for acknowledging that recovery-only or correction-only data is unbalanced, and for showing that HG-DAgger-style data skews heavily toward corrections. The paper is clearly written, the interface design is practical, and the DROID analysis (only 3.68% of episodes contain recovery) is a nice motivating datapoint.\n\nThe soft spots are real but not fatal. First, the headline \"10x less data\" compares against ALOHA Unleashed and Seed GR-3 on different robots, model sizes, and evaluation protocols. The 89-hour estimate for prior work is based on an assumed 1 minute per trajectory, which is stated but is still an estimate. The in-house comparison against batched full demonstrations and HG-DAgger shows a 2x improvement, not 10x, and the 10x claim should be presented as a cross-paper rough comparison, not a controlled result. Second, the \"test-time scaling\" claim in Figure 10 is the weakest part. The linear relationship is computed across training rounds, which confounds recovery count with training progress. Later checkpoints both succeed more and happen to recover more, so the correlation does not demonstrate that adding recoveries at deployment causes higher success. The paper itself weakens this to a correlation in Section 6, but the abstract still sells it as a central contribution. This could be settled from existing raw data—check whether, within a fixed checkpoint, rollouts with more recoveries succeed more—so the authors should be pushed to provide that analysis. The verification-generation gap argument (recovery is easier than correction) is plausible but not directly tested; it is an intuition supported by the empirical wins, not a measured mechanism.\n\nWho should read this: anyone working on imitation learning for long-horizon manipulation, especially on data collection protocols. It is a useful practical contribution even if the scaling-law framing overreaches. I would happily referee a revised version; the core result is probably real and the fixes are tractable. The main ask is to reframe the quantitative claims and fix the confounded scaling analysis before publication.","headline":"RaC is a solid, well-executed dataset-composition paper with real robot experiments, but its boldest claims—10x data efficiency and linear test-time scaling—are partly resting on cross-paper apples-to-oranges comparisons and a confounded correlation.","tokens_in":23531,"tokens_out":623,"would_cite":true,"duration_ms":9276,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training on interventions that rewind the robot to a familiar state and then finish the sub-task lifts shirt-hanging success to 78.3% in about five hours of data, versus 75% from roughly 89 hours, and success scales with recovery count.","keywords":["recovery and correction","imitation learning","human-in-the-loop intervention","long-horizon manipulation","data efficiency","bimanual manipulation","test-time scaling","flow matching policy"],"falsifier":"Record the gripper pose right after each human recovery segment and measure its distance to the nearest demonstration-prefix state (via the visitation heatmap or a learned density model); if recovered states are routinely out-of-distribution, the recovery premise is violated. A complementary test: deploy the trained RaC policy while blocking its learned rewind actions and compare success — if success does not drop, recovery is not what carries the gain.","tokens_in":22610,"feed_emoji":"🤖","tokens_out":14575,"duration_ms":147811,"temperature":0.7,"pith_summary":"The paper's claim is that the plateau of imitation learning on long-horizon, contact-rich robot tasks is a data-composition problem: expert teleoperation records clean successes, so the policy never learns what to do after its own mistakes. RaC adds a second training phase in which a human operator takes over during policy rollouts and each intervention is structured into two parts — a recovery segment that rewinds the robot to a familiar, in-distribution state, and a corrective segment that completes the interrupted sub-task — after which the episode ends. Training a flow-matching diffusion transformer on this mixture, the paper reports, gives the policy a retry skill: on shirt hanging, airtight-lid sealing, takeout-box packing, and a simulated assembly task, RaC beats both batched full demonstrations and HG-DAgger-style interventions (corrections that push the task forward without a structured rewind) with steeper success-versus-data curves and roughly an order of magnitude less human data. The headline number is 78.3% success on shirt hanging from about five hours of data, where prior reported results needed eighty-nine hours for 75%. A reader should care because the protocol changes only how data is collected — the imitation objective and network are untouched — so the gains could ride on top of any demonstration-trained policy.","feed_headline":"Five hours of retry training beat 89 hours of robot demos","feed_subtitle":"Human rewinds plus corrections lift shirt-hanging success to 78 percent.","key_machinery":"The load-bearing mechanism is the structured intervention trajectory, defined by two rules. Rule 1 ('recover then correct') makes every human takeover start with a recovery segment — rewinding the grippers to a familiar state visited earlier in demonstrations — followed by a corrective segment that pushes the sub-task forward. Rule 2 ('termination after intervention') ends the episode once the intervention finishes. The rationale is a verification–generation gap: recovery targets a broad set of familiar states and is cheaper to learn than the precise corrective action, so retries amplify success. The trained object is a ~368M-parameter multimodal diffusion transformer (MM-DiT) fitted with a","core_discovery":"Central claim: recovery segments — trajectories that rewind the robot to a previously visited state — are a learnable skill, not wasted data. Because familiar past states form a broad target while goal states are narrow, recovery is cheaper to learn than the corrective action (a 'verification–generation gap'), so a policy that can recover can retry, and retries make failure decay exponentially. Two rules shape each human intervention: recover-then-correct, then terminate the episode. RaC beats batched full demonstrations and HG-DAgger-style interventions with steeper scaling and about an order of magnitude less data — 78.3% shirt-hanging success in ~5 hours versus 75% from ~89 hours — and su","pith_inferences":["The mechanism suggests a partially automatable loop the paper leaves open: near-miss states from a policy's own rollouts could seed recovery data, with the human operator reduced to labelling when a rewind is warranted — the heatmap guidance already points toward such a detector.","Figure 10's linear trend is measured across training rounds, not imposed at deployment; the exponential-retry argument in Section 4.1 predicts success should also rise with an explicit test-time retry budget N, which could be tested by granting a fixed number of rewind allowances per episode.","Rule 2 concentrates data on early sub-tasks, and the ablations show termination alone improves scaling; the paper attributes this to avoiding mixed human-policy state distributions, but early-subtask concentration is a plausible partial alternative explanation worth disentangling.","If recovery exploits a genuine verification–generation gap, RaC's advantage should narrow in tasks where the set of valid past states is as tight as the goal set — a testable boundary on where the protocol helps."],"forward_implications":["Imitation learning's long-horizon plateau is partly a data-composition effect: success-only demonstrations leave failure and out-of-distribution states under-covered, and rebalancing the trajectory mix toward recovery-and-correction changes the slope of success versus data.","Specialist long-horizon skills become dramatically cheaper: comparable shirt-hanging success (78.3%) is reached with roughly five hours of human data where prior reported results needed about eighty-nine.","RaC policies exhibit an action-space analogue of chain-of-thought scaling: performance is linearly correlated with the number of recovery maneuvers in a rollout, and successful rollouts are longer on average.","Because the imitation objective and architecture are unchanged, the recovery-and-correction data recipe can be layered onto other demonstration-trained policies, including generalist vision-language-action models, and the paper hypothesizes it gives online reinforcement learning better starting points."],"supporting_citations":[{"why":"Reports 75% shirt-hanging success from roughly 89 hours (5,345 demonstrations); this is the data-efficiency baseline RaC claims to beat with about five hours.","marker":"[50]"},{"why":"The human-in-the-loop intervention baseline RaC is compared against and departs from: HG-DAgger-style corrections push the task forward without a structured recovery phase.","marker":"[22]"},{"why":"Closest prior protocol — operators backtrack to earlier visual states to collect failure-recovery-success sequences; RaC extends the idea from offline RL to scaling imitation learning.","marker":"[3]"},{"why":"Supplies the verification–generation gap concept used to argue that learning to recover is cheaper than learning the corrective action.","marker":"[38]"},{"why":"Provides the o1-style test-time scaling analogy that frames RaC's claimed linear scaling of success with recovery count.","marker":"[35]"},{"why":"Dataset analysis showing only 3.68% of episodes contain a recovery segment, motivating the explicit collection and scaling of recovery behavior.","marker":"[23]"},{"why":"The multi-modal diffusion transformer (MM-DiT) architecture used to train all policies in the experiments.","marker":"[12]"},{"why":"Provides the flow-matching training objective (Equation 4.1) used for imitation learning in RaC.","marker":"[28]"}],"fun_headline_variants":["Rewind-and-correct training cuts robot data needs 10x","Teach robots to recover, not just imitate: 10x less data","78% success from 5 hours of retry vs 75% from 89 hours of demos","Recovery maneuvers: the key to scaling robot skills","Human rewinds turn robot failures into skill gains"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a human operator's rewind returns the robot to a state that genuinely looks like the demonstration data, and that learning to recover is materially easier than learning the correction itself; if either gives way, the data-efficiency gain and the retry mechanism do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Rewind-and-correct training cuts robot data needs 10x","Teach robots to recover, not just imitate: 10x less data","78% success from 5 hours of retry vs 75% from 89 hours of demos","Recovery maneuvers: the key to scaling robot skills","Human rewinds turn robot failures into skill gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3285,"prompt_tokens":800,"completion_tokens":2485,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2391}},"tokens_in":544,"tokens_out":2485,"duration_ms":23195,"temperature":1.0,"reasoning_tokens":2391,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:25:46.865799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the gripper pose right after each human recovery segment and measure its distance to the nearest demonstration-prefix state (via the visitation heatmap or a learned density model); if recovered states are routinely out-of-distribution, the recovery premise is violated. A complementary test: deploy the trained RaC policy while blocking its learned rewind actions and compare success — if success does not drop, recovery is not what carries the gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dataset analysis showing only 3.68% of episodes contain a recovery segment, motivating the explicit collection and scaling of recovery behavior."}],"review_version":1}