Pith. sign in

REVIEW 4 major objections 5 minor 16 cited by

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.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 21:25 UTC pith:2SZANLND

load-bearing objection 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. the 4 major comments →

arxiv 2509.07953 v1 pith:2SZANLND submitted 2025-09-09 cs.RO cs.LG

RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

Zheyuan Hu , Robyn Wu , Naveen Enock , Jasmine Li , Riya Kadakia , Zackory Erickson , Aviral Kumar This is my paper
classification cs.RO cs.LG
keywords recovery and correctionimitation learninghuman-in-the-loop interventionlong-horizon manipulationdata efficiencybimanual manipulationtest-time scalingflow matching policy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 5.3, Result 2 / Figure 10] 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.
  2. [Table 1 / Appendix D] 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.
  3. [Section 4.1 / 4.2 / 4.3] 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.
  4. [Algorithm 1 / Section 4.4] 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.
minor comments (5)
  1. [Section 5.4, first paragraph] 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.
  2. [Figure 10] 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.
  3. [Section 5.3, Result 3 / Figure 11] 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.
  4. [Algorithm 1] Line 8 has a typo: "Pair each recovery a correction" should read "Pair each recovery with a correction."
  5. [Section 4.2, DROID analysis] 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).

Circularity Check

1 steps flagged

Figure 10's 'test-time scaling' is a fitted correlation confounded with training round, not an independent prediction.

specific steps
  1. fitted input called prediction [Section 5.3, Result 2 and Figure 10; abstract]
    "To do so, we analyze the subset of evaluation rollouts that successfully solve all sub-tasks across different rounds, and annotate each rollout with the number of recovery attempts it contains. In Figure 10, we show the average number of recovery segments observed against the task success rates. The correlation coefficients r indicate a linear relationship between the task success and recovery frequency."

    The 'test-time scaling' claim is not an independent prediction: the x-axis (average recoveries per successful trajectory) and the y-axis (success rate) are both computed from the same evaluation rollouts, and the plotted points come from different training rounds. Recovery count is never controlled at deployment; it is an emergent property of later checkpoints. The linear slope in Figure 10 is fitted to these same points and then restated as 'RaC enables test-time scaling.' Section 6 concedes this is only a correlation. Thus the scaling law reduces to a fit of the data it is supposed to explain, and the causal claim that more recovery maneuvers cause higher success is unsupported by the presented evidence.

full rationale

The main data-efficiency result (Table 1, Figure 8) is an independent empirical comparison against external prior work (ALOHA Unleashed, Seed GR-3) and in-house baselines; it is not circular. The method is a data-collection protocol, not a derivation from its own outputs. The only significant circularity concern is the 'test-time scaling' claim: the scaling line is fitted to the same data that provides both axes and is confounded with training round, so the central 'o1-style' scaling contribution partially reduces to a fit presented as a prediction. The VG-gap citation [38] is a self-citation (Kumar co-author) but is used as an analogy rather than as a load-bearing theorem, so it does not independently raise the score.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central mechanism of RaC is not a mathematical derivation; it rests on empirical assumptions about human-guided recovery data and on a fitted linear-scaling description. No new physical or algorithmic entities are introduced. The improvements claimed over prior work depend on a cross-paper comparison that assumes task similarity and a data-length conversion.

free parameters (3)
  • linear scaling fit coefficients for Figure 10 = y = 1.169x - 0.387 (shirt), y = 0.179x + 0.099 (lid), y = 0.253x + 0.111 (box)
    Least-squares lines fit to the same success-rate vs recovery-count data used to claim test-time scaling; the linear law is a fitted description, not an independent prediction.
  • assumed demo length conversion for ALOHA Unleashed = 1 minute per trajectory
    Used in Appendix D to estimate 89 hours of ShirtEasy training data from 5345 trajectories; if average trajectory length differs, the 10x comparison shifts.
  • per-task round budgets K and base budget R0 = K = 6 (shirt), 10 (lid), 9 (box); R0 varies by task
    Experimental design choices selected by hand; they shape the scaling curves but are not fitted to a target result.
axioms (5)
  • domain assumption Human operators can reliably identify and demonstrate recovery to in-distribution states during rollouts
    Sections 4.2 and 4.3; the heatmap overlay is the only guide and no quantitative check verifies that recovered states match the demonstration distribution.
  • domain assumption Learning recovery is easier than learning correction for these tasks (verification-generation gap)
    Section 4.1 argues heuristically from broad initial states and narrow goal states; this is not proven or independently measured.
  • domain assumption Gripper-visitation frequency from an overhead camera is a faithful proxy for the in-distribution state set
    Section 4.3 uses SAM2-based heatmaps to guide operators; assumes visitation frequency captures what the policy needs.
  • standard math Flow-matching / MM-DiT imitation learning is a reliable policy optimization procedure
    Section 4.4; standard flow matching objective Equation 4.1 is adopted without new justification.
  • domain assumption The ShirtEasy setup from ALOHA Unleashed is similar enough for a direct data-efficiency comparison
    Appendix D asserts similarity but the robot, model size, and evaluation details differ from RaC's setup.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction." pith.science (2026). https://pith.science/paper/2SZANLND

@misc{pith2026250907953,
  author       = {Pith},
  title        = {Pith review of: RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SZANLND}},
  note         = {Machine review of arXiv:2509.07953}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Modern paradigms for robot imitation train expressive policy architectures on large amounts of human demonstration data. Yet performance on contact-rich, deformable-object, and long-horizon tasks plateau far below perfect execution, even with thousands of expert demonstrations. This is due to the inefficiency of existing ``expert'' data collection procedures based on human teleoperation. To address this issue, we introduce RaC, a new phase of training on human-in-the-loop rollouts after imitation learning pre-training. In RaC, we fine-tune a robotic policy on human intervention trajectories that illustrate recovery and correction behaviors. Specifically, during a policy rollout, human operators intervene when failure appears imminent, first rewinding the robot back to a familiar, in-distribution state and then providing a corrective segment that completes the current sub-task. Training on this data composition expands the robotic skill repertoire to include retry and adaptation behaviors, which we show are crucial for boosting both efficiency and robustness on long-horizon tasks. Across three real-world bimanual control tasks: shirt hanging, airtight container lid sealing, takeout box packing, and a simulated assembly task, RaC outperforms the prior state-of-the-art using 10$\times$ less data collection time and samples. We also show that RaC enables test-time scaling: the performance of the trained RaC policy scales linearly in the number of recovery maneuvers it exhibits. Videos of the learned policy are available at https://rac-scaling-robot.github.io/.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 16 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Closing the Lab-to-Store Gap: A Data-Efficient Post-Training and Experience-Driven Learning VLA Framework for Retail Humanoids

    cs.RO 2026-07 conditional novelty 6.0

    On a real humanoid restocking chips, a data-efficient post-training pipeline raised success from 0% to 32%, one experience-driven refinement pass to 42%, and a second pass dropped it to 22%.

  2. EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration

    cs.RO 2026-07 conditional novelty 6.0

    EgoRecovery co-trains robot recovery policies from egocentric human recovery videos by sharing a gated corrective-intent bottleneck, raising average closed-loop recovery success from 52.5% (robot-only) to 85% on four ...

  3. VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon

    cs.RO 2026-07 unverdicted novelty 6.0

    VLA-Corrector adds a detect-and-correct inference layer using a latent vision monitor and online gradient guidance to enable adaptive action horizons in chunked VLA policies.

  4. ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies

    cs.LG 2026-06 unverdicted novelty 6.0

    ReGuide is a self-improving framework that uses phase-conditioned guidance to generate corrective rollouts and absorbs successful ones back into diffusion policy training, yielding 1.3-7.7x success gains on Robomimic tasks.

  5. Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through Corrections

    cs.RO 2026-06 unverdicted novelty 6.0

    SDP constructs sets of desired action-chunks from human correction pairs and trains diffusion policies to align with those sets, yielding better performance and robustness than standard behavior cloning on robotic tasks.

  6. Hand-in-the-Loop: Improving VLA Policies for Dexterous Manipulation via Seamless Hand-Arm Intervention

    cs.RO 2026-05 unverdicted novelty 6.0

    HandITL blends human intent with policy execution to eliminate gesture jumps in dexterous VLA interventions, cutting jitter by 99.8%, grasp failures by 87.5%, and yielding 19% better refined policies.

  7. Hand-in-the-Loop: Improving VLA Policies for Dexterous Manipulation via Seamless Hand-Arm Intervention

    cs.RO 2026-05 unverdicted novelty 6.0

    HandITL enables seamless human intervention in VLA policies for bimanual dexterous manipulation, cutting jitter by 99.8% and improving refined policies by 19% over standard teleoperation.

  8. UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation

    cs.RO 2026-05 conditional novelty 6.0

    By inverting a frozen flow-matching robot policy with fixed-point iteration, UniSteer converts human corrections into noise-space supervision and lifts real-world VLA adaptation success from 20% to 90% in about 66 min...

  9. UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation

    cs.RO 2026-05 unverdicted novelty 6.0

    UniSteer unifies human corrective actions and noise-space RL for VLA adaptation by inverting actions to noise targets, raising success rates from 20% to 90% in 66 minutes across four real-world manipulation tasks.

  10. TAMEn: Tactile-Aware Manipulation Engine for Closed-Loop Data Collection in Contact-Rich Tasks

    cs.RO 2026-04 unverdicted novelty 6.0

    TAMEn supplies a cross-morphology wearable interface and pyramid-structured visuo-tactile data regime that raises bimanual manipulation success rates from 34% to 75% via closed-loop collection.

  11. ARM: Advantage Reward Modeling for Long-Horizon Manipulation

    cs.RO 2026-04 unverdicted novelty 6.0

    ARM trains reward models on Progressive/Regressive/Stagnant labels to enable adaptive reweighting in offline RL, reaching 99.4% success on towel-folding with minimal human intervention.

  12. RISE: Self-Improving Robot Policy with Compositional World Model

    cs.RO 2026-02 unverdicted novelty 6.0

    RISE combines a controllable dynamics model and progress value model into a closed-loop self-improving pipeline that updates robot policies entirely in imagination, reporting over 35% absolute gains on three real-world tasks.

  13. $\pi^{*}_{0.6}$: a VLA That Learns From Experience

    cs.LG 2025-11 unverdicted novelty 6.0

    RECAP enables a generalist VLA to self-improve via advantage-conditioned RL on mixed real-world data, more than doubling throughput and halving failure rates on hard manipulation tasks.

  14. DexPIE: Stable Dexterous Policy Improvement from Real-World Experience

    cs.RO 2026-06 unverdicted novelty 5.0

    DexPIE improves dexterous manipulation success rates by 37% over demo policies via real-world experience collection with adapted intervention, multi-stage DAgger, asynchronous relative-action inference, and optimality...

  15. DeMaVLA: A Vision-Language-Action Foundation Model for Generalizable Deformable Manipulation

    cs.RO 2026-05 unverdicted novelty 5.0

    DeMaVLA is a VLA foundation model using a pruned action expert and flow matching, pre-trained on 5000 hours of real demonstrations and post-trained on multi-task folding data with human-in-the-loop correction, reporti...

  16. FAR: Failure-Aware Retry for Test-Time Recovery and Continual Policy Improvement

    cs.RO 2026-07 unverdicted novelty 4.0

    FAR combines failure-contrastive preference adaptation with action perturbations for test-time recovery and continual policy improvement, reporting 17.6% and 11.7% success gains over diffusion policies in simulation a...

Reference graph

Works this paper leans on

53 extracted references · 13 canonical work pages · cited by 14 Pith papers · 3 internal anchors

  1. [1]

    JUICER: Data-Efficient Imitation Learning for Robotic Assembly

    Lars Ankile, Anthony Simeonov, Idan Shenfeld, and Pulkit Agrawal. Juicer: Data-efficient imitation learning for robotic assembly, 2024. URLhttps://arxiv.org/abs/2404.03729

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    URLhttps://arxiv.org/abs/2410.24164

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.𝜋0: A visio...

    Pith/arXiv arXiv 2024

  3. [3]

    Visual backtracking teleoperation: A data collection protocol for offline image-based reinforcement learning

    David Brandfonbrener, Stephen Tu, Avi Singh, Stefan Welker, Chad Boodoo, Nikolai Matni, and Jake Varley. Visual backtracking teleoperation: A data collection protocol for offline image-based reinforcement learning. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11336–11342, 2023. doi: 10.1109/ICRA48891.2023.10161096

    arXiv 2023

  4. [4]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, De...

    Pith/arXiv arXiv 2022

  5. [5]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gon- zalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Y...

    Pith/arXiv arXiv 2023

  6. [6]

    Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems

    Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Xindong He, Xu Huang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025

    2025

  7. [7]

    Gr-3 technical report, 2025

    Chilam Cheang, Sijin Chen, Zhongren Cui, Yingdong Hu, Liqun Huang, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Xiao Ma, Hao Niu, Wenxuan Ou, Wanli Peng, Zeyu Ren, Haixin Shi, Jiawen Tian, Hongtao Wu, Xin Xiao, Yuyang Xiao, Jiafeng Xu, and Yichu Yang. Gr-3 technical report, 2025. URL https://arxiv.org/abs/2507.15493

    Pith/arXiv arXiv 2025

  8. [8]

    RMPflow: A Computational Graph for Automatic Motion Policy Generation

    Ching-An Cheng, Mustafa Mukadam, Jan Issac, Stan Birchfield, Dieter Fox, Byron Boots, and 16 RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction Nathan Ratliff. Rmpflow: A computational graph for automatic motion policy generation, 2019. URLhttps://arxiv.org/abs/1811.07049

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  9. [9]

    Diffusionpolicy: Visuomotorpolicylearningviaactiondiffusion.TheInternational Journal of Robotics Research, 2024

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, andShuranSong. Diffusionpolicy: Visuomotorpolicylearningviaactiondiffusion.TheInternational Journal of Robotics Research, 2024

    work page 2024

  10. [10]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots

    Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. InProceedings of Robotics: Science and Systems (RSS), 2024

    work page 2024

  11. [11]

    Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Mad- dukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov, Anikait Singh, Animesh Garg, ...

    Pith/arXiv arXiv 2023

  12. [12]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URLhttps://arxiv.org/abs/2403.03206

    Pith/arXiv arXiv 2024

  13. [13]

    D4rl: Datasets for deep data-driven reinforcement learning.arXiv preprint arXiv:2004.07219, 2020

    Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4rl: Datasets for deep data-driven reinforcement learning.arXiv preprint arXiv:2004.07219, 2020

    Pith/arXiv arXiv 2004

  14. [14]

    Efficientdatacollectionforrobotic manipulation via compositional generalization, 2024

    JensenGao, AnnieXie, TedXiao, ChelseaFinn, andDorsaSadigh. Efficientdatacollectionforrobotic manipulation via compositional generalization, 2024. URLhttps://arxiv.org/abs/2403.05110

    Pith/arXiv arXiv 2024

  15. [15]

    Why generalization in rl is difficult: Epistemic pomdps and implicit partial observability.Advances in neural information processing systems, 34:25502–25515, 2021

    Dibya Ghosh, Jad Rahme, Aviral Kumar, Amy Zhang, Ryan P Adams, and Sergey Levine. Why generalization in rl is difficult: Epistemic pomdps and implicit partial observability.Advances in neural information processing systems, 34:25502–25515, 2021

    2021

  16. [16]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  18. [18]

    Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

    Pith/arXiv arXiv 2022

  19. [19]

    Intervengen: In- terventional data generation for robust and data-efficient robot imitation learning, 2024

    Ryan Hoque, Ajay Mandlekar, Caelan Garrett, Ken Goldberg, and Dieter Fox. Intervengen: In- terventional data generation for robust and data-efficient robot imitation learning, 2024. URL https://arxiv.org/abs/2405.01472

    Pith/arXiv arXiv 2024

  20. [20]

    Scaling laws for neural language models.arXiv preprint, 2020

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint, 2020. 18 RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

    work page 2020

  21. [21]

    Ccil: Continuity-based data augmentation for corrective imitation learning.arXiv preprint arXiv:2310.12972, 2023

    Liyiming Ke, Yunchu Zhang, Abhay Deshpande, Siddhartha Srinivasa, and Abhishek Gupta. Ccil: Continuity-based data augmentation for corrective imitation learning.arXiv preprint arXiv:2310.12972, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Hg-dagger: Interactive imitation learning with human experts

    Michael Kelly, Chelsea Sidrane, Katherine Driggs-Campbell, and Mykel J Kochenderfer. Hg-dagger: Interactive imitation learning with human experts. In2019 International Conference on Robotics and Automation (ICRA), pages 8077–8083. IEEE, 2019

    2019

  23. [23]

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, Peter David Fagan, Joey Hejna, Masha Itkina, Marion Lepert, Yecheng Jason Ma, Patrick Tree Miller, Jimmy Wu, Suneel Belkhale, Shivin Dass, Huy Ha, Arhan Jain, Abraham Lee, You...

    work page 2024

  24. [24]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    How to train your robots? the impact of demonstration modality on imitation learning.arXiv preprint arXiv:2503.07017, 2025

    Haozhuo Li, Yuchen Cui, and Dorsa Sadigh. How to train your robots? the impact of demonstration modality on imitation learning.arXiv preprint arXiv:2503.07017, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Taming VR Teleoperation and Learning from Demonstration for Multi-Task Bimanual Table Service Manipulation

    Weize Li, Zhengxiao Han, Lixin Xu, Xiangyu Chen, Harrison Bounds, Chenrui Zhang, and Yifan Xu. Taming vr teleoperation and learning from demonstration for multi-task bimanual table service manipulation, 2025. URLhttps://arxiv.org/abs/2508.14542

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Data scaling laws in imitation learning for robotic manipulation.arXiv preprint arXiv:2410.18647, 2024

    Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation.arXiv preprint arXiv:2410.18647, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Robot learning on the job: Human-in-the-loop autonomy and learning during deployment.The International Journal of Robotics Research, page 02783649241273901, 2022

    Huihan Liu, Soroush Nasiriany, Lance Zhang, Zhiyao Bao, and Yuke Zhu. Robot learning on the job: Human-in-the-loop autonomy and learning during deployment.The International Journal of Robotics Research, page 02783649241273901, 2022. 19 RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

    work page 2022

  30. [30]

    Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Serl: A software suite for sample-efficient robotic reinforcement learning

    Jianlan Luo, Zheyuan Hu, Charles Xu, You Liang Tan, Jacob Berg, Archit Sharma, Stefan Schaal, Chelsea Finn, Abhishek Gupta, and Sergey Levine. Serl: A software suite for sample-efficient robotic reinforcement learning. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16961–16969, 2024. doi: 10.1109/ICRA57147.2024.10610040

    arXiv 2024

  32. [32]

    Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2025

    Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2025. URLhttps://arxiv.org/abs/2410.21845

    Pith/arXiv arXiv 2025

  33. [33]

    Human-in-the-loop imitation learning using remote teleoperation.arXiv preprint arXiv:2012.06733, 2020

    Ajay Mandlekar, Danfei Xu, Roberto Martín-Martín, Yuke Zhu, Li Fei-Fei, and Silvio Savarese. Human-in-the-loop imitation learning using remote teleoperation.arXiv preprint arXiv:2012.06733, 2020

    Pith/arXiv arXiv 2012

  34. [34]

    Octo: An open-source generalist robot policy

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Charles Xu, Jianlan Luo, Tobias Kreiman, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. InProceedings of Robotics: Science and...

    2024

  35. [35]

    Learning to reason with llms, 2024

    OpenAI. Learning to reason with llms, 2024. URL https://openai.com/index/ learning-to-reason-with-llms/

    2024

  36. [36]

    Sam 2: Segment anything in images and videos, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. URLhttps://arxiv.org/a...

    Pith/arXiv arXiv 2024

  37. [37]

    Areductionofimitationlearningandstructured prediction to no-regret online learning

    StephaneRoss, GeoffreyGordon, andDrewBagnell. Areductionofimitationlearningandstructured prediction to no-regret online learning. In Geoffrey Gordon, David Dunson, and Miroslav Dudík, editors,Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 ofProceedings of Machine Learning Research, pages 627–635...

    work page 2011

  38. [38]

    Scaling test-time compute without verification or rl is suboptimal.arXiv preprint arXiv:2502.12118, 2025

    Amrith Setlur, Nived Rajaraman, Sergey Levine, and Aviral Kumar. Scaling test-time compute without verification or rl is suboptimal.arXiv preprint arXiv:2502.12118, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Scaling LLM test-time compute optimally can be more effective than scaling model parameters.ICLR, 2025

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling model parameters.ICLR, 2025

    work page 2025

  40. [40]

    Latent policy barrier: Learning robust visuomotor policies by staying in-distribution, 2025

    Zhanyi Sun and Shuran Song. Latent policy barrier: Learning robust visuomotor policies by staying in-distribution, 2025. URLhttps://arxiv.org/abs/2508.05941

    Pith/arXiv arXiv 2025

  41. [41]

    Alex Hofer, Jan Humplik, Atil Iscen, Mithun George Jacob, Deepali Jain, Ryan Julian, Dmitry Kalashnikov, M

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gon- zalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, Steven Bohez, Konstantinos Bousmalis, Anthony Brohan, Thomas Buschmann, Arunkumar Byra- van, Serkan Cabi, Ken Caluwaerts, Federico Casarini, Oscar Chang, Jose Enri...

    Pith/arXiv arXiv 2025

  42. [42]

    A careful examination of large behavior models for multitask dexterous manipulation,

    TRI LBM Team, Jose Barreiros, Andrew Beaulieu, Aditya Bhat, Rick Cory, Eric Cousineau, Hongkai Dai, Ching-Hsin Fang, Kunimatsu Hashimoto, Muhammad Zubair Irshad, Masha Itkina, Naveen Kuppuswamy, Kuan-Hui Lee, Katherine Liu, Dale McConachie, Ian McMahon, Haruki Nishimura, Calder Phillips-Grafflin, Charles Richter, Paarth Shah, Krishnan Srinivasan, Blake Wu...

    work page

  43. [43]

    Bridgedata v2: A dataset for robot learning at scale

    Homer Walke, Kevin Black, Abraham Lee, Moo Jin Kim, Max Du, Chongyi Zheng, Tony Zhao, Philippe Hansen-Estruch, Quan Vuong, Andre He, Vivek Myers, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata v2: A dataset for robot learning at scale. InConference on Robot Learning (CoRL), 2023

    2023

  44. [44]

    Genie centurion: Accelerating scalable real-world robot training with human rewind-and-refine guidance, 2025

    Wenhao Wang, Jianheng Song, Chiming Liu, Jiayao Ma, Siyuan Feng, Jingyuan Wang, Yuxin Jiang, Kylin Chen, Sikang Zhan, Yi Wang, Tong Meng, Modi Shi, Xindong He, Guanghui Ren, Yang Yang, and Maoqing Yao. Genie centurion: Accelerating scalable real-world robot training with human rewind-and-refine guidance, 2025. URLhttps://arxiv.org/abs/2505.18793

    work page arXiv 2025

  45. [45]

    Octothinker: Mid-training incentivizes reinforcement learning scaling, 2025

    Zengzhi Wang, Fan Zhou, Xuefeng Li, and Pengfei Liu. Octothinker: Mid-training incentivizes reinforcement learning scaling, 2025. URLhttps://arxiv.org/abs/2506.20512. 21 RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

    Pith/arXiv arXiv 2025

  46. [46]

    Robocopilot: Human-in-the-loopinteractiveimitationlearningforrobotmanipulation,

    Philipp Wu, Yide Shentu, Qiayuan Liao, Ding Jin, Menglong Guo, Koushil Sreenath, Xingyu Lin, and PieterAbbeel. Robocopilot: Human-in-the-loopinteractiveimitationlearningforrobotmanipulation,

    work page

  47. [47]

    Compliant residual dagger: Improving real-world contact-rich manipulation with human corrections, 2025

    Xiaomeng Xu, Yifan Hou, Zeyi Liu, and Shuran Song. Compliant residual dagger: Improving real-world contact-rich manipulation with human corrections, 2025. URLhttps://arxiv.org/ abs/2506.16685

    arXiv 2025

  48. [48]

    URLhttps://arxiv.org/abs/2503.07771

    Pith/arXiv arXiv

  49. [49]

    Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    Ren, Dhruv Shah, and Anirudha Majumdar

    Lihan Zha, Apurva Badithela, Michael Zhang, Justin Lidard, Jeremy Bao, Emily Zhou, David Snyder, Allen Z. Ren, Dhruv Shah, and Anirudha Majumdar. Guiding data collection via factored scaling curves, 2025. URLhttps://arxiv.org/abs/2505.07728

    Pith/arXiv arXiv 2025

  51. [51]

    needs to rotate its mobile base from the table to the drying rack to hang the clothes

    Zhaxizhuoma, Kehui Liu, Chuyue Guan, Zhongjie Jia, Ziniu Wu, Xin Liu, Tianyu Wang, Shuai Liang, Pengan Chen, Pingrui Zhang, Haoming Song, Delin Qu, Dong Wang, Zhigang Wang, Nieqing Cao, Yan Ding, Bin Zhao, and Xuelong Li. Fastumi: A scalable and hardware-independent universal manipulation interface with dataset, 2025. URLhttps://arxiv.org/abs/2409.19499. ...

    Pith/arXiv arXiv 2025

  52. [52]

    Zhao, Jonathan Tompson, Danny Driess, Pete Florence, Kamyar Ghasemipour, Chelsea Finn, and Ayzaan Wahid

    Tony Z. Zhao, Jonathan Tompson, Danny Driess, Pete Florence, Kamyar Ghasemipour, Chelsea Finn, and Ayzaan Wahid. Aloha unleashed: A simple recipe for robot dexterity, 2024. URL https://arxiv.org/abs/2410.13126

    Pith/arXiv arXiv 2024

  53. [2025]

    URLhttps://arxiv.org/abs/2507.05331

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.