Pith. sign in

REVIEW 4 major objections 5 minor 56 references

This paper shows that focusing dense robotic data collection and offline reinforcement learning on the short contact-rich segment of a task, while using motion planning elsewhere, raises average real-world success to 96% from as little as 2

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 →

A modular framework combining autonomous dense data collection at critical contact segments with offline RL and motion planning for free-space achieves 96% average success across four real-world tasks with only 2–2.5 h of data.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A solid, practical modular pipeline for contact-rich manipulation, with a real but addressable concern about how task success is defined and measured. the 4 major comments →

arxiv 2607.15982 v1 pith:273JA2TU submitted 2026-07-17 cs.RO

Data and Learning Where it Matters for Contact-Rich Manipulation

classification cs.RO
keywords contact-rich manipulationoffline reinforcement learningautonomous data collectionmotion planningpose estimationQ-function success detectiondata efficiencygeneralization
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 reading

The paper is trying to establish that the reason end-to-end robot learning is brittle at high-precision, contact-rich tasks is not primarily data scale, but data structure: those tasks fail almost entirely inside a short critical segment of tight contact, while the surrounding free-space motion is easy. The authors propose collecting dense autonomous data only inside that segment and training an offline deep-RL policy on it, then chaining that policy with off-the-shelf pose estimation and motion planning for the rest. Across four real tasks—shelf stocking, Lego stacking, and two fan-cover assemblies—this achieves 94–98% success (96% average) from only 2 to 2.5 hours of autonomous collection, versus at most 55% for the strongest end-to-end baselines, and it stays above 85% in out-of-distribution setups. A sympathetic reader would care because it suggests a cheap, teleoperator-free recipe for high-precision manipulation, and a shift in where the robot-learning community should spend its data budget.

Core claim

On the paper's own terms, the central claim is that success in contact-rich manipulation is determined by where data is collected, not just how much. By detecting the pose-estimation uncertainty to define the critical segment, replaying a single demonstration to its start, and then running a mixed greedy-random exploration inside that segment, the authors autonomously generate dense, high-success-rate datasets for offline RL. The resulting policy—trained offline with SAC, an ensemble of layer-normalized critics, and a sparse pose-distance reward—solves the insertion, and the Q-function itself doubles as a success detector for switching back to planning. The paper reports that this pipeline a

What carries the argument

The load-bearing mechanism is the decomposition of the task into a free-space planning phase and a critical contact segment, with the boundary set by pose-estimation uncertainty. Inside that segment, an explorative policy mixes uniform random actions with greedy goal-seeking actions (with a randomized grasp offset and a safety-filter sphere) to generate dense, balanced offline data. Learning is offline SAC with randomized ensembled double Q-learning and layer normalization on the critics; a sparse reward (R_success when the end-effector reaches the reference goal pose) trains the Q-function, which is then thresholded (with hysteresis) to detect successful insertion during deployment. Hybrid

Load-bearing premise

The load-bearing premise is that the sparse reward—triggered when the end-effector pose reaches the reference trajectory's goal pose—actually means the task succeeded (full insertion of a deformable cover or box); if that pose proxy mislabels transitions, the Q-function, the success detector, and the reported 96% all inherit the error.

What would settle it

Relabel the offline dataset's success/failure using a ground-truth check (e.g., a human label or a force-signature insertion detector) instead of the pose-distance reward, retrain the same offline RL pipeline, and measure success on the same real tasks. If the gap between proxy-labeled and ground-truth-labeled performance is large, the pose-proxy reward is doing the work, not the data collection.

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

If this is right

  • Robot data collection can be focused on the few seconds of contact that decide task success, cutting collection time from many hours to 2–2.5 hours per task and removing dependence on teleoperator skill.
  • A Q-function trained with a sparse success reward can double as a task-success classifier, eliminating the need for separately trained reward or success detectors.
  • Local vision policies (wrist camera cropped to the contact region) remain robust under scene-level distractors and OOD changes because the learned part sees only a narrow window.
  • Fusing force-torque sensing with vision in the critical segment improves robustness and reduces maximum applied forces by about 49% compared to baselines.
  • The framework turns data collection into a largely autonomous process: a single kinesthetic demonstration plus a fixed scene suffices, with only occasional manual resets.

Where Pith is reading between the lines

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

  • The same focus principle could extend beyond assembly: any task with a short, contact-bound critical phase (cable insertion, snap-fitting, medical manipulation) may inherit the 2-hour data efficiency if the segment can be localized.
  • As pose estimation gets more accurate, the critical segment shrinks, implying the data requirement should drop further than the paper's already low 2–2.5 hours; this is a quantitative prediction one could test by varying pose quality.
  • The Q-function success detector is trained entirely on the sparse reward; one could generalize the switching idea to chain several learned critical segments, making the framework applicable to longer-horizon tasks.
  • The paper leaves open whether the same recipe works when free-space motion is not easy (e.g., cluttered scenes or moving targets); an obvious stress-test is to randomize the pre-contact approach more aggressively than the OOD scenarios shown.
Share X Bluesky LinkedIn Reddit HN

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 a compositional pipeline for contact-rich manipulation in which a single kinesthetic demonstration is replayed up to the contact-rich 'critical segment'; an autonomous greedy-random exploration policy then collects dense data at that segment while a planned free-space phase is handled by pose estimation and motion planning. Offline deep RL (SAC with ensembled critics) trains a policy for the critical segment from a sparse Cartesian-pose proximity reward, and in deployment the same learned Q-function is used to detect insertion success and terminate the policy. Across four real-world tasks (shelf stocking, Lego stacking, fan cover, fan cover hard), 2-2.5 hours of data collection yield an average success rate of 96% versus 55% for the strongest baseline, with improved out-of-distribution robustness and lower interaction forces. The paper also provides a simulation ablation of exploration probability, pose-estimation uncertainty, and sensor modalities.

Significance. If the headline result holds, this is a practically important contribution: it shows that targeted, largely autonomous data collection at the contact-rich segment can match or exceed end-to-end methods with far less data, and that offline RL with Q-function-based termination can be a credible alternative to hand-designed success detectors. The paper is refreshingly explicit about its data collection scheme and parameter settings, and the promise to release datasets and videos would make the comparison reproducible. However, the central success metric rests on an unvalidated Cartesian-pose reward proxy that also trains the Q-function used for success classification, and the real-world evaluations lack confidence intervals and multiple training seeds. These issues must be addressed before the quantitative claims can be considered established.

major comments (4)
  1. [§3.2, §3.4, Appendix A.4] The only task-completion signal during data collection is the sparse reward r_t = R_success * 1(||s_t - (s_g + Δs_grasp)||_2 ≤ ε), a Cartesian end-effector pose proximity to the reference trajectory plus a randomized offset. In deployment, the Q-function trained on exactly this reward is used as the success classifier (Eq. (4) and Eqs. (6)-(9)). The reported success rates are therefore not validated against an external ground-truth definition of task completion (e.g., full seating of the fan cover or full insertion of the salt box). Appendix A.4, Fig. 6 shows only Q-value separation for trajectories labeled successful/truncated using this same proxy, so it does not break the circularity. Please (a) report ε and the distribution of Δs_grasp, (b) compare the Q-function success labels against manual/human labels on held-out trials, and (c) report false-positive and false-negative rates for
  2. [Tables 2 and 3] All real-world policy evaluations use 50 (ID) or 20 (OOD) rollouts, but no confidence intervals are reported and only one trained policy per task is evaluated. The abstract claim of 96% vs 55% should be accompanied by binomial confidence intervals, and the paper should report variance over at least three policy training seeds. The simulated ablation (Fig. 4) uses three seeds, so the same standard should be applied to the hardware results that support the headline claim.
  3. [Abstract and Table 2] The statement 'compared to the strongest baseline at 55%' is based on task-averaged success. Table 2 shows MP&PE reaches 98% on shelf stocking, exactly matching Ours, while the average advantage comes mainly from fan cover and Lego. This should be stated honestly: the per-task comparison is not uniformly favorable. Please report per-task matched comparisons and clarify that the 55% is the strongest baseline on average.
  4. [§4.3, Table 1] The abstract and contributions describe the data collection as 'autonomous', but §4.3 says it 'required a few interventions when the scene was not reset correctly.' The number, duration, and type of interventions are not quantified. Please report the operator intervention rate and the wall-clock time actually spent by the human during the 2-2.5 hours, so the reader can judge the claim of eliminating reliance on a teleoperator.
minor comments (5)
  1. [Appendix A.7] The statement 'for fair comparison, we applied the same procedure to all baselines and our method' is followed by 'We simplified the shelf stocking task slightly by assuming the yellow salt package was pregrasped.' Clarify whether the pregrasped simplification applies only to HIL-SERL or to all methods in that task.
  2. [§3.4 and Appendix A.4] The main text introduces λ_success in Eq. (4) and says 'hysteresis-style detection' is used, but the threshold values λ_high = 0.8R_success and λ_low = 0.6R_success appear only in Appendix A.4. The notation and combination logic should be aligned in the main text.
  3. [Table 1] The Fan Cover (hard) row is incomplete: the entry '3' under Ours does not match the other columns, and the footnote says no additional data were collected. The table should be formatted consistently.
  4. [References] References [29] and [31] appear to be the same paper ('Force policy: Learning hybrid force-position control policy under interaction frame for contact-rich manipulation'). Please deduplicate.
  5. [§3.2 and Table 4] The success threshold ε and the force threshold F_threshold used for policy switching are never reported. Table 4 lists many hyperparameters but omits these two, which are needed to reproduce the experiments.

Circularity Check

0 steps flagged

No significant circularity: headline success rates are externally evaluated by physical task completion, not by the trained Q-function, and the only overlapping self-citation is implementation infrastructure.

full rationale

The paper is an empirical system paper rather than a derivation, and its central 96% claim is benchmarked against physical definitions of success. Table 2's footnote states: 'Partial success is when the objects are placed correctly but not fully inserted or assembled,' which means the evaluator distinguishes full insertion from mere placement by observing the physical scene, not by reading the Q-function threshold. The sparse reward r_t = R_success * 1(||s_t - (s_g + Δs_grasp)||_2 ≤ ε) is a proxy for task completion and is used to train the Q-function; the same Q-function is then used for deployment termination and success classification (Eq. 4, Appendix A.4). This creates a legitimate validation risk—the proxy is never checked against human labels—but it is a correctness/measurement concern, not circularity, because the reported success and failure counts are not produced by that same Q-function. The only overlapping self-citation, CRISP controllers [57], is infrastructure for low-level torque control, not a load-bearing element of the central argument. The ablation studies, OOD evaluations, and comparisons against independent baselines (DiTFlow, DP, π0.5, HIL-SERL, PE&MP) provide external evidence for the main claim. No step in the paper reduces a claimed result to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method rests on several design choices and domain assumptions: a fixed scene, a pose-distance reward, a force-position controller, and tuned thresholds. The most load-bearing is the pose-distance reward proxy, which is not validated against manual success labels.

free parameters (4)
  • Random action probability p = 0.8
    Probability of uniform random action during autonomous data collection; selected via ablation (Figure 4b) as optimal in 0.6–0.85 range.
  • Q-function success threshold λ_success = 0.93·R_success
    Deployment threshold for terminating the learned policy; paper states it 'needs to be tuned for deployment' (Appendix A.4). Hysteresis thresholds λ_high=0.8R_success, λ_low=0.6R_success are also tuned.
  • Safety filter sphere center and radius (s_safe, r_safe) = Not reported
    Constrains exploration to the critical segment; central to data collection but values not given in paper.
  • Success reward distance epsilon ε = Not reported
    Defines sparse reward r_t via ||s_t - (s_g + Δs_grasp)||_2 ≤ ε; controls what counts as success during training.
axioms (6)
  • domain assumption Contact-rich tasks contain a well-defined critical segment where failures concentrate; free-space motion is easy to plan.
    Stated in the Introduction and used to divide the task into planning and learning; if false, the modular decomposition loses its advantage.
  • domain assumption Scene layout remains fixed during data collection and the single demonstration is replayable.
    Section 3.2 requires replaying the demonstration and keeping objects at fixed locations; any scene changes require human intervention.
  • domain assumption Sparse reward based on Cartesian end-effector distance to a goal pose is a valid success indicator.
    Eq. (2) and reward definition in Section 3.2; not validated per task; affects all learned success signals.
  • domain assumption Off-the-shelf pose estimation provides accurate enough 6D poses for motion planning.
    Appendix A.5 evaluates pose estimation only on the fan cover task; other tasks assume it transfers.
  • domain assumption Hybrid force-position control with a fixed desired force is appropriate for the critical segment.
    Used in both data collection and deployment (Section 3.2); no ablation varying this choice.
  • ad hoc to paper Q-function outputs predict physical success for termination and switching.
    They train Q-functions on the sparse reward and threshold them for success; Figure 6 shows separation but this is heuristic and task-specific.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Data and Learning Where it Matters for Contact-Rich Manipulation." pith.science (2026). https://pith.science/paper/273JA2TU

@misc{pith2026260715982,
  author       = {Pith},
  title        = {Pith review of: Data and Learning Where it Matters for Contact-Rich Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/273JA2TU}},
  note         = {Machine review of arXiv:2607.15982}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Learned policies trained end-to-end on large datasets often remain brittle in high-precision tasks and struggle with generalization. We find that these limitations largely stem from a lack of structure and focus in data collection. Our key insight is to leverage dense data collection only for the critical segment of contact-rich tasks and to rely on traditional planning during simple free-space motion. We propose an automated data-collection scheme in combination with offline deep reinforcement learning for the critical segment of the task, eliminating reliance on a teleoperator's skill and on online policy updates. Across four challenging real-world tasks, using only 2 to 2.5 hours of autonomous data collection, we achieve an average success rate of 96%, compared to the strongest baseline at 55%. Notably, performance remains high in out-of-distribution scenarios where end-to-end approaches struggle. Our results pave the way for targeted data collection for contact-rich tasks and for high success rates in precision applications.

Figures

Figures reproduced from arXiv: 2607.15982 by Angela P. Schoellig, Christian Dietz, Gabor Marko, Jim Yun-Jin Li, Johannes Hechtl, Linus Schwarz, Luka Hofer, Oliver Hausd\"orfer, Ralf R\"omer, Timo Class.

Figure 1
Figure 1. Figure 1: Method. (Left) First, we record a single demonstration in the scene using kinesthetic teaching. The scene layout remains unchanged for the subsequent data collection. (Middle) The demonstration is replayed until the critical segment is reached, and then we execute a mixed greedy￾random policy πexplore for dense data collection, yielding successful and failed rollouts. We combine data collection with offlin… view at source ↗
Figure 2
Figure 2. Figure 2: Tasks. We evaluate our method on four challenging real world tasks: shelf stocking, Lego stacking, mounting two PBT/PC-molded parts (fan cover), and a difficult version of the fan cover task that we train on scene distractors, similar objects, and changing ground surfaces. For all tasks, we evaluate the policies on the additional out-of-distribution scenarios shown in Appendix A.6. The task is solved once … view at source ↗
Figure 3
Figure 3. Figure 3: Maximum forces applied during policy rollouts. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation of data collection parameters. (a) Effect of random action sampling rate p on average successful rollout length, dataset size, and success rate during data collection. (b) Trained policy success rate for different random exploration rates p during data collection. (c) Scaling of SR and dataset size with pose estimation uncertainty. chose for our hardware experiments. F/T measurements provide usefu… view at source ↗
Figure 5
Figure 5. Figure 5: Data collection setups for our different scenes. Our setup requires a fixed scene for data [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Q-function predictions for truncated and successful rollouts. Successful episodes are [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Quantitative pose estimation evaluation. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative pose estimation results including partial occlusions and arbitrary poses. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overview of our out-of-distribution settings used for the quantitative evaluations. For each [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Further qualitative evaluations only performed with method. We place the parts at differ [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual inputs used for our policies. Policies for our method receive a single viewpoint [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Policy failure cases for our method that partially concern baselines as well. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Policy success rate for different combinations of sensing modalities. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Maximum torques applied during policy rollouts. [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

56 extracted references · 8 linked inside Pith

  1. [1]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

  2. [2]

    Black, N

    K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y . Galliker, D. Ghosh, L. Groom, K. Hausman, b. ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. Vuong, H. Walke...

  3. [3]

    Y . Li, H. Jiang, J. Xia, H. Zhang, J. Du, Y . Zhou, J. Zeng, C. Hao, J. Ren, Q. Yu, et al. Forcevla2: Unleashing hybrid force-position control with force awareness for contact-rich ma- nipulation.arXiv preprint arXiv:2603.15169, 2026

  4. [4]

    O’Neill, A

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x mod- els: Open x-embodiment collaboration 0. InIEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903, 2024

  5. [5]

    Khazatsky, K

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karamcheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis, et al. DROID: A large-scale in-the-wild robot manipulation dataset. InRobotics: Science and Systems, 2024

  6. [6]

    T. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.Robotics: Science and Systems (RSS), 2023

  7. [7]

    GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    NVIDIA et al. GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

  8. [8]

    Carion, L

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

  9. [9]

    B. Wen, W. Yang, J. Kautz, and S. Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17868–17879, 2024

  10. [10]

    H. Li, Y . Cui, and D. Sadigh. How to train your robots? the impact of demonstration modality on imitation learning. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 1113–1120, 2025

  11. [11]

    Belkhale, Y

    S. Belkhale, Y . Cui, and D. Sadigh. Data quality in imitation learning. InAdvances in Neural Information Processing Systems, volume 36, pages 80375–80395. Curran Associates, Inc., 2023

  12. [12]

    J. Luo, C. Xu, J. Wu, and S. Levine. Precise and dexterous robotic manipulation via human- in-the-loop reinforcement learning.Science Robotics, 10(105), 2025

  13. [13]

    B. Tang, I. Akinola, J. Xu, B. Wen, A. Handa, K. Van Wyk, D. Fox, G. S. Sukhatme, F. Ramos, and Y . S. Narang. Automate: Specialist and generalist assembly policies over diverse geome- tries. InRobotics: Science and Systems, 2024

  14. [14]

    Y . Tian, J. Xu, Y . Li, J. Luo, S. Sueda, H. Li, K. D. D. Willis, and W. Matusik. Assemble them all: Physics-based planning for generalizable assembly by disassembly.ACM Transactions on Graphics, 41(6):1–11, 2022. 9

  15. [15]

    Y . Tian, K. D. Willis, B. Al Omari, J. Luo, P. Ma, Y . Li, F. Javid, E. Gu, J. Jacob, S. Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380–4386. IEEE, 2024

  16. [16]

    Ankile, A

    L. Ankile, A. Simeonov, I. Shenfeld, and P. Agrawal. Juicer: Data-efficient imitation learning for robotic assembly. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5096–5103. IEEE, 2024

  17. [17]

    B. Fu, S. K. Leong, X. Lian, and X. Ji. 6d robotic assembly based on rgb-only object pose estimation. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4736–4742. IEEE, 2022

  18. [18]

    Morgan, B

    A. Morgan, B. Wen, J. Liang, A. Boularias, A. Dollar, and K. Bekris. Vision-driven compliant manipulation for reliable; high-precision assembly tasks. InRobotics: Science and Systems XVII, RSS2021. Robotics: Science and Systems Foundation, 2021

  19. [19]

    B. Fu, S. K. Leong, Y . Di, G. Wang, J. Tang, F. Tombari, and X. Ji. Lanpose: Language- instructed 6d object pose estimation for robotic assembly. InEuropean Conference on Com- puter Vision, pages 43–59. Springer, 2024

  20. [20]

    B. Tang, M. A. Lin, I. A. Akinola, A. Handa, G. S. Sukhatme, F. Ramos, D. Fox, and Y . S. Narang. IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality. InProceedings of Robotics: Science and Systems, 2023

  21. [21]

    Schoettler, A

    G. Schoettler, A. Nair, J. A. Ojea, S. Levine, and E. Solowjow. Meta-reinforcement learning for robotic industrial insertion tasks. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9728–9735. IEEE, 2020

  22. [22]

    Bi and R

    T. Bi and R. D’Andrea. Sample-efficient learning to solve a real-world labyrinth game using data-augmented model-based reinforcement learning. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7455–7460. IEEE, 2024

  23. [23]

    Ankile, A

    L. Ankile, A. Simeonov, I. Shenfeld, M. Torne, and P. Agrawal. From imitation to refinement- residual rl for precise assembly. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 01–08. IEEE, 2025

  24. [24]

    Goyal, V

    A. Goyal, V . Blukis, J. Xu, Y . Guo, Y .-W. Chao, and D. Fox. RVT-2: Learning Precise Manip- ulation from Few Demonstrations. InProceedings of Robotics: Science and Systems, 2024

  25. [25]

    H. L. Brown, G. Hollinger, and S. Lee. Learning hybrid-control policies for high-precision in-contact manipulation under uncertainty.arXiv preprint arXiv:2604.19677, 2026

  26. [26]

    F. Shao, S. Endo, S. Hirche, and F. Ficuciello. Interactive force-impedance control.IEEE Robotics and Automation Letters, 11(5):6488–6495, 2026

  27. [27]

    Liang, X

    J. Liang, X. Cheng, and O. Kroemer. Learning preconditions of hybrid force-velocity con- trollers for contact-rich manipulation. InProceedings of The 6th Conference on Robot Learn- ing, volume 205 ofProceedings of Machine Learning Research, pages 679–689. PMLR, 2023

  28. [28]

    W. Liu, J. Wang, Y . Wang, W. Wang, and C. Lu. Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 1105–1112. IEEE, 2025

  29. [30]

    M. A. Lee, C. Florensa, J. Tremblay, N. Ratliff, A. Garg, F. Ramos, and D. Fox. Guided uncertainty-aware policy optimization: Combining learning and model-based strategies for sample-efficient policy learning. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 7505–7512, 2020

  30. [31]

    H. Fang, S. Tang, M. Mei, H. Qin, Z. He, J. Chen, Y . Feng, C. Wang, W. Liu, Z. He, et al. Force policy: Learning hybrid force-position control policy under interaction frame for contact-rich manipulation.arXiv preprint arXiv:2602.22088, 2026

  31. [32]

    L. Wang, J. Zhao, Y . Du, E. H. Adelson, and R. Tedrake. PoCo: Policy composition from and for heterogeneous robot learning. InRobotics: Science and Systems, 2024

  32. [33]

    H. Chen, J. Xu, H. Chen, K. Hong, B. Huang, C. Liu, J. Mao, Y . Li, Y . Du, and K. Driggs- Campbell. Multi-modal manipulation via multi-modal policy consensus.arXiv preprint arXiv:2509.23468, 2025

  33. [34]

    Y . Wang, Y . Zhang, M. Huo, R. Tian, X. Zhang, Y . Xie, C. Xu, P. Ji, W. Zhan, M. Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. 2024

  34. [35]

    R ¨omer, Y

    R. R ¨omer, Y . Zhang, Y . Li, and A. P. Schoellig. Clare: Continual learning for vision-language- action models via autonomous adapter routing and expansion.IEEE Robotics and Automation Letters, 2026

  35. [36]

    Mandlekar, S

    A. Mandlekar, S. Nasiriany, B. Wen, I. Akinola, Y . Narang, L. Fan, Y . Zhu, and D. Fox. Mim- icgen: A data generation system for scalable robot learning using human demonstrations. In Proceedings of The 7th Conference on Robot Learning, volume 229 ofProceedings of Machine Learning Research, pages 1820–1864. PMLR, 2023

  36. [37]

    Mandlekar, C

    A. Mandlekar, C. R. Garrett, D. Xu, and D. Fox. Human-in-the-loop task and motion planning for imitation learning. InProceedings of The 7th Conference on Robot Learning, volume 229 ofProceedings of Machine Learning Research, pages 3030–3060. PMLR, 2023

  37. [38]

    Yeh, T.-S

    C.-H. Yeh, T.-S. Nan, R. Vuorio, W. Hung, H. Y . Wu, S.-H. Sun, and P.-C. Hsieh. Action- constrained imitation learning. InForty-second International Conference on Machine Learn- ing, 2025

  38. [39]

    T. Lee, A. Wagenmaker, K. Pertsch, P. Liang, S. Levine, and C. Finn. Roboreward: General- purpose vision-language reward models for robotics.arXiv preprint arXiv:2601.00675, 2026

  39. [40]

    Liang, Y

    A. Liang, Y . Korkmaz, J. Zhang, M. Hwang, A. Anwar, S. Kaushik, A. Shah, A. S. Huang, L. Zettlemoyer, D. Fox, et al. Robometer: Scaling general-purpose robotic reward models via trajectory comparisons.arXiv preprint arXiv:2603.02115, 2026

  40. [41]

    P. J. Ball, L. Smith, I. Kostrikov, and S. Levine. Efficient online reinforcement learning with offline data. InInternational Conference on Machine Learning, pages 1577–1594. PMLR, 2023

  41. [42]

    Haarnoja, A

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. InInternational conference on machine learning, pages 1861–1870. Pmlr, 2018

  42. [43]

    J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

  43. [44]

    X. Chen, C. Wang, Z. Zhou, and K. W. Ross. Randomized ensembled double q-learning: Learning fast without a model. InInternational Conference on Learning Representations, 2021

  44. [45]

    Collins and A

    T. Collins and A. Bartoli. Infinitesimal plane-based pose estimation.International Journal of Computer Vision, 109(3):252–286, 2014. 11

  45. [46]

    Strangh ¨oner, P

    J. Strangh ¨oner, P. Hartmann, M. Braun, S. Wrede, and K. Neumann. Share-rl: Structured, interactive reinforcement learning for contact-rich industrial assembly tasks.arXiv preprint arXiv:2509.13949, 2025

  46. [47]

    Zhang, S

    X. Zhang, S. Jin, C. Wang, X. Zhu, and M. Tomizuka. Learning insertion primitives with discrete-continuous hybrid action space for robotic assembly tasks. In2022 International conference on robotics and automation (ICRA), pages 9881–9887. IEEE, 2022

  47. [48]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jegou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 9630–9640, 2021

  48. [49]

    Dasari, O

    S. Dasari, O. Mees, S. Zhao, M. K. Srirama, and S. Levine. The ingredients for robotic dif- fusion transformers. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 15617–15625. IEEE, 2025

  49. [50]

    Y . Chen, S. Tian, S. Liu, Y . Zhou, H. Li, and D. Zhao. ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy. InProceedings of Robotics: Science and Systems, 2025

  50. [51]

    Nikishin, M

    E. Nikishin, M. Schwarzer, P. D’Oro, P.-L. Bacon, and A. Courville. The primacy bias in deep reinforcement learning. InProceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Research, pages 16828–16847. PMLR, 2022

  51. [52]

    Coleman, I

    D. Coleman, I. Sucan, S. Chitta, and N. Correll. Reducing the barrier to entry of complex robotic software: a moveit! case study.arXiv preprint arXiv:1404.3785, 2014

  52. [53]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–

  53. [54]

    Lepert, J

    M. Lepert, J. Fang, and J. Bohg. Phantom: Training robots without robots using only human videos. InProceedings of The 9th Conference on Robot Learning, volume 305 ofProceedings of Machine Learning Research, pages 4545–4565. PMLR, 2025

  54. [55]

    Maddukuri, Z

    A. Maddukuri, Z. Jiang, L. Y . Chen, S. Nasiriany, Y . Xie, Y . Fang, W. Huang, Z. Wang, Z. Xu, N. Chernyadev, S. Reed, K. Goldberg, A. Mandlekar, L. Fan, and Y . Zhu. Sim-and-real co- training: A simple recipe for vision-based robotic manipulation. InProceedings of Robotics: Science and Systems (RSS), 2025

  55. [56]

    Cadene, S

    R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, S. Palma, P. Kooijmans, M. Ar- actingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, J. Moss, and T. Wolf. Lerobot: State-of-the-art machine learning for real-world robotics in pytorch. https://github.com/huggingface/lerobot, 2024

  56. [57]

    D. S. J. Pro, O. Hausd ¨orfer, R. R ¨omer, M. D ¨osch, M. Schuck, and A. P. Schoellig. Crisp- compliant ros2 controllers for learning-based manipulation policies and teleoperation.IEEE Robotics and Automation Practice, 2026. 12 A Appendix A.1 Data collection setups Figure 5: Data collection setups for our different scenes. Our setup requires a fixed scene...

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.