Pith. sign in

REVIEW 2 major objections 5 minor 39 references

To hold success rate fixed, required demonstrations grow super-exponentially as target precision nears a system-defined limit c.

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 · deepseek-v4-flash

2026-08-01 03:33 UTC pith:SB55EJJS

load-bearing objection An interesting empirical scaling law for precision in manipulation, but the central claim that c is target-SR-independent is enforced by the fitting protocol rather than tested. the 2 major comments →

arxiv 2607.23108 v1 pith:SB55EJJS submitted 2026-07-25 cs.RO

The Curse of Precision: A Data Scaling Law for High-Precision Robotic Manipulation

Cuijie Xu , Yuanfan Xu , Min Xue , Jianjie Lin , Jian Wang , Xudong Zhang , Yu Wang , Jincheng Yu This is my paper
classification cs.RO
keywords precision scaling lawimitation learninghigh-precision manipulationbehavior cloninglimit precisionsystem capability metricsuper-exponential scalingdata scaling
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.

This paper tries to establish a quantitative law connecting the amount of demonstration data needed by an imitation-learning robot to the precision demanded by a closed-world manipulation task. The claim is that to hold a fixed success rate, the number of demonstrations N grows super-exponentially as the required precision P approaches a system-specific ceiling c, captured by log N ∝ 1/(P−c). The paper further claims that c is not a fixed physical limit but an emergent property of the whole agent—sensors, expert demonstrations, and task complexity—so it can serve as a measurable capability metric. If true, engineers could predict data costs before scaling up, and use c as a diagnostic for debugging high-precision systems. The evidence comes from simulated peg insertion, cuboid stacking, and ball rolling tasks using a standard behavior-cloning policy.

Core claim

Across three simulated manipulation tasks, the paper reports the precision scaling law log(N) = m/(P−c)+n: at a fixed target success rate, the number of demonstrations N needed to reach precision P grows super-exponentially as P approaches the system's limit precision c. The authors find that a single shared c fits data for three different target success rates in each task, leading them to interpret c not as a fixed property of the task geometry but as an emergent capability of the whole agent—sensors, expert data, and task complexity. Ablations on a peg-insertion task support this: removing the wrist camera raises c from 2.35 mm to 3.85 mm; switching to a less ambiguous expert demonstration

What carries the argument

The paper's central object is the precision scaling identity log(N) = m/(P−c)+n, in which P is the task precision (a tolerance in millimeters), N is the number of demonstrations, and c is the 'limit precision'—the best precision the system can ever reach, in the same units as P. The identity does two kinds of work: it converts the intuitive 'curse of precision' (each extra decimal of accuracy costs vastly more data) into a quantitative prediction, and it turns c into a scalar that summarizes the entire agent's capability. The empirical method couples two fits: a per-P power law for failure rate versus data, and then a single global c chosen to maximize the coefficient of determination across

Load-bearing premise

The paper assumes that the limit precision c is a single constant for the system regardless of the target success rate, because it fits one shared c across all target success-rate curves; if c actually shifts with the target success rate, the interpretation of c as an intrinsic capability metric collapses.

What would settle it

Estimate c independently for two different target success rates, e.g., 0.5 and 0.9, using fresh data at several precision levels; if the two fitted c values differ beyond statistical uncertainty, the claim that c is a single system-level constant fails.

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

If this is right

  • With c fitted from a few experiments, engineers can predict the demonstration budget needed for any target precision before running a large data collection.
  • Because c changes with system components, it provides a quantitative, directly comparable capability metric for sensor suites, expert policies, and task formulations.
  • The super-exponential growth near c implies that brute-force data scaling is an inefficient path to sub-millimeter reliability; improving system components is the lever that moves the limit.
  • The law holds across image-based assembly and a state-based dynamic task, indicating it may generalize beyond a single task family.

Where Pith is reading between the lines

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

  • If c is truly independent of the target success rate, a short data sweep at one moderate precision could be used to extrapolate the full data/precision frontier, letting teams A/B test hardware or data changes without exhaustive full-scale training.
  • The form log N ∝ 1/(P−c) suggests a 'precision frontier' analogous to compute-efficient frontiers: for any budget of demonstrations, there is an optimal precision to target, and pushing past it wastes data.
  • A direct testable extension: on physical hardware, c should track measurable system noise—sensor resolution, control latency, and calibration error—so it could become a standard benchmark number reported alongside robot setups.
  • Demonstration clarity lowering c more than expert success rate hints that automatic data cleaning or demonstration editing could be a cheaper route to precision than collecting more or better teleoperation data.

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

2 major / 5 minor

Summary. The paper proposes a 'precision scaling law' for imitation learning in closed-world manipulation: for a fixed target success rate, the number of demonstrations N required grows super-exponentially as the task precision P approaches a limit c, modeled as log(N) = m/(P − c) + n. The authors further claim that c is not a fixed physical constant but an intrinsic, target-SR-independent property of the whole agent system (sensors, expert, task complexity). The law is validated in ManiSkill3 on Peg Insertion, Stack Cuboid, and Roll Ball, with power-law SR-vs-N fits used to interpolate N at target SRs, and a shared-c grid search used to fit the precision law. Ablations on Peg Insertion show c varying with observation modality, expert strategy, and task complexity.

Significance. If the central claim holds, c would be a practically useful, quantitative system-capability metric and the scaling law would give a predictive tool for data requirements in high-precision tasks. The paper also makes a useful conceptual shift from open-world scaling-for-breadth to closed-world scaling-for-precision. Positive features include high in-sample R² values across three tasks and several ablations, a clear demonstration that the law fails for an under-capacity model (R² = 0.22 vs 0.99), and a falsifiable functional form. However, the load-bearing claim that c is independent of target success rate is not actually tested—it is imposed by the fitting protocol—and the supporting statistics rest on single training runs and on N values obtained by interpolation/extrapolation from prior fits. These issues are fixable, but they currently weaken the empirical support for the paper's headline interpretation.

major comments (2)
  1. [IV-A, 'Policy Training Sweep' and 'Curve Fitting'] Each (N, P) data point corresponds to a single training run, yet the precision-law fits use N values obtained by interpolating or extrapolating the SR power-law fits from those same single runs. No seed variance is reported, and uncertainties from the first-stage fits are never propagated into the second-stage estimates of N or c. With only 4–6 precision levels per task, a single outlier run could materially change a, b, and hence the inferred N and c. The authors should either run multiple seeds (at least 3) or use a bootstrap/joint-fitting procedure to report confidence intervals on c and on the predicted N. They should also state explicitly which target-SR N values are interpolated versus extrapolated; extrapolation beyond the observed SR range (e.g., SR = 0.9 when the largest N only reaches lower SRs) would be a serious concern.
  2. [IV-B and IV-C, Tables II and III] The reported R² values are in-sample and are computed on the same data used to select c by grid search. This is not a predictive validation of the law, and it does not distinguish the proposed 1/(P−c) form from other flexible two- or three-parameter curves. Since the stated practical value of the law is prediction (Section V), the authors should provide a held-out test, e.g., fit on a subset of precision levels and predict N at the remaining levels, or report leave-one-precision-out cross-validated error. Such a test would also strengthen the ablations in Table III, which otherwise share the same circular fitting protocol.
minor comments (5)
  1. [IV-C, 'Impact of Observation Modality'] Typo: 'disabling the wrist-mounted camer' should be 'camera'.
  2. [II-A] The phrase 'these two axes of scaling, Breadth and Depth, are largely orthogonal' is clear, but 'Depth' is used informally. Consider defining it once in the introduction to avoid confusion with the U-Net model capacity discussion in IV-D.
  3. [IV-A, Table I] For Roll Ball, P is described as the radius of the target region, but Table I uses mm. That is fine, but the range 35–200 mm is very different from the peg/stack ranges; the text should make explicit that the absolute scale of c is task-dependent and not comparable across tasks.
  4. [V, Limitations] The simulation-only scope is appropriately acknowledged. The paper would be stronger if the authors also noted that Diffusion Policy is a single algorithm choice and that c may depend on the learning algorithm, not only on sensors and expert.
  5. [Figure 2] Panel (d)–(f) would benefit from showing the raw (P, N) points with error bars, rather than only the linearized fit lines, so readers can assess scatter and extrapolation regions.

Circularity Check

2 steps flagged

Shared-c grid-search fit enforces the target-SR independence of c; the central 'validation' restates the fitting constraint rather than testing it.

specific steps
  1. self definitional [Section IV-A 'Curve Fitting'; interpreted in Section IV-B 'Precision Scaling Law']
    "We then performed a grid search for a single c value that maximized the sum of the coefficients of determination (R2) across all three targetSR curves. This enforces the hypothesis that c is a constant property of the system, independent of the target success rate."

    The paper's load-bearing claim is that c is an intrinsic, target-SR-independent system property. That claim is not tested: the fitting protocol chooses the single c that makes all three target-SR curves as linear as possible in the (P, log N) plane, and the near-perfect R² values in Table II are computed on the same curves used for this grid search. The data are never allowed to exhibit an SR-dependent c; 'validation' is a restatement of the enforced constraint. A genuine test would estimate c separately for each target SR (or on held-out precision levels) and check whether the values agree.

  2. fitted input called prediction [Section IV-A 'Curve Fitting' → Section IV-B 'Precision Scaling Law']
    "we first estimated the data size N required to reach different target success rates (e.g., 0.5, 0.7, 0.9) for eachP by interpolating or extrapolating from the fitted linear models of the Success Rate Scaling Law."

    The N values entering the precision-law fit are not direct measurements; they are themselves predictions of the success-rate power-law fits from the same training runs. Fitting log(N) = m/(P-c) + n to these derived N values and then reporting the resulting R² as a validation of the precision law is an in-sample consistency check, not an independent confirmation. The empirical content lies mainly in the fact that the assumed functional form can fail (as in the standard-capacity Roll Ball case), not in the high R² values reported after the free parameter c and derived N are optimized.

full rationale

The derivation is not circular in the sense of being mathematically identical to its inputs: the proposed reciprocal form log(N) ∝ 1/(P−c) is not tautological, and the paper reports a case (standard-capacity Roll Ball) where it fails, so the law has residual empirical content. However, the paper's most novel interpretive claim—that c is an intrinsic system property independent of target success rate—is enforced by construction. A single c is grid-searched to maximize the sum of R² across all three target-SR curves, and the same curves are then cited as validation. The paper also fits the precision law to N values that were already interpolated/extrapolated from the success-rate power laws, so the 'perfect' precision-law fits are largely an in-sample description of previously fitted quantities. These features make the central quantitative claim partially circular, though the cross-configuration ordering of c values (wrist camera, expert strategy, task complexity) remains an independently meaningful empirical pattern. No load-bearing self-citation or imported uniqueness theorem is present. Score 6: one central 'prediction' reduces by construction to the fitting constraint, but the law's functional form is not forced.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 1 invented entities

The central law rests on three fitted parameter sets (power-law a,b; precision-law m,n; limit precision c), an assumed functional form, an imposed SR-independence of c, and simulation fidelity. No theoretical derivation is provided.

free parameters (3)
  • Per-(task, precision) power-law slope a and intercept b = e.g., Peg Insertion P=4: a=-0.19, b=1.35; P=10: a=-0.72, b=5.08 (Table I)
    Fitted via linear regression of log(1−SR) against log N at each precision.
  • Per-(task, target-SR) precision-law slope m and intercept n = e.g., Peg Insertion SR=0.5: m=8.59, n=7.62; SR=0.9: m=28.62, n=7.42 (Table II)
    Fitted by linear regression of log N against 1/(P−c).
  • Limit precision c = Peg baseline 2.35 mm; Stack Cuboid 2.75 mm; Roll Ball 20.3 mm; ablations: Expert-2 1.27, no-wrist 3.85, low-randomizatio
    Chosen by grid search maximizing summed R² across the three target-SR curves; this is a fitted parameter, not independently measured.
axioms (6)
  • domain assumption Failure rate scales as a power law with data at fixed precision: log(1−SR)=a log N+b.
    Adapted from prior generalization scaling laws [5]; no derivation given; all downstream N estimates rely on it.
  • ad hoc to paper The precision scaling law has the functional form log(N)=m/(P−c)+n with a single c shared across target success rates.
    Posited as Hypothesis 2; validity is judged only by in-sample R² after fitting c.
  • ad hoc to paper c is independent of the target success rate.
    Imposed by the grid-search protocol that fits one c across all target-SR curves; cited as validation rather than tested.
  • domain assumption SAPIEN/ManiSkill3 simulation faithfully represents precision-relevant contact physics.
    All experiments are in simulation; authors list real-hardware validation as future work.
  • domain assumption The fixed Diffusion Policy architecture with sufficient capacity is an adequate learner for all tasks.
    Section IV-D shows the law only emerges after increasing model capacity post hoc; capacity is a hidden precondition.
  • domain assumption Training only on successful expert trajectories is a valid data-generation choice.
    Data pool consists only of successful trajectories; for the aggressive expert (raw SR ~50%) this selects a biased subset.
invented entities (1)
  • Limit precision c no independent evidence
    purpose: Fitted parameter claimed to represent the intrinsic precision ceiling of the whole agent system (sensors, expert policy, task complexity).
    c is obtained by in-sample grid search and has no external or out-of-sample falsifiable handle; it is currently a fitted capability metric, not an independently verified property.

pith-pipeline@v1.3.0-alltime-deepseek · 11214 in / 15714 out tokens · 133718 ms · 2026-08-01T03:33:51.350127+00:00 · methodology

0 comments
read the original abstract

While scaling laws for imitation learning have primarily focused on generalization in open-world settings, the relationship between data and precision in closed-world tasks like robotic assembly remains largely unexplored. This paper systematically investigates this relationship and introduces a novel scaling law. We find that to achieve a fixed success rate, the required number of demonstrations $N$ grows super-exponentially as the target precision $P$ approaches a limit $c$. This relationship is accurately captured by the model $\log(N) \propto 1/(P-c)$. Crucially, we reveal that the limit precision $c$ is not a static physical constant of the task but an emergent property of the entire agent system, including its sensors and expert policy. Through experiments on canonical manipulation tasks, we validate this law and demonstrate that improving system components, such as adding a wrist camera or using a more effective expert, measurably lowers $c$, thus expanding the system's achievable precision. Our work provides a new theoretical framework for precision in robotics and a quantitative metric to evaluate system capabilities. Furthermore, these findings provide a practical methodology for guiding the development and debugging of high-precision manipulation systems.

Figures

Figures reproduced from arXiv: 2607.23108 by Cuijie Xu, Jianjie Lin, Jian Wang, Jincheng Yu, Min Xue, Xudong Zhang, Yuanfan Xu, Yu Wang.

Figure 1
Figure 1. Figure 1: Illustration of the diverse, high-precision manipulation tasks investigated in this study. Peg Insertion: A contact￾rich assembly task requiring precise alignment to overcome tight clearances. Cuboid Stacking: A precise placement task where stability is challenged by stacking slender cuboids with a small support base. Roll Ball: An underactuated dynamics task requiring control over a rolling ball to reach … view at source ↗
Figure 2
Figure 2. Figure 2: Empirical Validation of the Proposed Scaling Laws. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of the key system configurations investigated in our ablation studies. (a-b) Observation: The policy’s visual input is composed of a (a) base camera and (b) wrist camera. The baseline system utilizes both, while the ablation relies solely on the base camera. (c-d) Task Complexity: We compare two levels of domain randomization. (c) Low Randomization varies only the initial XY positions of … view at source ↗
Figure 4
Figure 4. Figure 4: Impact of model capacity on the Success Rate Scaling [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

39 extracted references · 14 linked inside Pith

  1. [1]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    Pith/arXiv arXiv 2001

  2. [2]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhari- wal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. A...

    Pith/arXiv arXiv 2005

  3. [3]

    Scaling vision transformers,

    X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer, “Scaling vision transformers,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), June 2022, pp. 12 104– 12 113

    2022

  4. [4]

    Deep learning scaling is predictable, empirically,

    J. Hestness, S. Narang, N. Ardalani, G. Diamos, H. Jun, H. Kianinejad, M. A. Patwary, Y . Yang, and Y . Zhou, “Deep learning scaling is predictable, empirically,”arXiv preprint arXiv:1712.00409, 2017

    Pith/arXiv arXiv 2017

  5. [5]

    Data scaling laws in imitation learning for robotic manipulation,

    F. Lin, Y . Hu, P. Sheng, C. Wen, J. You, and Y . Gao, “Data scaling laws in imitation learning for robotic manipulation,”arXiv preprint arXiv:2410.18647, 2024

    Pith/arXiv arXiv 2024

  6. [6]

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

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu, J. Ibarz, B. Ichter, A. Irpan, T. Jackson, S. Jesmonth, N. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. Leal, K.-H. Lee, S. Levine, Y . Lu, U. Malla, D. Manjunath, , I. Mordatch, O. Nachum, C. Parada, J. Peralta, E. Perez, K. Pertsch, J....

    Pith/arXiv arXiv 2022

  7. [7]

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

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choro- manski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florenceet al., “Rt- 2: Vision-language-action models transfer web knowledge to robotic control,”arXiv preprint arXiv:2307.15818, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Open x-embodiment: Robotic learning datasets and rt-x models,

    A. Padalkaret al., “Open x-embodiment: Robotic learning datasets and rt-x models,”arXiv preprint arXiv:2310.08864, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Siciliano and O

    B. Siciliano and O. Khatib, Eds.,Springer Handbook of Robotics, 2nd ed. Springer, 2016

    2016

  10. [10]

    A unified approach for motion and force control of robot manipulators: The operational space formulation,

    O. Khatib, “A unified approach for motion and force control of robot manipulators: The operational space formulation,”IEEE Journal on Robotics and Automation, vol. 3, no. 1, pp. 43–53, 1987

    1987

  11. [11]

    Impedance control: An approach to manipulation, part i: Theory,

    N. Hogan, “Impedance control: An approach to manipulation, part i: Theory,”Journal of Dynamic Systems, Measurement, and Control, vol. 107, no. 1, pp. 1–7, 1985

    1985

  12. [12]

    Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,

    S. Tao, F. Xiang, A. Shukla, Y . Qin, X. Hinrichsen, X. Yuan, C. Bao, X. Lin, Y . Liu, T. kai Chan, Y . Gao, X. Li, T. Mu, N. Xiao, A. Gurha, V . N. Rajesh, Y . W. Choi, Y .-R. Chen, Z. Huang, R. Calandra, R. Chen, S. Luo, and H. Su, “Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,”Robotics: Science and Systems, 2025

    2025

  13. [13]

    Socratic models: Composing zero-shot mul- timodal reasoning with language,

    A. Zeng, M. Attarian, B. Ichter, K. Choromanski, A. Wong, S. Welker, F. Tombari, A. Purohit, M. S. Ryoo, V . Sindhwani, J. Lee, V . Van- houcke, and P. Florence, “Socratic models: Composing zero-shot mul- timodal reasoning with language,”arXiv preprint arXiv:2204.00598, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Octo: An open-source generalist robot policy,

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, C. Xu, J. Luo, T. Kreiman, Y . Tan, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine, “Octo: An open-source generalist robot policy,”arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Robonet: Large-scale multi-robot learning,

    S. Dasari, F. Ebert, S. Tian, S. Nair, B. Bucher, K. Schmeckpeper, S. Singh, S. Levine, and C. Finn, “Robonet: Large-scale multi-robot learning,” inProceedings of the Conference on Robot Learning, ser. Proceedings of Machine Learning Research, L. P. Kaelbling, D. Kragic, and K. Sugiura, Eds., vol. 100. PMLR, 30 Oct–01 Nov 2020, pp. 885–897

    2020

  16. [16]

    A generalist agent,

    S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez, Y . Sulsky, J. Kay, J. T. Springenberg, T. Eccles, J. Bruce, A. Razavi, A. Edwards, N. Heess, Y . Chen, R. Hadsell, O. Vinyals, M. Bordbar, and N. de Freitas, “A generalist agent,”arXiv preprint arXiv:2205.06175, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    R. M. Murray, Z. Li, and S. S. Sastry,A mathematical introduction to robotic manipulation. CRC press, 1994

    1994

  18. [18]

    Hybrid position/force control of manipulators,

    M. H. Raibert and J. J. Craig, “Hybrid position/force control of manipulators,”Journal of dynamic systems, measurement, and control, vol. 103, no. 2, pp. 126–133, 1981

    1981

  19. [19]

    Reinforcement learning in robotics: A survey,

    J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,”The International Journal of Robotics Research, vol. 32, no. 11, pp. 1238–1274, 2013

    2013

  20. [20]

    Learning from demonstration,

    S. Schaal, “Learning from demonstration,” inAdvances in Neural Information Processing Systems 9. Cambridge, MA: MIT Press, 1997, pp. 1040–1046

    1997

  21. [21]

    A survey of robot learning from demonstration,

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,”Robotics and autonomous systems, vol. 57, no. 5, pp. 469–483, 2009

    2009

  22. [22]

    Deep reinforcement learning for high precision assembly tasks,

    T. Inoue, G. De Magistris, A. Munawar, T. Yokoya, and R. Tachibana, “Deep reinforcement learning for high precision assembly tasks,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 819–825

    2017

  23. [23]

    Learning latent plans from play,

    C. Lynch, M. Khansari, T. Xiao, V . Kumar, J. Tompson, S. Levine, and P. Sermanet, “Learning latent plans from play,” inProceedings of the Conference on Robot Learning (CoRL), 2019

    2019

  24. [24]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 1126–1135

    2017

  25. [25]

    Waypoint-based imitation learning for robotic manipulation,

    L. X. Shi, A. Sharma, T. Z. Zhao, and C. Finn, “Waypoint-based imitation learning for robotic manipulation,”7th Conference on Robot Learning (CoRL 2023), 2023

    2023

  26. [26]

    Perceiver-actor: A multi-task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: A multi-task transformer for robotic manipulation,” inCoRL, 2022

    2022

  27. [27]

    Vima: General robot manipulation with multimodal prompts,

    Y . Jiang, A. Gupta, Z. Zhang, G. Wang, Y . Dou, Y . Chen, L. Fei- Fei, A. Anandkumar, Y . Zhu, and L. Fan, “Vima: General robot manipulation with multimodal prompts,” inInternational Conference on Machine Learning, 2023

    2023

  28. [28]

    Imitation learning for high precision peg-in-hole tasks,

    S. Gubbi, S. N. Y . Kolathaya, and B. Amrutur, “Imitation learning for high precision peg-in-hole tasks,” in2020 6th International Conference on Control, Automation and Robotics (ICCAR), 2020, pp. 1–6

    2020

  29. [29]

    From imitation to refinement – residual rl for precise visual assembly,

    L. L. Ankile, A. Simeonov, I. Shenfeld, M. T. Villasevil, and P. Agrawal, “From imitation to refinement – residual rl for precise visual assembly,”arXiv preprint arXiv:2407.16677, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 23–30

    2017

  31. [31]

    Sim-to-real via sim- to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks,

    S. James, P. Wohlhart, M. Kalakrishnan, D. Kalashnikov, A. Irpan, J. Ibarz, S. Levine, R. Hadsell, and K. Bousmalis, “Sim-to-real via sim- to-sim: Data-efficient robotic grasping via randomized-to-canonical adaptation networks,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  32. [32]

    Fast robust peg-in- hole insertion with continuous visual servoing,

    R. Haugaard, J. Langaa, C. Sloth, and A. Buch, “Fast robust peg-in- hole insertion with continuous visual servoing,” inProceedings of the 2020 Conference on Robot Learning, vol. 155, PMLR. PMLR, 2021, pp. 1696–1705

    2020

  33. [33]

    Juicer: Data- efficient imitation learning for robotic assembly,

    L. Ankile, A. Simeonov, I. Shenfeld, and P. Agrawal, “Juicer: Data- efficient imitation learning for robotic assembly,”arXiv preprint arXiv:2402.14833, 2024

    Pith/arXiv arXiv 2024

  34. [34]

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

    J. Luo, C. Xu, J. Wu, and S. Levine, “Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning,”arXiv preprint arXiv:2410.21845, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    SAPIEN: A simulated part-based interactive environment,

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang, L. Yi, A. Chang, L. J. Guibas, and H. Su, “SAPIEN: A simulated part-based interactive environment,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  36. [36]

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,” inProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023

    2023

  37. [37]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016

    2016

  38. [38]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, ser. Lecture Notes in Computer Science, vol. 9351. Springer International Publishing, 2015, pp. 234–241

    2015

  39. [39]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”arXiv preprint arXiv:2006.11239, 2020

    Pith/arXiv arXiv 2006