Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DemoSpeedup shows that entropy-guided re-timing of human demonstrations can make visuomotor policies run up to 3x faster without sacrificing, and sometimes improving, task success.

desk verdict DemoSpeedup is a solid, honest empirical paper with a real but unproven causal premise: entropy-driven selection probably helps, but the missing control means we never see it isolated. read the letter →

arxiv 2506.05064 v2 pith:MLXPN4RZ submitted 2025-06-05 cs.RO

Lingxiao Guo , Zhengrong Xue , Zijing Xu , Huazhe Xu This is my paper
classification cs.RO
keywords imitationlearningroboticmanipulationdemonstrationaccelerationactionentropyvisuomotorpolicydatacurationgenerativetimeefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that slow robot execution comes from slow human demonstrations, and that you can accelerate the demonstrations themselves rather than the policy at test time. DemoSpeedup trains a proxy policy on the original demonstrations, uses the policy's per-frame action entropy to tell which moments demand precision and which are casual, and down-samples the casual moments more aggressively. The accelerated dataset then trains the final policy, which executes up to three times faster while matching or sometimes beating the success rate of the same policy trained on full-speed data. If that claim holds, re-timing demonstrations becomes a cheap, annotation-free way to speed up visuomotor policies across tasks and robot platforms.

What carries the argument

The central mechanism is the entropy-guided precision measurement. A proxy policy $\pi_\theta(A_t|o_t)$ is trained on the source demonstrations; then $N$ action-chunk samples are drawn per observation, a Gaussian kernel density estimate gives $\hat{p}(a_t|o_t)$ (Equation 1), and the conditional action entropy $\hat{H}(a_t|o_t)$ is computed per frame (Equation 2). The entropy values, concatenated with time index, are cleaned by Isolation Forest, normalized, and clustered with HDBSCAN to label low-entropy clusters as precision set $P$ and the rest as casual set $C$. Acceleration then applies piecewise down-sampling at low ratio $r_{\text{low}}=2$ and high ratio $r_{\text{high}}=3$ or $4$, with a replicate-before-downsample strategy that creates $N$ offset sub-chunks to retain visited-state diversity, and a geometrically consistent chunk length so the accelerated policy only fits far fewer action labels per segment.

What would settle it

Collect a dataset where ground-truth precision labels are available (e.g., contact-state annotations or human-annotated critical moments). If DemoSpeedup's low-entropy segments fail to coincide with the ground-truth high-precision moments at rates clearly better than chance, the entropy signal is not carrying the precision information the method needs. A second check: train DemoSpeedup on a task with a fast approach and a slow, variable insertion; if the approach segment is heavily downsampled because it has high entropy and success drops specifically from misaligned insertion, the premise is contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, DemoSpeedup is a self-supervised, task-agnostic demonstration-acceleration method. Its central claim is that action entropy estimated by an arbitrary generative behavior-cloning policy (ACT or Diffusion Policy) serves as a proxy for the precision level a demonstration segment requires. Frames with low estimated entropy correspond to high-precision operations that must be preserved; frames with high entropy correspond to casual motion that can be safely down-sampled. By segmenting trajectories with HDBSCAN clustering of entropy points, labeling low-entropy clusters as the precision set and the rest as casual, and then down-sampling at rates that grow with entropy (with a replicate-before-downsample trick to preserve state diversity and a shortened action chunk to keep geometric distance consistent), the resulting accelerated datasets train policies that execute roughly 1.7x-3x faster. In the paper's experiments, these policies maintain success rates on par with, and sometimes above, the same policy trained on the original-speed demonstrations; test-time down-sampling at the same speed causes a drop of over 8% on average.

Load-bearing premise

The load-bearing premise is that the action entropy estimated by a proxy policy trained on the original demonstrations correctly flags which frames can be accelerated without hurting task success; if the entropy reflects model uncertainty or dataset noise rather than the precision a moment demands, the accelerated data could silently drop causally necessary frames.

Editorial extensions

If this is right

  • Training on entropy-accelerated demonstrations can cut execution time by 1.7x-3x across ACT and Diffusion Policy on simulation and real-world tasks, without needing new hardware or annotation.
  • The same pipeline could be applied to any generative policy that can sample diverse action chunks, so future visuomotor architectures inherit the speedup mechanism.
  • Because episode length shrinks, compounding error in imitation learning decreases, which can translate into higher success rates on long-horizon tasks.
  • Test-time acceleration baselines that naively skip action chunks suffer average performance drops over 8%, suggesting demonstration-time re-timing is a safer route to speed.
  • The method's speedup is bounded by the 'casual' fraction of the demonstrations; datasets collected by slower operators stand to gain more.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The entropy signal likely conflates human choice variability with task-required precision, so in tasks where high-precision moments are also high-variability (e.g., multiple valid strategies), DemoSpeedup may over-accelerate critical frames; a version that separates entropy into aleatoric vs epistemic parts could be more robust.
  • A natural extension is to set the desired speedup first and solve for the per-segment down-sampling rates under a budget, instead of manually choosing $r_{\text{low}}$ and $r_{\text{high}}$ per task.
  • The method could be combined with data filtering (e.g., discarding unsuccessful demos) to simultaneously improve speed and success, since the proxy policy's entropy estimates depend on the quality of the source dataset.
  • If entropy correlates with human reaction time rather than precision, the method may over-accelerate approach phases that actually need visual feedback; this is testable by comparing against contact-oracle segmentation on contact-rich tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. DemoSpeedup proposes an entropy-guided pipeline for accelerating visuomotor policies trained from slow human demonstrations. A proxy policy (ACT or Diffusion Policy) is trained on the original-speed dataset and used to estimate a per-frame action entropy via kernel density estimation over sampled action chunks. The entropy values are preprocessed with Isolation Forest, clustered with HDBSCAN, and used to label frames as high-precision or casual; the demonstrations are then piecewise downsampled at two rates, with a replicate-before-downsample strategy and a halved, geometrically consistent action chunk. The accelerated dataset is used to train the final policy. Experiments cover 11 simulated tasks (Aloha, BiGym) and 5 real-world tasks on Galaxea R1, reporting 1.7x-3x speedups with success rates that are commonly on par with or higher than policies trained on original demonstrations. Ablations address replicate-before-downsample, chunk length, and controller gain, and comparisons include constant-rate downsampling, a contact oracle, and an AWE*-style baseline.

Significance. Demonstration acceleration is practically important because slow teleoperated demonstrations limit the deployment speed of imitation policies. The paper's empirical scope is substantial: 11 simulated tasks and 5 real-world tasks, two policy classes, comparison against constant-rate downsampling, an oracle contact heuristic, and a dynamic-programming downsampling baseline, plus ablations of the main design choices. The method is self-supervised and makes a falsifiable claim, namely that entropy-guided segmentation yields speedup with maintained success, and the direction of the results is consistent across most tasks. However, the load-bearing premise, that per-frame action entropy of a proxy policy is a reliable proxy for task-relevant precision, is not directly validated, and the entropy estimator itself is under-specified. The current comparisons do not fully isolate the causal contribution of the entropy signal. Given the breadth and generally positive results, the contribution is promising, but the missing validation and a few reproducibility issues should be addressed before the central claim is established.

major comments (4)
  1. [§3.1, Eqs. (1)-(2)] The estimator in Eqs. (1)-(2) is not established as a conditional action entropy. Eq. (1) averages kernel contributions over j = t-K+1,...,t, which mixes action samples generated from the proxy policy conditioned on earlier observations o_j rather than on the current observation o_t, so \hat p(a_t|o_t) is not a conditional density in the usual sense. No multivariate treatment of the action vector a_t is provided, and Eq. (2) sums p log p over the same samples without a normalization or volume factor that would make it a well-defined entropy estimate. Because the entire segmentation in §3.2 depends on this scalar H, the paper should either correct the estimator (e.g., use only samples from \pi_\theta(A_t|o_t) with a proper multivariate KDE and a defined entropy estimator) or empirically show that the computed quantity orders frames by task-relevant precision.
  2. [§7.1, Table 5] The comparison against Constant 2x and Constant 3x shows that DemoSpeedup balances success and speed better than uniform downsampling, but it does not isolate the entropy signal. The baselines differ from DemoSpeedup in the segmentation pattern, and no control with randomly permuted or reversed entropy labels, matched per-trajectory overall speed, and otherwise identical RBD and chunk settings is reported. Such a control is needed to rule out the possibility that the gains come primarily from a variable-rate two-speed structure rather than from the entropy-based selection of which frames to discard. This is the key missing experiment for the paper's central claim.
  3. [§4, Tables 1-2] No error bars, confidence intervals, or statistical tests are reported. Table 1 is averaged over 3 seeds but presents only point estimates, and Table 2 reports small success counts (e.g., 7/27 vs. 6/27, or 6/27 vs. 11/27) as if they were reliable differences. Because the central claim is that success is maintained or improved under speedup, the absence of uncertainty quantification makes it difficult to distinguish true parity from noise. I recommend reporting per-seed standard deviations for the simulation results and Wilson intervals or individual trial counts for the real-world results.
  4. [§3.2 and Appendix 8] Several free parameters of the pipeline are not reported: the KDE bandwidth h, the number of action samples N, the HDBSCAN hyperparameters, the Isolation Forest contamination setting, and the normalization/clustering threshold. Appendix 8 lists only the downsampling ratios r_low/r_high and the network hyperparameters. Since the precision/casual labeling depends directly on these choices, the method as described is not fully reproducible, and the sensitivity of the final success rate to h and N is unknown. Please report the values used and, ideally, include a sensitivity study for the entropy-estimation parameters.
minor comments (5)
  1. [§4.1] There is a typo in 'Fig 3.F or Aloha' which should read 'Fig. 3. For Aloha'; also, the text says the quantitative results are presented in Table 4, but the simulation results are in Table 1.
  2. [§4.1] For BiGym, the paper reports the maximum success rate over 50 evaluations during training; please state explicitly that the same checkpoint-selection protocol is applied to all compared methods, and consider also reporting results from the final checkpoint to avoid selection effects.
  3. [§4.4] The sentence 'Most Motions that approaches an object' is ungrammatical and should be rewritten, e.g., 'Most motions that approach an object'.
  4. [Throughout] Terminology is inconsistent: 'Hdbscan' and 'HDBSCAN' are both used, and the baseline 'A WE*' in Appendix 7.1 should be written consistently as 'AWE*'.
  5. [Appendix 6] The pseudocode formatting contains broken line breaks, e.g., 'def accelerate demos(Hlist,D):' and the label 'D speedup'; please clean up the algorithm listing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the speedup is a designed mechanism validated against external baselines, and no load-bearing claim reduces to its own inputs by construction.

full rationale

The Derivation chain in DemoSpeedup is: (1) train a proxy policy on original demonstrations; (2) estimate per-frame action entropy via a kernel-density surrogate (Eqs. 1 and 2); (3) cluster entropy into precision and casualness sets; (4) downsample casual segments at higher rates while preserving precision segments, using replicate-before-downsample; and (5) train an accelerated policy deployed at the original control frequency. The reported speedup is a measured outcome of rollouts, not a predicted quantity, and it is compared against the same policy trained on non-accelerated data, ACT-2x/DP-2x, Constant 2x/3x, Contact Oracle, and AWE* baselines (Tables 1, 2, 4, 5). The central empirical claim of maintained or improved success rate is therefore externally benchmarked, so the fact that the speedup is partly by construction does not make the evaluation circular. The load-bearing premise that low action entropy corresponds to high-precision segments is introduced as 'The key insight' and 'Our key observation' rather than derived from the target result; it is an unvalidated assumption and a correctness/validity risk, but not a definitional circularity. The under-specified kernel estimate in Eq. 1, including the missing bandwidth normalization and mixing actions conditioned on earlier observations, is an implementation and validity concern, not a circular step. The Limitations section explicitly admits that 'the desired acceleration rate in DemoSpeedup needs to be manually determined' and that DemoSpeedup 'occasionally causes minor performance drops, probably because of the dynamics mismatch'; these honest limitations confirm that the empirical claims are not forced by definition. The only overlapping-author citations (e.g., [4, 5]) are used for contextual statements about common practice and are not load-bearing for the entropy-segmentation argument. No equation is defined in terms of its own output, no fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on the empirical entropy-precision premise, manually selected acceleration rates, and unreported hyperparameters. No new physical entities are introduced.

free parameters (5)
  • r_low, r_high downsampling ratios = {2,4} for most tasks, {1,3} or {2,3} for some (Appendix Tables 6 and 7)
    Manually chosen per task and algorithm; the authors state the desired acceleration rate must be manually determined. These values directly set how much faster the final policy runs.
  • KDE bandwidth h = not reported
    Gaussian kernel bandwidth in Eq. 1 controls the entropy estimate; no value or selection procedure is given.
  • Number of action samples N = not reported
    Sample count used in Eq. 1 and 2 for entropy estimation; no value or sampling rule is reported.
  • HDBSCAN clustering parameters = not reported
    Minimum cluster size and related parameters for precision segmentation in Section 3.2 are not specified.
  • Accelerated policy chunk length = half the original chunk length (ACT 0.5s vs 1s, DP 24 vs 48 and similar)
    Set to half because observed speedups are around 2x, not derived from first principles.
assumptions (4)
  • domain assumption Action entropy estimated from a proxy policy is a reliable indicator of precision requirements.
    Central premise of Section 3.1: low-entropy frames correspond to high-precision moments and high-entropy frames to casual sections. It is supported only by downstream experiment success.
  • domain assumption Slowness of human teleoperated demonstrations is the primary cause of slow policy execution.
    Section 1 asserts tardy demonstrations are the primary cause. If inference or controller delays dominate, the acceleration gains would shrink.
  • domain assumption Skipping high-entropy frames preserves task-relevant information.
    Section 3.2 assumes casual sections can be safely downsampled; ablations and baselines provide indirect evidence, not a guarantee.
  • standard math Standard KDE and plug-in entropy estimation over sampled action chunks approximate the true conditional entropy.
    Equations 1 and 2 rely on Gaussian KDE and sample-based entropy, which are standard but depend on bandwidth and sample count.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration." pith.science (2026). https://pith.science/paper/MLXPN4RZ

@misc{pith2026250605064,
  author       = {Pith},
  title        = {Pith review of: DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLXPN4RZ}},
  note         = {Machine review of arXiv:2506.05064}
}
read the original abstract

Imitation learning has shown great promise in robotic manipulation, but the policy's execution is often unsatisfactorily slow due to commonly tardy demonstrations collected by human operators. In this work, we present DemoSpeedup, a self-supervised method to accelerate visuomotor policy execution via entropy-guided demonstration acceleration. DemoSpeedup starts from training an arbitrary generative policy (e.g., ACT or Diffusion Policy) on normal-speed demonstrations, which serves as a per-frame action entropy estimator. The key insight is that frames with lower action entropy estimates call for more consistent policy behaviors, which often indicate the demands for higher-precision operations. In contrast, frames with higher entropy estimates correspond to more casual sections, and therefore can be more safely accelerated. Thus, we segment the original demonstrations according to the estimated entropy, and accelerate them by down-sampling at rates that increase with the entropy values. Trained with the speedup demonstrations, the resulting policies execute up to 3 times faster while maintaining the task completion performance. Interestingly, these policies could even achieve higher success rates than those trained with normal-speed demonstrations, due to the benefits of reduced decision-making horizons. Project Page: https://demospeedup.github.io/

Figures

Figures reproduced from arXiv: 2506.05064 by the authors.

Figure 1
Figure 1. Manipulation speed is crucial for improving the productivity and ensuring the success [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DemoSpeedup utilizes a generative policy trained from original demonstrations to estimate conditional action entropy. Actions with high entropy (red points) are down-sampled at a higher rate while actions with low entropy (green points) are down-sampled at a lower rate. The problem still exists in large datasets [19, 20, 21, 22] collected by teleoperation. While VLAs trained with those large datasets [19, 23, 24, 25… view at source ↗
Figure 3
Figure 3. Simulation tasks. The environments are from Aloha and Bigym, featuring bimanual and mobile manipulation from human-collected datasets. Method Transfer Cube Insertion Sandwich Remove Move Plate Load Cups Put Cups success rate(↑) episode len(↓) success rate(↑) episode len(↓) success rate(↑) episode len(↓) success rate(↑) episode len(↓) success rate(↑) episode len(↓) success rate(↑) success rate(↑) ACT 72% 291 21% 452 … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Real-world Setup. We consider five real-world challenging tasks. Sort, Kitchenware emphasize long-horizon manipulation that require multiple skills. Bomb Deposal requires precise manipulation. Conveyer and its variation Conveyer Fast is sensitive to manipulation speed.…
Figure 5
Figure 5. Figure 5: Entropy Visualization. We showcase snapshots from the replayed demonstration and the corresponding normalized entropy curve of Sort(upper row) and Kitchenware(lower row). The green of the curve and the background stands for segmented precision set while red represents …

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Predicting continuous B-spline action curves instead of discrete chunks lets visuomotor policies finish real and simulated manipulation tasks substantially faster without large success-rate loss.

  2. TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    TempoVLA learns a single VLA policy with controllable execution speed via variable-speed trajectory augmentation and explicit speed conditioning.

Reference graph

Works this paper leans on

40 extracted references · 8 canonical work pages · cited by 2 Pith papers

  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, page 02783649241273668, 2023

    2023

  2. [2]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π0: A vision-language-action flow model for general robot control, 2024.URL https://arxiv. org/abs/2410.24164

    arXiv 2024

  3. [3]

    T. Z. Zhao, J. Tompson, D. Driess, P. Florence, K. Ghasemipour, C. Finn, and A. Wahid. Aloha unleashed: A simple recipe for robot dexterity.arXiv preprint arXiv:2410.13126, 2024

    arXiv 2024

  4. [4]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.arXiv preprint arXiv:2403.03954, 2024

    arXiv 2024

  5. [5]

    Z. Xue, S. Deng, Z. Chen, Y . Wang, Z. Yuan, and H. Xu. Demogen: Synthetic demonstration generation for data-efficient visuomotor policy learning.arXiv preprint arXiv:2502.16932, 2025. [6]https://www.figure.ai/news/helix-logistics. Accessed: 2025-2-26

    arXiv 2025

  6. [7]

    J. Xie, Z. Wang, J. Tan, H. Lin, and X. Ma. Subconscious robotic imitation learning.arXiv preprint arXiv:2412.20368, 2024

    arXiv 2024

  7. [8]

    Cheng, J

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang. Open-television: Teleoperation with immer- sive active visual feedback.arXiv preprint arXiv:2407.01512, 2024

    arXiv 2024

  8. [9]

    R. Ding, Y . Qin, J. Zhu, C. Jia, S. Yang, R. Yang, X. Qi, and X. Wang. Bunny- visionpro: Real-time bimanual dexterous teleoperation for imitation learning.arXiv preprint arXiv:2407.03162, 2024

    arXiv 2024

Show all 40 references
  1. [10]

    Y . Ze, Z. Chen, W. Wang, T. Chen, X. He, Y . Yuan, X. B. Peng, and J. Wu. Generalizable humanoid manipulation with improved 3d diffusion policies.arXiv preprint arXiv:2410.10803, 2024

    2024 arXiv

  2. [11]

    Z. Fu, T. Z. Zhao, and C. Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation.arXiv preprint arXiv:2401.02117, 2024

    2024 arXiv

  3. [12]

    P. Wu, Y . Shentu, Z. Yi, X. Lin, and P. Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163. IEEE, 2024

    2024

  4. [13]

    S. Yang, M. Liu, Y . Qin, R. Ding, J. Li, X. Cheng, R. Yang, S. Yi, and X. Wang. Ace: A cross-platform visual-exoskeletons system for low-cost dexterous teleoperation.arXiv preprint arXiv:2408.11805, 2024

    2024 arXiv

  5. [14]

    Jiang, R

    Y . Jiang, R. Zhang, J. Wong, C. Wang, Y . Ze, H. Yin, C. Gokmen, S. Song, J. Wu, and L. Fei- Fei. Behavior robot suite: Streamlining real-world whole-body manipulation for everyday household activities.arXiv preprint arXiv:2503.05652, 2025. 9

    2025 arXiv

  6. [15]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    2023 arXiv

  7. [16]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.\pi {0.5}: a vision-language-action model with open-world general- ization.arXiv preprint arXiv:2504.16054, 2025

    2025 arXiv

  8. [17]

    D. Qu, H. Song, Q. Chen, Y . Yao, X. Ye, Y . Ding, Z. Wang, J. Gu, B. Zhao, D. Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model.arXiv preprint arXiv:2501.15830, 2025

    2025 arXiv

  9. [18]

    J. Liu, H. Chen, P. An, Z. Liu, R. Zhang, C. Gu, X. Li, Z. Guo, S. Chen, M. Liu, et al. Hy- bridvla: Collaborative diffusion and autoregression in a unified vision-language-action model. arXiv preprint arXiv:2503.10631, 2025

    2025 arXiv

  10. [19]

    Q. Bu, J. Cai, L. Chen, X. Cui, Y . Ding, S. Feng, S. Gao, X. He, X. Huang, S. Jiang, et al. Agi- bot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025

    2025 arXiv

  11. [20]

    K. Wu, C. Hou, J. Liu, Z. Che, X. Ju, Z. Yang, M. Li, Y . Zhao, Z. Xu, G. Yang, et al. Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation.arXiv preprint arXiv:2412.13877, 2024

    2024 arXiv

  12. [21]

    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.arXiv preprint arXiv:2403.12945, 2024

    2024 arXiv

  13. [22]

    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 models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automati...

    2024

  14. [23]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    2025 arXiv

  15. [24]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    2024 arXiv

  16. [25]

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, Y . Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2411.19650, 2024

    2024 arXiv

  17. [26]

    A. S. Chen, A. M. Lessing, Y . Liu, and C. Finn. Curating demonstrations using online experi- ence.arXiv preprint arXiv:2503.03707, 2025

    2025 arXiv

  18. [27]

    L. X. Shi, A. Sharma, T. Z. Zhao, and C. Finn. Waypoint-based imitation learning for robotic manipulation.arXiv preprint arXiv:2307.14326, 2023

    2023 arXiv

  19. [28]

    Belkhale, Y

    S. Belkhale, Y . Cui, and D. Sadigh. Hydra: Hybrid robot actions for imitation learning. In Conference on Robot Learning, pages 2113–2133. PMLR, 2023

    2023

  20. [29]

    Belkhale, Y

    S. Belkhale, Y . Cui, and D. Sadigh. Data quality in imitation learning.Advances in neural information processing systems, 36:80375–80395, 2023

    2023

  21. [30]

    L. Lai, A. Z. Huang, and S. J. Gershman. Action chunking as policy compression. 2022

    2022

  22. [31]

    J. Heer. Fast & accurate gaussian kernel density estimation. In2021 IEEE Visualization Conference (VIS), pages 11–15. IEEE, 2021. 10

    2021

  23. [32]

    D. Xu, Y . Wang, Y . Meng, and Z. Zhang. An improved data anomaly detection method based on isolation forest. In2017 10th international symposium on computational intelligence and design (ISCID), volume 2, pages 287–291. IEEE, 2017

    2017

  24. [33]

    McInnes, J

    L. McInnes, J. Healy, S. Astels, et al. hdbscan: Hierarchical density based clustering.J. Open Source Softw., 2(11):205, 2017

    2017

  25. [34]

    Chernyadev, N

    N. Chernyadev, N. Backshall, X. Ma, Y . Lu, Y . Seo, and S. James. Bigym: A demo-driven mobile bi-manual manipulation benchmark.arXiv preprint arXiv:2407.07788, 2024

    2024 arXiv

  26. [35]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InPro- ceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  27. [36]

    Y . Seo, J. Kim, S. James, K. Lee, J. Shin, and P. Abbeel. Multi-view masked world models for visual robotic manipulation. InInternational Conference on Machine Learning, pages 30613–30632. PMLR, 2023

    2023

  28. [37]

    Y . S. P. Abbeel. Coarse-to-fine q-network with action sequence for data-efficient robot learning. Target, 2(4e9):6e9. [38]https://docs.galaxea-ai.com/zh/Guide/R1/R1_VR_Teleop_Usage_Tutorial/

  29. [39]

    Pertsch, K

    K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    2025 arXiv

  30. [40]

    Prasad, K

    A. Prasad, K. Lin, J. Wu, L. Zhou, and J. Bohg. Consistency policy: Accelerated visuomotor policies via consistency distillation.arXiv preprint arXiv:2405.07503, 2024

    2024 arXiv

  31. [41]

    G. Lu, Z. Gao, T. Chen, W. Dai, Z. Wang, W. Ding, and Y . Tang. Manicm: Real-time 3d diffu- sion policy via consistency model for robotic manipulation.arXiv preprint arXiv:2406.01586, 2024

    2024 arXiv

  32. [42]

    Shridhar, L

    M. Shridhar, L. Manuelli, and D. Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. InConference on Robot Learning, pages 785–799. PMLR, 2023. 11 Appendix 6DemoSpeeduppseudocode We provide the complete pseudocode ofDemoSpeedupin Algorithm 1. Algorithm 1:D...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.