Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Object-Focus Actor for Data-efficient Robot Generalization Dexterous Manipulation

T0 review · 4 major / 6 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 The paper claims that a single fixed pre-manipulation pose per object category, combined with a policy trained only on the final hand-focused trajectory, lets dexterous manipulation be learned from 10 demonstrations and generalized to new…

desk verdict A promising but under-validated system paper: the pipeline idea has real legs, yet the hand-set pre-manipulation offset is an unexamined load-bearing assumption. read the letter →

arxiv 2505.15098 v1 pith:RV3UOJDF submitted 2025-05-21 cs.RO cs.AI

Yihang Li , Tianle Zhang , Xuelong Wei , Jiayi Li , Lin Zhao , Dongchi Huang , Zhirui Fang , Minhua Zheng
show 2 more authors
Wenjun Dai Xiaodong He
This is my paper
classification cs.ROcs.AI
keywords dexterousmanipulationimitationlearningdataefficiencygeneralizationobject-centricpolicyhand-focusimagespre-manipulationposeactionchunking
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 argues that the main obstacle to generalizable dexterous manipulation — the need for large amounts of data covering many object positions and backgrounds — can be sidestepped by exploiting a simple regularity: once the robot hand reaches a fixed 'pre-manipulation pose' relative to the object, the remaining wrist and finger motion is nearly the same regardless of where the object started. The proposed Object-Focus Actor (OFA) therefore learns only this final consistent phase, using hand-focused camera images and poses and actions expressed relative to the pre-manipulation pose. Real-world experiments across seven tasks (six single-hand and one bimanual) report that this approach outperforms the ACT baseline, generalizes to out-of-distribution object placements and background changes, and retains high success rates with as few as 10 demonstrations. If the claims hold, imitation learning for dexterous hands becomes practical and data-efficient in settings that previously seemed to require VLA-style large-scale pretraining.

What carries the argument

The object-focus end trajectory is the paper's central object: the observation that the final part of a dexterous manipulation is nearly invariant once the hand reaches a canonical pose relative to the object. On top of this, the pre-manipulation pose — a hand-set 6D offset from the estimated object pose, defined per object category by Eq. (1) — is the mechanism that converts arbitrary start states into a common policy input. The learning pipeline then uses hand-focus images (the projected hand bounding box enlarged to twice its area so the object is included) and relative representations (proprioception and action chunks expressed relative to the pre-manipulation pose) to train a conditional VAE policy that mimics the consistent end trajectory.

What would settle it

Re-run OFA with controlled amounts of noise added to the estimated object pose (for example, 1–5 cm translation and 5–10 degree rotation errors) before computing the pre-manipulation pose; if success rate drops sharply even at small perturbations, the fixed-offset assumption is the bottleneck. Alternatively, test objects of the same category but noticeably different size or shape; if success collapses on the same-category variants, the category-level offset is too rigid.

Watch

Extended reading notes

Core claim

The central discovery is that dexterous manipulation trajectories converge: regardless of the object's initial position, when the hand approaches the object the subsequent wrist and finger motions follow nearly the same path. OFA exploits this by decomposing the task into a geometric stage — estimate the object's 6D pose with FoundationPose [22], compute a fixed category-level pre-manipulation offset, and plan a collision-free path with CuRobo [23] — and a learned stage, an ACT-style CVAE policy that takes as input a local 'hand-focus' image (the projected hand region enlarged to include the object) plus relative proprioception, and outputs relative action chunks. The paper reports that this two-stage design outperforms ACT and its own ablations across seven real-world tasks, generalizes to out-of-distribution positions and backgrounds, and retains competence with only 10 demonstrations.

Load-bearing premise

The method assumes that a single, hand-set pre-manipulation pose per object category is reachable and correct for every object instance and placement, and that the 6D pose estimator is accurate enough to deliver it.

Editorial extensions

If this is right

  • A dexterous manipulation policy trained from 10–30 demonstrations can succeed on object placements and backgrounds never seen during training, without requiring large-scale robot data.
  • The pipeline separates perception and planning (pose estimation plus motion planning) from imitation learning, so the learned policy itself is simplified and more robust.
  • Using relative pose/action representations and hand-focus images reduces the action space the policy must learn, which is why the method outperforms the ACT baseline in the reported tests.
  • The approach extends to bimanual tasks, as demonstrated by the tray-lifting task, suggesting the same convergence insight applies to coordinated dual-hand motion.
  • Because the policy is conditioned on a canonical pre-manipulation pose, the same learned end trajectory can be reused for objects of the same category placed anywhere in the workspace.

Reading between the lines

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

  • A testable extension is to learn the pre-manipulation offset from data rather than setting it by hand; if the offset is learned for each object category, the method might tolerate larger shape variation within a category.
  • The core convergence assumption suggests that any task with a stereotyped final phase — for example tool use or assembly — could be decomposed the same way, with the learned policy only responsible for the contact phase.
  • One risk not explored in the paper is sensitivity to pose-estimation noise: if FoundationPose's error grows on shiny or textureless objects, the policy could receive an inconsistent start state and fail, which could be checked by adding controlled pose perturbations during evaluation.
  • The hand-focus image is built from the robot's kinematics rather than from the object, so the approach is naturally limited to objects visible near the hand; a camera-fixed or object-centric cropping might extend it to cases where the hand is off-screen at the start.
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 / 6 minor

Summary. The paper introduces Object-Focus Actor (OFA), a hierarchical system for dexterous robot manipulation that aims to generalize to new object positions and backgrounds from very few demonstrations. The pipeline first estimates the object's 6D pose, computes a category-specific pre-manipulation pose by adding a fixed relative offset, plans a collision-free motion to that pose with cuRobo, and then executes a learned CVAE policy that uses hand-focused image crops and poses/actions expressed relative to the pre-manipulation pose. The authors evaluate on seven real-world tasks, comparing against ACT and three ablations, and report higher success rates, better out-of-distribution position/background robustness, and data efficiency down to 10 demonstrations.

Significance. The object-focus and relative-action formulation is a sensible way to reduce the effective variance seen by the policy, and the idea of exploiting a convergent pre-manipulation pose is plausible and potentially valuable. The paper includes real-robot experiments on a bimanual platform and a reasonable baseline comparison, and it explicitly targets the under-studied problem of positional generalization in dexterous manipulation. However, the strength of the claims is currently limited by the small number of trials and by the unexamined sensitivity of the hand-set pre-manipulation offset; if the authors provide the missing analyses, the approach could constitute a meaningful step toward data-efficient generalization.

major comments (4)
  1. [§3.3, Eq. (1)] The pre-manipulation pose Tm is obtained by applying a hand-selected category-level offset (∆R, ∆t) to the estimated object pose. This offset is a free parameter with no reported tuning procedure, sensitivity analysis, or validation across object instances. Because the policy's hand-focus image crop and relative proprioception/action chunks are all defined with respect to Tm, a wrong or noisy Tm places the policy outside its training distribution. The paper does not report FoundationPose estimation errors, the spread of actually reached Tm values, or any ablation over offset perturbations. Please add this analysis or explicitly discuss the limits it places on the generalization claims.
  2. [§3.4 and §4.1.3] The training data collection protocol is underspecified. The paper states that demonstrations are collected with the object placed randomly and the initial hand position fixed, but it does not say whether demonstrations begin at the pre-manipulation pose or how Tm is obtained for each demonstration. If the policy is trained on full trajectories while inference always starts at Tm, there is a distribution shift that could explain some of the reported success. Please specify the exact procedure for constructing the training (state, action) pairs relative to Tm, and confirm that the same Tm computation is used in training and inference.
  3. [§4.2, Tables 1 and 2, and Fig. 5] All quantitative results are reported from 10 evaluations per cell with no confidence intervals or statistical tests. Differences such as 'Catch Loopy' OFA 80 vs OFA w/o of 90 lie within plausible sampling noise, and the 10-demonstration claim for Grasp Cup (40% success) is not 'robust performance' as stated. Please provide per-trial data or confidence intervals (e.g., Wilson intervals) for the key comparisons, and temper the text accordingly.
  4. [§4.3 'Different manipulating-object position'] The out-of-distribution position experiment compares only OFA with ACT. This leaves the mechanism of generalization unclear: how much comes from the relative action/proprioception design (OFA w/o rel), how much from the hand-focus image (OFA w/o of), and how much from the hierarchical pre-manipulation pipeline itself? The contributions claimed for the policy design are best isolated by running the same OOD tests for the ablations.
minor comments (6)
  1. [§3.4] The enlargement factor for hand-focus images is set to 'twice' the hand's bounding box; the paper does not justify this choice or test its sensitivity.
  2. [Throughout] There are numerous typographical and formatting issues: 'enviorments' (Related Work), 'Appedndix A' (Section 4.1.2), 'theOF A' (Section 4.2), and inconsistent spacing in inline citations.
  3. [Figure 5] Figure 5 is referenced but its content (actual success rates and any error bars) is not described in the text; please provide numeric values for the position and background generalization tests.
  4. [Conclusion] The statement 'exceeding performance by more than 50%' is ambiguous (absolute vs relative improvement) and not clearly supported by the tables; please rephrase with explicit numbers.
  5. [Appendix A] The task 'Catch Loopy' is described as 'catch loopy from the environment' which is unclear; please define the object or use a standard name.
  6. [General] The paper does not include an explicit limitations discussion; a short paragraph on the dependencies on FoundationPose, cuRobo, and the hand-set offset would improve reproducibility and scholarly balance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: OFA's claims are validated by external real-world task success, and no predicted result reduces by construction to a fitted or self-cited input.

full rationale

OFA is an empirical systems paper whose central claims are tested against real-world task completion, not derived from or reduced to fitted parameters. The 'object-focus end trajectory' insight motivates the architecture (hand-focus images, relative proprioception, relative action chunks, and the pre-manipulation pose of Eq. 1), but the paper never presents a fitted constant as a prediction; the learned policy is trained with imitation loss (Eq. 2) and evaluated by success rate on held-out positions and backgrounds, which is an external benchmark. The hand-set category-level offset in Eq. 1 is a system input, and the absence of pose-error or offset-sensitivity analysis is a correctness/robustness concern, not circularity. The related-work citations to Open-TeleVision and HumanPlus include authors who overlap with this paper (Xiaodong He, Jiayi Li), but these are background references and are not load-bearing for the proposed method or its evaluation. No uniqueness theorem is imported from the authors' prior work, and no equation equates an output with an input by construction. Therefore no circular steps are present.

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

The paper does not derive any constants or laws; its free parameters are design choices in a real-robot pipeline. The most consequential is the manually-set pre-manipulation pose offset, which the learned policy assumes as its start. The main axioms are domain assumptions about task structure and the reliability of off-the-shelf perception and planning.

free parameters (4)
  • Category-specific pre-manipulation pose offset (delta R, delta t) = Not specified in paper
    Hand-set per object category in Sec. 3.3; determines the start of the learned policy and is load-bearing for all results.
  • Hand-focus image enlargement factor = 2x
    Chosen by hand in Sec. 3.4 to include both hand and object in the cropped observation; affects the visual information available to the policy.
  • CVAE loss weight eta = Not specified
    Hyperparameter in Eq. (2) balancing reconstruction and KL divergence; no ablation or sensitivity analysis is reported.
  • Action chunk length k = Not specified
    Controls the prediction horizon in the action chunk; not documented or swept.
assumptions (4)
  • domain assumption Manipulation trajectory convergence
    Sec. 1 and Fig. 1 assume the final phase of a dexterous manipulation is consistent across starting positions, which is the basis for training only on the post-pre-manipulation segment.
  • domain assumption Category-level pose offset validity
    Sec. 3.3 assumes one unified pre-manipulation pose per object category works for all objects and placements in that category.
  • domain assumption Hand-focus image sufficiency
    Sec. 3.4 assumes the enlarged projected hand area contains enough visual information for the policy to complete the task.
  • domain assumption External module reliability
    Sec. 3.2 and 3.3 rely on FoundationPose for accurate 6D pose and CuRobo for collision-free planning, without reporting pose error or planning failure rates in the results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Object-Focus Actor for Data-efficient Robot Generalization Dexterous Manipulation." pith.science (2026). https://pith.science/paper/RV3UOJDF

@misc{pith2026250515098,
  author       = {Pith},
  title        = {Pith review of: Object-Focus Actor for Data-efficient Robot Generalization Dexterous Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RV3UOJDF}},
  note         = {Machine review of arXiv:2505.15098}
}
read the original abstract

Robot manipulation learning from human demonstrations offers a rapid means to acquire skills but often lacks generalization across diverse scenes and object placements. This limitation hinders real-world applications, particularly in complex tasks requiring dexterous manipulation. Vision-Language-Action (VLA) paradigm leverages large-scale data to enhance generalization. However, due to data scarcity, VLA's performance remains limited. In this work, we introduce Object-Focus Actor (OFA), a novel, data-efficient approach for generalized dexterous manipulation. OFA exploits the consistent end trajectories observed in dexterous manipulation tasks, allowing for efficient policy training. Our method employs a hierarchical pipeline: object perception and pose estimation, pre-manipulation pose arrival and OFA policy execution. This process ensures that the manipulation is focused and efficient, even in varied backgrounds and positional layout. Comprehensive real-world experiments across seven tasks demonstrate that OFA significantly outperforms baseline methods in both positional and background generalization tests. Notably, OFA achieves robust performance with only 10 demonstrations, highlighting its data efficiency.

Figures

Figures reproduced from arXiv: 2505.15098 by the authors.

Figure 1
Figure 1. The grasping trajectories of the dexterous robot hand. The four green points denote random [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall design of OFA 3.2 Manipulating-Object Perception and Pose Estimation Object locating: To conveniently extract the interested object from origin image, the open-set grounding object detector, GroundingDINO [36] is exploited. It passes the object prior to subsequent segmentation model, SAM [37] for pixel-level location of the object. Pose estimation: The pre-manipulation pose of robot hand depends on a relativ… view at source ↗
Figure 3
Figure 3. The extraction of robot hand focus area. The left image and the right image are captured by [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The bimanual robot used in our experiment. The mainly used modules for this paper are [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Generalization testing of different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 10 canonical work pages

  1. [1]

    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

    arXiv 2023

  2. [2]

    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

    arXiv 2024

  3. [3]

    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

  4. [4]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  5. [5]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    arXiv 2023

  6. [6]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  7. [7]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  8. [8]

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024

    2024

Show all 39 references
  1. [9]

    Brohan, N

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

    2023 arXiv

  2. [10]

    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

  3. [11]

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024

    2024 arXiv

  4. [12]

    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.arXiv preprint arXiv:2410.24164, 2024

    2024 arXiv

  5. [13]

    A. M. Okamura, N. Smaby, and M. R. Cutkosky. An overview of dexterous manipulation. InProceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065), volume 1, pages 255–262. IEEE, 2000

    2000

  6. [14]

    C. Wang, H. Shi, W. Wang, R. Zhang, L. Fei-Fei, and C. K. Liu. Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

    2024 arXiv

  7. [15]

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn. Humanplus: Humanoid shadowing and imitation from humans.arXiv preprint arXiv:2406.10454, 2024

    2024 arXiv

  8. [16]

    Cheng, J

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang. Open-television: Teleoperation with immersive active visual feedback.arXiv preprint arXiv:2407.01512, 2024. 9

    2024 arXiv

  9. [17]

    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

  10. [18]

    H.-S. Fang, C. Wang, M. Gou, and C. Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11444–11453, 2020

    2020

  11. [19]

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains.IEEE Transactions on Robotics, 2023

    2023

  12. [20]

    P. Li, T. Liu, Y . Li, Y . Geng, Y . Zhu, Y . Yang, and S. Huang. Gendexgrasp: Generalizable dexterous grasping. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8068–8074. IEEE, 2023

    2023

  13. [21]

    R. Wang, J. Zhang, J. Chen, Y . Xu, P. Li, T. Liu, and H. Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359–11366. IEEE, 2023

    2023

  14. [22]

    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

    2024

  15. [23]

    Sundaralingam, S

    B. Sundaralingam, S. K. S. Hari, A. Fishman, C. Garrett, K. Van Wyk, V . Blukis, A. Millane, H. Oleynikova, A. Handa, F. Ramos, et al. Curobo: Parallelized collision-free minimum-jerk robot motion generation. arXiv preprint arXiv:2310.17274, 2023

    arXiv

  16. [24]

    B. Fang, S. Jia, D. Guo, M. Xu, S. Wen, and F. Sun. Survey of imitation learning for robotic manipulation. International Journal of Intelligent Robotics and Applications, 3:362–369, 2019

    2019

  17. [25]

    S. P. Arunachalam, S. Silwal, B. Evans, and L. Pinto. Dexterous imitation made easy: A learning-based framework for efficient dexterous manipulation. In2023 ieee international conference on robotics and automation (icra), pages 5954–5961. IEEE, 2023

    2023

  18. [26]

    S. P. Arunachalam, I. Güzey, S. Chintala, and L. Pinto. Holo-dex: Teaching dexterity with immersive mixed reality. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5962–5969. IEEE, 2023

    2023

  19. [27]

    Y . Zhu, Z. Jiang, P. Stone, and Y . Zhu. Learning generalizable manipulation policies with object-centric 3d representations.arXiv preprint arXiv:2310.14386, 2023

    2023 arXiv

  20. [28]

    Y . Zhu, A. Lim, P. Stone, and Y . Zhu. Vision-based manipulation from single human video with open-world object graphs.arXiv preprint arXiv:2405.20321, 2024

    2024 arXiv

  21. [29]

    Valassakis, G

    E. Valassakis, G. Papagiannis, N. Di Palo, and E. Johns. Demonstrate once, imitate immediately (dome): Learning visual servoing for one-shot imitation learning. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8614–8621. IEEE, 2022

    2022

  22. [30]

    Huang, C

    W. Huang, C. Wang, Y . Li, R. Zhang, and L. Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation.arXiv preprint arXiv:2409.01652, 2024

    2024 arXiv

  23. [31]

    J. Kerr, C. M. Kim, M. Wu, B. Yi, Q. Wang, K. Goldberg, and A. Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction.arXiv preprint arXiv:2409.18121, 2024

    2024 arXiv

  24. [32]

    M. Pan, J. Zhang, T. Wu, Y . Zhao, W. Gao, and H. Dong. Omnimanip: Towards general robotic manipulation via object-centric interaction primitives as spatial constraints.arXiv preprint arXiv:2501.03841, 2025

    2025 arXiv

  25. [33]

    E. Zhou, Q. Su, C. Chi, Z. Zhang, Z. Wang, T. Huang, L. Sheng, and H. Wang. Code-as-monitor: Constraint-aware visual programming for reactive and proactive robotic failure detection.arXiv preprint arXiv:2412.04455, 2024

    2024 arXiv

  26. [34]

    J. Pari, N. M. Shafiullah, S. P. Arunachalam, and L. Pinto. The surprising effectiveness of representation learning for visual imitation.arXiv preprint arXiv:2112.01511, 2021

    2021 arXiv

  27. [35]

    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

    2024 arXiv

  28. [36]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023. 10

    2023 arXiv

  29. [37]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023

    2023

  30. [38]

    De Haan, D

    P. De Haan, D. Jayaraman, and S. Levine. Causal confusion in imitation learning.Advances in neural information processing systems, 32, 2019

    2019

  31. [39]

    K. Sohn, H. Lee, and X. Yan. Learning structured output representation using deep conditional generative models.Advances in neural information processing systems, 28, 2015. 11 A Appendix Table 1: Descriptions of robot dexterous manipulation tasks Task Task Description Grasp Cu...

    2015

Pith tools

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