Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

MimicFunc: Imitating Tool Manipulation from a Single Human Video via Functional Correspondence

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A robot can learn a tool-use skill from a single RGB-D video of a human and perform the same function with novel, visually different tools, by matching a function-centered frame from three keypoints instead of geometric similarity.

desk verdict Solid one-shot tool-imitation paper; the function-frame representation is a real contribution, but the evaluation misses grasp-point accuracy and confidence intervals. read the letter →

arxiv 2508.13534 v1 pith:GB32RU2R submitted 2025-08-19 cs.RO cs.AIcs.CV

classification cs.ROcs.AIcs.CV
keywords toolmanipulationimitationfromhumanvideofunctionalcorrespondencefunctionframekeypointabstractionone-shotgeneralizationvisuomotorpolicylearningRGB-D
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

The paper claims that one human demonstration video is enough for a robot to generalize a tool-use skill to novel tools that look very different but share the same function. Its central move is to replace geometric or visual correspondence with functional correspondence: each tool is reduced to three 3D keypoints—where it must touch the target, where the hand grasps it, and a center anchor—which define a 'function frame' that captures the invariant motion pattern of the task. The robot transfers the skill by aligning this frame between the demonstrated tool and the test tool, using a vision-language model with the demo keypoints as in-context references to propose candidate regions and a dense semantic correspondence model to refine the keypoints. In real-robot experiments across pour, cut, scoop, brush, and pound, the paper reports 79.5% success on novel-tool generalization, and shows that successful rollouts can be recycled to train visuomotor policies, replacing teleoperation data collection for new objects. If correct, this offers a scalable way to teach robots from human videos rather than from robot demonstrations.

What carries the argument

The function frame: a local coordinate frame built from the three keypoints, with the function point as origin and the unit vector from the center point to the function point as the principal axis; the plane normal comes from the cross product of the center-to-function and grasp-to-function vectors. This frame carries the argument because it encodes the spatiotemporal pattern of tool use—approach, grasp, direct the functional part toward the target—independent of shape. Skill transfer becomes a constrained frame-alignment optimization: the function keyframe must satisfy point, axis, and plane interaction primitives, and the full trajectory is warped and optimized to fit the test tool's frame

What would settle it

On a held-out set of tools with large intra-function variation (e.g., teapot vs. ladle for pouring), compare end-to-end success when the VLM region proposal is replaced by random regions: if success stays near the reported 79.5%, the proposed perception cascade is not the reason for the result; the paper's own ablation predicts a large drop.

Watch

Extended reading notes

Core claim

MimicFunc's discovery is that functional correspondence between tools can be established through a structured, function-centric skeleton rather than through shape matching. The skeleton consists of three keypoints: the function point where the tool interacts with the target, the grasp point where the hand holds the tool, and the center point serving as an object-agnostic anchor. These define the function frame, a local coordinate frame whose origin is the function point and whose principal axis points from the center to the function point. The paper reports that aligning this frame at the function keyframe—enforcing point, axis, and plane constraints, plus a vision-language state evaluator t

Load-bearing premise

The whole transfer collapses if the perception stack—VLM region proposals plus dense correspondence plus depth back-projection—cannot put the function and grasp points in the right 3D places on tools that look very different from the demonstrated one.

Editorial extensions

If this is right

  • One RGB-D human video can supply the training signal for a new tool in about 5.1 seconds of capture, versus roughly 48 seconds of teleoperation per demonstration.
  • A skill learned from one tool transfers to multiple levels of novelty: new spatial layouts, new instances, new categories, and even different robot embodiments and environments.
  • Rollouts generated by the method, filtered by a VLM success check, can train visuomotor policies that improve novel-instance success by 41.6% and novel-category success by 50.0% over policies trained only on teleoperation data.
  • Long-horizon tasks can be assembled by chaining function plans, with 76.0% task success and 80.0% step completion; the reported bottleneck is arm reachability, not correspondence.
  • Swapping out the perception cascade (VLM region proposal plus dense correspondence) for either component alone degrades keypoint accuracy and downstream success, so the function frame's power depends on the full transfer stack.

Reading between the lines

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

  • The function-frame representation should extend to bimanual or multi-finger manipulation by adding one grasp point per hand; the paper names this as future work, and the frame construction suggests the extension is natural.
  • Because the frame is deliberately insensitive to function-irrelevant geometry, the method may transfer across tool categories that share only an affordance—any flat rigid object for pounding, any concave object for scooping—rather than only across visually similar instances; the category-level experiments only begin to test this boundary.
  • The render-and-ask VLM evaluator used at keyframe alignment could also run during execution as a closed-loop corrective signal, which might address the contact-rich failure cases the paper attributes to trajectory generation.
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

3 major / 5 minor

Summary. The paper proposes MimicFunc, a framework for one-shot imitation of tool manipulation from a single RGB-D human video. It extracts three functional keypoints (function, grasp, center), constructs a function-centric local frame, transfers these keypoints to novel test tools via VLM-based region proposal and dense semantic correspondence, and synthesizes an SE(3) trajectory through constrained optimization. The method is evaluated on five functions (Pour, Cut, Scoop, Brush, Pound) with a real robot, reporting a 79.5% average success rate for novel-tool generalization, outperforming three baselines. The generated rollouts are additionally used as data augmentation for training ACT and Diffusion Policy visuomotor policies. The evaluation metric is external task success, so the central claim is not circular; however, the strength of the claim is undercut by missing grasp-point evaluation and lack of statistical significance testing.

Significance. If the reported results are reliable, the function-frame abstraction is a valuable contribution: it provides an interpretable, function-centric representation that can transfer manipulation skills across tools with substantial intra-function variation, going beyond geometric/visual similarity. The paper's strengths include real-robot evaluation with defined success conditions, a modular pipeline with ablations, a perception-level evaluation of function-point transfer, and planned code/video release. The data-generation experiments for visuomotor policies point to a scalable alternative to teleoperation, though the evidence is preliminary. The main weaknesses are the absence of any evaluation of grasp-point transfer accuracy—despite grasping being the largest failure category—and the lack of confidence intervals or significance tests for all success-rate claims. These issues are load-bearing for the central 79.5% generalization claim and need to be addressed.

major comments (3)
  1. [§A.2.1, Table 2; §3.3 Algorithm 1] The paper evaluates functional keypoint transfer only for the function point (AKD 18.54 px, AP@45 94.44%), never for the grasp point. Algorithm 1 transfers both q_func and q_grasp through the same cascade, and Figure 8 attributes the largest share of failures to grasping. Without measuring grasp-point localization accuracy on test tools, it is impossible to tell whether the 20.5% failure rate stems from perception error or gripper/execution limits. Please report grasp-point AKD/AP, or analyze failure cases by comparing predicted vs. ground-truth grasp points. This is a required missing link for the novel-tool generalization claim.
  2. [§4.2, Figure 3] Success rates are reported as point estimates (e.g., 79.5% overall, per-task rates) with no confidence intervals, error bars, or statistical tests. The claim that MimicFunc 'significantly outperforms' baselines is not supported without such analysis, especially given only 10 trials per task. Please report per-task counts, Wilson intervals or bootstrapped CIs, and pairwise significance tests (or at least effect sizes with uncertainty) for the comparisons in Figure 3 and Table 1.
  3. [§4.2, Figure 4 and §A.4] The visuomotor-policy data-augmentation experiment does not isolate the benefit of MimicFunc-generated data from the increased number of demonstrations. ACT is trained on 50 teleoperation demos, while ACT+DA adds MimicFunc rollouts on the novel objects; there is no control with an equivalent number of additional teleoperation demos on the same novel objects. The claim that MimicFunc data 'surpasses the variability and limited precision typically observed in human teleoperation' is also supported only by a single 50-sample comparison (53.85% vs 46.15% in §A.4) without significance testing. Please add a teleoperation matching control or clearly state the comparison as 'vs. no additional teleoperation data' and temper the corresponding claims.
minor comments (5)
  1. [§4.1 Baselines] Typo: 'DINOB OT' should be 'DINOBot'. Also 'DINOBOT' appears elsewhere; please standardize.
  2. [§3.4 Eq. (1)] The summation notation appears as 'N −1X' due to missing formatting; please fix to \sum_{t=0}^{N-1}.
  3. [§A.2.1 Table 2] Caption says 'Demo+VLM+DCS' while the text uses 'DSC' (dense semantic correspondence). Use one consistent abbreviation.
  4. [§A.1.2 Figure 7] Figure 7 shows human demonstrations but no robot executions; the qualitative comparison would be stronger if paired with corresponding robot rollouts in the same figure.
  5. [§3.4] The constraints Π_init and Π_func are defined only informally; please explicitly state how Π_init is constructed from the test tool's initial state and how the optimization is initialized.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central success metric is external task success, and the keypoint-transfer pipeline is the method itself, not a fitted prediction.

full rationale

The paper's central claim is that MimicFunc achieves 79.5% success on novel-tool generalization. This is measured by physical task-success conditions (Appendix A.1.1: 'The particles within the tool are transferred into the target container', 'The blade of the tool makes contact with the target from above', etc.), which are external to the method's internal representations. The functional keypoints, function frames, and trajectory optimization are the mechanism of the method, not circular validation targets. The functional-keypoint transfer accuracy is evaluated against human-annotated ground-truth keypoints (Table 2), and the ablation in Figure 6 compares component variants; this is standard method evaluation, not a fitted input renamed as prediction. The VLM-based state evaluator in Section 3.3 and A.2.3 is used inside the pipeline to accept or refine function-frame alignments; although A.4 also uses a VLM to filter generated rollouts, the main baseline comparison and reported success rates are grounded in physical task conditions, so this is at most an evaluation-design concern for the data-generation experiment rather than a circular derivation of the core claim. No load-bearing self-citation chain is present; the cited prior works on keypoints, correspondence models, and VLM prompting are external building blocks. The absence of grasp-point transfer accuracy in the perception evaluation is a correctness/completeness gap, not a circularity: it does not show that any claimed prediction is equivalent to its inputs by construction. Overall, the derivation chain is self-contained against external task success, and no step reduces to its own inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the representational adequacy of the function frame and on the reliability of several pre-trained components (VideoCLIP, HaMeR, CoTracker, VLM, DSC). These are assumed working tools, not derived or shipped with the paper. Free parameters such as cost relaxation, weights, scaling, and resampling are hand-picked and not swept. The invented entity is the function frame itself, which has no independent evidence outside the paper's own success metrics.

free parameters (4)
  • Trajectory cost relaxation fraction = 0.3 (initial 30% of trajectory)
    Chosen by hand to allow approach-phase flexibility in the optimization; not swept or justified.
  • Function frame optimization cost weights = implicit equal weighting of point and rotation terms
    The objective in Eq. (1) sums an L2 point distance and a Log-map rotation distance with no reported weights; the balance affects generated trajectories.
  • Trajectory warping scale factor s = not specified
    Optional scaling of the demonstration trajectory to the test tool; the paper does not state how s is set.
  • VLM refinement resampling parameters = not specified
    Number of resampling iterations and search radius around initial constraints are implementation details not reported.
assumptions (6)
  • domain assumption The function frame defined by function point, grasp point, and center point is a sufficient invariant representation of tool-use function across tools.
    Core representational premise of the method, introduced in Sections 3.2-3.3; motivated by cognitive anthropology but not derived or validated independently of the paper's success metrics.
  • domain assumption The target object frame estimated via PCA of the segmented point cloud is a stable and unambiguous reference frame.
    All trajectories are expressed in this frame (Section 3.2, Appendix A.1.4); sign flips or symmetry ambiguities would break the transfer.
  • domain assumption VideoCLIP keyframe discovery correctly identifies the initial, grasping, and function keyframes in the human video.
    Section 3.2; erroneous keyframes would misalign the entire trajectory.
  • domain assumption The demonstration-guided VLM region proposal plus dense semantic correspondence produces accurate function and grasp keypoints on test tools.
    Algorithm 1 and Section 3.3; the ablation in Fig 6 shows the cascade is critical to performance.
  • domain assumption The VLM state evaluator can correctly judge functional validity of rendered interactions and identify failing primitives.
    Section 3.3; the refinement stage depends on VLM feedback, which may be unreliable.
  • standard math Nonlinear optimization solved with CasADi and IPOPT converges to a useful local solution of the stated trajectory cost.
    Appendix A.3; standard numerical optimization is assumed reliable.
invented entities (1)
  • Function frame
    purpose: Function-centric local coordinate frame (origin at function point, function axis from center to function point) used to establish correspondences and transfer trajectories across tools.
    New representational construct introduced by this paper; there is no falsifiable handle outside the paper, and its adequacy is an assumption rather than a derived result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MimicFunc: Imitating Tool Manipulation from a Single Human Video via Functional Correspondence." pith.science (2026). https://pith.science/paper/GB32RU2R

@misc{pith2026250813534,
  author       = {Pith},
  title        = {Pith review of: MimicFunc: Imitating Tool Manipulation from a Single Human Video via Functional Correspondence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GB32RU2R}},
  note         = {Machine review of arXiv:2508.13534}
}
read the original abstract

Imitating tool manipulation from human videos offers an intuitive approach to teaching robots, while also providing a promising and scalable alternative to labor-intensive teleoperation data collection for visuomotor policy learning. While humans can mimic tool manipulation behavior by observing others perform a task just once and effortlessly transfer the skill to diverse tools for functionally equivalent tasks, current robots struggle to achieve this level of generalization. A key challenge lies in establishing function-level correspondences, considering the significant geometric variations among functionally similar tools, referred to as intra-function variations. To address this challenge, we propose MimicFunc, a framework that establishes functional correspondences with function frame, a function-centric local coordinate frame constructed with keypoint-based abstraction, for imitating tool manipulation skills. Experiments demonstrate that MimicFunc effectively enables the robot to generalize the skill from a single RGB-D human video to manipulating novel tools for functionally equivalent tasks. Furthermore, leveraging MimicFunc's one-shot generalization capability, the generated rollouts can be used to train visuomotor policies without requiring labor-intensive teleoperation data collection for novel objects. Our code and video are available at https://sites.google.com/view/mimicfunc.

Figures

Figures reproduced from arXiv: 2508.13534 by the authors.

Figure 1
Figure 1. Given a single human video, MimicFunc enables the robot to manipulate novel tools [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of MimicFunc Pipeline. MimicFunc consists of three stages: (1) Functional keypoint extraction from human video, (2) Functional correspondence establishment with function frame, and (3) Function frame-based action generation. represents the 6-DoF end-effector pose at timestep t, where Rt ∈ SO(3) and Tt ∈ R 3 denote 3D orientation and translation, respectively. 3.2 Functional Keypoint Extraction from Human Vi… view at source ↗
Figure 3
Figure 3. Quantitative comparison to baselines. Highlighted tools are used in human videos. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Performance evaluation of visuomotor policy training. Evaluation on Data Generation for Visuo￾motor Policy Training. To support the claim that the rollout data generated by MimicFunc can be leveraged for visuomotor policy train￾ing, we conduct experiments using the BC …
Figure 5
Figure 5. Figure 5: Visualization of grasping and function keyframes of human demonstrations and robot [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study results. Ablation Study. We conduct an ablation study on the functional keypoint transfer strat￾egy. Three strategies are evaluated: (1) Demo+VLM+DSC (ours); (2) Demo+DSC, which relies solely on a dense semantic corre￾spondence model for keypoint transfe…
Figure 7
Figure 7. Figure 7: Visualization of grasping and function keyframes of human demonstrations. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Failure analysis of system components. The modular design of MimicFunc facilitates the interpretation and in-depth analysis of fail￾ure cases. The result of the failure analysis is reported in [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results of real-robot executions. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Intermediate rendering results of function axis refinement. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Continue with ORCID 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. AffordGen: Generating Diverse Demonstrations for Generalizable Object Manipulation with Afford Correspondence

    cs.RO 2026-04 unverdicted novelty 6.0 of 10

    AffordGen synthesizes large-scale affordance-aware manipulation trajectories via keypoint correspondence on 3D meshes, enabling zero-shot visuomotor policies for unseen objects from few source demos.

  2. EAvatar: Expression-Aware Head Avatar Reconstruction with Generative Geometry Priors

    cs.CV 2025-08 unverdicted novelty 5.0 of 10

    EAvatar reconstructs expression-aware head avatars by combining a sparse key-Gaussian control mechanism with generative 3D geometry priors, reporting finer, more controllable facial detail.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages · cited by 2 Pith papers

  1. [1]

    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. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

  2. [2]

    Cheang, S

    C. Cheang, S. Chen, Z. Cui, Y . Hu, L. Huang, T. Kong, H. Li, Y . Li, Y . Liu, X. Ma, et al. Gr-3 technical report. arXiv preprint arXiv:2507.15493, 2025

  3. [3]

    Vitiello, K

    P. Vitiello, K. Dreczkowski, and E. Johns. One-shot imitation learning: A pose estimation perspective. In Conference on Robot Learning, pages 943–970. PMLR, 2023

  4. [4]

    Heppert, M

    N. Heppert, M. Argus, T. Welschehold, T. Brox, and A. Valada. Ditto: Demonstration imitation by trajectory transformation. arXiv preprint arXiv:2403.15203, 2024

  5. [5]

    Di Palo and E

    N. Di Palo and E. Johns. Dinobot: Robot manipulation via retrieval and alignment with vision foundation models. arXiv preprint arXiv:2402.13181, 2024

  6. [6]

    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

  7. [7]

    J. Li, Y . Zhu, Y . Xie, Z. Jiang, M. Seo, G. Pavlakos, and Y . Zhu. Okami: Teaching humanoid robots manipulation skills through single video imitation. In 8th Annual Conference on Robot Learning

  8. [8]

    O. Biza, S. Thompson, K. R. Pagidi, A. Kumar, E. van der Pol, R. Walters, T. Kipf, J.-W. van de Meent, L. L. Wong, and R. Platt. One-shot imitation learning via interaction warping. In Conference on Robot Learning, pages 2519–2536. PMLR, 2023

Show all 44 references
  1. [9]

    Zhang and A

    X. Zhang and A. Boularias. One-shot imitation learning with invariance matching for robotic manipulation. arXiv preprint arXiv:2405.13178, 2024

  2. [10]

    J. Zhu, Y . Ju, J. Zhang, M. Wang, Z. Yuan, K. Hu, and H. Xu. Densematcher: Learning 3d se- mantic correspondence for category-level manipulation from a single demo. In The Thirteenth International Conference on Learning Representations. 9

  3. [11]

    S. L. Washburn. Tools and human evolution. Scientific American, 203(3):62–75, 1960

  4. [12]

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469–483, 2009

  5. [13]

    Ravichandar, A

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard. Recent advances in robot learning from demonstration. Annual review of control, robotics, and autonomous systems, 3 (1):297–330, 2020

  6. [14]

    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

  7. [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

  8. [16]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024

  9. [17]

    C. Finn, T. Yu, T. Zhang, P. Abbeel, and S. Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357–368. PMLR, 2017

  10. [18]

    Y . Duan, M. Andrychowicz, B. Stadie, O. Jonathan Ho, J. Schneider, I. Sutskever, P. Abbeel, and W. Zaremba. One-shot imitation learning. Advances in neural information processing systems, 30, 2017

  11. [19]

    T. Yu, C. Finn, S. Dasari, A. Xie, T. Zhang, P. Abbeel, and S. Levine. One-shot imitation from observing humans via domain-adaptive meta-learning. Robotics: Science and Systems XIV , 2018

  12. [20]

    Gleicher

    M. Gleicher. Retargetting motion to new characters. In Proceedings of the 25th annual con- ference on Computer graphics and interactive techniques, pages 33–42, 1998

  13. [21]

    K. Hu, C. Ott, and D. Lee. Online human walking imitation in task and joint space based on quadratic programming. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3458–3464. IEEE, 2014

  14. [22]

    Manuelli, W

    L. Manuelli, W. Gao, P. Florence, and R. Tedrake. kpam: Keypoint affordances for category- level robotic manipulation. In The International Symposium of Robotics Research, pages 132–

  15. [23]

    F. Liu, K. Fang, P. Abbeel, and S. Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024

  16. [24]

    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. In 8th Annual Conference on Robot Learning

  17. [25]

    J. Gao, Z. Tao, N. Jaquier, and T. Asfour. K-vil: Keypoints-based visual imitation learning. IEEE Transactions on Robotics, 2023

  18. [26]

    J. Gao, X. Jin, F. Krebs, N. Jaquier, and T. Asfour. Bi-kvil: Keypoints-based visual imitation learning of bimanual manipulation tasks. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16850–16857. IEEE, 2024

  19. [27]

    Nasiriany, F

    S. Nasiriany, F. Xia, W. Yu, T. Xiao, J. Liang, I. Dasgupta, A. Xie, D. Driess, A. Wahid, Z. Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. In Forty-first International Conference on Machine Learning. 10

  20. [28]

    H. Xu, G. Ghosh, P.-Y . Huang, D. Okhonko, A. Aghajanyan, F. Metze, L. Zettlemoyer, and C. Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Process- ing, pa...

  21. [29]

    Pavlakos, D

    G. Pavlakos, D. Shan, I. Radosavovic, A. Kanazawa, D. Fouhey, and J. Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826–9836, 2024

  22. [30]

    Karaev, I

    N. Karaev, I. Rocco, B. Graham, N. Neverova, A. Vedaldi, and C. Rupprecht. Cotracker: It is better to track together. In European Conference on Computer Vision, pages 18–35. Springer, 2025

  23. [31]

    Zhang, C

    J. Zhang, C. Herrmann, J. Hur, L. Polania Cabrera, V . Jampani, D. Sun, and M.-H. Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024

  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. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17359–17369, 2025

  25. [33]

    J. Sola, J. Deray, and D. Atchuthan. A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537, 2018

  26. [34]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1–31, 2024

  27. [35]

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021

  28. [36]

    C. R. Garrett, T. Lozano-P´erez, and L. P. Kaelbling. Pddlstream: Integrating symbolic planners and blackbox samplers via optimistic adaptive planning. In Proceedings of the international conference on automated planning and scheduling, volume 30, pages 440–448, 2020

  29. [37]

    use scrubber to brush the plate

    Y . Ju, K. Hu, G. Zhang, G. Zhang, M. Jiang, and H. Xu. Robo-abc: Affordance general- ization beyond categories via semantic correspondence for robot manipulation. In European Conference on Computer Vision, pages 222–239. Springer, 2024. 11 A Appendix A.1 Experimental Setup an...

  30. [38]

    It points from the center point pt center to the function point pt func at t

    Function axis • Definition: vt H = pt func − pt center ∥pt func − pt center∥ • Description: vt H is a normalized vector that defines the function axis. It points from the center point pt center to the function point pt func at t. This axis represents the principal direction al...

  31. [39]

    Grasp vector • Definition: ut H = pt grasp − pt func ∥ptgrasp − pt func∥ • Description: ut H is a normalized vector that points from the function pointpt func to the grasp point pt grasp at t

  32. [40]

    Unit normal vector • Definition: nt H = ut H × vt H ∥ut H × vt H ∥ • Description: nt H is the unit normal vector of the function plane P t H

  33. [41]

    Similarly, vt R, ut R, nt R, and P t R are defined for Πt R

    Function plane • Definition: P t H : (p − pt func) · nt H = 0 • Description: P t H is defined by the function point and its normal vector, describing the tool’s spatial configuration at t. Similarly, vt R, ut R, nt R, and P t R are defined for Πt R. 17 A.2.3 Function Frame Ali...

  34. [42]

    Let Rsym ∈ SO(3) denote a symmetry rotation

    Symmetry-Based Repositioning: If the target object exhibits geometric symmetries (e.g., rotational symmetry about one of the principal axes), we exploit this property to reposition the demonstration so that the test tool can adopt a more feasible approach direction. Let Rsym ∈...

  35. [43]

    Function Frame Trajectory Pre-processing: We pre-process the demonstration’s function frame trajectory by first applying a rotation around one of the principal axes (e.g., x-, y-, or z-axis). The alignment angle θ is computed based on the angular difference between the initial...

  36. [44]

    Beyond the trajectory cost and keyframe constraints detailed in the manuscript, we introduce the following enhancements: • Early Trajectory Cost Relaxation

    Function Frame Trajectory Transformation: To account for differences in position and scale between the demonstration and the test tool, we apply a translation t ∈ R3 and an optional scaling factor s ∈ R: Πt warp = s · Πt align + t 18 Optimization Constraints and Costs. Beyond ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.