Pith. sign in

REVIEW 3 major objections 6 minor 28 references

Dense correspondences that stay both semantic and smooth let a robot reuse one demo on new object shapes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 09:09 UTC pith:IJYKIKMT

load-bearing objection Solid training-free assembly that actually improves geometric coherence for one-shot skill transfer; the robot gains are real but not cleanly isolated to the map. the 3 major comments →

arxiv 2607.28382 v1 pith:IJYKIKMT submitted 2026-07-30 cs.RO

SemAnCorr: Semantic Anchored Correspondence for Zero-Shot Manipulation Skill Transfer

Xiaoxiang Dong , William Baron , Hongyi Chen , Uksang Yoo , Jeffrey Ichnowski , Weiming Zhi This is my paper
classification cs.RO
keywords dense correspondencefunctional mapszero-shot skill transfersemantic anchorsarticulated manipulationgeometric coherenceobject-centric skills
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Robots often learn a skill on one object and fail on the next mug, bottle, or pair of scissors even when the function is the same, because geometry changed. Nearest-neighbor matching of visual features finds roughly the right part but scatters matches so badly that the local frames needed for how to grasp and move are lost. SemAnCorr is a training-free pipeline that first picks a few mutually consistent semantic anchor regions by jointly optimizing pose and part matching, then spreads those constraints across the whole surface with functional maps. The resulting vertex map is both part-correct and spatially coherent. On a PartNet-Mobility correspondence benchmark it reaches 90.8 percent semantic accuracy and more than doubles geometric coherence over strong baselines; in five real robot tasks a single demonstration transfers more reliably to unseen objects than prior correspondence methods.

Core claim

Correspondences that are simultaneously semantically consistent and geometrically coherent are both necessary and achievable without training, and they translate directly into more reliable zero-shot transfer of object-centric manipulation skills from one demonstration to geometrically different instances.

What carries the argument

Semantic Anchored Correspondence (SemAnCorr): joint pose-correspondence optimization selects a small set of high-confidence semantic anchors, then a constrained functional map propagates those anchors into a dense, smooth vertex-level map that recovers local geometric frames.

Load-bearing premise

Each object mesh must be one connected piece, and a few mutually correct semantic part pairs must exist and be recoverable as anchors; if the anchors are wrong or the mesh is broken, the smooth map spreads the error.

What would settle it

On the same PartNet-Mobility pairs or the five real robot tasks, measure whether forcing wrong or missing anchors, or using disconnected real-world meshes, collapses geometric coherence and task success while semantic part labels still look plausible.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single kinesthetic demonstration can be replayed on new object instances that share function but not shape, without category-specific training.
  • Semantic accuracy alone is insufficient for skill transfer; geometric coherence is equally required to recover interaction frames.
  • Anchor confidence scores can serve as a lightweight prior for whether a demonstrated skill is likely to transfer across categories.
  • Successful transferred executions can be collected as extra demonstrations to bootstrap learned manipulation policies.

Where Pith is reading between the lines

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

  • The same anchor-plus-functional-map pattern may extend to multi-contact bimanual or dexterous skills if multiple anchors stay mutually consistent.
  • Noisy single-view mesh reconstructions are likely the practical bottleneck; better surface completion could raise real-world success without changing the correspondence method.
  • Cross-category anchor confidence could be turned into an automatic curriculum or retrieval signal for which past demos to attempt on a novel object.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. SemAnCorr is a training-free pipeline for dense vertex-level correspondence between object meshes that share functional structure but differ in geometry. It lifts multi-view SigLIP2 features into 3D, clusters them into semantic parts, selects a small set of anchor pairs via bilateral-margin ranking and joint rigid-pose / soft-correspondence optimization, then propagates those anchors with a ZoomOut-style functional map under local neighborhood and anchor-pinning constraints. The authors introduce a PartNet-Mobility dense-correspondence benchmark scoring both semantic accuracy (sAcc) and a Geometric Coherence Score (GCS = harmonic mean of continuity and coverage), report 90.8% average within-category sAcc and roughly 2× GCS over the strongest baseline (D3Fields), show limited cross-category pairs and component ablations, and transfer single kinesthetic demonstrations to five real object pairs via Procrustes-guided GPG grasps, outperforming D3Fields on the harder tasks (Table V).

Significance. Skill transfer across geometrically diverse but functionally related instances is a central open problem in robot learning. The paper’s contribution is practical and well-scoped: a training-free recipe that couples pretrained semantic features with classical functional maps, plus an explicit manipulation-oriented geometric metric (GCS) that separates “where” from “how.” Strengths include a multi-category benchmark with reported means and standard deviations (Tables I–III), a clean three-way ablation (Table IV), and real-robot trials rather than simulation-only claims. If the causal link from geometric coherence to transfer success holds under tighter controls, the work would be a useful reference method and evaluation protocol for object-centric zero-shot manipulation.

major comments (3)
  1. [§V-G, Table V] Abstract and §V-G claim that sAcc/GCS gains “translate directly” into more reliable zero-shot skill transfer (Table V). The transfer stack is not pure map application: contact regions are Procrustes-aligned, GPG samples grasps ranked by quality and SE(3) proximity to T_prior, and motions are relative to the executed grasp (§V-G, Skill Transfer). Tasks 1–2 already succeed for both methods; the 4–6 trial gaps on Tasks 3–5 could come from better semantic part localization feeding GPG rather than from functional-map continuity that GCS measures. Please isolate correspondence as the sole variable (same reconstructed meshes, same GPG candidates, swap only the map) and/or report Cont/Cov/GCS on the five real object pairs so the causal claim is supported rather than inferred.
  2. [§III, §IV-C, §V-G] §III and §IV-C require each mesh to be a single connected component for a well-posed Laplace–Beltrami basis, and that α mutually consistent semantic region pairs exist and are recovered by joint pose-correspondence optimization. Real experiments use SAM3D meshes aligned to depth (§V-G); residual failures are attributed to “mesh-to-workspace alignment errors rather than correspondences,” but no quantitative correspondence metrics on those meshes are given, nor is robustness to disconnected/noisy components characterized. A short failure analysis (when anchors are wrong or the mesh is fragmented) and GCS on the real pairs would make the load-bearing premise falsifiable outside clean PartNet-Mobility assets.
  3. [§V-E, Table III, Fig. 5] Cross-category evidence is thin relative to the claim that anchor selection “generalizes beyond category identity” (§V-E): only Scissors→Pliers and Kettle→Bottle appear in Table III, while Fig. 5’s confidence matrix is descriptive and not linked to transfer success. Either expand the cross-category correspondence table (and, if feasible, one cross-category robot task) or narrow the claim to within-category / closely related pairs so the abstract does not oversell generalization.
minor comments (6)
  1. [Title, §II–III] Title line and several headings lack spaces (“SemAnCorr:SemanticAnchoredCorrespondence”, “RELATEDWORK”, “PROBLEMSTATEMENT”). Clean typography throughout.
  2. [§IV-A] Eq. (3) text reads “which is as a compact descriptor”; missing word (“used as”).
  3. [§V-A] Hyperparameters K=6, α=3, k=30, λ_g/λ_s, B, σ, τ=2 are fixed with no sensitivity plot. A brief appendix sweep (especially α and k) would strengthen reproducibility claims in §V-A.
  4. [§V-C] DenseMatcher is a learned, category-exposed method; state clearly whether it was used zero-shot on PartNet-Mobility categories or adapted, so the comparison in Tables I–III is fair.
  5. [Fig. 3] Fig. 3b baseline color transfers are hard to read in grayscale print; consider edge overlays or a quantitative callout next to each panel.
  6. [Header] arXiv ID / date in the header (2607.28382, Jul 2026) looks like a placeholder; verify before camera-ready.

Circularity Check

0 steps flagged

No significant circularity: empirical systems method evaluated on external labels and held-out robot trials, not self-defined predictions.

full rationale

SemAnCorr is a training-free engineering pipeline (SigLIP2 lifting → clustering → joint pose-anchor selection → functional-map propagation). Its claims are comparative empirical results (sAcc, Cont/Cov/GCS on PartNet-Mobility part labels; real-world success rates vs D3Fields), not first-principles derivations or fitted quantities renamed as predictions. GCS is an explicitly defined harmonic mean of continuity and coverage used as an evaluation metric, not an objective the method optimizes into tautology. Anchor confidence (Fig. 5) is presented as a descriptive compatibility signal, not as proof that transfer must succeed. Classical functional maps and ZoomOut are cited from external literature; there is no load-bearing self-citation uniqueness chain, no parameter fit that forces the reported accuracy, and no renaming of a known law as a new derivation. Experimental confounds (GPG ranking, mesh alignment) affect causal attribution of robot gains, which is a correctness/evidence issue, not circularity of the derivation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

The central claim rests on standard spectral geometry plus the domain bet that pretrained 2D semantic features, once lifted and clustered, yield a few bijective part anchors good enough to pin a smooth functional map—and that mesh connectivity and reconstruction quality match that spectral model. Free parameters are few but hand-fixed; invented quantities are mostly method/metric names rather than physical entities.

free parameters (6)
  • K (number of semantic clusters) = 6
    Fixed to 6 for all experiments (§V-A) without reported sweep; controls part granularity and anchor pool.
  • α (number of semantic anchors) = 3
    Fixed to 3 (§V-A); directly limits how many part constraints pin the functional map.
  • k (Laplace-Beltrami basis size) = 30
    Fixed to 30 (§V-A); sets spectral smoothness/resolution of the functional map.
  • λ_g, λ_s (geo vs semantic weights in joint score)
    Appear in Score(c,c';R,t) (§IV-B) with cosine schedule between semantic and geometric emphasis; values not numerically specified in text.
  • τ (continuity edge tolerance in source edge lengths) = 2
    Set to 2 for Cont metric (§V-B); affects GCS and thus geometric claims.
  • B (pose optimization initializations) and σ (chamfer kernel width)
    Control joint pose-correspondence optimization (§IV-B); B mentioned, σ in geo(·) exponential; exact values underspecified.
axioms (6)
  • domain assumption Each object mesh is a single connected component so the Laplace–Beltrami eigenbasis for functional maps is well-posed.
    Stated explicitly in §III; required by §IV-C spectral construction.
  • domain assumption There exist α pairs of semantically corresponding surface regions between the two instances that can serve as anchors.
    §III equation (1) and surrounding text; without this, anchor selection cannot ground the map.
  • domain assumption Low-frequency functional maps plus local neighbor-restricted ZoomOut yield spatially smooth pointwise correspondences suitable for recovering local interaction frames.
    Classical FM/ZoomOut prior [2,26] adopted in §IV-C; smoothness is assumed to equal geometric coherence needed for skill transfer.
  • domain assumption Pretrained SigLip2 patch features, multi-view lifted and clustered with position, encode part-level semantics transferable across instances and related categories.
    §IV-A feature extraction and clustering; foundation of training-free claim.
  • standard math Functional maps act as linear operators in the first k eigenfunctions and recover pointwise maps by nearest neighbors in the spectral embedding.
    Standard FM machinery cited from Ovsjanikov et al.; used throughout §IV-C.
  • ad hoc to paper Real-world SAM3D meshes aligned to depth are accurate enough that correspondence error, not reconstruction error, is the main variable—except where authors attribute residual failures to alignment.
    §V-G experimental pipeline; partially contradicted by their own failure analysis.
invented entities (4)
  • SemAnCorr pipeline (semantic anchors + joint pose-correspondence + anchored FM) independent evidence
    purpose: Name the full training-free correspondence method claimed to jointly optimize semantic and geometric quality.
    Composite method entity; evaluated empirically against baselines rather than postulated as a physical object.
  • Bilateral margin confidence for anchor pairs no independent evidence
    purpose: Rank mutually exclusive cluster matches before/after joint optimization (§IV-B).
    Method-specific scoring rule; not independently measured outside this selection procedure.
  • Geometric Coherence Score (GCS = harmonic mean of Continuity and Coverage) no independent evidence
    purpose: Single scalar tying correspondence geometry to local-frame recoverability for manipulation (§V-B).
    Paper-defined metric; useful but not an external standard; drives “2× baseline” geometric claim.
  • Relative cosine similarity after subtracting per-object mean cluster embedding no independent evidence
    purpose: Isolate part-discriminative signal from category-level feature collapse (§IV-B).
    Ad hoc normalization choice central to anchor ranking.

pith-pipeline@v1.2.0-daily-grok45 · 16821 in / 4188 out tokens · 71255 ms · 2026-07-31T09:09:47.074357+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SemAnCorr: Semantic Anchored Correspondence for Zero-Shot Manipulation Skill Transfer." pith.science (2026). https://pith.science/paper/IJYKIKMT

@misc{pith2026260728382,
  author       = {Pith},
  title        = {Pith review of: SemAnCorr: Semantic Anchored Correspondence for Zero-Shot Manipulation Skill Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJYKIKMT}},
  note         = {Machine review of arXiv:2607.28382}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Transferring manipulation skills across object instances that share functionality but differ in geometry remains a fundamental challenge in robot learning. While recent correspondence methods leverage dense visual descriptors and 3D feature fields, nearest-neighbor feature matching often produces spatially incoherent correspondences that fail to recover the local geometric frames required for reliable skill transfer. We introduce SemAnCorr, a training-free framework that establishes dense correspondence by selecting semantically consistent anchor regions through joint pose-correspondence optimization and propagating these constraints over the object surface using functional maps. The resulting correspondences preserve both semantic consistency and geometric coherence, enabling object-centric manipulation skills to transfer across geometrically diverse instances. We evaluate SemAnCorr on a dense correspondence benchmark built on PartNet-Mobility, achieving 90.8% semantic accuracy in our benchmark evaluation while improving geometric coherence over recent state-of-the-art baselines. Finally, we show that these improvements translate directly into real-world manipulation performance: using a single demonstration, SemAnCorr enables substantially more reliable zero-shot manipulation skill transfer to previously unseen objects than existing correspondence methods. Videos and additional visualizations are available at [https://semancorr.github.io](https://semancorr.github.io) .

Figures

Figures reproduced from arXiv: 2607.28382 by Hongyi Chen, Jeffrey Ichnowski, Uksang Yoo, Weiming Zhi, William Baron, Xiaoxiang Dong.

Figure 1
Figure 1. Figure 1: We propose SemAnCorr, a dense vertex-level correspondence across geometrically diverse object instances that respects semantically similar parts between objects while smoothly spanning object surface (left). The resulting correspondence enables zero-shot transfer of manipulation skills across instances that share functional structure but differ in geometry (right). Abstract—Transferring manipulation skills… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed SemAnCorr framework. To capture rich semantic representations, SemAnCorr extracts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Inner-Category Correspondence Visualization. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-Category Correspondence Visualization. On the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Category-level anchor confidence scores. Within [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real-world manipulation visualization. From left to right: Task 1, 2, 3, 4, and 5. For each task, a manipulation [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Correspondence visualizations for real-world objects. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

28 extracted references · 4 linked inside Pith

  1. [1]

    SAPIEN: A simulated part-based interactive environment,

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

    2020

  2. [2]

    Functional maps: a flexible representation of maps between shapes,

    M. Ovsjanikov, M. Ben-Chen, J. Solomon, A. Butscher, and L. Guibas, “Functional maps: a flexible representation of maps between shapes,” vol. 31, no. 4, 2012

    2012

  3. [3]

    The wave kernel signature: A quantum mechanical approach to shape analysis,

    M. Aubry, U. Schlickewei, and D. Cremers, “The wave kernel signature: A quantum mechanical approach to shape analysis,”2011 IEEE Inter- national Conference on Computer Vision Workshops (ICCV Workshops), pp. 1626–1633, 2011

    2011

  4. [4]

    Deep geometric func- tional maps: Robust feature learning for shape correspondence,

    N. Donati, A. Sharma, and M. Ovsjanikov, “Deep geometric func- tional maps: Robust feature learning for shape correspondence,”2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8589–8598, 2020

    2020

  5. [5]

    Densematcher: Learning 3d semantic correspondence for category- level manipulation from a single demo,

    J. Zhu, Y . Ju, J. Zhang, M. Wang, Z. Yuan, K. Hu, and H. Xu, “Densematcher: Learning 3d semantic correspondence for category- level manipulation from a single demo,”International Conference on Learning Representations (ICLR) 2025, 2024

    2025

  6. [6]

    Dense object nets: Learn- ing dense visual object descriptors by and for robotic manipulation,

    P. R. Florence, L. Manuelli, and R. Tedrake, “Dense object nets: Learn- ing dense visual object descriptors by and for robotic manipulation,” in Conference on Robot Learning, 2018

    2018

  7. [7]

    Emergent correspondence from image diffusion,

    L. Tang, M. Jia, Q. Wang, C. P. Phoo, and B. Hariharan, “Emergent correspondence from image diffusion,” inThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  8. [8]

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence,

    J. Zhang, C. Herrmann, J. Hur, L. P. Cabrera, V . Jampani, D. Sun, and M. Yang, “A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence,”Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  9. [9]

    Robo-abc: Af- fordance generalization beyond categories via semantic correspondence for robot manipulation,

    Y . Ju, K. Hu, G. Zhang, G. Zhang, M. Jiang, and H. Xu, “Robo-abc: Af- fordance generalization beyond categories via semantic correspondence for robot manipulation,” inEuropean Conference on Computer Vision, pp. 222–239, Springer, 2025

    2025

  10. [10]

    Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation,

    Y . Kuang, J. Ye, H. Geng, J. Mao, C. Deng, L. J. Guibas, H. Wang, and Y . Wang, “Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation,” inConference on Robot Learning, 2024

    2024

  11. [11]

    Uad: Unsupervised affordance distillation for generalization in robotic manipulation,

    Y . Tang, W. Huang, Y . Wang, C. Li, R. Yuan, R. Zhang, J. Wu, and F.- F. Li, “Uad: Unsupervised affordance distillation for generalization in robotic manipulation,”2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 3822–3831, 2025

    2025

  12. [12]

    Adapting by analogy: Ood generalization of visuomotor policies via functional correspondence,

    P. Gupta, H. Admoni, and A. V . Bajcsy, “Adapting by analogy: Ood generalization of visuomotor policies via functional correspondence,” ArXiv, vol. abs/2506.12678, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Nerf-supervision: Learning dense object descriptors from neural radiance fields,

    L. Yen-Chen, P. R. Florence, J. T. Barron, T.-Y . Lin, A. Rodriguez, and P. Isola, “Nerf-supervision: Learning dense object descriptors from neural radiance fields,”2022 International Conference on Robotics and Automation (ICRA), pp. 6496–6503, 2022

    2022

  14. [14]

    D3fields: Dynamic 3d descriptor fields for zero- shot generalizable rearrangement,

    Y . Wang, M. Zhang, Z. Li, T. Kelestemur, K. Driggs-Campbell, J. Wu, L. Fei-Fei, and Y . Li, “D3fields: Dynamic 3d descriptor fields for zero- shot generalizable rearrangement,” in8th Annual Conference on Robot Learning, 2024

    2024

  15. [15]

    O3afford: One-shot 3d object-to- object affordance grounding for generalizable robotic manipulation,

    T. Tian, X. Kang, and Y .-L. Kuo, “O3afford: One-shot 3d object-to- object affordance grounding for generalizable robotic manipulation,” ArXiv, vol. abs/2509.06233, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Where2act: From pixels to actions for articulated 3d objects,

    K. Mo, L. J. Guibas, M. Mukadam, A. K. Gupta, and S. Tul- siani, “Where2act: From pixels to actions for articulated 3d objects,” 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6793–6803, 2021

    2021

  17. [17]

    3d affordancenet: A benchmark for visual object affordance understanding,

    S. Deng, X. Xu, C. Wu, K. Chen, and K. Jia, “3d affordancenet: A benchmark for visual object affordance understanding,”2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1778–1787, 2021

    2021

  18. [18]

    Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts,

    H. Geng, H. Xu, C. Zhao, C. Xu, L. Yi, S. Huang, and H. Wang, “Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts,”2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7081–7091, 2022

    2023

  19. [19]

    Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions,

    Y . Wang, R. Wu, K. Mo, J. Ke, Q. Fan, L. J. Guibas, and H. Dong, “Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions,” inEuropean Conference on Computer Vision, 2021

    2021

  20. [20]

    Keto: Learning keypoint representations for tool manipulation,

    Z. Qin, K. Fang, Y . Zhu, L. Fei-Fei, and S. Savarese, “Keto: Learning keypoint representations for tool manipulation,”2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 7278–7285, 2019

    2020

  21. [21]

    Gift: Generalizable interaction-aware functional tool affordances without la- bels,

    D. Turpin, L. Wang, S. Tsogkas, S. J. Dickinson, and A. Garg, “Gift: Generalizable interaction-aware functional tool affordances without la- bels,”ArXiv, vol. abs/2106.14973, 2021

    Pith/arXiv arXiv 2021

  22. [22]

    Tool-as- interface: Learning robot policies from observing human tool use,

    H. Chen, C. Zhu, S. Liu, Y . Li, and K. R. Driggs-Campbell, “Tool-as- interface: Learning robot policies from observing human tool use,” in Conference on Robot Learning (CoRL), 2025

    2025

  23. [23]

    Robust task-based grasping as a service,

    J. Song, A. K. Tanwani, J. Ichnowski, M. Danielczuk, K. Sanders, J. Chui, J. A. Ojea, and K. Goldberg, “Robust task-based grasping as a service,”2020 IEEE 16th International Conference on Automation Science and Engineering (CASE), pp. 22–28, 2020

    2020

  24. [24]

    Grim: Task-oriented grasping with conditioning on generative examples,

    Shailesh, A. Raj, N. Kumar, P. Shukla, A. Melnik, M. Beetz, and G. C. Nandi, “Grim: Task-oriented grasping with conditioning on generative examples,”ArXiv, vol. abs/2506.15607, 2025

    arXiv 2025

  25. [25]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,”2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 11941–11952, 2023

    2023

  26. [26]

    Zoomout: spectral upsampling for efficient shape correspondence,

    S. Melzi, J. Ren, E. Rodolà, A. Sharma, P. Wonka, and M. Ovsjanikov, “Zoomout: spectral upsampling for efficient shape correspondence,” vol. 38, no. 6, 2019

    2019

  27. [27]

    Sam 3d: 3dfy anything in images,

    S. D. Team, X. Chen, F.-J. Chu, P. Gleize, K. J. Liang, A. Sax, H. Tang, W. Wang, M. Guo, T. Hardin, X. Li, A. Lin, J. Liu, Z. Ma, A. Sagar, B. Song, X. Wang, J. Yang, B. Zhang, P. Dollár, G. Gkioxari, M. Feiszli, and J. Malik, “Sam 3d: 3dfy anything in images,” 2025

    2025

  28. [28]

    Grasp pose detection in point clouds,

    A. ten Pas, M. Gualtieri, K. Saenko, and R. Platt, “Grasp pose detection in point clouds,”arXiv preprint arXiv:1706.09911, 2017

    Pith/arXiv arXiv 2017