Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Reactive In-Air Clothing Manipulation with Confidence-Aware Dense Correspondence and Visuotactile Affordance

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A dual-arm robot folds and hangs crumpled shirts directly in midair by deferring actions when visual confidence is low.

desk verdict Solid systems paper with a real in-air capability; thin stats and a hand-tuned confidence threshold make the headline claim provisional. read the letter →

arxiv 2509.03889 v1 pith:DG4CNK7O submitted 2025-09-04 cs.RO cs.AIcs.LG

classification cs.ROcs.AIcs.LG
keywords deformableobjectmanipulationclothingdensevisualcorrespondencedistributionallossconfidence-awareplanningvisuotactileaffordanceroboticfoldinghanging
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

Most robotic clothing systems flatten the garment first or assume key features are visible. This paper argues that a robot can fold and hang crumpled shirts directly—on a table and suspended in air—if it knows when its visual correspondences are trustworthy and can touch-check its grasps. The system pairs a dense correspondence network trained with a distributional loss, which outputs a probability heatmap rather than a single forced match, with a visuotactile grasp affordance network and a tactile classifier that validates contact. When correspondence confidence falls below a threshold, the robot rotates the shirt and re-evaluates instead of committing to a guess. The reported result is folding success in 6 of 10 trials and hanging success in 7 of 10, from highly occluded configurations without flattening or full-state estimation.

What carries the argument

The distributional dense correspondence network and the confidence-based state machine it feeds. The network maps pixels to descriptors and, via a softmax over descriptor distances, outputs a probability heatmap over the canonical shirt; a KL loss against a multimodal Gaussian target lets symmetric regions count as multiple valid matches, producing confidence estimates. Those estimates gate every grasp: only above-threshold correspondence and affordance lead to an action, otherwise the shirt rotates and is re-evaluated. The tactile classifier plays a complementary role, supervising the affordance network, validating grasp success in real time, and signaling when the shirt is under tension.

What would settle it

Collect many real suspended-shirt queries, record the predicted peak confidence for each match, and compare whether accepted-by-threshold matches land in the correct region much more often than rejected ones. If the high-confidence set is not clearly more accurate than the low-confidence set, or no single threshold separates them reliably on real data, confidence-aware deferral is not doing the causal work. Alternatively, run the same folding task with deferral disabled by forcing a grasp at the first candidate regardless of confidence: if success does not drop, deferral is not load-bearing.

Watch

Extended reading notes

Core claim

The claim is that confidence-aware dense correspondence plus visuotactile affordance lets a dual-arm robot manipulate garments in air directly from crumpled states. A ResNet-based network maps each pixel of a deformed shirt to a descriptor and predicts a probability distribution over matches on a canonical shirt, trained with a KL divergence to a multimodal Gaussian target that allows symmetric valid matches; the result is per-pixel confidence heatmaps. A state machine gates every grasp: both correspondence confidence and grasp affordance must clear thresholds, otherwise the shirt is rotated 30 degrees and re-evaluated, with a fallback grab to unfold it. High-resolution tactile sensing super

Load-bearing premise

The load-bearing premise is that the confidence numbers produced by the correspondence network are trustworthy on real, unseen garments, so that a single fixed threshold separates dependable matches from unreliable ones; if the network systematically overestimates its certainty, the deferral strategy loses its reason to wait and the claimed in-air robustness collapses.

Editorial extensions

If this is right

  • Folding and hanging become feasible directly from crumpled and suspended states, removing flattening as a required preprocessing step.
  • Confidence thresholds make the system robust to self-occlusion: when features are hidden, the robot rotates instead of guessing, converting some hard perceptual errors into recoverable events.
  • The same task-agnostic grasp affordance module serves both folding and hanging, so a new task does not require retraining the grasp selector from scratch.
  • Tactile validation turns empty grasps, excess fabric, and slipped cloth into recoverable failures rather than terminal errors.
  • The dense descriptor map can act as a shared interface for other planners, such as translating human video demonstrations into robot grasp targets.

Reading between the lines

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

  • An implicit design choice is a single fixed confidence threshold; a runtime-adaptive threshold that reads the shape of the whole probability heatmap would likely improve the low-confidence cases the paper reports most, such as bottom regions and heavily folded sleeves.
  • The deferral principle is not garment-specific: any deformable object with symmetric or ambiguous correspondences, such as ropes, bags, or upholstery, could use the same confidence-gated state machine provided a canonical template exists.
  • Because the tactile classifier supervises affordance and validates grasps online, the system could collect large amounts of self-supervised grasp data during normal operation, plausibly improving both affordance and confidence calibration over time.
  • The human-video pipeline is limited by hand occlusion and skeleton errors, but it points to a scalable data loop: correspondence plus confidence could automatically filter unreliable human demonstrations.
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 presents a dual-arm robotic system for garment folding and hanging that operates directly on crumpled table-top and in-air garments. It couples a ResNet-based dense visual correspondence network trained with a distributional (KL) loss to output pixel-wise correspondence probabilities and confidence, a visuotactile grasp-affordance network fine-tuned via tactile labels, and a reactive state machine that defers grasping when confidence or affordance are below thresholds. Experiments report simulation correspondence accuracy, real-world region-level classification (e.g., 73.3% for suspended net, 68.9% safe decisions for combined), tactile classifier accuracy (99.7%), affordance precision@80 (76.3%), and end-to-end folding (6/10) and hanging (7/10) success. The paper also demonstrates extracting canonical grasp points from human videos.

Significance. If the reported results hold, the work would demonstrate a meaningful new capability: reactive in-air garment manipulation from crumpled states without flattening or full-state estimation, enabled by confidence-gated deferral and tactile feedback. The distributional loss for garment correspondences and the simulated dataset with realistic seams/hem details are useful contributions, and the tactile-grounded affordance fine-tuning is a sensible approach to mitigate sim2real gaps. However, the paper's central claim hinges on the calibration of the learned confidence estimates, which the authors themselves note is imperfect, and the empirical evaluation is thin (10 trials per task, no baselines). The contributions are potentially significant but require additional validation to be convincing.

major comments (4)
  1. [§6, App. 7.4, Eq. (1)] The central mechanism is confidence-based deferral, but the paper does not establish that the confidence values from Eq. (1) are calibrated. Eq. (1) is a softmax over descriptor distances; training by KL divergence to a Gaussian target does not guarantee the resulting probabilities track empirical correctness. Section 6 concedes the network 'occasionally overestimates its certainty in challenging configurations,' and App. 7.4 reports that real-world transfer of thresholds 'introduced high variability,' with the deployed 6e-6 threshold chosen post hoc as a 'clear inflection point.' As a result, the reported safe-decision rates (68.9%, 70.8%, 80%) may reflect threshold tuning rather than calibrated confidence. Please provide calibration curves, a threshold sweep, and an ablation without deferral to substantiate the load-bearing claim.
  2. [§4, Table 1, App. 7.2] The main task claims rest on 10 trials per task, with no confidence intervals or significance tests. Folding success 6/10 and hanging 7/10 change by 10% with one trial. The comparison 'without affordance fine-tuning' is also a single 10-trial run. There is no real-world comparison to prior correspondence-based folding ([1], [28]) or to a non-reactive baseline using the same network. For a journal-level claim of robust in-air manipulation, please report more trials, statistical treatment, and at least a no-deferral ablation.
  3. [§4, App. 7.4] The real-world correspondence evaluation conflates correct rejection with calibration. A 'safe decision' is defined as either correct or below the 6e-6 threshold, but that threshold's calibration is exactly what is at issue. Region-level labels (Fig. 13) are coarser than pixel correspondences, and the inverse-direction evaluation only includes human-verifiable points, potentially biasing the sample. Please report precision-recall as a function of threshold, include all query points, and separate correct classifications from low-confidence rejections in each region.
  4. [§3.3, §4, App. 7.5] The tactile classifier is both the supervision source for fine-tuning the affordance network and the in-loop validator for grasp success during execution. App. 7.5 acknowledges that this classifier cannot determine whether the grasped region corresponds to the intended visual target, so some positive labels are noisy (grasping cloth in front of the target). While the end-to-end folding/hanging metrics are independent, the affordance quality metric and the in-loop recovery both rely on the same classifier, creating potential bias. Please provide a human-verified evaluation of grasp success on a subset of trials, and report the tactile classifier's confusion matrix with respect to the intended-target criterion.
minor comments (6)
  1. [Sec. 3.2] The notation for the predicted distribution switches between \hat{p}_b and \hat{p}_{b_i}; clarify the subscript convention.
  2. [Sec. 3.3 vs App. 7.5] The grasp count is given as 8,500 in Sec. 3.3 but 8,000 in App. 7.5; reconcile the numbers.
  3. [Table 1 caption] The definition of 'correspondence success' for low-confidence outcomes is unclear; a trial with no grasp attempt is counted as correspondence success in Table 1. Explain how a non-grasp is scored as a correct correspondence.
  4. [Fig. 4(a) caption] The axis label says 'cumulative fraction of images' but the text says 'cumulative fraction of pixels'; make consistent.
  5. [Sec. 4, folding results] The relationship between '6 out of 10 trials' and '30 total grasps attempted' is unclear; clarify whether the 6 trials are those in which at least one graspable keypoint was found.
  6. [App. 7.6] The human-video demonstration is described as 'outside the primary focus' yet is listed as a contribution in the abstract; consider moving it to future work or clearly labeling it as a preliminary illustration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are grounded in external simulation ground truth, independent human-labeled evaluations, and physical task success.

full rationale

The paper's central derivation chain is self-contained against external evidence. Dense correspondence is trained with a distributional loss against simulated ground-truth pixel correspondences (Blender vertex indexing and canonical template alignment), not against the robot's own success; the confidence estimates are softmax probabilities defined by Equation 1 and supervised by an explicit multimodal Gaussian target, so they are not defined in terms of the downstream manipulation outcome. The grasp affordance network is fine-tuned using the tactile classifier, but its offline evaluation uses 125 human-labeled grasp points with precision@80, and final folding/hanging success is measured by physical task completion (6/10 and 7/10 trials), not by the classifier's own outputs. The confidence threshold of 6e-6 is an empirically selected operating point (Appendix 7.4), and while this raises a genuine calibration/robustness concern—especially given the paper's own admission that the network 'occasionally overestimates its certainty' (Section 6) and that threshold transfer 'introduced high variability' (Appendix 7.4)—choosing a threshold on real data is not equivalent to fitting the reported task success; it is a hyperparameter, not a circular prediction. Self-citations to [19] provide the U-Net architecture and data-generation pipeline, but those are externally published prior results with their own experiments, and the present contribution extends them rather than reducing to them. No equation is defined in terms of the result it is supposed to establish, and no fitted parameter is renamed as a prediction. The limitations flagged in the manuscript relate to calibration and generalization, not to circularity.

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

The system relies on several domain assumptions about sim-to-real transfer, target distribution modeling, and tactile classifier validity. The free parameters are mostly hyperparameters and thresholds, with the correspondence confidence threshold being the most consequential for the central claim. No new physical entities are introduced.

free parameters (5)
  • Correspondence confidence threshold = 6e-6 (chosen from real-world inflection point, Appendix 7.4)
    Used in state machine to defer low-confidence grasps; threshold tuned on real data after ROC in simulation showed high variability.
  • Affordance confidence threshold = not reported (described as 'predefined')
    Used to gate grasp execution; value not disclosed, making replication harder.
  • Gaussian sigma for distributional loss target = not explicitly stated; selected from {1,2,10,20} (larger values favored for generalization)
    Controls target distribution sharpness; affects correspondence confidence calibration.
  • Descriptor dimension d = 16
    Selected via hyperparameter search over {3,9,16,25}; affects representation capacity.
  • Maximum fabric layers in affordance labeling = 2
    Ad hoc constraint defining 'successful' grasp in simulation; affects affordance training labels.
assumptions (4)
  • domain assumption Blender 4.2 cloth simulation with realistic seams and hems is sufficiently faithful to real cloth dynamics and appearance for sim-to-real transfer of visual correspondence.
    Central to training the correspondence network; the paper notes that without seams, real-world performance drops (Appendix 7.3).
  • domain assumption The multimodal Gaussian target distribution with manually chosen sigma correctly models correspondence uncertainty arising from garment symmetries.
    Used in the distributional loss; sigma is a hyperparameter, and the paper reports that different sigma values trade off simulation accuracy vs. real-world generalization.
  • domain assumption The tactile classifier's three-way categorization (success, too little fabric, excess layers) is a valid proxy for grasp quality that correlates with downstream task success.
    This classifier supervises affordance fine-tuning and validates grasps during execution; the paper acknowledges it cannot verify whether the intended target region was grasped (Section 3.3).
  • domain assumption Confidence estimates from the correspondence network are calibrated enough that a fixed threshold (6e-6) reliably separates trustworthy matches on real, unseen garments.
    The state machine's deferral behavior depends on this; the paper admits the network 'occasionally overestimates its certainty' (Section 6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reactive In-Air Clothing Manipulation with Confidence-Aware Dense Correspondence and Visuotactile Affordance." pith.science (2026). https://pith.science/paper/DG4CNK7O

@misc{pith2026250903889,
  author       = {Pith},
  title        = {Pith review of: Reactive In-Air Clothing Manipulation with Confidence-Aware Dense Correspondence and Visuotactile Affordance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DG4CNK7O}},
  note         = {Machine review of arXiv:2509.03889}
}
read the original abstract

Manipulating clothing is challenging due to complex configurations, variable material dynamics, and frequent self-occlusion. Prior systems often flatten garments or assume visibility of key features. We present a dual-arm visuotactile framework that combines confidence-aware dense visual correspondence and tactile-supervised grasp affordance to operate directly on crumpled and suspended garments. The correspondence model is trained on a custom, high-fidelity simulated dataset using a distributional loss that captures cloth symmetries and generates correspondence confidence estimates. These estimates guide a reactive state machine that adapts folding strategies based on perceptual uncertainty. In parallel, a visuotactile grasp affordance network, self-supervised using high-resolution tactile feedback, determines which regions are physically graspable. The same tactile classifier is used during execution for real-time grasp validation. By deferring action in low-confidence states, the system handles highly occluded table-top and in-air configurations. We demonstrate our task-agnostic grasp selection module in folding and hanging tasks. Moreover, our dense descriptors provide a reusable intermediate representation for other planning modalities, such as extracting grasp targets from human video demonstrations, paving the way for more generalizable and scalable garment manipulation.

Figures

Figures reproduced from arXiv: 2509.03889 by the authors.

Figure 1
Figure 1. Overview of visuotactile garment manipulation system. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Generating a simulated shirt dataset. Blender 4.2 is used to simulate deformed shirts. Our animation pipeline allows flexibility in shirt geometries with the addition of realistic, key features like seams and hems often found on real shirts. A consistent vertex indexing across the shirt dataset is used, allowing alignment with a canonical template. We use Blender 4.2 [33] to simulate a wide variety of shirt geometri… view at source ↗
Figure 3
Figure 3. Training dense correspondence in simulation. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Comparison of contrastive vs. distributional training, with and without symmetric supervision [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Correspondence and affordance heatmaps for real images. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Extracting grasp points from human video demonstrations. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Confidence-based state machine for folding strategy. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Irrecoverable Failure Modes of Folding. Though the confidence-based state machine is able to recover from mistakes in folding, some cases are unaccounted for and irrecoverable in the system. Incorrect correspondence grasps, picking the correct feature but on the wrong …
Figure 9
Figure 9. Figure 9: Cumulative pixel match error across different descriptor dimensions ( [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Cumulative pixel match error for different Gaussian [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Cumulative pixel match error on hanging shirts for networks trained on hanging and com [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Cumulative pixel match error on shirts on a table for networks trained on table, hanging, [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Shirt region classification used for real-world evaluation. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Examples of out-of-distribution shirts tested. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Visuotactile grasp affordance training in simulation. [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Fine-tuned visuotactile grasp affordance compared to baselines. [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 31 canonical work pages

  1. [1]

    Ganapathi, P

    A. Ganapathi, P. Sundaresan, B. Thananjeyan, A. Balakrishna, D. Seita, J. Grannen, M. Hwang, R. Hoque, J. E. Gonzalez, N. Jamali, K. Yamane, S. Iba, and K. Goldberg. Learning to smooth and fold real fabric using dense object descriptors trained on synthetic color images. arXiv, 2020

  2. [28]

    R. Wu, H. Lu, Y . Wang, Y . Wang, and H. Dong. Unigarmentmanip: A unified framework for category-level garment manipulation via dense visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2024

  3. [2]

    Learning to Fold Real Garments with One Arm: A Case Study in Cloud-Based Robotics Research

    R. Hoque, K. Shivakumar, S. Aeron, G. Deza, A. Ganapathi, A. Wong, J. Lee, A. Zeng, V . Van- houcke, and K. Goldberg. Learning to fold real garments with one arm: A case study in cloud-based robotics research. arXiv preprint arXiv:2204.10297, 2022

  4. [3]

    Ha and S

    H. Ha and S. Song. Flingbot: The unreasonable effectiveness of dynamic manipulation for cloth unfolding. In Conference on Robot Learning, pages 24–33. PMLR, 2022

  5. [4]

    Cloth Funnels: Canonicalized-Alignment for Multi-Purpose Garment Manipulation

    A. Canberk, C. Chi, H. Ha, B. Burchfiel, E. Cousineau, S. Feng, and S. Song. Cloth fun- nels: Canonicalized-alignment for multi-purpose garment manipulation, 2022. URL https: //arxiv.org/abs/2210.09347

  6. [5]

    GarmentNets: Category-Level Pose Estimation for Garments via Canonical Space Shape Completion

    C. Chi and S. Song. Garmentnets: Category-level pose estimation for garments via canonical space shape completion. CoRR, abs/2104.05177, 2021. URL https://arxiv.org/abs/ 2104.05177

  7. [6]

    Chi and D

    C. Chi and D. Berenson. Occlusion-robust deformable object tracking without physics simula- tion. In Intelligent Robots and Systems (IROS), 2019 IEEE International Conference on. IEEE, 2019

  8. [7]

    Doumanoglou, A

    A. Doumanoglou, A. Kargakos, T.-K. Kim, and S. Malassiotis. Autonomous active recognition and unfolding of clothes using random decision forests and probabilistic planning. In 2014 IEEE international conference on robotics and automation (ICRA) , pages 987–993. IEEE, 2014

Show all 38 references
  1. [8]

    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

  2. [9]

    Maitin-Shepard, M

    J. Maitin-Shepard, M. Cusumano-Towner, J. Lei, and P. Abbeel. Cloth grasp point detection based on multiple-view geometric cues with application to robotic towel folding. In2010 IEEE International Conference on Robotics and Automation , pages 2308–2315. IEEE, 2010

  3. [10]

    Zhang and Y

    F. Zhang and Y . Demiris. Learning garment manipulation policies toward robot-assisted dress- ing. Science robotics, 7(65):eabm6010, 2022

  4. [11]

    Z. Sun, Y . Wang, D. Held, and Z. Erickson. Force-constrained visual policy: Safe robot- assisted dressing via multi-modal sensing. IEEE Robotics and Automation Letters , 2024

  5. [12]

    W. Chen, D. Lee, D. Chappell, and N. Rojas. Learning to grasp clothing structural regions for garment manipulation tasks. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4889–4895. IEEE, 2023

  6. [13]

    Y . Chen, S. Wei, B. Xiao, J. Lyu, J. Chen, F. Zhu, and H. Wang. Robohanger: Learning generalizable robotic hanger insertion for diverse garments. arXiv preprint arXiv:2412.01083, 2024. 10

  7. [14]

    W. Chen, K. Li, D. Lee, X. Chen, R. Zong, and P. Kormushev. Graphgarment: Learning garment dynamics for bimanual cloth manipulation tasks. arXiv preprint arXiv:2503.05817 , 2025

  8. [15]

    Y . Wu, W. Yan, T. Kurutach, L. Pinto, and P. Abbeel. Learning to manipulate deformable objects without demonstrations, 2019

  9. [16]

    Hoque, D

    R. Hoque, D. Seita, A. Balakrishna, A. Ganapathi, A. K. Tanwani, N. Jamali, K. Yamane, S. Iba, and K. Goldberg. VisuoSpatial foresight for multi-step, multi-task fabric manipulation, 2020

  10. [17]

    X. Lin, Y . Wang, Z. Huang, and D. Held. Learning visible connectivity dynamics for cloth smoothing. In Conference on Robot Learning, pages 256–266. PMLR, 2022

  11. [18]

    J. Qian, T. Weng, L. Zhang, B. Okorn, and D. Held. Cloth region segmentation for robust grasp selection. In IEEE International Conference on Intelligent Robots and Systems , 2020. ISBN 9781728162126. doi:10.1109/IROS45743.2020.9341121

  12. [19]

    Sunil, S

    N. Sunil, S. Wang, Y . She, E. Adelson, and A. R. Garcia. Visuotactile affordances for cloth manipulation with local control. In 6th Annual Conference on Robot Learning , 2022. URL https://openreview.net/forum?id=s6NEzqZKaP-

  13. [20]

    Yamazaki, K

    K. Yamazaki, K. Nagahama, and M. Inaba. Daily clothes observation from visible surfaces based on wrinkle and cloth-overlap detection. In MVA, pages 275–278, 2011

  14. [21]

    Willimon, S

    B. Willimon, S. Birchfield, and I. Walker. Model for unfolding laundry using interactive per- ception. In IEEE International Conference on Intelligent Robots and Systems , 2011. ISBN 9781612844541. doi:10.1109/IROS.2011.6048796

  15. [22]

    Cusumano-Towner, A

    M. Cusumano-Towner, A. Singh, S. Miller, J. F. O’Brien, and P. Abbeel. Bringing clothing into desired configurations with limited perception. In 2011 IEEE international conference on robotics and automation, pages 3893–3900. IEEE, 2011

  16. [23]

    T. Tang, Y . Fan, H.-C. Lin, and M. Tomizuka. State estimation for deformable objects by point registration and dynamic simulation. In Intelligent Robots and Systems (IROS), 2017 IEEE International Conference on. IEEE, 2017

  17. [24]

    C. B. Choy, J. Gwak, S. Savarese, and M. Chandraker. Universal correspondence network. CoRR, abs/1606.03558, 2016. URL http://arxiv.org/abs/1606.03558

  18. [25]

    Schmidt, R

    T. Schmidt, R. A. Newcombe, and D. Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters , 2017

  19. [26]

    P. R. Florence, L. Manuelli, and R. Tedrake. Dense object nets: Learning dense visual object descriptors by and for robotic manipulation. CoRR, abs/1806.08756, 2018. URL http:// arxiv.org/abs/1806.08756

  20. [27]

    Sundaresan, J

    P. Sundaresan, J. Grannen, B. Thananjeyan, A. Balakrishna, M. Laskey, K. Stone, J. E. Gonzalez, and K. Goldberg. Learning rope manipulation policies using dense object de- scriptors trained on synthetic depth data. CoRR, abs/2003.01835, 2020. URL https: //arxiv.org/abs/2003.01835

  21. [29]

    Ganapathi, P

    A. Ganapathi, P. Sundaresan, B. Thananjeyan, A. Balakrishna, D. Seita, R. Hoque, J. E. Gonza- lez, and K. Goldberg. MMGSD: Multi-Modal Gaussian Shape Descriptors for Correspondence Matching in 1D and 2D Deformable Objects. In Intelligent Robots and Systems (IROS), 2020 IEEE In...

  22. [30]

    Florence

    P. Florence. Dense visual learning for robot manipulation. PhD thesis, Massachusetts Institute of Technology, 01 2020

  23. [31]

    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

  24. [32]

    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. arXiv preprint arXiv:2304.07193, 2023

  25. [33]

    Blender Online Community. Blender. https://www.blender.org/, 2025. Version 4.2

  26. [34]

    Albisser

    A. Albisser. Procedural cloth sewing toolbox for blender 4.2+. https:// alexandrealbisser.gumroad.com/l/ProceduralSewingToolbox, 2024. Software tool

  27. [35]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical im- age segmentation. In International Conference on Medical image computing and computer- assisted intervention, pages 234–241. Springer, 2015

  28. [36]

    S. Wang, Y . She, B. Romero, and E. Adelson. Gelsight wedge: Measuring high-resolution 3d contact geometry with a compact robot finger. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages 6468–6475. IEEE, 2021

  29. [37]

    Sanderson

    M. Sanderson. Test collection based evaluation of information retrieval systems . Now Pub- lishers Inc, 2010

  30. [38]

    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. 12 7 Appendix 7.1 Blender Simulation Para...

Pith tools

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