Pith. sign in

REVIEW 4 major objections 5 minor 59 references

High Fidelity Capture, Reconstruction, and Transfer of Human Demonstrations for Robot-Assisted Bathing

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that contact regions—the surface patches where a caregiver's hand presses on a patient's body—can serve as the organizing primitive for capturing, reconstructing, and transferring human bathing demonstrations to a robot.

desk verdict A valuable dataset and a coherent contact-centric pipeline, undercut by a circular reconstruction metric and a thin transfer evaluation; deserves review, but not at face value. read the letter →

arxiv 2608.09127 v1 pith:6E6YU5QW submitted 2026-08-10 cs.RO cs.GR

Arjun S. Lakshmipathy , Jonathan P. King , Ethan Zuo , Rohit Satishkumar , Hongyi Chen , Jeffrey Ichnowski , Dan Ding , Zackory Erickson
show 1 more author
Nancy S. Pollard
This is my paper · ORCID
classification cs.ROcs.GR
keywords robot-assistedbathingphysicalhuman-robotinteractioncontact-richmanipulationtactilesensinghumanmotioncapturedemonstrationtransfersoftrobotichandretargeting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that the skilled, sustained physical contact in a human bathing task can be recorded and transferred to a robot if contact regions—not just joint angles or center-of-mass targets—are used as the organizing primitive. It builds a dataset of 128 clinician demonstrations, about 257,000 frames, with synchronized motion, body shape, tactile contact, and force, from three clinicians and six subjects. On top of that data, it reconstructs the demonstrations, uses them to choose tendon routings and motor commands for a soft anthropomorphic hand, and retargets the wrist motion to a seven-degree-of-freedom arm, then replays the task on a mannequin. A closed-loop version uses live tactile pressure to correct the arm trajectory, bringing applied pressure closer to the human original than an open-loop replay. If correct, this supplies a public dataset and a working pipeline for learning physical human-robot interaction from real caregiver technique.

What carries the argument

The contact match pass is the mechanism. Starting from calibrated taxel positions on the tactile glove, the pipeline queries the closest point on the SMPL-X body mesh for each active taxel and stores the trajectory in barycentric coordinates. It then adds $\lambda_c \sum_{i=0}^{C} \|\mathbf{c}_{h_i} - \mathbf{c}_{b_i}\|_2^2$ to the constrained hand pose optimization, pulling the reconstructed hand's corresponding surface points onto the reported body contact points. Contact trajectories expressed in barycentric coordinates are the same representation later used for retargeting to the mannequin, since they are agnostic to target body shape and pose.

What would settle it

Attach a dense set of small visual markers to the skin of the body region being bathed, record a capture, and compare the reconstructed contact points from the contact match pass against the marker positions touched at the same time. If the median distance from active taxel to true touched surface does not improve when the contact match term is active, even though the L2 distance to the fitted mesh improves, then the closest-point correspondence is measuring mesh alignment rather than contact, and the central primitive fails.

Watch

Extended reading notes

Core claim

The central claim is that contact regions are a sufficient primitive for high-fidelity capture and transfer of bathing demonstrations. By replacing bare-skin hand fitting with a constrained skeleton hand and adding a contact match objective, the authors report contact distance to the body mesh drops from a median of 1.714 cm to 0.536 cm compared with the baseline reconstruction. The same primitive—contact trajectories stored in barycentric coordinates—then carries through hand pose selection, tendon routing, arm retargeting, and closed-loop pressure control. The paper's strongest assertion is that its dataset is the first to provide synchronized motion, shape, contact, and force during sustained contact-rich human-human interaction, and that its transfer strategies use that data at capture, reconstruction, hand design, and arm control layers.

Load-bearing premise

The reconstruction assumes that the point where an active glove taxel presses on the body is the point on the fitted body mesh closest to that taxel's calibrated location; if the mesh is shifted even slightly—by clothing, soft-tissue deformation, or mannequin proxy error—the optimizer and the reported L2 metric both measure the wrong contact.

Editorial extensions

If this is right

  • Bathing demonstrations can be digitized with synchronized motion, shape, contact, and force, making real caregiver technique available as data rather than idealized protocol.
  • Using contact regions as the processing primitive reduces hand-body gap artifacts and contorted finger poses enough that the reconstruction is usable for downstream design and control.
  • Contact trajectories encoded in barycentric coordinates transfer to a different body (mannequin proxy) regardless of shape or pose differences, because the retargeting step does not require matching source and target geometry.
  • A closed-loop controller that adjusts end-effector position until tactile pressure matches the human reference keeps applied pressure closer to human levels than open-loop replay, at the cost of spatial distribution and dynamic range.

Reading between the lines

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

  • The same contact-region pipeline could extend to other sustained skin-contact care tasks, such as dressing, repositioning, and therapeutic wiping, where sliding contact and shape uncertainty dominate.
  • The paper's observation that the stabilizing hand carries most of the pressure suggests a hardware corollary: asymmetric bimanual robots with a high-torque, low-DOF support arm and a more agile wiping hand may be a better platform than a symmetric dual-arm design.
  • A likely next step is taxel-wise distribution control: the closed-loop controller here matches aggregate pressure, so replacing the pressure-sum target with a per-taxel or learned distribution target would address the distributional gap the paper reports.
  • The closest-point assumption implies that deployment on human subjects, which the paper says is not yet ready, will require a runtime body-shape estimator with sub-frame latency; without such perception, the transfer primitive may not survive patient motion.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper presents a capture, reconstruction, and transfer pipeline for robot-assisted bathing. The authors record clinicians performing bathing motions on human subjects using optical markers and a 65-taxel tactile glove, reconstruct body and hand shape/motion with an SMPL-X/MANO pipeline, add a contact-match stage that pulls active taxels toward closest-point body contacts, and use the resulting contact regions to design poses for a tendon-driven DexKit soft hand and to retarget wrist trajectories to an xArm. They contribute a dataset of 128 captures (about 257,000 frames) and demonstrate open- and closed-loop mannequin bathing rollouts. The central claims are that contact regions serve as an effective processing primitive and that the resulting reconstructions are high fidelity.

Significance. If the reconstruction and transfer claims held, this would be a valuable resource: it is the first dataset with synchronized motion, shape, contact, and force during sustained human-human bathing interactions, and the hardware demonstrations show a plausible route from human demonstration to robot deployment on a soft hand/arm. The paper is also unusually honest about limitations, including the sim-to-real gap, the inability to achieve distributed force control, the lack of shear sensing, and the fact that the system is not ready for deployment on human subjects. However, the quantitative support for the 'high fidelity' claim is currently weak because the headline metric is the same objective term minimized by the optimizer, and the transfer evaluation lacks statistical grounding. With independent validation and additional trials, the dataset and pipeline could be a meaningful contribution to pHRI research.

major comments (4)
  1. [III-B (Eq. 2), IV-C (Table II)] The primary quantitative evidence for reconstruction accuracy is circular. Equation (2) adds λ_c Σ_{i=0}^{C} ||c_hi − c_bi||² to the pose objective, where c_bi are the closest-point body contacts for active taxels. Table II then reports the same L2 distance between estimated body contacts and glove taxels as the accuracy metric. Reducing this term relative to a MoSh++ baseline is a necessary consequence of optimizing it, so the comparison primarily demonstrates that the contact-match optimizer ran. To support the 'high fidelity' claim, please provide an independent evaluation, for example by holding out a subset of taxels during fitting, using external contact ground truth (e.g., marked or inked contact regions, or a mannequin with known geometry), or comparing against manually labeled contact locations. As written, Table II is not evidence of reconstruction correctness.
  2. [III-B (closest-point association), III-D (barycentric retargeting), Appendix C] The reconstruction pipeline assumes that, for each active taxel, the true body contact point is the closest point on the SMPL-X body mesh. This correspondence is never independently validated, and it can be systematically wrong: the grip-material layer, glove deformation, soft-tissue compression, clothing, and the explicitly stated proxy-mesh offset in Appendix C all shift the closest-point target. Because the optimizer is rewarded for reducing exactly this closest-point distance, it will pull the hand toward the biased surface while the evaluation metric improves. These barycentric contact trajectories are later rolled out on the robot arm (Section III-D), so correspondence errors propagate into the transfer results. Please validate the contact association or quantify its sensitivity to the proxy offset and to the layered hand/body geometry.
  3. [IV-E, Fig. 14] The transfer evaluation in Fig. 14 does not support the strength of the claims made about it. The figure appears to show a single rollout per condition with no error bars, no number of trials, and no statistical tests, yet the text states that open-loop pressures are 'significantly higher' and that closed-loop pressures 'reasonably track' the human demonstration. Moreover, the closed-loop controller is designed to drive the tactile sum toward the recorded human value, so agreement in the normalized sum is partly by construction; the more informative deviations are in absolute pressure and taxel-wise distribution, where the paper concedes substantial mismatch. Please report repeated trials with variance and include a per-taxel or distributional error metric in addition to the controlled sum.
  4. [V-A, V-C, Abstract] The abstract's claim of 'high quality synchronized motion, shape, contact, and force' should be tempered by the paper's own statements in Section V-A that mild-pressure recordings are 'largely indistinguishable from noise' and in Section V-C that normal and shear forces are not differentiated. These are honest and welcome acknowledgments, but they are in tension with the unqualified 'high fidelity' language used in the title and abstract. Please either add explicit qualifiers to the headline claims or provide evidence that the dataset's force channel is informative for the regimes that matter in the reported experiments.
minor comments (5)
  1. [IV-A] The text 'a) beck, neck, and obliques' contains a typo; 'beck' should be 'back'.
  2. [V-A] The word 'assymetry' should be 'asymmetry'.
  3. [III-B, IV-C] The notation c_hi/c_bi and the units of the Table II metric should be defined explicitly; the current text does not state whether the reported distances are per-taxel or aggregated across the 65 taxels.
  4. [III-B] The weighting hyperparameters λ_s, λ_j, and λ_c are listed as fixed for all captures, but no sensitivity analysis is provided; a brief perturbation study would help the reader assess robustness to these choices.
  5. [Appendix A] The appendix states that re-projection MAE losses are normally distributed, but Fig. 15 and the related text do not show error bars or trial counts; adding these would improve interpretability.

Circularity Check

2 steps flagged · score 7.0 of 10

Reconstruction accuracy metric is the contact-match objective term itself, and closed-loop pressure tracking is the controller's own setpoint.

  1. self definitional [Section III-B Eq. (2) and Section IV-C Table II]
    "We then perform a second contact match pass by adding a single additional objective term to Eq. 1: λc Σ ||c_hi − c_bi||^2_2 (2) ... We quantitatively evaluate the quality of our reconstruction pipeline using two metrics: L2 distances between estimated body contacts and glove taxels, and smoothness of body contact trajectories. Table II tabulates distance comparisons of our method against the original MoSh++ [42] reconstruction across 10 sample bathing demonstrations."

    The Table II metric is exactly the Eq. (2) term that the contact match pass minimizes. Since the body contact points c_bi are generated as closest-point queries from taxel positions onto the SMPL-X body (Section III-B), the optimization directly drives the measured quantity to small values. The comparison against MoSh++ therefore only proves that minimizing a contact distance reduces that same contact distance; it does not independently validate that the taxel-to-body correspondences are physically correct, nor that the reconstruction is 'high fidelity' in any external sense. A non-circular evaluation would need held-out taxels, independent contact ground truth, or a different metric not contained in the optimized objective.

  2. other [Section III-D and Section IV-E, Figure 14]
    "Corrective movements are applied toward the body until the online sum of tactile pressures matches the original measurement, then reversed if that value is exceeded. ... Closed-loop pressures also reasonably track the original human demonstration, particularly when viewed on a normalized scale."

    The closed-loop controller's stopping rule is equality between the online tactile sum and the original human measurement, and Figure 14 evaluates exactly that absolute/normalized tactile sum. Hence the reported 'tracking' is the controller's own target, not an independent outcome of the transfer pipeline. The paper's limitation paragraph concedes this by noting that 'optimizing for an aggregate pressure sum does not necessitate parity in the underlying taxel-wise distribution,' confirming that only the constructed aggregate signal is compared.

full rationale

The paper's headline claim of 'high fidelity' reconstruction rests primarily on Table II's L2 distance between estimated body contacts and glove taxels. This quantity is identical to the contact-match objective added in Eq. (2), with body contacts defined by closest-point queries from the same taxel data. Consequently, the quantitative reconstruction evaluation is circular: it validates the optimization by its own loss. The closed-loop pressure comparison is also partly tautological because the controller drives the tactile sum to the human recording and then reports that sum tracks. Other contributions—dataset capture, synchronized multimodal data, soft-hand design, and retargeting—are not circular, and the paper's self-citations are ordinary uses of prior work rather than load-bearing circularity. The central reconstruction metric, however, is forced by the objective it minimizes, warranting a partial-circularity score of 7 rather than a clean bill.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central reconstruction claim rests on three hand-selected hyperparameters, the closest-point correspondence assumption, and the validity of SMPL-X/MANO/MoSh++ fits. The transfer claim additionally relies on manual motor calibration and a static-body assumption. No new physical entities are introduced; the main free parameters are optimization weights and manually calibrated motor positions.

free parameters (4)
  • lambda_s, lambda_j, lambda_c (reconstruction weighting hyperparameters) = 1, 50, 40
    Hand-selected relative weights in Eq. 1 and Eq. 2, chosen to offset dimensionality differences between vertices (776), taxels (65), and DOFs (29). They directly influence the reconstructed poses and the reported contact distance metric.
  • Pose candidate counts |S| and |S'| = 50 and 20
    |S|=50 selected from a sweep showing diminishing returns on MAE loss; |S'|=20 chosen based on a six-motor budget. These determine the DexKit hand's pose coverage and motor blending quality.
  • Manually determined motor positions phi(s) for each candidate pose = Not reported numerically
    The paper states 'we manually determine motor positions for each candidate pose' (Section IV-D). This manual calibration is a fitted set of values that the closed-loop control depends on.
  • Closed-loop pressure tolerance bounds = Not reported numerically
    Tolerance bounds around the target pressure sum were introduced to mitigate arm oscillations and to implement the temporal safety timeout, as described in Section IV-E and the discussion.
assumptions (7)
  • domain assumption SMPL-X and MANO parametric models faithfully represent human body and hand shape and pose.
    Used in Section III-B to fit subject bodies and clinician hands from ROM tests and MoSh++.
  • domain assumption MoSh++ provides reasonable initial shape and pose fits for the marker data.
    The entire reconstruction pipeline starts from MoSh++ output (Section III-B), and the paper's own figures show severe artifacts in that baseline, meaning the accuracy of the final result depends on the quality of the initialization.
  • domain assumption All forces measured by the tactile glove are normal to each taxel.
    Explicitly stated in Section III-A: 'We assume all forces are normal to each taxel.' Shear forces are ignored, which the paper acknowledges in limitations.
  • domain assumption A reduced-DOF hand model with anatomical joint limits captures the relevant hand motion for bathing.
    The shape match pass replaces the MANO hand mesh with a constrained skeleton-driven model (Section III-B), relying on the assumption that the reduced DOF set can represent clinician hand poses.
  • domain assumption The closest point on the body mesh to a taxel is the true contact point for that taxel.
    This is the core assumption of the contact match pass (Eq. 2) and the L2 evaluation metric. It fails if the body mesh is misaligned due to clothing, soft tissue deformation, or mannequin proxies, which are all present in this paper.
  • domain assumption TRAM's SMPL-X fitting of the mannequin is an adequate proxy for the mannequin's actual geometry.
    Used in Section III-D to retarget contact trajectories. The appendix notes the need for clothing and arm attachment for TRAM to identify the mannequin as human, and the proxy mesh is intentionally offset.
  • domain assumption The subject body and mannequin are static during deployment.
    The transfer rollout assumes a static body; the paper explicitly states in limitations that 'our current assumption of a static body will not extend to a human subject.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of High Fidelity Capture, Reconstruction, and Transfer of Human Demonstrations for Robot-Assisted Bathing." pith.science (2026). https://pith.science/paper/6E6YU5QW

@misc{pith2026260809127,
  author       = {Pith},
  title        = {Pith review of: High Fidelity Capture, Reconstruction, and Transfer of Human Demonstrations for Robot-Assisted Bathing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6E6YU5QW}},
  note         = {Machine review of arXiv:2608.09127}
}
read the original abstract

Despite the demand for robots in high-value clinical tasks like bathing, contemporary systems still lack the safety and reliability required for complex, sustained physical interaction with humans. A key challenge hindering the development of such systems is that collecting, understanding, and effectively transferring highly dynamic, contact-rich human bathing demonstrations is difficult, even with modern motion and tactile sensing equipment. We present a straightforward, but effective framework for doing so with high fidelity by utilizing contact regions as a key processing primitive. We use our framework to build a dataset of bathing demonstrations performed by trained clinicians on human subjects. We then use this dataset to design and control an arm-mounted dexterous soft hand to perform bathing tasks on a mannequin using open- and closed-loop strategies. Our dataset is the first to provide high quality synchronized motion, shape, contact, and force during sustained, contact-rich human-human interaction, and our transfer strategies demonstrate effective use of these data across multiple levels of the robotics stack. All relevant materials will be publicly released to enable further advancements in physical human-robot interaction (pHRI) research.

Figures

Figures reproduced from arXiv: 2608.09127 by the authors.

Figure 1
Figure 1. Our contributions enable straightforward, yet high [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our pipeline. Starting with captures of motion and tactile forces from human bathing demonstrations, we [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Optical marker configurations used on (a) human [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Comparison film strip of a leg bathing reconstruction [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Visualizations of improved surface contact following [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: MoSh++ [42] shape reconstructions of all subjects. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Contact coverage (yellow) over all captures of a [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Visualizations of the most commonly (a) contacted [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 10
Figure 10. Figure 10: Visualization of a bathing “stroke” from the (green) [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 12
Figure 12. Figure 12: Sample poses used to build Mϕ(S ′ ) . Active contacts in each pose (pink) are simulated by pressing against the hand [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]
Figure 13
Figure 13. Figure 13: Film strip of back bathing demonstration retargeted [PITH_FULL_IMAGE:figures/full_fig_p007_13.png]
Figure 17
Figure 17. Figure 17: Illustrations of the (a) mannequin input image and (b) [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]
Figure 16
Figure 16. Figure 16: The DexKit hand is comprised of a cast foam hand [PITH_FULL_IMAGE:figures/full_fig_p012_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 48 canonical work pages

  1. [1]

    D. D. Dunlop, S. L. Hughes, and L. M. Manheim. Disability in activities of daily living: patterns of change and a hierarchy of disability.American journal of public health, 87(3):378–383, 1997

    work page 1997

  2. [2]

    S. Katz, A. B. Ford, R. W. Moskowitz, B. A. Jackson, and M. W. Jaffe. Studies of illness in the aged: the index of adl: a standardized measure of biological and psychosocial function.Journal of the American Medical Association, 185(12):914–919, 1963

    work page 1963

  3. [3]

    M Gill, A

    T. M Gill, A. Guo, and H. G. Allore. The epidemiology of bathing disability in older persons.Journal of the American Geriatrics Society, 54(10):1524–1530, 2006

    work page 2006

  4. [4]

    L. R. Czuba, C. M. Sommerich, and S. A. Lavender. Ergonomic and safety risk factors in home health care: Exploration and assessment of alternative interventions. Work, 42(3):341–353, 2012

    work page 2012

  5. [5]

    A. R. Darragh, C. M. Sommerich, S. A. Lavender, K. J. Tanner, K. V ogel, and M. Campo. Musculoskeletal discomfort, physical demand, and caregiving activities in informal caregivers.Journal of Applied Gerontology, 34(6):734–760, 2015

    work page 2015

  6. [6]

    Erickson, H

    Z. Erickson, H. M. Clever, V . Gangaram, G. Turk, C. K. Liu, and C. C. Kemp. Multidimensional capacitive sens- ing for robot-assisted dressing and bathing. InIEEE In- ternational Conference on Rehabilitation Robotics, pages 224–231, 2019

    work page 2019

  7. [7]

    F. Liu, V . Patil, Z. Erickson, and Z. Temel. Charac- terization of a meso-scale wearable robot for bathing assistance. InIEEE International Conference on Robotics and Biomimetics, pages 2146–2152. IEEE, 2022

    work page 2022

  8. [8]

    Madan, S

    R. Madan, S. Valdez, D. Kim, S. Fang, L. Zhong, D. T. Virtue, and T. Bhattacharjee. Rabbit: A robot- assisted bed bathing system with multimodal perception and integrated compliance. InProceedings of the 2024 ACM/IEEE international conference on human-robot in- teraction, pages 472–481, 2024

    work page 2024

Show all 59 references
  1. [9]

    Huang, D

    I. Huang, D. Chow, and R. Bajcsy. Soft tactile contour following for robot-assisted wiping and bathing. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 7797–7802, 2022

    2022

  2. [10]

    A. C. Dometios, X. S. Papageorgiou, A. Arvanitakis, C. S. Tzafestas, and P. Maragos. Real-time end-effector motion behavior planning approach using on-line point- cloud data towards a user adaptive assistive bath robot. In IEEE/RSJ International Conference on Intelligent Robot...

    2017

  3. [11]

    Zlatintsi, I

    A. Zlatintsi, I. Rodomagoulakis, P. Koutras, A. C. Dome- tios, V . Pitsikalis, C. S. Tzafestas, and P. Maragos. Multimodal signal processing and learning aspects of human-robot interaction for an assistive bathing robot. InIEEE International Conference on Acoustics, Speech and...

    2018

  4. [12]

    C. H. King, T. L. Chen, A. Jain, and C. C. Kemp. Towards an assistive robot that autonomously performs bed baths for patient hygiene. InIEEE/RSJ International Conference on Intelligent Robots and Systems, pages 319–324, 2010

    2010

  5. [13]

    Sundaram, P

    S. Sundaram, P. Kellnhofer, Y . Li, J. Y . Zhu, A. Torralba, and W. Matusik. Learning the signatures of the human grasp using a scalable tactile glove.Nature, pages 698– 702, 2019

    2019

  6. [14]

    Y . Luo, C. Liu, Y . J. Lee, J. DelPreto, K. Wu, M. Foshey, D. Rus, T. Palacios, Y . Li, a. Torralba, et al. Adaptive tactile interaction transfer via digitally embroidered smart gloves.Nature Communications, 15(1):868, 2024

    2024

  7. [15]

    TactileGlove: Hand pres- sure and force measurement

    Pressure Profile Systems. TactileGlove: Hand pres- sure and force measurement. https://pressureprofile.com/ body-pressure-mapping/tactile-glove, 2026. Accessed: 2026-01-26

    2026

  8. [16]

    H. S. Fang, B. Romero, Y . Xie, A. Hu, B. R. Huang, J. Alvarez, M. Kim, G. Margolis, K. Anbarasu, M. Tomizuka, et al. Dexop: A device for robotic transfer of dexterous human manipulation.arXiv preprint arXiv:2509.04441, 2025

    2025 arXiv

  9. [17]

    Adeniji, Z

    A. Adeniji, Z. Chen, V . Liu, V . Pattabiraman, R. Bhirangi, S. Haldar, P. Abbeel, and L. Pinto. Feel the force: Contact-driven learning from humans.arXiv preprint arXiv:2506.01944, 2025

    2025 arXiv

  10. [18]

    J. Yin, H. Qi, Y . Wi, S. Kundu, M. Lambeta, W. Yang, C. Wang, T. Wu, J. Malik, and T. Hellebrekers. Osmo: Open-source tactile glove for human-to-robot skill trans- fer.arXiv preprint arXiv:2512.08920, 2025

    2025

  11. [19]

    Y . Mao, U. Yoo, Y . Yao, S. N. Syed, L. Bondi, J. Francis, J. Oh, and J. Ichnowski. Visuo-acoustic hand pose and contact estimation.arXiv preprint arXiv:2508.00852, 2025

    2025 arXiv

  12. [20]

    C. Xing, H. Li, Y . L. Wei, T. A. Ren, T. Tu, Y . Lin, E. Schumann, W. S. Zheng, and M. R. Cutkosky. Taccap: A wearable fbg-based tactile sensor for seamless human- to-robot skill transfer.arXiv preprint arXiv:2503.01789, 2025

    2025 arXiv

  13. [21]

    A. S. Lakshmipathy, N. Feng, Y . X. Lee, M. Mahler, and N. S. Pollard. Contact edit: Artist tools for intuitive modeling of hand-object interactions.ACM Transactions on Graphics, 42(4), 2023

    2023

  14. [22]

    Turpin, L

    D. Turpin, L. Wang, E. Heiden, Y . C. Chen, M. Macklin, S. Tsogkas, S. Dickinson, and A. Garg. Grasp’d: Differ- entiable contact-rich grasp synthesis for multi-fingered hands. InEuropean Conference on Computer Vision, pages 201–221. Springer, 2022

    2022

  15. [23]

    Brahmbhatt, A

    S. Brahmbhatt, A. Handa, J. Hays, and D. Fox. Con- tactgrasp: Functional multi-finger grasp synthesis from contact. In2019 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2386–2393. IEEE, 2019

    2019

  16. [24]

    Christen, M

    S. Christen, M. Kocabas, E. Aksan, J. Hwangbo, J. Song, and O. Hilliges. D-grasp: Physically plausible dy- namic grasp synthesis for hand-object interactions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20577–20586, 2022

    2022

  17. [25]

    Y . Wu, J. Wang, Y . Zhang, S. Zhang, O. Hilliges, F. Yu, and S. Tang. Saga: Stochastic whole-body grasping with contact. InEuropean Conference on Computer Vision, pages 257–274. Springer, 2022

    2022

  18. [26]

    A. S. Lakshmipathy, J. K. Hodgins, and N. S. Pollard. Kinematic motion retargeting for contact-rich anthropo- morphic manipulations.ACM Transactions on Graphics, 44(2):1–20, 2025

    2025

  19. [27]

    Mandi, Y

    Z. Mandi, Y . Hou, D. Fox, Y . Narang, A. Mandlekar, and S. Song. Dexmachina: Functional retargeting for bimanual dexterous manipulation.arXiv preprint arXiv:2505.24853, 2025

    2025 arXiv

  20. [28]

    C. Pan, C. Wang, H. Qi, Z. Liu, H. Bharadhwaj, A. Sharma, T. Wu, G. Shi, J. Malik, and F. Hogan. Spider: Scalable physics-informed dexterous retargeting. arXiv preprint arXiv:2511.09484, 2025

    2025

  21. [29]

    T. Pang, H. J. T. Suh, L. Yang, and R. Tedrake. Global planning for contact-rich manipulation via local smooth- ing of quasi-dynamic contact models.IEEE Transactions on robotics, 39(6):4691–4711, 2023

    2023

  22. [30]

    Cheng, E

    X. Cheng, E. Huang, Y . Hou, and M. T. Mason. Contact mode guided motion planning for quasidynamic dexter- ous manipulation in 3d. InInternational Conference on Robotics and Automation, pages 2730–2736. IEEE, 2022

    2022

  23. [31]

    Taheri, N

    O. Taheri, N. Ghorbani, M. J. Black, and D. Tzionas. Grab: A dataset of whole-body human grasping of ob- jects. InEuropean conference on computer vision, pages 581–600. Springer, 2020

    2020

  24. [32]

    Z. Fan, O. Taheri, D. Tzionas, M. Kocabas, M. Kauf- mann, M. J. Black, and O. Hilliges. Arctic: A dataset for dexterous bimanual hand-object manipulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943–12954, 2023

    2023

  25. [33]

    DelPreto, C

    J. DelPreto, C. Liu, Y . Luo, M. Foshey, Y . Li, A. Torralba, W. Matusik, and D. Rus. Actionsense: A multimodal dataset and recording framework for human activities us- ing wearable sensors in a kitchen environment.Advances in Neural Information Processing Systems, 35:13800– 1...

    2022

  26. [34]

    Y . R. Song, J. Li, R. Fu, D. Murphy, K. Zhou, R. Shiv, Y . Li, H. Xiong, C. E. Owens, Y . Du, et al. Opentouch: Bringing full-hand touch to real-world interaction.arXiv preprint arXiv:2512.16842, 2025

    2025

  27. [35]

    UFACTORY xArm 7 Robotic Arm

    UFACTORY. UFACTORY xArm 7 Robotic Arm. https: //www.ufactory.us/product/ufactory-xarm-7, 2026. Ac- cessed: 2026-01-26

    2026

  28. [36]

    J. P. King, H. Ahluwalia, I. B. Heredia-Marin, K. McClary, E. Zuo, M. Baratz, J. E. Colgate, J. Clark, M. Orta Martinez, and N. S. Pollard. Dexkit: A hard- ware platform for multi-finger dexterous robotics. In Proceedings of the 19th International Symposium on Experimental Rob...

    2025

  29. [37]

    Gripping Material TB641

    3M™. Gripping Material TB641. https://www.3m.com/ 3M/en US/p/d/b40069682/, 2026. Accessed: 2026-01- 26

    2026

  30. [38]

    L. K. Baxter. Capacitive sensors.IEEE Design and Applications, 1997

    1997

  31. [39]

    V16 camera specifications - Vantage documentation

    Vicon Motion Systems. V16 camera specifications - Vantage documentation. https://help.vicon.com/space/ Vantage/15041618/V16+camera+specifications, 2026. Accessed: 2026-01-26

    2026

  32. [40]

    Pavlakos, V

    G. Pavlakos, V . Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J Black. Expressive body capture: 3d hands, face, and body from a single image. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10975–10985, 2019

    2019

  33. [41]

    Romero, D

    J. Romero, D. Tzionas, and M. J. Black. Embodied hands: Modeling and capturing hands and bodies to- gether.ACM Transactions on Graphics, 36(6), 2017

    2017

  34. [42]

    Mahmood, N

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black. Amass: Archive of motion capture as surface shapes. InIEEE International Conference on Computer Vision, pages 5442–5451, 2019

    2019

  35. [43]

    D. J. Magee.Orthopedic Physical Assessment. Elsevier Health Sciences, 2013

    2013

  36. [44]

    Magnenat-Thalmann, R

    N. Magnenat-Thalmann, R. Laperri `ere, and D. Thal- mann. Joint-dependent local deformations for hand animation and object grasping. InGraphics Interface, pages 26–33, 1989

    1989

  37. [45]

    R. Sawhney. Fcpw: Fastest closest points in the west, 2021

    2021

  38. [46]

    C. C. Kuo, R. Glover, and K. S. Dhir. Analyzing and modeling the maximum diversity problem by zero- one programming.Decision Sciences, 24(6):1171–1185, 1993

    1993

  39. [47]

    J. B. Ghosh. Computational aspects of the maxi- mum diversity problem.Operations Research Letters, 19(4):175–181, 1996

    1996

  40. [48]

    Sivakumar, K

    A. Sivakumar, K. Shaw, and D. Pathak. Robotic telekine- sis: Learning a robotic hand imitator by watching humans on youtube.arXiv preprint arXiv:2202.10448, 2022

    2022 arXiv

  41. [49]

    Y . Wang, Z. Wang, L. Liu, and K. Daniilidis. Tram: Global trajectory and motion of 3d humans from in- the-wild videos. InEuropean Conference on Computer Vision, pages 467–487, 2024

    2024

  42. [50]

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

    2024

  43. [51]

    M. T. Mason.Mechanics of robotic manipulation. MIT press, 2001

    2001

  44. [52]

    Ferrari, J

    C. Ferrari, J. Canny, et al. Planning optimal grasps. In Proceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., volume 3, pages 2290–

    1992

  45. [53]

    K. M. Lynch and F. C. Park.Modern robotics. Cambridge University Press, 2017

    2017

  46. [54]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. InIEEE/RSJ Interna- tional conference on intelligent robots and systems, pages 5026–5033, 2012

    2012

  47. [55]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    2021 arXiv

  48. [56]

    Lakshmipathy, D

    A. Lakshmipathy, D. Bauer, C. Bauer, and N. S. Pollard. Contact transfer: A direct, user-driven method for human to robot transfer of grasps and manipulations. In2022 International Conference on Robotics and Automation (ICRA), pages 6195–6201. IEEE, 2022

    2022

  49. [57]

    Elandt, E

    R. Elandt, E. Drumwright, M. Sherman, and A. Ruina. A pressure field model for fast, robust approximation of net contact force and moment between nominally rigid objects. InIEEE/RSJ International Conference on Intelligent Robots and Systems, pages 8238–8245. IEEE, 2019

    2019

  50. [58]

    Y . Wi, J. Yin, E. Xiang, A. Sharma, J. Malik, M. Mukadam, N. Fazeli, and T. Hellebrekers. Tactalign: Human-to-robot policy transfer via tactile alignment. arXiv:2602.13579, 2026

    2026

  51. [59]

    boundaries

    W. Yuan, S. Dong, and E. H. Adelson. Gelsight: High- resolution robot tactile sensors for estimating geometry and force.Sensors, 17(12):2762, 2017. Fig. 15: Complete MAE loss curve over the parameter sweep of|S|. Fig. 16: The DexKit hand is comprised of a cast foam hand with a...

    2017

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.