Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Imitation Learning for Adaptive Control of a Virtual Soft Exoglove

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · 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 a tendon-driven virtual glove, trained by reinforcement learning to assist a musculoskeletal hand model weakened by simulated sarcopenia, restores an average of 90.5% of healthy manipulation success.

desk verdict A coherent simulation proof-of-concept for personalized exoglove control, but the headline 90.5% restoration needs a specified actuation model and per-condition stats before it can be interpreted. read the letter →

arxiv 2505.09099 v1 pith:UFWISYBI submitted 2025-05-14 cs.RO cs.LG

Shirui Lyu , Vittorio Caggiano , Matteo Leonetti , Dario Farina , Letizia Gionfrida This is my paper · ORCID
classification cs.ROcs.LG
keywords imitationlearningreinforcementmusculoskeletalmodelsoftexoglovehandrehabilitationsarcopeniaobjectmanipulationvideodemonstration
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 establish that an assistive hand glove can be designed and tuned entirely in simulation, starting from ordinary video of a healthy person grasping objects and ending with a controller that compensates for a specific simulated muscle impairment. Using a biologically detailed musculoskeletal hand model, the authors weaken muscle strength to mimic sarcopenia and train a virtual tendon-driven glove with reinforcement learning, using the healthy subject's behavior as the reference. The learned glove controller restores, on average, 90.5% of the healthy hand's success in moving three everyday objects along their reference trajectories. If this chain holds together, it offers a route to personalized wearable rehabilitation in which the patient's individual muscle deficit pattern, rather than a generic assistance program, determines the glove's control law.

What carries the argument

The load-bearing machinery is a chain of three trained components. A behavior prior, learned by imitation from video demonstrations of a single healthy subject, gives the musculoskeletal hand model the ability to follow the subject's reference joint trajectories. An object-centric reinforcement learning manipulator then fine-tunes that prior using a reward that combines trajectory-following with object-position and orientation matching, producing a healthy baseline for each object. The system then switches to a weaker hand by applying a muscle-weakening model, and trains a separate exoglove controller on top: a virtual tendon-driven glove with three control dimensions (index-finger flexion-extension, middle-finger flexion-extension, and passive thumb internal-rotation support), modeled on a physical prototype. The key design choice is that the glove shares the hand's observation space and reward function, so the two agents optimize the same objective, matching the object to the demonstrated trajectory, rather than acting as separate systems.

What would settle it

Build the physical glove and drive it with the same controller while a healthy participant's grip strength is temporarily reduced, then measure whether success on the same three objects reaches at least 90.5% of that participant's baseline; a shortfall would show the simulated assistance does not transfer to hardware.

Watch

Extended reading notes

Core claim

The central discovery is that imitating a healthy subject's hand kinematics from video, then fine-tuning on object-manipulation tasks, produces a musculoskeletal control policy that, when the same hand's muscles are weakened by a simulated sarcopenia model, can be coupled with a three-degree-of-freedom tendon-driven exoglove controller to recover close to healthy performance. In experiments with three everyday objects, the unaided weakened hand drops objects and accrues large trajectory error, while the same weakened hand with the learned glove assistance follows the reference trajectory with an average of 90.5% of the healthy baseline success rate. The glove's effect is also visible at the biomechanical level: the distance between the middle finger's proximal interphalangeal joint and the wrist returns to a pattern matching the healthy hand's contraction and holding phases. The authors interpret this as evidence that the shared reward and observation design lets the glove controller learn compensatory actions aligned with the user's own behavior.

Load-bearing premise

The result stands only if the simulated weakening of hand muscles and the virtual glove's actuation faithfully capture a real impaired hand and a real tendon-driven glove; if either mapping is inaccurate, the 90.5% restoration is a simulator artifact.

Editorial extensions

If this is right

  • If the simulation chain is correct, assistive glove control can be personalized without hand-crafted therapy protocols: a patient's muscle deficit pattern, which muscles are weak, becomes a parameter of the simulation, and the controller is retrained for that pattern.
  • Because the glove only actuates the index and middle fingers plus thumb support, the 90.5% restoration suggests that partial mechanical assistance can carry most of the manipulation load for power- and precision-grasp-style tasks of this type.
  • The imitation-learned behavior prior speeds convergence on unseen trajectories from the same subject, meaning new objects or activities could be added to a patient's training set without rebuilding the controller from scratch.
  • The success criterion used, keeping the object's center of mass within 2.5 cm of the reference trajectory, mirrors the structure of clinical dexterity tests, so the training objective connects directly to how rehabilitation outcomes are measured.

Reading between the lines

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

  • The paper does not test across impairment severities; a natural extension would be to vary the weakening factor and find the deficit level at which the three-dimensional glove can no longer restore 90.5%, which would define the device's clinical coverage.
  • The same pipeline could translate to other neurological impairments, such as spasticity, tremor, or partial paralysis, by substituting a different simulated muscle model, but the paper only demonstrates pure strength reduction, and the glove's limited actuation points may not suit those conditions.
  • If the shared observation space transfers to physical hardware, a patient-specific glove setting could be generated from a short smartphone video without instrumented motion capture, because all states used by the controller are vision-estimable joint and object positions.
  • The paper leaves the sim-to-real gap open; the strongest test of the claimed mechanism would be building the physical glove and checking whether its measured assistive torques match the simulated ones, rather than only comparing end-task success.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The manuscript presents a simulation pipeline for personalized assistive exoglove control. It extracts hand joint trajectories of one subject from the DexYCB dataset, trains a MyoSuite musculoskeletal hand model to track demonstrations via RL-based imitation learning, fine-tunes it on object-centric manipulation tasks for three YCB objects, introduces a simulated sarcopenia-weakened hand, and adds a virtual tendon-driven exoglove with a 3D action space. The exoglove controller is trained with the same object-centric reward function and, according to the abstract and Fig. 7, restores 90.5% of the original manipulation proficiency for the weakened hand. Evaluation uses success rates computed over 30 trials and accumulated object-position error curves.

Significance. If substantiated, the framework would be a useful proof-of-concept for combining imitation learning with musculoskeletal simulation to personalize exoglove assistance for individual muscle deficits. The paper builds on appropriate prior work (MyoDex, MyoSuite, DexYCB) and is commendable for using a biologically plausible hand model and publicly available data and simulators. However, the headline quantitative claim is supported only by an unbroken-down average and an underspecified virtual glove model; as it stands, the contribution is a coherent pipeline demonstration rather than a validated assistive-control result.

major comments (4)
  1. [Section III-C and Fig. 2] The exoglove actuation model is never formalized. The text states only that the glove is based on a real prototype [20] and has a 3D action space (index flexion, middle flexion, thumb support), but no equations or parameter values describe how the 3D action vector maps to tendon tensions, applied forces, or position targets in MuJoCo, nor are force limits or attachment points given. Because the 90.5% restoration figure in the Abstract and Fig. 7 depends critically on the strength and effort of this virtual glove, the result cannot be reproduced or interpreted; the paper must specify the actuation mapping, the sarcopenia weakness factor (severity of maximum-force reduction), and per-object success rates for the healthy, weakened-without-glove, and weakened-with-glove conditions.
  2. [Section IV-C and Fig. 7] The quantitative backbone of the central claim is missing. Table I reports only healthy success rates (89%, 97.6%, 93.8%); there is no table or text reporting success rates for the weakened hand without the glove or with the glove, no per-object restoration percentages, and no error bars or confidence intervals for the 30 trials. The '90.5% of the original manipulation proficiency' is therefore an unverifiable average, and Fig. 7's accumulated-error curves, while visually suggestive, do not support the claimed success-rate restoration without the corresponding numbers.
  3. [Section III-B, Eq. (2)] The definition of R_demo is inconsistent with its prose. The text says it matches musculoskeletal hand joint positions with human-demonstration joints, but Eq. (2) compares x_t^(p) (object position) with q(t) (reference joints); additionally, the notation for the hand joints used in this reward term is never introduced. As written, the equation cannot be implemented as described, and since R_demo is one of the two core rewards for all three training stages, this is a load-bearing ambiguity that must be corrected.
  4. [Section III-C and IV-C] The evaluation is circular with respect to the training objective. The exoglove controller is trained with the same reward terms (Rdemo + Robj) and evaluated on the same objects and reference trajectories used for training, so the 90.5% restoration largely measures how well the policy re-optimizes its own training objective rather than its ability to generalize to new weakness profiles, objects, or tasks. The authors should include a held-out trajectory or object, a different sarcopenia severity, or an explicit statement that this is a closed-loop re-optimization result rather than a generalization result.
minor comments (5)
  1. [Throughout] There are several typos and grammatical errors: 'Results shows' in the Abstract, 'succssess' in Section IV-C, 'V alidation' in Section IV-D, and 'for enhanced graspin' in Section IV-D; a careful proofread is needed.
  2. [Fig. 7] The caption says the glove 'restores an average of 90.5% of the original success rate,' but the figure actually plots accumulated error; the caption should be reworded to avoid equating error curves with success rates.
  3. [Table I] The table omits units or a legend for 'Performances' and does not explicitly state whether these are healthy-baseline numbers; the caption should be made explicit.
  4. [Section IV-A] The text says the prior is trained on 16 environments with 6 objects, while only 3 objects appear in the downstream evaluation; clarify the relationship between the environmental objects and the selected YCB objects.
  5. [Section III-B] The reward notation mixes p/o superscripts and q(t) without a table of symbols; adding one would improve reproducibility. The success threshold of 2.5 cm is also a free parameter with no sensitivity analysis.

Circularity Check

1 steps flagged · score 6.0 of 10

The 90.5% restoration metric is a thresholded version of the reward the exoglove controller was trained to maximize on the same objects, making the central validation in-distribution rather than independent.

  1. fitted input called prediction [Section III-B 'Reward function'; Section IV-B 'Training of hand-object manipulation tasks'; Fig. 7 caption.]
    "The reward function for matching the object position and orientation to the reference trajectory is defined by: Robj := λ1 exp(−α1||x(p)_t − x̂(p)_t||2 − β|∠x(o)_t − ∠x̂(o)_t|) ... the manipulation model is trained on a reward function defined by R = R_demo + R_obj ... For each step during evaluation, the object’s centre of mass has to be within a 2.5-cm range from the reference position to be considered a successful step. ... The glove controller restores an average of 90.5% of the original success rate."

    The exoglove controller is trained with the same reward R = R_demo + R_obj, and the evaluation success criterion ('object's centre of mass has to be within a 2.5-cm range from the reference position') is a thresholded version of the Robj position-matching term. Since training and evaluation use the same YCB objects and reference trajectories, the reported 90.5% restoration measures how well the optimizer re-attained its own training objective on its training distribution, not an independent generalization or prediction. The frozen weakened hand and limited 3-DoF glove action space add a physical constraint, so the result is not entirely forced, but the central quantitative claim is in-distribution by construction.

full rationale

The main claimed result, 90.5% restoration of manipulation proficiency, is not validated against an independent objective. The exoglove policy is trained with reward R = R_demo + R_obj, and the evaluation success rate is defined by the object centre of mass staying within 2.5 cm of the reference trajectory, which is a discretized Robj term. Training and evaluation also use the same object-trajectory pairs (chef can, sugar box, tomato can from DexYCB), so the restoration figure is a measure of the policy's performance on its own training objective. This is a genuine partial circularity in the validation of the central claim, though not a total logical collapse: the weakened hand is frozen and the glove has only three actuation dimensions, so reaching the objective is not guaranteed a priori. The self-citations to MyoSuite [19] and MyoDex [26] are not treated as circular here: they are external, published simulators and methods used for parameters and baselines, not an invoked uniqueness theorem. The underspecified exoglove actuation mapping (no equations for tendon routing, application points, or force limits) and the missing sarcopenia severity are serious reproducibility and correctness gaps, but under the hard rules they are not themselves evidence of circularity. Score 6 reflects that the central claim reduces to an in-distribution training objective rather than an independent prediction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The central claim depends on simulation-fidelity assumptions about the musculoskeletal hand model, the sarcopenia impairment model, the DexYCB reference data, and the virtual exoglove actuation. The only introduced component is the virtual glove, which is not hardware-validated here. The reported 90.5% is a training-task score, not a generalization benchmark.

free parameters (2)
  • Success threshold distance = 2.5 cm
    Section IV-B defines a successful evaluation step as the object center of mass being within 2.5 cm of the reference position. This threshold directly determines the reported success rates and the headline 90.5% restoration figure.
  • Sarcopenia weakness scaling = not reported
    Section IV-C introduces muscle weakness via the MyoSuite sarcopenia model, but the paper does not state the exact reduction in maximum muscle force. The severity of the simulated impairment and the glove compensation result depend on this unreported parameter.
assumptions (5)
  • domain assumption MyoSuite musculoskeletal hand model is a biologically accurate representation of human hand dynamics.
    Section III-B builds on MyoHand from MyoSuite [19]; the entire simulation evidence rests on this representational adequacy, and no human validation is presented.
  • domain assumption Sarcopenia muscle weakening in MyoSuite adequately represents neurological motor impairment.
    Section IV-C simulates impairment by reducing maximum muscle strength with the MyoSuite sarcopenia model; the authors do not compare against patient data.
  • domain assumption DexYCB video pose tracking provides accurate reference kinematics for imitation.
    Reference joint positions extracted from video (Section IV-A) are treated as ground truth for the healthy behavior prior; pose-estimation errors propagate into the learned controller.
  • domain assumption The virtual exoglove actuation is a valid proxy for a physical tendon-driven soft glove.
    The glove is modeled on a real prototype [20] with 3 control dimensions, but its simulated force output is not calibrated against hardware measurements.
  • domain assumption The 2.5 cm object-position success threshold and the 80% success-rate interpretation are appropriate task performance measures.
    Section IV-B defines success by a 2.5 cm distance threshold and cites prior work that 80% success reads as user-perceived success; these choices determine the headline 90.5%.
invented entities (1)
  • Virtual tendon-driven exoglove
    purpose: Provides 3-dimensional assistive actuation (finger contraction, extension, and thumb support) to compensate for weakened hand muscles in simulation.
    The virtual glove is modeled after a previously reported physical prototype [20], but its simulated actuation dynamics are not validated against the physical device in this paper, so there is no direct falsifiable handle outside the simulation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Imitation Learning for Adaptive Control of a Virtual Soft Exoglove." pith.science (2026). https://pith.science/paper/UFWISYBI

@misc{pith2026250509099,
  author       = {Pith},
  title        = {Pith review of: Imitation Learning for Adaptive Control of a Virtual Soft Exoglove},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFWISYBI}},
  note         = {Machine review of arXiv:2505.09099}
}
read the original abstract

The use of wearable robots has been widely adopted in rehabilitation training for patients with hand motor impairments. However, the uniqueness of patients' muscle loss is often overlooked. Leveraging reinforcement learning and a biologically accurate musculoskeletal model in simulation, we propose a customized wearable robotic controller that is able to address specific muscle deficits and to provide compensation for hand-object manipulation tasks. Video data of a same subject performing human grasping tasks is used to train a manipulation model through learning from demonstration. This manipulation model is subsequently fine-tuned to perform object-specific interaction tasks. The muscle forces in the musculoskeletal manipulation model are then weakened to simulate neurological motor impairments, which are later compensated by the actuation of a virtual wearable robotics glove. Results shows that integrating the virtual wearable robotic glove provides shared assistance to support the hand manipulator with weakened muscle forces. The learned exoglove controller achieved an average of 90.5\% of the original manipulation proficiency.

Figures

Figures reproduced from arXiv: 2505.09099 by the authors.

Figure 1
Figure 1. Pipeline overview illustrating the three main stages: (1) extraction of user behavior via imitation learning from healthy subject video data, (2) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the tendon-driven exoglove actuation in simulation, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Comparison of model training WITH (blue) and WITHOUT (red) the learned behaviour prior on previously unseen object and trajectory pair [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: Comparison of hand actuation behaviours for the tomato can, [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Accumulated error in object manipulation tasks with the exoglove [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 32 canonical work pages

  1. [20]

    Learning fingertip force to grasp deformable objects for soft wearable robotic glove with tsm,

    E. Rho, D. Kim, H. Lee, and S. Jo, “Learning fingertip force to grasp deformable objects for soft wearable robotic glove with tsm,”IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8126–8133, 2021

    work page 2021

  2. [1]

    Rehabilitation of motor function after stroke: a multiple systematic review focused on techniques to stimulate upper extremity recovery,

    S. M. Hatem, G. Saussez, M. Della Faille, V . Prist, X. Zhang, D. Dispa, and Y . Bleyenheuft, “Rehabilitation of motor function after stroke: a multiple systematic review focused on techniques to stimulate upper extremity recovery,”Frontiers in human neuroscience, vol. 10, p. 442, 2016

    work page 2016

  3. [2]

    Individualization of exosuit assistance based on measured muscle dynamics during versatile walking,

    R. W. Nuckols, S. Lee, K. Swaminathan, D. Orzel, R. D. Howe, and C. J. Walsh, “Individualization of exosuit assistance based on measured muscle dynamics during versatile walking,”Science robotics, vol. 6, no. 60, p. eabj1362, 2021

    work page 2021

  4. [3]

    Rehabilitation robots for the treatment of sensorimotor deficits: a neurophysiological perspective,

    R. Gassert and V . Dietz, “Rehabilitation robots for the treatment of sensorimotor deficits: a neurophysiological perspective,”Journal of neuroengineering and rehabilitation, vol. 15, pp. 1–15, 2018

    work page 2018

  5. [4]

    Stroke unit care, inpatient rehabilitation and early supported discharge,

    H. Rodgers and C. Price, “Stroke unit care, inpatient rehabilitation and early supported discharge,”Clinical Medicine, vol. 17, no. 2, pp. 173– 177, 2017

    work page 2017

  6. [5]

    Stroke rehabilitation: clinical picture, assessment, and therapeutic challenge,

    C. L. Richards, F. Malouin, and S. Nadeau, “Stroke rehabilitation: clinical picture, assessment, and therapeutic challenge,”Progress in brain research, vol. 218, pp. 253–280, 2015

    work page 2015

  7. [6]

    How does stroke affect skeletal muscle? state of the art and rehabilitation perspective,

    V . Azzollini, S. Dalise, and C. Chisari, “How does stroke affect skeletal muscle? state of the art and rehabilitation perspective,”Frontiers in Neurology, vol. 12, p. 797559, 2021

    work page 2021

  8. [7]

    Probability of regaining dexterity in the flaccid upper limb: impact of severity of paresis and time since onset in acute stroke,

    G. Kwakkel, B. J. Kollen, J. van der Grond, and A. J. Prevo, “Probability of regaining dexterity in the flaccid upper limb: impact of severity of paresis and time since onset in acute stroke,”Stroke, vol. 34, no. 9, pp. 2181–2186, 2003

    work page 2003

Show all 39 references
  1. [8]

    Time-dependent central compensatory mechanisms of finger dexterity after spinal cord injury,

    Y . Nishimura, H. Onoe, Y . Morichika, S. Perfiliev, H. Tsukada, and T. Isa, “Time-dependent central compensatory mechanisms of finger dexterity after spinal cord injury,”Science, vol. 318, no. 5853, pp. 1150–1155, 2007

    2007

  2. [9]

    Adaptive model-based myoelectric control for a soft wearable arm exosuit: A new generation of wearable robot control,

    N. Lotti, M. Xiloyannis, G. Durandau, E. Galofaro, V . Sanguineti, L. Masia, and M. Sartori, “Adaptive model-based myoelectric control for a soft wearable arm exosuit: A new generation of wearable robot control,”IEEE Robotics & Automation Magazine, vol. 27, no. 1, pp. 43–53, 2020

    2020

  3. [10]

    Real-time assistive control via imu locomotion mode detection in a soft exosuit: An effective approach to enhance walking metabolic efficiency,

    X. Zhang, E. Tricomi, F. Missiroli, N. Lotti, and L. Masia, “Real-time assistive control via imu locomotion mode detection in a soft exosuit: An effective approach to enhance walking metabolic efficiency,” IEEE/ASME Transactions on Mechatronics, 2023

    2023

  4. [11]

    Data-driven predictive control for robust exoskeleton locomotion,

    K. Li, J. Kim, X. Xiong, K. A. Hamed, Y . Yue, and A. D. Ames, “Data-driven predictive control for robust exoskeleton locomotion,” arXiv preprint arXiv:2403.15658, 2024

    2024 arXiv

  5. [12]

    A wearable gait lab powered by sensor-driven digital twins for quan- titative biomechanical analysis post-stroke,

    D. Simonetti, M. Hendriks, B. Koopman, N. Keijsers, and M. Sartori, “A wearable gait lab powered by sensor-driven digital twins for quan- titative biomechanical analysis post-stroke,”Wearable Technologies, vol. 5, p. e13, 2024

    2024

  6. [13]

    Task-agnostic exoskeleton control via biological joint moment estimation,

    D. D. Molinaro, K. L. Scherpereel, E. B. Schonhaut, G. Evangelopou- los, M. K. Shepherd, and A. J. Young, “Task-agnostic exoskeleton control via biological joint moment estimation,”Nature, vol. 635, no. 8038, pp. 337–344, 2024

    2024

  7. [14]

    Soft robotic shorts improve outdoor walking efficiency in older adults,

    E. Tricomi, F. Missiroli, M. Xiloyannis, N. Lotti, X. Zhang, M. Ste- fanakis, M. Theisen, J. Bauer, C. Becker, and L. Masia, “Soft robotic shorts improve outdoor walking efficiency in older adults,”Nature Machine Intelligence, pp. 1–11, 2024

    2024

  8. [15]

    Shared autonomy locomotion synthesis with a virtual powered prosthetic ankle,

    B. K. Hodossy and D. Farina, “Shared autonomy locomotion synthesis with a virtual powered prosthetic ankle,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 4738–4748, 2023

    2023

  9. [16]

    Experiment-free exoskeleton assistance via learning in simulation,

    S. Luo, M. Jiang, S. Zhang, J. Zhu, S. Yu, I. Dominguez Silva, T. Wang, E. Rouse, B. Zhou, H. Yuk,et al., “Experiment-free exoskeleton assistance via learning in simulation,”Nature, vol. 630, no. 8016, pp. 353–359, 2024

    2024

  10. [17]

    ViViDex: Learn- ing vision-based dexterous manipulation from human videos,

    Z. Chen, S. Chen, C. Schmid, and I. Laptev, “ViViDex: Learn- ing vision-based dexterous manipulation from human videos,” in arXiv:2404.15709, 2024

    2024 arXiv

  11. [18]

    DexYCB: A benchmark for capturing hand grasping of objects,

    Y .-W. Chao, W. Yang, Y . Xiang, P. Molchanov, A. Handa, J. Tremblay, Y . S. Narang, K. Van Wyk, U. Iqbal, S. Birchfield, J. Kautz, and D. Fox, “DexYCB: A benchmark for capturing hand grasping of objects,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  12. [19]

    Myosuite: A contact-rich simulation suite for musculoskeletal motor control,

    V . Caggiano, H. Wang, G. Durandau, M. Sartori, and V . Kumar, “Myosuite: A contact-rich simulation suite for musculoskeletal motor control,” inProceedings of The 4th Annual Learning for Dynamics and Control Conference(R. Firoozi, N. Mehr, E. Yel, R. Antonova, J. Bohg, M. Schw...

    2022

  13. [21]

    GRAB: A dataset of whole-body human grasping of objects,

    O. Taheri, N. Ghorbani, M. J. Black, and D. Tzionas, “GRAB: A dataset of whole-body human grasping of objects,” inEuropean Conference on Computer Vision (ECCV), 2020

    2020

  14. [22]

    Hoi4d: A 4d egocentric dataset for category- level human-object interaction,

    Y . Liu, Y . Liu, C. Jiang, K. Lyu, W. Wan, H. Shen, B. Liang, Z. Fu, H. Wang, and L. Yi, “Hoi4d: A 4d egocentric dataset for category- level human-object interaction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21013–21022,...

    2022

  15. [23]

    Learning joint reconstruction of hands and ma- nipulated objects,

    Y . Hasson, G. Varol, D. Tzionas, I. Kalevatykh, M. J. Black, I. Laptev, and C. Schmid, “Learning joint reconstruction of hands and ma- nipulated objects,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11807–11816, 2019

    2019

  16. [24]

    Learning video-conditioned policies for unseen manipulation tasks,

    E. Chane-Sane, C. Schmid, and I. Laptev, “Learning video-conditioned policies for unseen manipulation tasks,” in2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 909–916, IEEE, 2023

    2023

  17. [25]

    Learning dexterous manipulation from exemplar object trajectories and pre-grasps,

    S. Dasari, A. Gupta, and V . Kumar, “Learning dexterous manipulation from exemplar object trajectories and pre-grasps,” in2023 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp. 3889– 3896, 2023

    2023

  18. [26]

    Myodex: a generalizable prior for dexterous manipulation,

    V . Caggiano, S. Dasari, and V . Kumar, “Myodex: a generalizable prior for dexterous manipulation,” inInternational Conference on Machine Learning, pp. 3327–3346, PMLR, 2023

    2023

  19. [27]

    Dexmv: Imitation learning for dexterous manipulation from human videos,

    Y . Qin, Y .-H. Wu, S. Liu, H. Jiang, R. Yang, Y . Fu, and X. Wang, “Dexmv: Imitation learning for dexterous manipulation from human videos,” inEuropean Conference on Computer Vision, pp. 570–587, Springer, 2022

    2022

  20. [28]

    Dexvip: Learning dexterous grasping with human hand pose priors from video,

    P. Mandikal and K. Grauman, “Dexvip: Learning dexterous grasping with human hand pose priors from video,” inConference on Robot Learning, pp. 651–661, PMLR, 2022

    2022

  21. [29]

    Videodex: Learning dexterity from internet videos,

    K. Shaw, S. Bahl, and D. Pathak, “Videodex: Learning dexterity from internet videos,” inConference on Robot Learning, pp. 654–665, PMLR, 2023

    2023

  22. [30]

    Unidexgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learn- ing,

    W. Wan, H. Geng, Y . Liu, Z. Shan, Y . Yang, L. Yi, and H. Wang, “Unidexgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learn- ing,” inProceedings of the IEEE/CVF International Conference on Computer Visio...

    2023

  23. [31]

    Human-in-the-loop control of soft exosuits using impedance learning on different terrains,

    Z. Li, X. Li, Q. Li, H. Su, Z. Kan, and W. He, “Human-in-the-loop control of soft exosuits using impedance learning on different terrains,” IEEE Transactions on Robotics, vol. 38, no. 5, pp. 2979–2993, 2022

    2022

  24. [32]

    Wearable robots for the real world need vision,

    L. Gionfrida, D. Kim, D. Scaramuzza, D. Farina, and R. D. Howe, “Wearable robots for the real world need vision,”Science Robotics, vol. 9, no. 90, p. eadj8812, 2024

    2024

  25. [33]

    Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view,

    D. Kim, B. B. Kang, K. B. Kim, H. Choi, J. Ha, K.-J. Cho, and S. Jo, “Eyes are faster than hands: A soft wearable robot learns user intention from the egocentric view,”Science Robotics, vol. 4, no. 26, p. eaav2949, 2019

    2019

  26. [34]

    The ycb object and model set: Towards common benchmarks for manipulation research,

    B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar, “The ycb object and model set: Towards common benchmarks for manipulation research,” in2015 international conference on ad- vanced robotics (ICAR), pp. 510–517, IEEE, 2015

    2015

  27. [35]

    The grasp taxonomy of human grasp types,

    T. Feix, J. Romero, H.-B. Schmiedmayer, A. M. Dollar, and D. Kragic, “The grasp taxonomy of human grasp types,”IEEE Transactions on human-machine systems, vol. 46, no. 1, pp. 66–77, 2015

    2015

  28. [36]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012

    2012

  29. [37]

    Adult norms for the box and block test of manual dexterity,

    V . Mathiowetz, G. V olland, N. Kashman, and K. Weber, “Adult norms for the box and block test of manual dexterity,”The American journal of occupational therapy, vol. 39, no. 6, pp. 386–391, 1985

    1985

  30. [38]

    A standardized approach to performing the action research arm test,

    N. Yozbatiran, L. Der-Yeghiaian, and S. C. Cramer, “A standardized approach to performing the action research arm test,”Neurorehabili- tation and neural repair, vol. 22, no. 1, pp. 78–90, 2008

    2008

  31. [39]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”Advances in Neural Informa- tion Processing Systems, 2017

    2017

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.