Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Dense tactile signals and motor-current torque estimates let multi-finger grasp-force and reorientation policies train entirely in simulation and run zero-shot on real hardware.

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

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

T0 review · grok-4.5

2026-07-11 11:09 UTC pith:4KLLHYIS

load-bearing objection Solid systems package for zero-shot force-command grasping and in-hand rotation on a multi-finger hand; the inverted-catch training distribution is the main caveat on the headline claim. the 3 major comments →

arxiv 2607.04940 v1 pith:4KLLHYIS submitted 2026-07-06 cs.RO

Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulation

Zhe Zhao , Zhibin Li , Yilin Ou , Mengshi Qi This is my paper
classification cs.RO
keywords sim-to-realdexterous manipulationtactile sensingforce controlin-hand reorientationactuator modelingreinforcement learningmulti-finger hands
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Training multi-finger robot hands to grasp with commanded force and to reorient objects in the palm is hard because real contact physics and imperfect motors do not match simulation. This paper shows that three practical ingredients close that gap: a fast distance-based approximation of dense fingertip touch, a calibration that turns ordinary motor current into usable joint-torque estimates, and randomized actuator models that capture torque-speed limits and backlash. Policies are trained only in simulation with an asymmetric actor-critic method and then deployed directly on a five-finger hand. The resulting controllers track commanded grip strength across objects and keep rotating a cube without any real-world fine-tuning. The work therefore supplies a concrete, sensor-and-actuator-aware recipe for force-aware dexterous skills that previously required either real-robot data or specialized torque sensors.

Core claim

A full-state reinforcement-learning policy that observes dense tactile contact centers/forces together with joint torques (derived from motor current) can be trained entirely in simulation and transferred zero-shot to a real five-finger hand, producing both command-based controllable grasp-force tracking and continuous in-hand object reorientation—the first such demonstration of controllable grasping on a multi-finger dexterous hand.

What carries the argument

Parallel-forward-kinematics distance fields that turn dense virtual tactile units into high-rate contact forces and force-weighted centers (approximated by a Mooney-Rivlin stress model), paired with a quasi-static current-to-torque map and episode-wise randomization of PD gains, backlash deadbands, and torque-speed envelopes; these three elements jointly supply the observation space and dynamics that let the policy learn force regulation transferable to hardware.

Load-bearing premise

The distance-based virtual touch sensors, the simple current-to-torque curve, and the randomized motor model together match real contact forces and actuator behavior closely enough that a policy trained under inverted catch-and-hold never needs real-world fine-tuning.

What would settle it

Measure real fingertip contact forces and joint currents while the zero-shot policy executes commanded force levels and cube reorientations; if the force-tracking error systematically exceeds the simulated ranges or the hand drops the object far more often than in simulation, the fidelity claim fails.

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

If this is right

  • Commercial multi-finger hands that report only motor current can obtain usable torque feedback for force-aware policies without adding joint torque sensors.
  • Force-commanded grasping becomes a trainable skill rather than a hand-tuned impedance controller, enabling grip strength to be set by a single scalar input.
  • In-hand reorientation policies that rely on tactile centers and forces rather than vision can run at high rate and remain robust under lighting or occlusion changes.
  • The same simulation stack can be reused for other contact-rich finger skills once the actuator randomization ranges are identified for a new hand.

Where Pith is reading between the lines

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

  • If the distance-field tactile approximation generalizes, similar lightweight contact models could replace expensive soft-body or FEM tactile simulators for other high-DoF manipulators.
  • The inverted catch-and-hold training setup may be a transferable curriculum for any under-actuated or high-friction hand where free-space exploration is wasteful.
  • Once force tracking is reliable, the same observation stack could support gentler tasks such as handling fragile or deformable objects by simply lowering the force command.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents a zero-shot sim-to-real RL pipeline for a 12-DoF five-finger hand (XHand) that uses dense tactile and joint-torque/current observations to learn two contact-rich skills: command-based controllable grasp-force tracking and in-hand object reorientation. To close the reality gap it contributes (i) a fast tactile approximation that places dense virtual tactile units on the fingertips and detects contact via parallel forward-kinematics distances plus a Mooney-Rivlin stress model (Eqs. 11–14, Fig. 3), (ii) a quasi-static current-to-torque calibration that maps motor current to simulated joint torque without hardware torque sensors (Fig. 2), and (iii) a randomized actuator model that includes PD gains, backlash dead-band and torque–speed saturation (Eqs. 15–18). Policies are trained with asymmetric actor–critic PPO under an inverted “catch-the-object” regime for grasping and continuous 90° cube rotation for reorientation, then deployed zero-shot. Real-robot ablations (Tables IV–V, Figs. 4–9) show that tactile contact centers/forces, 6-D orientation, and both force and torque rewards improve consecutive successes and force-command tracking.

Significance. If the zero-shot force-tracking claim holds under ordinary contact conditions, the work supplies a practical, sensor-light recipe for multi-finger force-aware manipulation that many commercial hands can adopt. The combination of a computationally cheap dense-tactile surrogate, current-based torque estimation, and actuator randomization is a concrete engineering contribution that other groups can re-implement. Real-hardware demonstrations of both force-modulated grasps on unseen objects and continuous in-hand reorientation, together with systematic observation and reward ablations, constitute useful empirical evidence for the value of full-state tactile/torque feedback. The inverted-catch training distribution and quasi-static calibration, however, leave open how far the headline “first controllable grasping” claim generalizes beyond the specialized regime used for training.

major comments (3)
  1. Sec. III-C.1 and Fig. 7: the force-adaptive grasping policy is trained exclusively under an inverted catch-and-hold regime in which objects are dropped into a fixed palm-up hand. The policy never approaches, closes around, or stabilizes an object resting on a surface. Consequently the zero-shot force-command tracking results (Figs. 5–6, Table V) and the abstract’s “first demonstration of controllable grasping” claim are demonstrated only for a contact distribution that is statistically different from ordinary table-top grasping. The manuscript should either (a) report real-robot force-tracking trials under a conventional approach-and-grasp protocol or (b) explicitly scope the claim to the inverted-catch setting.
  2. Fig. 2 and Sec. III-E.3: the current-to-torque map is obtained under quasi-static fingertip loading. No dynamic validation (varying velocity, simultaneous multi-finger contact, or free-space motion) is provided. Because the same map is used both to align simulation torque with real current and as a policy observation, any velocity-dependent discrepancy directly affects the claimed force controllability. A short dynamic calibration or residual-error plot would strengthen the load-bearing transfer argument.
  3. Tables IV–V and the real-world evaluation protocol: each ablation condition is evaluated on only ten trials, reported as sorted consecutive-success counts without standard errors or confidence intervals. With such small N the ranking of observation components (force-weighted center, contact force, 6-D orientation) is suggestive but not statistically robust. Reporting mean ± std or bootstrap intervals, or increasing the number of trials, is needed before the ablation conclusions can be treated as definitive.
minor comments (5)
  1. Abstract and Introduction: the phrase “to our knowledge, this is the first demonstration…” should be tempered or supported by a short related-work comparison that explicitly distinguishes prior force-aware or tactile in-hand results (e.g., torque-controlled tactile reorientation, binary-touch rotation).
  2. Eqs. (2)–(6) and (15)–(18): several free parameters (Mooney-Rivlin coefficients, reward weights, actuator randomization ranges) are stated without the numerical intervals used in training; listing them in a table or appendix would aid reproducibility.
  3. Fig. 4 caption and surrounding text: “object selection task” appears to be a residual phrase; the figure actually shows contact traces during in-hand rotation.
  4. Table I vs. real-world sensing: object Z is estimated via SAM mask distance while X/Y are fixed; the resulting observation noise is never quantified, yet the policy is claimed to transfer zero-shot. A brief noise characterization would clarify the perceptual gap.
  5. Typographical inconsistencies: “sim-to-real reinforcement learning that leverages” (missing “method”), “robustin-hand”, “full-state feedback—tactile”, and mixed “XHand/Xhand” spelling.

Circularity Check

0 steps flagged

No circularity: empirical sim-to-real systems paper whose force-tracking and reorientation claims are hardware-validated, not true by construction of calibrations or rewards.

full rationale

The paper is an engineering/systems contribution that trains PPO policies in IsaacLab under an inverted catch-and-hold regime, using a distance-field tactile approximation (Eqs. 11–14), a quasi-static current-to-torque map fitted from fingertip loading data (Fig. 2), Mooney-Rivlin parameters fitted to the sensor rubber (Fig. 3b), and a randomized PD/backlash/torque-speed actuator model (Eqs. 15–18). These are standard domain-randomization and calibration steps; none of them is then re-labeled as a first-principles prediction of the same quantities. The load-bearing claims—commanded grasp-force tracking (Figs. 5–6, Tab. V) and zero-shot in-hand reorientation (Tab. IV, Fig. 8)—are evaluated by direct deployment on the physical XHand and are therefore falsifiable by hardware failure. Related-work self-citations (Qi et al. on video/pose tasks) are explicitly scoped as non-central and do not underwrite any uniqueness or derivation step. No equation reduces to its own fitted input by definition, and no uniqueness theorem is imported from the authors. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on engineering approximations and fitted maps rather than free physical constants. Load-bearing pieces are the virtual tactile distance model, Mooney-Rivlin coefficients for the fingertip rubber, the empirical current-to-torque/force alignment, randomized actuator envelope parameters, and numerous reward weights. These are domain modeling choices and free parameters that make zero-shot transfer possible in the authors' setup; they are not derived from first principles.

free parameters (5)
  • Mooney-Rivlin C1,C2,C3 for tactile rubber
    Stress-strain coefficients used to map penetration distance to simulated contact force; given as C1=-0.03215, C2=0.14808, C3=0.00217 MPa without independent validation protocol in the text.
  • Current-to-torque / force normalization (I_max, τ_max, F_max)
    Quasi-static calibration that aligns real motor current with simulated joint torque and contact force onto [0,1]; fitted from scatter data in Fig. 2 and required for policy observations.
  • Actuator randomization ranges (kp, kd, backlash ε, stall torque τ0, η, q̇_max)
    Resampled each episode; values and ranges are not fully tabulated but are essential to the claimed sim-to-real bridge.
  • Reward weights (w_torque, w_force, w_diff, w_outter, w_action, w_vel, σ=0.10)
    Hand-chosen scalars that shape force tracking and finger coordination; ablation shows they change force/torque ranges, so the demonstrated skill depends on them.
  • Force command scale and contact force validity bounds (0.01–200, thumb torque 0.01–1.1)
    Indicator thresholds and command range [0,1] define what 'controllable grasp force' means in training and evaluation.
axioms (5)
  • ad hoc to paper Penetration of virtual tactile points (v_oij · n_o < 0) plus distance-weighted Mooney-Rivlin stress adequately approximates real dense tactile force and contact centers for RL transfer.
    Core tactile simulation assumption in §III.E.1–2 and Fig. 3–4; fidelity is argued by qualitative real/sim contact plots, not a formal error bound.
  • domain assumption Motor current under quasi-static loading is a sufficiently linear proxy for joint torque and contact force on the XHand for policy feedback.
    Standard in current-based force estimation literature; instantiated via the paper's calibration (§III.E.3, Fig. 2).
  • domain assumption Randomizing PD gains, backlash deadband, and torque-speed saturation covers the dominant unmodeled actuator discrepancies for zero-shot transfer.
    Actuator model §III.E.4; common domain-randomization premise, not proven complete for all hardware variation.
  • ad hoc to paper Inverted catch-the-object training distribution yields policies that remain valid for real grasping and force modulation of varied objects.
    Task setup §III.C.1 deliberately simplifies exploration; generalization is shown on several objects but is a structural premise of the evaluation.
  • standard math Asymmetric actor-critic PPO with the listed actor/critic observation splits is a valid training algorithm for these continuous control skills.
    Standard RL practice; not novel math, but assumed to produce the reported policies.
invented entities (2)
  • Dense virtual tactile units with parallel-FK distance-field contact detection no independent evidence
    purpose: Provide high-rate, high-resolution touch signals (600 units) cheap enough for massive RL without full soft-body FEM.
    Engineering construct specific to this pipeline; independent evidence is limited to qualitative real/sim contact alignment (Fig. 4), not external sensor benchmarks.
  • Four-finger consistency / outer-joint penalty rewards for human-like coordination no independent evidence
    purpose: Encourage correlated flexion of homologous fingers to avoid unnatural independent motor postures.
    Reward design choice introduced for this hand morphology; not an external physical entity.

pith-pipeline@v1.1.0-grok45 · 18643 in / 3836 out tokens · 34638 ms · 2026-07-11T11:09:58.332525+00:00 · methodology

0 comments
read the original abstract

Human-like dexterous hands with multiple fingers offer human-level manipulation capabilities but remain difficult to train the control policies that can deploy on real hardware due to contact-rich physics and imperfect actuation. We present a sim-to-real reinforcement learning method that leverages dense tactile feedback combined with joint torque sensing to explicitly regulate physical interactions. To enable effective sim-to-real transfer, we introduce (i) a computationally fast tactile simulation that computes distances between dense virtual tactile units and the object via parallel forward kinematics, providing high-rate, high-resolution touch signals needed by RL; (ii) a current-to-torque calibration that eliminates the need for torque sensors on dexterous hands by mapping motor current to joint torque; and (iii) actuator dynamics modeling with randomization to account for non-ideal torque-speed effects and bridge the actuation gaps. Using an asymmetric actor-critic PPO pipeline, we train policies entirely in simulation and deploy them directly to a five-finger hand. The resulting policies demonstrate two essential human-hand skills: (1) command-based controllable grasp force tracking and (2) reorientation of objects in the hand, both of which are robustly executed without fine-tuning on the robot. By combining tactile and torque in the observation space with scalable sensing and actuation modeling, our system provides a practical solution to achieve reliable dexterous manipulation. To our knowledge, this is the first demonstration of controllable grasping on a multi-finger dexterous hand trained entirely in simulation and transferred zero-shot on real hardware.

Figures

Figures reproduced from arXiv: 2607.04940 by Mengshi Qi, Yilin Ou, Zhe Zhao, Zhibin Li.

Figure 1
Figure 1. Figure 1: Learning full-state policy with tactile sensing and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Calibration and alignment of current-force (real robot) versus torque-force (simulation) properties. Once the fingers are in contact with the object, we employ torque command rewards Rtorque to encourage the agent to apply specific joint torques. The target torque τtarget is a con￾tinuous value provided by the instruction of the environment, and the reward function has a special handling for the thumb comp… view at source ↗
Figure 3
Figure 3. Figure 3: Contact point modeling and material properties. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of real-world and simulated contact data [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Grasping objects with controllable magnitudes of grasping forces – from low to high strength. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Varying joint torque and contact forces with force [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization results of force-adaptive grasping tasks in Real-world and simulation environments [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization results of in-hand manipulation tasks in real-world and simulation environments [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Force-adaptive grasping of irregularly shaped objects [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

45 extracted references · 7 linked inside Pith

  1. [1]

    Learning dexterous in-hand manipulation,

    OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. J ´ozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba, “Learning dexterous in-hand manipulation,”arXiv preprint arXiv:1808.00177, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    Solving rubik’s cube with a robot hand,

    I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng, Q. Yuan, W. Zaremba, L. Zhang, and OpenAI, “Solving rubik’s cube with a robot hand,” arXiv preprint arXiv:1910.07113, 2019

    Pith/arXiv arXiv 1910

  3. [3]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” inProc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), Vancouver, Canada, 2017

    2017

  4. [4]

    Dextreme: Transfer of agile in-hand manipulation from simulation to reality,

    A. Handa, A. Allshire, V . Makoviychuk, A. Petrenko, R. Singh, J. Liu, D. Makoviichuk, K. Van Wyk, A. Zhurkevich, B. Sundaralingam, Y . Narang, J.-F. Lafleche, D. Fox, and G. State, “Dextreme: Transfer of agile in-hand manipulation from simulation to reality,” inProc. IEEE Int. Conf. on Robotics and Automation (ICRA), 2023, see also arXiv:2210.13702

    Pith/arXiv arXiv 2023

  5. [5]

    Robot synesthesia: In-hand manipulation with visuotactile sensing,

    Y . Yuan, H. Che, Y . Qin, B. Huang, Z.-H. Yin, K.-W. Lee, Y . Wu, S.-C. Lim, and X. Wang, “Robot synesthesia: In-hand manipulation with visuotactile sensing,”arXiv preprint arXiv:2312.01853, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Rotating without seeing: Towards in-hand dexterity through touch,

    Z.-H. Yin, B. Huang, Y . Qin, Q. Chen, and X. Wang, “Rotating without seeing: Towards in-hand dexterity through touch,” inRobotics: Science and Systems (RSS), Daegu, Republic of Korea, 2023

    2023

  7. [7]

    Robustdex- grasp: Robust dexterous grasping of general objects from single-view perception,

    H. Zhang, Z. Wu, L. Huang, S. Christen, and J. Song, “Robustdex- grasp: Robust dexterous grasping of general objects from single-view perception,”arXiv preprint arXiv:2504.05287, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Sim-to-real reinforcement learning for vision-based dexterous manipulation on humanoids,

    T. Lin, K. Sachdev, L. Fan, J. Malik, and Y . Zhu, “Sim-to-real reinforcement learning for vision-based dexterous manipulation on humanoids,”arXiv preprint arXiv:2502.20396, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Dexterous in-hand manipulation of slender cylindrical objects through deep reinforcement learning with tactile sensing,

    W. Hu, B. Huang, W. W. Lee, S. Yang, Y . Zheng, and Z. Li, “Dexterous in-hand manipulation of slender cylindrical objects through deep reinforcement learning with tactile sensing,”Robotics and Autonomous Systems, vol. 186, p. 104904, 2025

    2025

  10. [10]

    DC- SAM: In-context segment anything in images and videos via dual consistency,

    M. Qi, P. Zhu, X. Li, X. Bi, L. Qi, H. Ma, and M.-H. Yang, “DC- SAM: In-context segment anything in images and videos via dual consistency,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 4, pp. 4642–4656, 2026

    2026

  11. [11]

    Robust disentangled counterfactual learning for physical audiovisual commonsense reasoning,

    M. Qi, C. Lv, and H. Ma, “Robust disentangled counterfactual learning for physical audiovisual commonsense reasoning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 3, pp. 2514– 2527, 2026

    2026

  12. [12]

    Explainable action form assessment by exploiting multimodal chain-of-thoughts reasoning,

    M. Qi, Y . Wu, W. Yun, X. Zhang, and H. Ma, “Explainable action form assessment by exploiting multimodal chain-of-thoughts reasoning,” IEEE Transactions on Image Processing, vol. 35, pp. 6371–6384, 2026

    2026

  13. [13]

    Action quality assessment via hierarchical pose-guided multi-stage contrastive regression,

    M. Qi, H. Ye, J. Peng, and H. Ma, “Action quality assessment via hierarchical pose-guided multi-stage contrastive regression,”IEEE Transactions on Image Processing, vol. 34, pp. 6461–6474, 2025

    2025

  14. [14]

    Towards balanced multi- modal learning in 3D human pose estimation,

    M. Qi, J. Peng, X. Zhang, and H. Ma, “Towards balanced multi- modal learning in 3D human pose estimation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 21 231–21 241

    2026

  15. [15]

    Weakly-supervised temporal action localization by inferring salient snippet-feature,

    W. Yun, M. Qi, C. Wang, and H. Ma, “Weakly-supervised temporal action localization by inferring salient snippet-feature,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, pp. 6908–6916, 2024

    2024

  16. [16]

    Semantics-aware spatial- temporal binaries for cross-modal video retrieval,

    M. Qi, J. Qin, Y . Yang, Y . Wang, and J. Luo, “Semantics-aware spatial- temporal binaries for cross-modal video retrieval,”IEEE Transactions on Image Processing, vol. 30, pp. 2989–3004, 2021

    2021

  17. [17]

    Few- shot ensemble learning for video classification with SlowFast memory networks,

    M. Qi, J. Qin, X. Zhen, D. Huang, Y . Yang, and J. Luo, “Few- shot ensemble learning for video classification with SlowFast memory networks,” inProceedings of the 28th ACM International Conference on Multimedia, 2020, pp. 3007–3015

    2020

  18. [18]

    Sports video captioning via attentive motion representation and group relationship modeling,

    M. Qi, Y . Wang, A. Li, and J. Luo, “Sports video captioning via attentive motion representation and group relationship modeling,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 8, pp. 2617–2633, 2020

    2020

  19. [19]

    Unsupervised domain adaptation with temporal-consistent self-training for 3D hand-object joint reconstruction,

    M. Qi, E. Remelli, M. Salzmann, and P. Fua, “Unsupervised domain adaptation with temporal-consistent self-training for 3D hand-object joint reconstruction,” 2020

    2020

  20. [20]

    Decomposed vector-quantized variational autoencoder for human grasp generation,

    Z. Zhao, M. Qi, and H. Ma, “Decomposed vector-quantized variational autoencoder for human grasp generation,” inComputer Vision – ECCV

  21. [21]

    Springer Nature Switzerland, 2024, pp. 447–463

    2024

  22. [22]

    Human grasp generation for rigid and deformable objects with decomposed VQ-V AE,

    M. Qi, Z. Zhao, and H. Ma, “Human grasp generation for rigid and deformable objects with decomposed VQ-V AE,” 2025

    2025

  23. [23]

    Tacsl: A library for visuotactile sensor simulation and learning,

    I. Akinola, J. Xu, J. Carius, D. Fox, and Y . Narang, “Tacsl: A library for visuotactile sensor simulation and learning,”IEEE Transactions on Robotics, vol. 41, p. 2645–2661, 2025

    2025

  24. [24]

    Tacex: Gelsight tactile simulation in isaac sim – combining soft-body and visuotactile simulators,

    D. H. Nguyen, T. Schneider, G. Duret, A. Kshirsagar, B. Belousov, and J. Peters, “Tacex: Gelsight tactile simulation in isaac sim – combining soft-body and visuotactile simulators,” 2024

    2024

  25. [25]

    Tacflex: Multimode tactile imprints simulation for visuotactile sensors with coating patterns,

    C. Zhang, S. Cui, J. Hu, T. Jiang, T. Zhang, R. Wang, and S. Wang, “Tacflex: Multimode tactile imprints simulation for visuotactile sensors with coating patterns,”IEEE Transactions on Robotics, vol. 41, pp. 3965–3985, 2025

    2025

  26. [26]

    General- purpose sim2real protocol for learning contact-rich manipulation with marker-based visuotactile sensors,

    W. Chen, J. Xu, F. Xiang, X. Yuan, H. Su, and R. Chen, “General- purpose sim2real protocol for learning contact-rich manipulation with marker-based visuotactile sensors,”IEEE Transactions on Robotics, vol. 40, pp. 1509–1526, 2024

    2024

  27. [27]

    3d vitac:learning fine-grained manipulation with visuo-tactile sensing,

    B. Huang, Y . Wang, X. Yang, Y . Luo, and Y . Li, “3d vitac:learning fine-grained manipulation with visuo-tactile sensing,” inProceedings of Robotics: Conference on Robot Learning(CoRL), 2024

    2024

  28. [28]

    Robot synesthesia: In-hand manipulation with visuotactile sensing,

    Y . Yuan, H. Che, Y . Qin, B. Huang, Z.-H. Yin, K.-W. Lee, Y . Wu, S.-C. Lim, and X. Wang, “Robot synesthesia: In-hand manipulation with visuotactile sensing,” 2024

    2024

  29. [29]

    Learning purely tactile in-hand ma- nipulation with a torque-controlled hand,

    L. Sievers, J. Pitz, and B. B ¨auml, “Learning purely tactile in-hand ma- nipulation with a torque-controlled hand,” inInternational Conference on Robotics and Automation, 2022, pp. 2745–2751

    2022

  30. [30]

    Dexterous in-hand manipulation of slender cylindrical objects through deep reinforcement learning with tactile sensing,

    W. Hu, B. Huang, W. W. Lee, S. Yang, Y . Zheng, and Z. Li, “Dexterous in-hand manipulation of slender cylindrical objects through deep reinforcement learning with tactile sensing,” p. 104904, 2025

    2025

  31. [31]

    Bi-touch: Bimanual tactile manipulation with sim-to-real deep reinforcement learning,

    Y . Lin, A. Church, M. Yang, H. Li, J. Lloyd, D. Zhang, and N. F. Lepora, “Bi-touch: Bimanual tactile manipulation with sim-to-real deep reinforcement learning,”IEEE Robotics and Automation Letters, vol. 8, no. 9, pp. 5472–5479, 2023

    2023

  32. [32]

    Fungrasp: Functional grasping for diverse dexterous hands,

    L. Huang, H. Zhang, Z. Wu, S. Christen, and J. Song, “Fungrasp: Functional grasping for diverse dexterous hands,”IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 6175–6182, 2025

    2025

  33. [33]

    Re- inforcement learning for non-prehensile manipulation: Transfer from simulation to physical system,

    K. Lowrey, S. Kolev, J. Dao, A. Rajeswaran, and E. Todorov, “Re- inforcement learning for non-prehensile manipulation: Transfer from simulation to physical system,”2018 IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR), pp. 35–42, 2018

    2018

  34. [34]

    Actuator-constrained reinforcement learning for high-speed quadrupedal locomotion,

    Y .-H. Shin, T.-G. Song, G. Ji, and H.-W. Park, “Actuator-constrained reinforcement learning for high-speed quadrupedal locomotion,” 2023

    2023

  35. [35]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbabu, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. J. Fan, Y . Zhu, C. Liu, and G. Shi, “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,”arXiv preprint arXiv:2502.01143, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Bridging the sim-to-real gap for athletic loco-manipulation,

    N. Fey, G. B. Margolis, M. Peticco, and P. Agrawal, “Bridging the sim-to-real gap for athletic loco-manipulation,” 2025

    2025

  37. [37]

    Motor-current-based estimation of cartesian contact forces and torques for robotic manipulators and its application to force control,

    A. Wahrburg, J. B ¨os, K. D. Listmann, F. Dai, B. Matthias, and H. Ding, “Motor-current-based estimation of cartesian contact forces and torques for robotic manipulators and its application to force control,”IEEE Transactions on Automation Science and Engineering, vol. 15, no. 2, pp. 879–886, 2018

    2018

  38. [38]

    Toward sensorless interaction force estimation for industrial robots using high-order finite-time observers,

    L. Han, J. Mao, P. Cao, Y . Gan, and S. Li, “Toward sensorless interaction force estimation for industrial robots using high-order finite-time observers,”IEEE Transactions on Industrial Electronics, vol. 69, no. 7, pp. 7275–7284, 2022

    2022

  39. [39]

    External torque estimation for an industrial robot arm using joint torsion and motor current measurements,

    T. Gold, A. V ¨olz, and K. Graichen, “External torque estimation for an industrial robot arm using joint torsion and motor current measurements,”IFAC-PapersOnLine, vol. 52, pp. 352–357, 01 2019

    2019

  40. [40]

    Current- based impedance control for interacting with mobile manipulators,

    J. de Wolde, L. Knoedler, G. Garofalo, and J. Alonso-Mora, “Current- based impedance control for interacting with mobile manipulators,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 753–760

    2024

  41. [41]

    Global identification of joint drive gains and dynamic parameters of parallel robots,

    S. Briot and M. Gautier, “Global identification of joint drive gains and dynamic parameters of parallel robots,”Multibody System Dynamics, vol. 136, 01 2013

    2013

  42. [42]

    Xhand product page,

    Generic, “Xhand product page,” https://www.robotera.com/en/goods1/ 4.html

  43. [43]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automa- tion Letters, vol. 8, no. 6, pp. 3740–3747, 2023

    2023

  44. [44]

    Segment anything,

    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, 2023, pp. 4015–4026

    2023

  45. [45]

    On the continuity of rotation representations in neural networks,

    Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li, “On the continuity of rotation representations in neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 5745–5753

    2019