Pith. sign in

REVIEW 4 major objections 5 minor 32 references

A tactile-only policy, TacRefineNet, refines in-hand object pose to roughly 5 mm and 3.5 degrees using only fingertip touch, with no vision.

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 · deepseek-v4-flash

2026-08-04 13:40 UTC pith:ZCZFWZT2

load-bearing objection Useful 4-DoF tactile regrasp system with a neat cross-combination training idea, but the 6-DoF/arbitrary-pose claims outrun what was actually trained and tested. the 4 major comments →

arxiv 2509.25746 v2 pith:ZCZFWZT2 submitted 2025-09-30 cs.RO

TacRefineNet: Goal-Conditioned Tactile Grasp Refinement for Edge-Prominent Objects

Shuaijun Wang , Haoran Zhou , Diyun Xiang , Yangwei You This is my paper
classification cs.RO
keywords tactile sensingdexterous manipulationgrasp refinementin-hand posegoal-conditioned policysim-to-realtactile servoingedge-prominent objects
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.

This paper tries to establish that a dexterous hand can correct an in-hand object's pose using only fingertip touch, without vision. Its policy, TacRefineNet, takes current and target tactile images plus finger joint angles and outputs a wrist motion; the hand repeatedly opens, moves, and regrasps until the touch pattern matches the target. If correct, this closes the "last-mile" of grasping—final pose errors—for thin, flat, edge-prominent objects like plates, discs, and rods, with mean errors near 5 mm and 3.5 degrees. The authors also claim this is the first tactile-only method for arbitrary in-hand pose adjustment for known objects.

Core claim

In the paper's terms, TacRefineNet is a tactile-only, goal-conditioned policy that regresses a 6-DoF wrist pose increment from multi-finger tactile images and hand proprioception. Trained with cross-combination sampling—randomly pairing current and target tactile images—it can steer an 11-DoF hand to any in-hand target within the sampled pose range without retraining. Real-world tests on an 11-DoF five-fingered hand with piezoresistive fingertip arrays report about 5.2 mm and 3.5° mean error after five steps, success rates up to 80.7% under a 10°/10 mm criterion for fixed targets, and continuous correction during long-horizon perturbations. The authors note the system samples only four of si

What carries the argument

The central mechanism is an iterative external-dexterity tactile servoing loop: after each grasp, fingertip tactile images (11×9 taxel arrays) and joint angles are fed to a multi-branch Siamese-style network that fuses per-finger tactile encodings with positional encoding and proprioception, then a three-layer MLP predicts wrist pose increments. The hand reopens, moves, and regrasps until the current tactile image matches the target. Cross-combination training—randomly pairing current and target samples from an N×N set—provides goal conditioning without per-pose retraining, and a physics-based simulation of the tactile sensors supplies large-scale training data before fine-tuning on real dat

Load-bearing premise

The central claim stands or falls on the assumption that x-translation and yaw errors are unnecessary to control directly—that they are absent, negligible, or correctable through coupling with the four sampled dimensions pitch, roll, y, and z.

What would settle it

Run the trained policy with an initial pose perturbed only in x-translation (or only in yaw), with all four sampled dimensions already at target; if the policy never outputs a corrective increment and the pose error in that dimension does not shrink, then the arbitrary-6-DoF claim is falsified in practice. A simpler check is to inspect the training data: if it contains no variation along x or yaw, the ground-truth increments for those dimensions are always zero.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, tactile-only refinement can replace vision in final grasp alignment for thin/planar objects, reducing error accumulation in long-horizon tasks.
  • Cross-combination training means a single demonstrated target tactile image is enough to specify any reachable in-hand pose, enabling user-specified goals without retraining.
  • The sim-to-real recipe (large simulated dataset plus small real fine-tuning) is what makes the zero-shot deployment work; Policy B outperforms simulation-only training.
  • The system can track a moving object over time using only touch, suggesting application in dynamic grasping and assembly.
  • Generalization to unseen objects is partial and mostly along roll, so the method is currently limited to known objects or similar-geometry instances.

Where Pith is reading between the lines

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

  • Because the training samples omit x-translation and yaw, the "6-DoF arbitrary pose" claim is only as strong as the assumption that those errors are negligible or coupled to the sampled dimensions; a deployment requiring explicit x or yaw correction would likely need additional sensing.
  • The same cross-combination idea could extend to other pose-observable contact signatures, but for objects with symmetry or weak tactile discriminability, ambiguity grows—so tactile-only refinement is best suited to objects with distinctive contact geometry.
  • A natural extension is to fuse sparse vision or tactile history over multiple regrasps to resolve the redundant dimensions, or to train directly on all six DOFs with data augmentation that breaks the x/roll and yaw/other couplings.
  • If the approach scales to wider object classes, it could serve as a plug-in "final alignment" module inside larger manipulation pipelines, reducing the need for high-precision arm controllers.

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

4 major / 5 minor

Summary. The paper presents TacRefineNet, a tactile-only, goal-conditioned framework for refining the in-hand pose of flat, edge-prominent objects. A multi-branch Siamese-like policy takes current and target fingertip tactile images plus joint positions and outputs a wrist pose increment; the hand iteratively opens, moves, and regrasps to converge to the target. The policy is trained in MuJoCo with a physics-based tactile model (156,007 simulated samples) and cross-combination pairing of current/target poses, then optionally fine-tuned on real data (Policy B). Real-world experiments report millimeter-level position accuracy and ~0.02 rad orientation errors on seen objects, with limited generalization to unseen objects in the same category. The paper claims 6-DoF arbitrary in-hand pose refinement, zero-shot sim-to-real deployment, and being the first tactile-only method for arbitrary in-hand pose adjustment.

Significance. If the claims were fully supported, the paper would be a valuable contribution: a practical tactile-only servoing loop for high-precision regrasping, a large simulated tactile dataset, and a clean cross-combination training scheme that avoids per-pose retraining. The real-robot evaluation and the sim-only vs. fine-tuned comparison are concrete strengths. However, the headline claims are substantially over-broad relative to the evidence: the system is trained and evaluated only on four pose dimensions (pitch, roll, y, z), and the best real-world policy is fine-tuned on real data, not zero-shot. The contribution is better characterized as 4-DoF tactile refinement along tactilely observable dimensions, with sim-to-real transfer improved by a small amount of real fine-tuning.

major comments (4)
  1. [§III-C; Eq. (1); Contributions] Data collection samples only pitch, roll, y, and z because tactile images along x overlap with roll and images along yaw overlap with other axes (§III-C). Yet Eq. (1) outputs Δx∈R^6 and the first contribution claims 6-DoF in-hand pose refinement. Section IV-C parameterizes test poses only along pitch, roll, y, z, and Table III contains no pure-x or pure-yaw rows. The policy is therefore never trained or evaluated on x/yaw increments, so the 6-DoF and 'arbitrary in-hand pose' claims are unsupported. Please either narrow the claims to 4-DoF refinement along the observable dimensions or supply training and evaluation data for x and yaw.
  2. [Abstract vs. §IV-B] The abstract states the policy is 'trained entirely in MuJoCo before zero-shot deployment to an 11-DoF five-fingered hand.' Section IV-B, however, defines Policy B as 'pretrained on simulated data and fine-tuned using real-world data,' and Table II shows that Policy B, not the simulation-only Policy A, achieves the best real-world results. The zero-shot sim-to-real claim is thus contradicted by the paper's own experimental protocol. The abstract and Section V should be revised to distinguish simulation-only results from fine-tuned results.
  3. [Abstract vs. §IV-A metrics] The abstract reports '80.7% and 59.3% success under the 10°/10 mm criterion for fixed and random targets.' Section IV-A defines success thresholds as ε_pos=0.005 m and ε_rot=0.05 rad, which are stricter than 10 mm/10°, and Table II lists Policy B success rates of 100% for all groups. No 80.7%/59.3% values appear anywhere in the body. The mismatch between the abstract's numbers/thresholds and the experimental section must be resolved, and the provenance of the abstract numbers clarified.
  4. [§IV-D, Fig. 8] The long-horizon tracking experiment is described as tracking motion along 'the yaw dimension' in the body text, while the Fig. 8 caption lists pitch, y, and roll. Since §III-C explicitly excludes yaw from training because of overlap with other axes, claiming yaw tracking without any yaw-specific quantitative result is inconsistent. The caption/body disagreement compounds the issue. Either provide yaw-specific evaluation and explain how the policy handles yaw, or remove the yaw claim from the experiment.
minor comments (5)
  1. [§IV-A, Eq. (3)] The quaternion distance formula δ_rot = 2 arccos(|⟨Q,Q_g⟩|) is standard but the notation Q, Q_g is not defined before use; please define the quaternion variables and the inner-product convention.
  2. [Table III] The table is described as having 'darker colors indicate higher success rates,' but the printed table appears grayscale and the darkening is not legible. Please use a numerical color scale or add explicit values.
  3. [§III-D.2] The amount of real-world fine-tuning data for Policy B is not specified. Please report the size of D_real and how many real poses were used.
  4. [References [21], [24]] Reference [21] contains a placeholder arXiv ID '2312.XXXX'; please update to the actual identifier. Also check whether [24] is cited with the correct venue/arXiv number.
  5. [§IV-E vs. §V] Section IV-E reports limited within-category generalization to unseen objects, while the Conclusion says 'the policy is currently effective on a single object.' These statements should be reconciled; the conclusion appears to undersell the generalization result or Section IV-E should clarify the failure modes.

Circularity Check

0 steps flagged

No circular derivation: supervised regression against ground-truth pose increments; the 6-DoF/4-DoF mismatch is an evidence gap, not a circular step.

full rationale

The paper's derivation chain is an empirical supervised-learning pipeline rather than an analytical derivation that reduces to its inputs. Equation (1) defines the policy as a function from current/target tactile images and joint states to a 6-DoF wrist increment; Equation (2) is an MSE loss against ground-truth pose increments collected in simulation and the real world; Equation (3) measures pose error relative to the wrist pose recorded at demonstration time. The target pose is self-consistent with the target tactile image, but it is not used to construct the training labels, so this is not circularity. Cross-combination training is a data-augmentation scheme, not a hidden fit of the evaluation quantity. Self-citations appear in the related-work survey (e.g., refs [2], [4], [7], [12]) but none is load-bearing for the central claim, and no uniqueness theorem or ansatz is imported from the authors' prior work. The paper's own Section III-C restricts training sampling to pitch, roll, y, and z, and Section IV-C evaluates only those dimensions; Section IV-D even labels the tracked axis inconsistently. This means the '6-DoF' and 'arbitrary in-hand pose' claims are not fully supported by the reported experiments, but that is an evidence/validity gap, not a circular step. No fitted parameter is renamed as a prediction, and no result is equivalent to its inputs by construction. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central result rests on the fidelity of the tactile simulator, on the discriminative power of the tactile images for the four sampled dimensions, and on the transferability of the learned increment mapping. No new physical entities are introduced.

free parameters (4)
  • Pose sampling bounds for pitch, roll, y, z
    Upper/lower bounds for systematic pose-space sampling are chosen by hand; they define the range of 'arbitrary' target poses and exclude x and yaw (Section III-C).
  • Success thresholds epsilon_pos, epsilon_rot = 0.005 m; 0.05 rad in Section IV-A (abstract states 10 mm/10 deg)
    Hand-chosen criteria for success; inconsistency between abstract and full text changes reported success rates.
  • Data augmentation parameters = Tactile scale ~mu(0.5, 1.0); joint noise ~mu(-0.04, 0.04) (Table I)
    Chosen by hand to improve robustness; affect final performance and were not optimized systematically.
  • Network hyperparameters (encoder, latent dim, MLP, learning rate, batch size) = not reported
    Not specified in the paper; they are hand-selected and needed for reproduction.
axioms (4)
  • domain assumption MuJoCo elastic-contact tactile model faithfully reproduces real piezoresistive fingertip responses.
    Used for large-scale training data (Section III-B); Policy A sim-only underperforms Policy B, so this assumption holds only partially.
  • ad hoc to paper Tactile images uniquely encode pose along pitch, roll, y, z for the object set; x/roll and yaw/other overlaps justify dropping x and yaw.
    Stated in Section III-C; if false, the policy cannot distinguish poses and refinement fails.
  • domain assumption Cross-combination training from random pairs generalizes to all target poses in the sampled range.
    Justifies the arbitrary-target claim without per-pose retraining; empirically tested on 16 combinations.
  • domain assumption Object pose is stable during each regrasp, so tactile images correspond to commanded wrist pose.
    The loop opens, moves, and re-grasps; if the object pose changes unpredictably, tactile feedback no longer corresponds to the commanded wrist pose.

pith-pipeline@v1.3.0-alltime-deepseek · 8979 in / 19403 out tokens · 146996 ms · 2026-08-04T13:40:15.580072+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of TacRefineNet: Goal-Conditioned Tactile Grasp Refinement for Edge-Prominent Objects." pith.science (2026). https://pith.science/paper/ZCZFWZT2

@misc{pith2026250925746,
  author       = {Pith},
  title        = {Pith review of: TacRefineNet: Goal-Conditioned Tactile Grasp Refinement for Edge-Prominent Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZCZFWZT2}},
  note         = {Machine review of arXiv:2509.25746}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate final grasp alignment remains challenging for edge-prominent objects such as thin plates, discs, and rods, whose sparse contacts are easily occluded and poorly resolved by depth sensing. We present TacRefineNet, a tactile-only, goal-conditioned framework for local refinement along tactilely observable pose dimensions. Given current and target multi-finger tactile images and their corresponding hand-joint configurations, a Siamese policy network directly predicts corrective wrist pose increments. The hand iteratively opens, moves, and regrasps, forming an external-dexterity tactile servoing loop. Cross-combination training pairs current and target samples, allowing targets within the sampled pose range to be specified without retraining. We collect 156,007 simulated samples from 15 plates, discs, and rods and train the policy entirely in MuJoCo before zero-shot deployment to an 11-DoF five-fingered hand with piezoresistive sensors. On seen objects, the real system achieves 80.7\% and 59.3\% success under the $10^\circ$/10\,mm criterion for fixed and random targets, respectively; after five steps, the mean errors are approximately 5.2\,mm and $3.5^\circ$. Experiments further show continuous correction under long-horizon perturbations and limited within-category transfer to unseen objects, with reduced performance for symmetric or weakly discriminative contacts. Project website is available at https://sites.google.com/view/tacrefinenet

Figures

Figures reproduced from arXiv: 2509.25746 by Diyun Xiang, Haoran Zhou, Shuaijun Wang, Yangwei You.

Figure 1
Figure 1. Figure 1: Fine-grained flat object grasping using only multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TacRefineNet, including (a) data collection, (b) network architecture and training, and (c) real-world [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pipeline for data collection combining simulation and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative grasp refinement: target (left), intermediate adjustments, and final tactile image (right). [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of pose error curves for the two policies. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Dexterous hand tracking a moving object using fingertip tactile sensing only. The object sequentially moves along the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Left: unseen objects. Right: snapshots of the dexterous hand grasping these objects in the desired poses. [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

32 extracted references · 5 linked inside Pith

  1. [1]

    Deep learning ap- proaches to grasp synthesis: A review,

    R. Newbury, M. Gu, L. Chumbley, A. Mousavian, C. Eppner, J. Leitner, J. Bohg, A. Morales, T. Asfour, D. Kragicet al., “Deep learning ap- proaches to grasp synthesis: A review,”IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3994–4015, 2023

    2023

  2. [2]

    Dydexhandover: Human-like bimanual dynamic dexterous handover using rgb-only perception,

    H. Zhou, Y . You, and S. Wang, “Dydexhandover: Human-like bimanual dynamic dexterous handover using rgb-only perception,”arXiv preprint arXiv:2509.17350, 2025

    arXiv 2025

  3. [3]

    A generic control method of manipulator based on optimization,

    S. Wang, X. Wang, B. Zhan, S. Wang, and F. Zha, “A generic control method of manipulator based on optimization,” in2017 2nd Interna- tional Conference on Advanced Robotics and Mechatronics (ICARM). IEEE, 2017, pp. 486–491

    2017

  4. [4]

    Learning adaptive reaching and pushing skills using contact information,

    S. Wang, L. Sun, F. Zha, W. Guo, and P. Wang, “Learning adaptive reaching and pushing skills using contact information,”Frontiers in Neurorobotics, vol. 17, p. 1271607, 2023

    2023

  5. [5]

    Integrated grasp planning and visual object localization for a humanoid robot with five-fingered hands,

    A. Morales, T. Asfour, P. Azad, S. Knoop, and R. Dillmann, “Integrated grasp planning and visual object localization for a humanoid robot with five-fingered hands,” in2006 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2006, pp. 5663–5668

    2006

  6. [6]

    Dexgrasp anything: Towards universal robotic dexterous grasping with physics awareness,

    Y . Zhong, Q. Jiang, J. Yu, and Y . Ma, “Dexgrasp anything: Towards universal robotic dexterous grasping with physics awareness,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 584–22 594

    2025

  7. [7]

    Learning adaptive grasping from human demonstrations,

    S. Wang, W. Hu, L. Sun, X. Wang, and Z. Li, “Learning adaptive grasping from human demonstrations,”IEEE/ASME Transactions on Mechatronics, vol. 27, no. 5, pp. 3865–3873, 2022

    2022

  8. [8]

    Tinyvla: Towards fast, data-efficient vision-language- action models for robotic manipulation,

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shenet al., “Tinyvla: Towards fast, data-efficient vision-language- action models for robotic manipulation,”IEEE Robotics and Automation Letters, 2025

    2025

  9. [9]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    2023

  10. [10]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023

    2023

  12. [12]

    Learning an image-based visual servoing controller for object grasping,

    S. Wang, L. Sun, M. Li, P. Wang, F. Zha, W. Guo, and Q. Li, “Learning an image-based visual servoing controller for object grasping,”Inter- national Journal of Humanoid Robotics, vol. 21, no. 05, p. 2350033, 2024

    2024

  13. [13]

    Tactile robotics: An outlook,

    S. Luo, N. F. Lepora, W. Yuan, K. Althoefer, G. Cheng, and R. Dahiya, “Tactile robotics: An outlook,”arXiv preprint arXiv:2508.11261, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    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,”arXiv preprint arXiv:2410.24091, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Multi-fingered in-hand manipulation with various object properties using graph convolutional networks and distributed tactile sensors,

    S. Funabashi, T. Isobe, F. Hongyi, A. Hiramoto, A. Schmitz, S. Sugano, and T. Ogata, “Multi-fingered in-hand manipulation with various object properties using graph convolutional networks and distributed tactile sensors,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2102– 2109, 2022

    2022

  16. [16]

    Tactile regrasp: Grasp adjustments via simulated tactile transformations,

    F. R. Hogan, M. Bauza, O. Canal, E. Donlon, and A. Rodriguez, “Tactile regrasp: Grasp adjustments via simulated tactile transformations,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 2963–2970

    2018

  17. [17]

    Fine-grained recognition of manipulation activities on objects via multi-modal sensing,

    X. Liu, B. Zhang, L. Wang, S. Chen, X. Xie, X. Tong, T. Gu, and K. Li, “Fine-grained recognition of manipulation activities on objects via multi-modal sensing,”IEEE Transactions on Mobile Computing, vol. 23, no. 10, pp. 9614–9628, 2024

    2024

  18. [18]

    Learning of grasp adaptation through experience and tactile sensing,

    M. Li, Y . Bekiroglu, D. Kragic, and A. Billard, “Learning of grasp adaptation through experience and tactile sensing,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. Ieee, 2014, pp. 3339–3346

    2014

  19. [19]

    Sim2real manipulation on unknown objects with tactile-based rein- forcement learning,

    E. Su, C. Jia, Y . Qin, W. Zhou, A. Macaluso, B. Huang, and X. Wang, “Sim2real manipulation on unknown objects with tactile-based rein- forcement learning,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 9234–9241

    2024

  20. [20]

    Fbi: Learning dexterous in-hand manipulation with dynamic visuotactile shortcut policy,

    Y . Chen, W. Xu, Z. Yu, T. Tang, Y . Li, S. Yao, and C. Lu, “Fbi: Learning dexterous in-hand manipulation with dynamic visuotactile shortcut policy,”arXiv preprint arXiv:2508.14441, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Vint-6d: Visuo-tactile dataset for in-hand manipulation via object reposing and regrasping,

    Z. Wanet al., “Vint-6d: Visuo-tactile dataset for in-hand manipulation via object reposing and regrasping,”arXiv preprint arXiv:2312.XXXX, 2023

    2023

  22. [22]

    Learning tactile grasping with tactile-driven reinforce- ment learning,

    B. Wenet al., “Learning tactile grasping with tactile-driven reinforce- ment learning,”IEEE Transactions on Robotics, vol. 36, no. 5, pp. 1523– 1538, 2020. 9

    2020

  23. [23]

    Vita: Visuo-tactile action dataset for contact-rich manipulation,

    V . Dikhaleet al., “Vita: Visuo-tactile action dataset for contact-rich manipulation,” inConference on Robot Learning (CoRL), 2022

    2022

  24. [24]

    Hydroelastictouch: Simulation of tactile sensors with hydroelastic contact surfaces,

    D. P. Leins, F. Patzelt, and R. Haschke, “Hydroelastictouch: Simulation of tactile sensors with hydroelastic contact surfaces,”arXiv preprint arXiv:2501.08077, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Learning to imitate human tactile-proprioceptive contact with transformers,

    O. Adenijiet al., “Learning to imitate human tactile-proprioceptive contact with transformers,” inRobotics: Science and Systems (RSS), 2023

    2023

  26. [26]

    Soft tactile sensing for object classification and fine grasping adjustment using a pneumatic hand with an inflatable palm,

    M. Su, D. Huang, Y . Guan, C. Xiang, H. Zhu, and Z. Liu, “Soft tactile sensing for object classification and fine grasping adjustment using a pneumatic hand with an inflatable palm,”IEEE Transactions on Industrial Electronics, vol. 71, no. 4, pp. 3873–3883, 2023

    2023

  27. [27]

    Tactile regrasp: Grasp adjustments via simulated tactile transformations,

    F. Hoganet al., “Tactile regrasp: Grasp adjustments via simulated tactile transformations,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020

    2020

  28. [28]

    Learning tactile feature encodings for reactive grasping and regrasping,

    H. Maoet al., “Learning tactile feature encodings for reactive grasping and regrasping,”Science Robotics, vol. 9, no. 71, 2024

    2024

  29. [29]

    Tactile-based grasping stability prediction based on human grasp demonstration for robot manipulation,

    Z. Zhao, W. He, and Z. Lu, “Tactile-based grasping stability prediction based on human grasp demonstration for robot manipulation,”IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2646–2653, 2024

    2024

  30. [30]

    Visuo-tactile feedback-based robot manipulation for object packing,

    W. Liang, F. Fang, C. Acar, W. Q. Toh, Y . Sun, Q. Xu, and Y . Wu, “Visuo-tactile feedback-based robot manipulation for object packing,” IEEE Robotics and Automation Letters, vol. 8, no. 2, pp. 1151–1158, 2023

    2023

  31. [31]

    Tacgnn: Learning in-hand manipulation from tactile graphs,

    F. Yanget al., “Tacgnn: Learning in-hand manipulation from tactile graphs,” inIEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  32. [32]

    Tactile goal-conditioned in-hand manipulation via reinforcement learning,

    M. Pitzet al., “Tactile goal-conditioned in-hand manipulation via reinforcement learning,”IEEE Transactions on Robotics, 2024

    2024