Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

A MEMS tactile sensor's digital twin—finite-element simulation plus a trained neural net—replicates readings within 5.95% average error and lifts a seven-shape classifier from 33.57% to 95.0% accuracy.

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 18:12 UTC pith:VH5C7XHN

load-bearing objection Useful hardware plus a plausible real-to-sim pipeline, but the digital twin's fidelity claim rests on thin, possibly in-sample validation; worth a serious look with revisions. the 4 major comments →

arxiv 2509.10063 v1 pith:VH5C7XHN submitted 2025-09-12 cs.RO cs.AI

TwinTac: A Wide-Range, Highly Sensitive Tactile Sensor with Real-to-Sim Digital Twin Sensor Model

Xiyan Huang , Zhe Xu , Chenxi Xiao This is my paper
classification cs.RO cs.AI
keywords tactile sensordigital twinfinite element methodneural networkMEMS barometersim-to-realdata augmentationrobot learning
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.

TwinTac pairs a physical tactile sensor (PhysTac) with a simulated copy (DigiTac) so that tactile data can be generated in a physics simulator rather than collected by hand. The authors claim the simulated sensor matches the real one to within 5.95% of the maximum contact pressure, and that synthetic data from the twin boosts a seven-shape classification task from 33.57% to 95.0% accuracy when a small real dataset is augmented with simulated trials. The work addresses a gap in robot learning: non-optical tactile sensors have lacked simulation models, so touch-driven reinforcement learning has had to rely on slow, contact-heavy physical data collection.

Core claim

The central claim is that a non-optical tactile sensor—an array of MEMS barometer taxels embedded under a silicone gel—can be faithfully simulated by combining finite-element analysis of the gel's deformation with a small neural network that maps von Mises stress at 609 surface nodes to the eight taxel readings. Training this map on synchronized real and simulated indentation data (aligned by dynamic time warping) yields predictions within 5.95% of the real sensor's maximum contact pressure. The paper further shows that simulated data from this digital twin can substitute for real data in training: a time-series transformer that classifies seven indenter shapes jumps from 33.57% to 95.0% acc

What carries the argument

The load-bearing mechanism is the learned mapping Φ: X → S. The input X is, for each of the eight taxels, the inverse-distance-weighted average of the von Mises stress over the surface tetrahedron nodes clustered around that taxel; the output S is the eight-dimensional taxel pressure vector. A compact MLP (three 128-unit hidden layers with batch normalization and dropout) learns this mapping from paired data collected by running the same indentation protocol in a finite-element soft-body simulator and on the physical sensor, with dynamic time warping aligning the two sequences via their axial force traces.

Load-bearing premise

The finite-element model of the silicone gel, with unstated material parameters and an added 1 mm gap between gel and shell, is faithful enough that the single learned stress-to-taxel mapping continues to hold for unseen indenters, depths, speeds, and contact locations.

What would settle it

Apply the same indentation protocol with an off-center or sharp-tipped indenter, or use lateral sliding contact, and compare DigiTac's predicted taxel signals against PhysTac's readings. If per-taxel prediction error under these conditions substantially exceeds the reported 5.95% average, then the digital twin's validity is limited to normal, centered indentation. Alternatively, a cross-validation that holds out entire indenter shapes during training would reveal whether the mapping generalizes beyond the calibrated set.

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

If this is right

  • DigiTac can generate large volumes of synthetic tactile data for training, reducing the cost and risk of physical contact-data collection.
  • Because the digital twin runs in a physics simulator, it can be embedded in reinforcement-learning loops for contact-rich manipulation policies.
  • The PhysTac hardware's demonstrated range (0.007 N to 212 N) and 7.24 kPa/N sensitivity mean a single sensor design can cover both delicate and forceful contacts.

Where Pith is reading between the lines

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

  • The inverse-distance stress aggregation acts as a spatial low-pass filter; the twin's fidelity under sliding or shear contacts, where the stress field is asymmetric, remains untested and is the most direct stress test of the learned mapping.
  • The 1 mm gap deliberately inserted between gel and shell in the simulation to keep the FEM solver stable likely changes the deformation field; if the real gel touches the shell, the mapping at large loads may be extrapolating beyond its training regime.
  • A baseline that trains the classifier directly on raw FEM stress fields (without the learned taxel mapping) would separate the digital twin's contribution from the simple benefit of more training data—a testable ablation the paper does not report.
  • The same real-to-sim recipe (FEM deformation plus a small learned mapping) could transfer to other MEMS- or barometer-based tactile skins, provided a soft-body model and synchronized data are available.

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 / 6 minor

Summary. The paper introduces TwinTac, a tactile sensing system comprising a physical 8-taxel MEMS-based sensor (PhysTac) and a digital twin (DigiTac) built from FEM simulations in Isaac Gym. A lightweight MLP maps aggregated von Mises stress from a 609-node surface subset to the eight taxel outputs, trained on paired real/simulated temporal frames after DTW alignment. The authors report that DigiTac replicates PhysTac with 5.95% average error of maximum contact pressure on two indentation samples, and that augmenting a 7-shape classifier with 96 simulated trials per shape improves test accuracy from 33.57% (or 39.29%, per the caption) to 95.0%. The central claims are that the digital twin is high-fidelity and useful for Sim2Real data generation for non-optical tactile sensors.

Significance. If the fidelity and Sim2Real augmentation results hold, this is a useful contribution to an underdeveloped area: simulation models for non-visuotactile sensors. The real-to-sim calibration approach is sensible, and the downstream classification experiment is an appropriate external benchmark. The paper also includes a genuinely characterized hardware sensor with a wide force range, which is valuable in itself. However, the current evidence for the digital twin's central claim is incomplete: the fidelity result is reported on only two samples without a stated train/test split, error bars, or statistics, and the simulation model's material parameters are not specified. These gaps are load-bearing because the paper's stated purpose is to generate training data for unseen tactile interactions.

major comments (4)
  1. [Sec. IV-B / Fig. 8] The key fidelity result is not established as a generalization result. The training description in Sec. III-C says the network was trained on 36k paired temporal frames, but it never states whether the two indentation samples in Fig. 8 were held out from that training set. If they were drawn from the training distribution, the 5.95% average error measures interpolation of the learned mapping, not prediction for unseen conditions. Please state explicitly whether the Fig. 8 samples were excluded from the 36k pairs, and report error statistics (standard deviation, per-sample errors, and a small held-out set of conditions) to support the generalization claim.
  2. [Sec. III-B / IV-B] The FEM elastomer model is not reproducible or verifiable as described. The material parameters for Ecoflex 00-31 are not reported; the text only says Isaac Gym's flexible material engine was used. Further, a 1 mm gap between the soft body and side shell was introduced for numerical stability, which can alter the mechanical response. Since the learned mapping is only as good as the simulated stress field, the absence of material parameters and a comparison of simulated deformation against physical deformation is a major gap. Please report the constitutive parameters and any calibration/validation of the FEM model.
  3. [Sec. IV-B / IV-A] The 'wide-range' claim is not validated for the digital twin. PhysTac is characterized up to 212 N and the data-collection protocol in Sec. III-B includes varied indenters, locations, and action sequences, but the DigiTac fidelity evaluation in Fig. 8 uses only two small sinusoidal indentations (D = 1 mm and D = 2 mm). The digital twin's accuracy at large loads, off-center contacts, and with different indenter shapes is unknown. Please add fidelity evaluations covering the intended operating range, or explicitly narrow the claim.
  4. [Sec. IV-C / Fig. 9] The classification result contains an unreconciled internal inconsistency: the text reports 33.57% accuracy for the real-only baseline, while the caption of Fig. 9(C) reports 39.29%. This must be corrected. Additionally, no variance or statistical significance is reported for the 33.57%/95% comparison (e.g., repeated training runs with different seeds), and the real-only baseline of 20 training trials per shape is very small. Please clarify the number and report means/std over multiple runs.
minor comments (6)
  1. [Eq. (1)] The notation is unclear: N is used for the number of taxels, but later N_i is used for cluster size. The role of lambda_i and p(x_i) should be defined more carefully.
  2. [Sec. III-B / Eq. (3)-(5)] The sub/superscript placement in Fsim and Freal definitions (e.g., simF_z^(t) vs F^{(t)}_{sim}) is inconsistent and should be harmonized.
  3. [Fig. 3] The caption appears to mismatch the panel labels: 'Indenters' is labeled (A) in the caption but (B) in the figure, and 'Robot motion' is labeled (B) but appears as (A). Please fix the panel references.
  4. [Sec. III-B] The text says 'we wrap the pressure sensor's readings using DTW results'; this should be 'warp'.
  5. [Sec. IV-A] Typo: 'untill unloaded' should be 'until unloaded'.
  6. [Fig. 8] The plots would be much more informative with error bars or shaded confidence intervals across repeated trials; currently only single traces are shown.

Circularity Check

0 steps flagged

No significant circularity: the digital-twin mapping is supervised calibration and the central utility claim is validated on an external held-out classification benchmark.

full rationale

The paper's derivation chain is: build PhysTac hardware, simulate the elastomer with FEM in Isaac Gym, collect synchronized real and simulated indentation data, and train a lightweight MLP (DigiTac) to map aggregated FEM stress features to the 8 taxel outputs (Sec. III). The fidelity evaluation in Sec. IV-B compares DigiTac's output to PhysTac's output and reports a 5.95% average error. This is a supervised regression result: the MLP is trained on paired real/sim frames, so on the training distribution the error reflects fit quality rather than independent prediction. The paper does not explicitly state whether the two showcased samples in Fig. 8 were held out, but Fig. 3 depicts a training/validation split and Sec. III-C mentions validation-based learning-rate reduction, so the fidelity result is at least plausibly a generalization check. Crucially, the paper's central practical claim is independently supported: in Sec. IV-C, a time-series Transformer trained with DigiTac-generated simulation data plus 20 real trials per indenter reaches 95.0% accuracy on held-out real-world trials, versus 33.57% when trained on real data alone. That downstream benchmark is external to the mapping's training objective and uses real-world test data the mapping never saw. No load-bearing self-citations, imported uniqueness theorems, or ansatz-by-citation steps appear; the FEM simulation parameters are not derived from the target outputs, and the RBF pressure visualization is a standard interpolation technique, not a disguised predictive claim. The main weakness—absent explicit language about the train/test split for the fidelity number—is a reporting rigor issue, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on a learned mapping from a simulated stress field to real sensor outputs. Key free parameters include the MLP weights, unstated material properties, hand-chosen mesh and gap settings, and hyperparameters. The axioms are the assumptions that the FEM model, the stress-to-signal relationship, and the data alignment are all valid enough for the mapping to generalize.

free parameters (6)
  • DigiTac MLP weights = learned from 36k paired frames
    The mapping from FEM stress features to taxel outputs is a neural network trained on synchronized real and simulated data; all of its weights are fitted.
  • Ecoflex 00-31 material parameters
    FEM simulation requires elastic modulus, Poisson ratio, and damping; values are not reported, so replication requires guessing or contacting authors.
  • Domain randomization noise sigma = 0.05
    Zero-mean Gaussian noise injected during training; chosen by hand.
  • Node subset size = 609
    The authors state they 'found' that a predefined subset of 609 nodes yields satisfactory performance; the exact selection criterion is not given.
  • 1 mm gap between gel and shell = 1 mm
    Introduced to handle large deformations and ensure numerical stability; a modeling choice that affects the stress field.
  • Learning rate schedule parameters = 1e-3 initial, 0.3 factor
    AdamW hyperparameters chosen by hand.
axioms (5)
  • domain assumption FEM soft-body simulation in Isaac Gym accurately represents the deformation of the Ecoflex gel under indentation.
    Used throughout Sec. III-B; if the simulated stress field deviates from physical deformation, the learned mapping has no physical basis.
  • domain assumption Von Mises stress at surface nodes near each taxel determines the taxel pressure reading.
    Eq. 6 aggregates only von Mises stress; no justification that other stress components or shear are irrelevant.
  • domain assumption DTW alignment of F_real and F_sim correctly synchronizes the pressure readings with simulated frames.
    Sec. III-B data preprocessing; DTW assumes monotonic correspondence and cannot fix missing contacts.
  • ad hoc to paper The 609-node subset is representative of the full stress field.
    Authors state they 'found' this subset yields satisfactory performance, but selection procedure is not described.
  • standard math RBF interpolation reconstructs a meaningful continuous pressure map from only 8 taxels.
    Eq. 1 is a standard interpolation, but the fidelity of the resulting spatial distribution is not quantitatively validated.

pith-pipeline@v1.3.0-alltime-deepseek · 10214 in / 11408 out tokens · 108609 ms · 2026-08-04T18:12:35.417627+00:00 · methodology

0 comments
read the original abstract

Robot skill acquisition processes driven by reinforcement learning often rely on simulations to efficiently generate large-scale interaction data. However, the absence of simulation models for tactile sensors has hindered the use of tactile sensing in such skill learning processes, limiting the development of effective policies driven by tactile perception. To bridge this gap, we present TwinTac, a system that combines the design of a physical tactile sensor with its digital twin model. Our hardware sensor is designed for high sensitivity and a wide measurement range, enabling high quality sensing data essential for object interaction tasks. Building upon the hardware sensor, we develop the digital twin model using a real-to-sim approach. This involves collecting synchronized cross-domain data, including finite element method results and the physical sensor's outputs, and then training neural networks to map simulated data to real sensor responses. Through experimental evaluation, we characterized the sensitivity of the physical sensor and demonstrated the consistency of the digital twin in replicating the physical sensor's output. Furthermore, by conducting an object classification task, we showed that simulation data generated by our digital twin sensor can effectively augment real-world data, leading to improved accuracy. These results highlight TwinTac's potential to bridge the gap in cross-domain learning tasks.

Figures

Figures reproduced from arXiv: 2509.10063 by Chenxi Xiao, Xiyan Huang, Zhe Xu.

Figure 1
Figure 1. Figure 1: We introduce TwinTac, a unified tactile system consisting of: (A) PhysTac, a physical tactile sensor with high sensitivity and a large measurement range, and (B) DigiTac, the digital twin of PhysTac. text, providing cost-effective and risk-free environments for generating interaction data. Advances in soft-body contact simulation techniques (e.g., Material Point Method (MPM) and Finite Element Method (FEM)… view at source ↗
Figure 2
Figure 2. Figure 2: Technical pipeline of TwinTac, including the procedures for fabricating the PhysTac sensor through gel casting, as well as the techniques used for data collection and the creation of the DigiTac sensor via Real2Sim. Elastomer. The sensor elastomer is fabricated using Ecoflex 00-31 silicone gel (Smooth-On Inc.). This gel is selected for its low cost, adequate elasticity for deformation recovery, and durabil… view at source ↗
Figure 4
Figure 4. Figure 4: DigiTac’s network design. (A) Aggregated stress features from neighboring tetrahedron nodes, followed by (B) MLP network that infers the sensor’s signal outputs. sensor’s output signals. Therefore, DigiTac’s sensor model is created through a mapping from the FEM output to the sensor signals: Φ : X 7→ S. This process can be achieved efficiently using a lightweight network structure, as shown in [PITH_FULL_… view at source ↗
Figure 6
Figure 6. Figure 6: characterizes the taxel’s measurement range in response to external contact forces, as well as its sensitivity. Specifically, [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Characterization of PhysTac using different indenters. (A) Shape of Indenters. (B) Locations where indentation applied. (C) Sensor’s reading (with RBF interpolation). using only 8 MEMS taxels, which provides insufficient resolution to reveal surface textures and subtle geometric differences. B. Characterization of DigiTac Sensor Model Next, we assess the fidelity of the DigiTac model by comparing its outpu… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of DigiTac’s output versus PhysTac’s output under the same indentation condition. At the same time, a real-world indentation experiment was conducted to collect a correlated data set, although with much smaller quantities. A classification neural network was then designed to classify the real-world indentation data using a time-series Transformer [43] ( [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Network structure and confusion matrices for the classification task. (A) Time-series transform network. (B) Indenters used for object classification. (C) Confusion matrix obtained by training solely on real-world data (test accuracy: 39.29%). (D) Confusion matrix obtained by training on augmented data generated via DigiTac (test accuracy: 95%). approach, aligning real-world indentation data with Isaac Gym… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. RGB-S: Image-Aligned Tactile Saliency for Robust Dexterous Manipulation

    cs.RO 2026-06 unverdicted novelty 6.0

    RGB-S projects tactile contacts onto images as force-modulated Gaussian saliency maps via kinematics and zero-initialized conditioning, raising real-world occluded dexterous manipulation success by 26.7 percentage poi...

  2. ETac: A Lightweight and Efficient Tactile Simulation Framework for Learning Dexterous Manipulation

    cs.RO 2026-04 unverdicted novelty 6.0

    ETac is a data-driven tactile simulation framework that matches FEM deformation accuracy at high speed, supporting 4096 parallel environments at 869 FPS and yielding 84.45% success in blind grasping across four object types.

Reference graph

Works this paper leans on

43 extracted references · 8 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Human tactile sensing and sensorimotor mech- anism: from afferent tactile signals to efferent motor control,

    Y . Wei, A. G. Marshall, F. P. McGlone, A. Makdani, Y . Zhu, L. Yan, L. Ren, and G. Wei, “Human tactile sensing and sensorimotor mech- anism: from afferent tactile signals to efferent motor control,”Nature Communications, vol. 15, no. 1, p. 6857, 2024

    2024

  2. [2]

    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, pp. 3339– 3346, Ieee, 2014

    2014

  3. [3]

    More than a feeling: Learning to grasp and regrasp using vision and touch,

    R. Calandra, A. Owens, D. Jayaraman, J. Lin, W. Yuan, J. Malik, E. H. Adelson, and S. Levine, “More than a feeling: Learning to grasp and regrasp using vision and touch,”IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. 3300–3307, 2018

    2018

  4. [4]

    Tac- tile regrasp: Grasp adjustments via simulated tactile transformations,

    F. R. Hogan, M. Bauza, O. Canal, E. Donlon, and A. Rodriguez, “Tac- tile regrasp: Grasp adjustments via simulated tactile transformations,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2963–2970, IEEE, 2018

    2018

  5. [5]

    Grasping in the dark: Compliant grasping using shadow dexterous hand and biotac tactile sensor,

    K. Ganguly, B. Sadrfaridpour, K. B. Kidambi, C. Fermller, and Y . Aloimonos, “Grasping in the dark: Compliant grasping using shadow dexterous hand and biotac tactile sensor,”arXiv e-prints, pages arXiv–2011, 2020

    2011

  6. [6]

    Tactile sensors for advanced intelligent systems,

    C. Wang, L. Dong, D. Peng, and C. Pan, “Tactile sensors for advanced intelligent systems,”Advanced Intelligent Systems, vol. 1, no. 8, p. 1900090, 2019

    2019

  7. [7]

    Gelsight svelte hand: A three-finger, two-dof, tactile-rich, low-cost robot hand for dexterous manipulation,

    J. Zhao and E. H. Adelson, “Gelsight svelte hand: A three-finger, two-dof, tactile-rich, low-cost robot hand for dexterous manipulation,” ArXiv, vol. abs/2309.10886, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Eyesight hand: Design of a fully-actuated dexterous robot hand with inte- grated vision-based tactile sensors and compliant actuation,

    B. Romero, H.-S. Fang, P. Agrawal, and E. H. Adelson, “Eyesight hand: Design of a fully-actuated dexterous robot hand with inte- grated vision-based tactile sensors and compliant actuation,”ArXiv, vol. abs/2408.06265, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Dextouch: Learning to seek and manipulate objects with tactile dexterity,

    K.-W. Lee, Y . Qin, X. Wang, and S.-C. Lim, “Dextouch: Learning to seek and manipulate objects with tactile dexterity,”arXiv preprint arXiv:2401.12496, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Progress on flexible tactile sensors in robotic applications on objects properties recognition, manipulation and human-machine interactions,

    J. Jin, S. Wang, Z. Zhang, D. Mei, and Y . Wang, “Progress on flexible tactile sensors in robotic applications on objects properties recognition, manipulation and human-machine interactions,”Soft Science, vol. 3, no. 1, pp. N–A, 2023

    2023

  11. [11]

    Tactile object recognition using deep learning and dropout,

    A. Schmitz, Y . Bansho, K. Noda, H. Iwata, T. Ogata, and S. Sugano, “Tactile object recognition using deep learning and dropout,” in2014 IEEE-RAS International Conference on Humanoid Robots, pp. 1044– 1050, IEEE, 2014

    2014

  12. [12]

    Improved learning of robot manipulation tasks via tactile intrinsic motivation,

    N. Vulin, S. Christen, S. Stev ˇsi´c, and O. Hilliges, “Improved learning of robot manipulation tasks via tactile intrinsic motivation,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2194–2201, 2021

    2021

  13. [13]

    Learning in-hand translation using tactile skin with shear and normal force sensing,

    J. Yin, H. Qi, J. Malik, J. Pikul, M. Yim, and T. Hellebrekers, “Learning in-hand translation using tactile skin with shear and normal force sensing,”arXiv preprint arXiv:2407.07885, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Difftactile: A physics-based differentiable tactile simulator for contact-rich robotic manipulation,

    Z. Si, G. Zhang, Q. Ben, B. Romero, Z. Xian, C. Liu, and C. Gan, “Difftactile: A physics-based differentiable tactile simulator for contact-rich robotic manipulation,”arXiv preprint arXiv:2403.08716, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Taxim: An example-based simulation model for gelsight tactile sensors,

    Z. Si and W. Yuan, “Taxim: An example-based simulation model for gelsight tactile sensors,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2361–2368, 2022

    2022

  16. [16]

    Sim-to-real transfer for robotic manipulation with tactile sensory,

    Z. Ding, Y .-Y . Tsai, W. W. Lee, and B. Huang, “Sim-to-real transfer for robotic manipulation with tactile sensory,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6778–6785, IEEE, 2021

    2021

  17. [17]

    Novel tactile sensor technology and smart tactile sensing systems: A review,

    L. Zou, C. Ge, Z. J. Wang, E. Cretu, and X. Li, “Novel tactile sensor technology and smart tactile sensing systems: A review,”Sensors, vol. 17, no. 11, p. 2653, 2017

    2017

  18. [18]

    Isaacgym: A high-performance simulation environment for reinforcement learning

    NVIDIA, “Isaacgym: A high-performance simulation environment for reinforcement learning.”https://junxnone.github.io/ isaacgymdocs/index.html, 2021. Accessed: 2021-01-17

    2021

  19. [19]

    Tactile sens- ing—from humans to humanoids,

    R. S. Dahiya, G. Metta, M. Valle, and G. Sandini, “Tactile sens- ing—from humans to humanoids,”IEEE transactions on robotics, vol. 26, no. 1, pp. 1–20, 2009

    2009

  20. [20]

    R. S. Dahiya and M. Valle,Robotic tactile sensing: technologies and system, vol. 1. Springer, 2013

    2013

  21. [21]

    Recent progress in tactile sensing and sensors for robotic manipulation: can we turn tactile sensing into vision?,

    A. Yamaguchi and C. G. Atkeson, “Recent progress in tactile sensing and sensors for robotic manipulation: can we turn tactile sensing into vision?,”Advanced Robotics, vol. 33, no. 14, pp. 661–673, 2019

    2019

  22. [22]

    Gelsight: High-resolution robot tactile sensors for estimating geometry and force,

    W. Yuan, S. Dong, and E. H. Adelson, “Gelsight: High-resolution robot tactile sensors for estimating geometry and force,”Sensors, vol. 17, no. 12, p. 2762, 2017

    2017

  23. [23]

    Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation,

    M. Lambeta, P.-W. Chou, S. Tian, B. Yang, B. Maloon, V . R. Most, D. Stroud, R. Santos, A. Byagowi, G. Kammerer,et al., “Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation,”IEEE Robotics and Automation Letters, vol. 5, no. 3, pp. 3838–3845, 2020

    2020

  24. [24]

    Visuotactile sensors with emphasis on gelsight sensor: A review,

    A. C. Abad and A. Ranasinghe, “Visuotactile sensors with emphasis on gelsight sensor: A review,”IEEE Sensors Journal, vol. 20, no. 14, pp. 7628–7638, 2020

    2020

  25. [25]

    Combined use of fsr sensor array and svm classifier for finger motion recognition based on pressure distribution map,

    N. Li, D. Yang, L. Jiang, H. Liu, and H. Cai, “Combined use of fsr sensor array and svm classifier for finger motion recognition based on pressure distribution map,”Journal of Bionic Engineering, vol. 9, no. 1, pp. 39–47, 2012

    2012

  26. [26]

    Recent progress in pressure and temperature tactile sensors: principle, classification, integration and outlook,

    J. Yu, K. Zhang, and Y . Deng, “Recent progress in pressure and temperature tactile sensors: principle, classification, integration and outlook,”Soft Sci, vol. 1, no. 6, 2021

    2021

  27. [27]

    The feel of mems barometers: Inexpensive and easily customized tactile array sensors,

    Y . Tenzer, L. P. Jentoft, and R. D. Howe, “The feel of mems barometers: Inexpensive and easily customized tactile array sensors,” IEEE Robotics & Automation Magazine, vol. 21, no. 3, pp. 89–95, 2014

    2014

  28. [28]

    Barotac: Barometric three-axis tactile sensor with slip detection capability,

    G. Kim and D. Hwang, “Barotac: Barometric three-axis tactile sensor with slip detection capability,”Sensors, vol. 23, no. 1, p. 428, 2022

    2022

  29. [29]

    Barometer-based tactile skin for anthropomorphic robot hand,

    R. K ˜oiva, T. Schwank, G. Walck, M. Meier, R. Haschke, and H. Ritter, “Barometer-based tactile skin for anthropomorphic robot hand,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 9821–9826, IEEE, 2020

    2020

  30. [30]

    Force measurement of hand and fingers,

    K. Bretz, A. Jobbagy, and K. Bretz, “Force measurement of hand and fingers,”Biomechanica Hungarica, 04 2010

    2010

  31. [31]

    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,”arXiv preprint arXiv:2408.06506, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    When vision meets touch: A contemporary review for visuotactile sensors from the signal processing perspective,

    S. Li, Z. Wang, C. Wu, X. Li, S. Luo, B. Fang, F. Sun, X.-P. Zhang, and W. Ding, “When vision meets touch: A contemporary review for visuotactile sensors from the signal processing perspective,”arXiv preprint arXiv:2406.12226, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Efficient tactile simulation with differentiability for robotic manipulation,

    J. Xu, S. Kim, T. Chen, A. R. Garcia, P. Agrawal, W. Matusik, and S. Sueda, “Efficient tactile simulation with differentiability for robotic manipulation,” inConference on Robot Learning, pp. 1488– 1498, PMLR, 2023

    2023

  34. [34]

    3d force and contact estimation for a soft-bubble visuotactile sensor using fem,

    J.-C. Peng, S. Yao, and K. Hauser, “3d force and contact estimation for a soft-bubble visuotactile sensor using fem,” in2024 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp. 5666– 5672, IEEE, 2024

    2024

  35. [35]

    Tirgel: A visuo-tactile sensor with total internal reflection mechanism for external observation and contact detection,

    S. Zhang, Y . Sun, J. Shan, Z. Chen, F. Sun, Y . Yang, and B. Fang, “Tirgel: A visuo-tactile sensor with total internal reflection mechanism for external observation and contact detection,”IEEE Robotics and Automation Letters, 2023

    2023

  36. [36]

    Tacipc: Intersection- and inversion-free fem-based elastomer simulation for optical tactile sensors,

    W. Du, W. Xu, J. Ren, Z. Yu, and C. Lu, “Tacipc: Intersection- and inversion-free fem-based elastomer simulation for optical tactile sensors,”IEEE Robotics and Automation Letters, 2024

    2024

  37. [37]

    Sim-to-real for robotic tactile sensing via physics-based simulation and learned latent projections,

    Y . Narang, B. Sundaralingam, M. Macklin, A. Mousavian, and D. Fox, “Sim-to-real for robotic tactile sensing via physics-based simulation and learned latent projections,” in2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 6444–6451, IEEE, 2021

    2021

  38. [38]

    Tactile sensibility in the human hand: relative and absolute densities of four types of mechanoreceptive units in glabrous skin.,

    R. S. Johansson and A. B. Vallbo, “Tactile sensibility in the human hand: relative and absolute densities of four types of mechanoreceptive units in glabrous skin.,”The Journal of physiology, vol. 286, no. 1, pp. 283–300, 1979

    1979

  39. [39]

    Injection injuries: seemingly minor injuries with major consequences,

    H. Dailiana, D. Kotsaki, S. Varitimidis, S. Moka, M. Bakarozi, K. Oikonomou, and N. Malizos, “Injection injuries: seemingly minor injuries with major consequences,”Hippokratia, vol. 12, no. 1, p. 33, 2008

    2008

  40. [40]

    Fast tetrahedral meshing in the wild,

    Y . Hu, T. Schneider, B. Wang, D. Zorin, and D. Panozzo, “Fast tetrahedral meshing in the wild,”ACM Trans. Graph., vol. 39, July 2020

    2020

  41. [41]

    On the yielding of materials,

    R. V on Mises, “On the yielding of materials,”Zeitschrift f ¨ur Ange- wandte Mathematik und Mechanik, vol. 13, pp. 161–209, 1913

    1913

  42. [42]

    Dynamic time warping,

    M. M ¨uller, “Dynamic time warping,”Information retrieval for music and motion, pp. 69–84, 2007

    2007

  43. [43]

    Transformers in time series: A survey,

    Q. Wen, T. Zhou, C. Zhang, W. Chen, Z. Ma, J. Yan, and L. Sun, “Transformers in time series: A survey,”arXiv preprint arXiv:2202.07125, 2022

    Pith/arXiv arXiv 2022