Pith. sign in

REVIEW 4 major objections 6 minor 26 references

This paper shows that shear deformation fields captured by an embedded camera in an elastomeric endcap can be mapped to six-axis contact forces and torques, giving tensegrity robots a low-cost path to contact-aware state estimation.

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 ↗

A 130 mm open-source visuotactile endcap uses optically tracked marker motion and a residual MLP to estimate 6-axis contact wrenches, enabling ground-contact detection on tensegrity robots.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A solid hardware paper with a genuinely useful bonding technique and clean static wrench results; the dynamic-generalization claim in the abstract overstates what the paper's own numbers show. the 4 major comments →

arxiv 2607.15633 v1 pith:X7TSCXKS submitted 2026-07-17 cs.RO

Scalable Open-Source Visuotactile Sensor for 6-Axis Contact Wrench Estimation in Tensegrity Robots

Wenzhe Tong , Jonathan Mi , Xili Yi , Nima Fazeli , Xiaonan Huang This is my paper
classification cs.RO
keywords visuotactile sensingtensegrity robotscontact wrench estimationshear vector fieldneural networkgyroid infill bondingendcap sensoropen-source hardware
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.

The reading

This paper aims to show that a cheap, reproducible visuotactile endcap can give tensegrity robots the one thing they lack: direct measurement of ground contact. The sensor is a 130 mm silicone shell with a camera inside watching a painted dot pattern; when the shell deforms, the pattern shifts, and the paper computes a two-channel shear field from those shifts. A small residual neural network then maps that shear field to a six-axis wrench, and the authors report that the mapping works with mean squared error 0.1531 on static validation data, tracks force trends under dynamic sliding, and correctly identifies which of six endcaps touch the ground on a 12 kg robot. The fabrication novelty—letting liquid silicone flow into the gyroid infill of a 3D-printed TPU part to form an adhesive-free bond—is what makes the design scalable and open-source. If the claims hold, contact-aware state estimation for tensegrity robots becomes a practical, low-cost option rather than a research luxury.

Core claim

The central claim is that the shear vector field of the endcap's inner surface encodes the full six-axis contact wrench well enough for a regression network to recover it. Static validation gives MSE 0.1531 (with force errors smaller than torque errors), and a dynamic out-of-distribution test gives MSE 2.67, which the authors attribute to latency and missing temporal context. The paper's positive result is that despite the torque degradation, the force estimates remain coherent enough for reliable binary contact detection, demonstrated by matching manual video labels across zero-, one-, two-, and three-contact configurations on the 12 kg robot. The authors are explicit that robot-level six-a

What carries the argument

The two load-bearing mechanisms are (1) the shear vector field, a two-channel map (height x width) of lateral deformation of the elastomer shell's inner surface obtained by optical flow on camera frames, which serves as the neural network's input; and (2) the gyroid-infill mechanical interlock, where liquid silicone cures inside the open lattice of a 3D-printed thermoplastic polyurethane interface, creating an adhesive-free bond claimed to be stronger and lower-variation than the best adhesive tested. A residual MLP with two residual blocks and a linear head performs the shear-to-wrench regression, and a simple threshold ∥F̂∥₂ > 0.2 N converts predicted force into a binary contact signal.

Load-bearing premise

The learned shear-to-wrench mapping transfers from flat-surface, quasi-static lab presses at fixed tilt angles to arbitrary ground contacts on the real tensegrity robot without retraining.

What would settle it

Place the instrumented 12 kg robot on a force plate, record predicted per-endcap wrenches during a rolling or stepping motion, and compare against the plate's ground-truth wrench; if predicted contact force magnitude deviates by more than a few newtons or the active-endcap set disagrees with video labels in more than a small fraction of frames, the central transfer claim is falsified.

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

If this is right

  • Contact-aware state estimation: knowing which endcaps are anchored lets a floating-base estimator be reformulated with explicit contact constraints, which the paper argues is a missing ingredient for stable tensegrity locomotion.
  • Reproducible hardware: the parametric mold and modular TPU interface mean the same camera and lighting hardware can be reused across robot sizes, with only two shell parameters changed.
  • Open-source adoption: the released design files and code are intended to let other groups replicate the sensor at low cost, enabling broader research into tactile feedback for compliant robots.
  • Wrench cues without a force plate: the sensor gives six-axis contact wrench estimates onboard, which could support feedback control and localization in field deployments where external motion capture is impractical.

Where Pith is reading between the lines

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

  • If dynamic torque error stays high, the practical contribution may be nearer to reliable three-axis force plus contact detection than to full six-axis wrench estimation; the paper's own dynamic MSE and phase lag point in that direction.
  • The transfer from scripted flat-surface presses to unstructured terrain is untested; a natural next experiment is ground-truth wrench comparison on a step or gravel surface, which would either confirm or break the central transfer assumption.
  • Because the architecture is agnostic to gel geometry, the same shear-field-to-wrench recipe could be reused on other soft endcaps or manipulation fingertips, provided the training distribution covers the new contact geometries.
  • The 0.2 N contact threshold is a heuristic; feeding continuous wrench estimates directly into a state estimator, with uncertainty, would likely outperform thresholding in practice.
Share X Bluesky LinkedIn Reddit HN

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 manuscript presents a 130 mm open-source visuotactile endcap sensor for tensegrity robots, integrating a cast elastomer shell, a 3D-printed TPU interface, a rigid base with an embedded camera and LED ring, and a gyroid-infill mechanical bonding method. A residual MLP maps a 2×30×30 shear vector field to a 6D contact wrench. The authors collect a 34,368-sample static press dataset using a KUKA arm and ATI F/T sensor, report static validation MSE of 0.1531, dynamic MSE of 2.67, and demonstrate binary ground-contact detection on a 12 kg tensegrity robot with six endcaps. The paper claims accurate and stable wrench estimation, out-of-domain dynamic generalization, and reliable contact detection, with open-source release.

Significance. The hardware contribution is timely and potentially valuable: a low-cost, reproducible, large-format visuotactile sensor with a novel mechanical interlock could benefit tensegrity robots and other compliant platforms. The open-source release and mechanical bonding characterization are explicit strengths. The static wrench-estimation result on held-out test data (MSE 0.1513, close to training loss) is credible evidence of a working shear-to-force mapping under the tested quasi-static conditions. However, the more prominent claims—dynamic generalization and robot-level 6-axis wrench capability—are not supported by the reported evidence, and the aggregate mixed-unit MSE is not a physically interpretable accuracy metric. The central value proposition therefore currently rests on static laboratory data plus an untested transfer assumption.

major comments (4)
  1. [Abstract; §VI-C] The abstract states that experimental results demonstrate 'out-of-domain generalization under dynamic motion,' but §VI-C reports dynamic MSE of 2.67, approximately 17× the static MSE of 0.1531, and explicitly says the model 'should not be interpreted as fully dynamic-calibrated.' Torque predictions show 'higher variability and larger phase lag.' This is a direct contradiction of the abstract's strongest claim. Please revise the abstract and conclusion to state the dynamic limitation quantitatively and qualify what 'generalization' means (e.g., force trends remain useful for contact detection, but 6-axis wrench accuracy degrades substantially).
  2. [§V-B, §VI-B, Eq. (4)] The headline static MSE of 0.1531 is computed with an unweighted Euclidean norm over a vector whose components mix forces in N and torques in N·m. This aggregate metric is not physically interpretable and may be dominated by the larger-magnitude force components. The brief caveat in §VI-B is insufficient; the paper should report per-axis RMSE/MAE for F_x, F_y, F_z, T_x, T_y, T_z on the held-out test set, and ideally separate force and torque errors. Without these, the 'accurate 6-axis wrench estimation' claim is not quantitatively supported even in the static case.
  3. [§VI-D, Eq. (6), Fig. 10] The robot experiment validates only binary contact labels: endcaps are visually classified from video for four configurations, and the predicted active set 'matches' these labels. No quantitative detection metrics (precision, recall, F1, number of trials, false positives/negatives) are reported, and the threshold τ_F selection in Eq. (6) is described only qualitatively. The abstract's claim that the system 'reliably identify ground contacts' is therefore anecdotal. Please report a confusion matrix or similar metrics and describe the no-contact calibration procedure quantitatively.
  4. [§V-A, §VI-D] The model is trained exclusively on flat-plate static presses with θ=10–50°, F_n=5–35 N, and evaluated dynamically on the same KUKA/flat-plate setup. The robot deployment uses six endcaps on ground contacts with arbitrary configurations, but no wrench ground truth is available there; §VI-D itself defers 'full robot-level wrench validation' to future work. Given the large dynamic degradation in §VI-C, the transfer from laboratory flat-plane data to arbitrary tensegrity ground contacts is a substantive assumption, not a demonstrated result. Either add robot-level wrench validation with independent ground truth or explicitly scope the title and abstract to 'static and quasi-static flat-surface wrench estimation with binary contact detection on the robot.'
minor comments (6)
  1. [§V-C, §VI-B] The model checkpoint is selected by lowest validation error, so the validation set is not fully held-out. The phrase 'held-out validation set' in §VI-B is inaccurate; the test set is the independent held-out set. Since the test MSE (0.1513) is close to the validation MSE (0.1531), this does not affect the static conclusion, but the wording should be corrected.
  2. [§V-B] The residual MLP architecture is not fully specified: the number of residual blocks K_blk is never given. Please state K_blk and any dropout or normalization layers for reproducibility.
  3. [§V-A] The ROS synchronization tolerance is 50 ms at a 30 Hz camera rate, which could introduce label noise during dynamic motion. The paper does not discuss the effect of this synchronization error on dynamic wrench estimates.
  4. [§VI-A.2] The subsection is titled 'Shear' but the test is a 180-degree peel test based on ASTM D1876. Peel strength is not the same as shear strength; please correct the terminology and clarify which quantity is reported.
  5. [§VI-A.1] The sentence 'For mechanical bonding, 25 mm of interference was utilized' is ambiguous; presumably this is a bonded overlap length. Please state the geometry explicitly.
  6. [Fig. 7 caption] Minor typo: 'TPU/' appears at the end of the caption. Also fix inconsistent 'Fig: 1' and 'Fig: 2b-d' colon formatting elsewhere.

Circularity Check

0 steps flagged

No significant circularity: wrench estimates are validated against independently measured ATI F/T labels; acknowledged dynamic and robot-level limitations are correctness concerns, not circularity.

full rationale

The paper's central claim is an empirical supervised mapping from shear vector fields to 6D wrench (Eq. 2-4), trained with MSE loss against ground-truth wrenches from an ATI Gamma F/T sensor and evaluated on held-out validation/test splits. The static validation MSE (0.1531) and dynamic MSE (2.67) are reported as measured outcomes, not as quantities that are equal to the training objective by construction. The binary contact indicator (Eq. 5-6) is a derived signal using a threshold tau_F chosen from a no-contact calibration interval; this is standard threshold calibration, and the robot experiment validates the resulting contact labels against manual video annotations rather than treating the threshold as a prediction. Self-citations such as [9] are used only for motivational context ('contact sensing has received limited attention' and ground contact occurring 'along the side of the endcap rather than at its tip'); they are not load-bearing evidence for the wrench-estimation result. The GelSlim 4.0 citation [24] supplies the optical-flow shear-field estimation method, not the force/torque mapping, so no 'ansatz smuggled via citation' pattern applies. The paper explicitly acknowledges its own limitations: dynamic MSE is 'significantly higher than the static contact prediction error of 0.1531,' and the model 'should not be interpreted as fully dynamic-calibrated'; moreover, 'full robot-level wrench validation remains future work.' These are admissions of limited generalization and should be weighed as correctness or overclaim concerns, but they do not make the derivation circular. No equation or fitted parameter reduces to the target result by construction, and no uniqueness theorem or prior self-result is invoked to force the model choice. Accordingly, the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central quantitative claims rest on a fully learned mapping and a calibrated threshold; the paper does not provide a physical forward model, so the fit parameters are the main 'axioms.' The domain assumptions about synchronization, transfer, and sufficiency of shear fields are load-bearing.

free parameters (4)
  • Residual MLP weights θ = not enumerated
    Learned from 34,368 labeled samples; the wrench estimator is entirely data-driven, so the central quantitative claim is inseparable from these fitted weights.
  • Contact detection threshold τF = 0.2 N
    Selected from a no-contact calibration interval (§V-D); used to produce binary contact cues on the 12 kg robot.
  • Training hyperparameters = lr=1e-3, batch=32, epochs=100, hidden=64
    Chosen by hand; no ablation is reported to show sensitivity of the MSE result to these choices.
  • Gyroid infill density = 10%
    Described as balancing bonding strength and flexibility; a hand-chosen design parameter central to the bonding claim.
axioms (6)
  • domain assumption Shear vector field computed from marker optical flow is a sufficient statistic for the 6D contact wrench at the endcap center.
    The entire network input is V∈R^{2×30×30}; if shear fields omit deformation modes that carry torque information, torque estimates cannot be recovered. The paper's own §VI-C notes torque is sensitive to boundary distortions.
  • domain assumption The ATI Gamma F/T measurement and camera frames are synchronized well enough (approx. 50 ms tolerance) to provide correct labels.
    Dynamic validation shows phase lag, so this assumption is strained under motion; §VI-C attributes lag to camera frame rate, latency, and lack of temporal history.
  • domain assumption The flat-plane training distribution (θ=10–50°, Fn=5–35 N, circular frictional draws) transfers to real tensegrity ground contacts on all six endcaps without retraining.
    Robot-level test only checks binary contact labels, not wrench accuracy; full robot-level wrench validation is left to future work (§VI-D).
  • domain assumption Optical-flow shear-field estimation methods from GelSlim 4.0 work on this sensor's marker design and camera optics.
    The input pipeline is inherited from [24]; the paper does not re-derive or calibrate it for the 130 mm shell.
  • domain assumption Mechanical bonding strength measured in static tensile and peel tests implies field durability under repeated impact and abrasion.
    Bond tests are quasi-static; fatigue and impact testing are listed as future work (§VII).
  • standard math Residual MLPs can approximate the shear-to-wrench mapping with the given data size.
    No formal guarantee; relies on standard deep-learning approximation assumptions.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Open-Source Visuotactile Sensor for 6-Axis Contact Wrench Estimation in Tensegrity Robots." pith.science (2026). https://pith.science/paper/X7TSCXKS

@misc{pith2026260715633,
  author       = {Pith},
  title        = {Pith review of: Scalable Open-Source Visuotactile Sensor for 6-Axis Contact Wrench Estimation in Tensegrity Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7TSCXKS}},
  note         = {Machine review of arXiv:2607.15633}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents a scalable, open-source visuotactile sensing system for tensegrity robots that enables six-axis wrench estimation and contact detection. The proposed endcap sensor integrates an elastomeric shell, a 3D-printed thermoplastic polyurethane (TPU) interface, and a rigid base housing an embedded camera and LED illumination ring. A novel gyroid-infill bonding technique is introduced to form a durable elastomer-TPU interface without adhesives, yielding a lightweight and modular design compatible with large-scale tensegrity structures. A tactile-to-wrench neural network maps shear vector fields to six-dimensional force and torque measurements. Experimental results demonstrate accurate and stable wrench estimation with a mean squared error (MSE) of 0.1531 on static validation data and out-of-domain generalization under dynamic motion. Furthermore, full-system integration on a 12 kg tensegrity robot confirms the sensor's ability to reliably identify ground contacts. The system substantially improves the practicality of tactile feedback for tensegrity robots, offering a low-cost, reproducible, and physically interpretable pathway toward contact-aware proprioception and state estimation. Open source files are available at \href{https://github.com/Jonathan-Twz/tensegrity-gelfoot}{github.com/Jonathan-Twz/tensegrity-gelfoot}

Figures

Figures reproduced from arXiv: 2607.15633 by Jonathan Mi, Nima Fazeli, Wenzhe Tong, Xiaonan Huang, Xili Yi.

Figure 1
Figure 1. Figure 1: Exploded view of the visuotactile endcap sensor and its integration [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of the manufacturing steps to produce one endcap. First, the (a) PLA mold and (b) TPU interface are 3D printed and assembled. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sectional view of gyroid infill from (a) 3D printing slicer software [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Endcap visuotactile sensor characterization with KUKA Med R820 [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Endcap camera image and the corresponding [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Histogram of training set wrench label over six axes. [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Box plot of (a) max tensile and (b) average peel forces endured [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Temporal comparison of predicted and measured 6D contact [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Validation of endcap visuotactile sensing on a tensegrity robot under varying ground-contact conditions. The robot is sequentially placed into [PITH_FULL_IMAGE:figures/full_fig_p007_10.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

26 extracted references · 3 linked inside Pith

  1. [1]

    R. E. Skelton and M. C. De Oliveira,Tensegrity systems. Springer, 2009, vol. 1

    2009

  2. [2]

    Design of a variable stiffness quasi-direct drive cable-actuated tensegrity robot,

    J. Mi, W. Tong, Y . Ma, and X. Huang, “Design of a variable stiffness quasi-direct drive cable-actuated tensegrity robot,” 2024. [Online]. Available: https://arxiv.org/abs/2409.05751

    Pith/arXiv arXiv 2024

  3. [3]

    Design and evolution of a modular tensegrity robot platform,

    J. Bruce, K. Caluwaerts, A. Iscen, A. P. Sabelhaus, and V . SunSpiral, “Design and evolution of a modular tensegrity robot platform,” in2014 IEEE International Conference on Robotics and Automation (ICRA). Hong Kong, China: IEEE, May 2014, pp. 3483–3489

    2014

  4. [4]

    Tensegrity Robotics,

    D. S. Shah, J. W. Booth, R. L. Baines, K. Wang, M. Vespignani, K. Bekris, and R. Kramer-Bottiglio, “Tensegrity Robotics,”Soft Robotics, vol. 9, no. 4, pp. 639–656, Aug. 2022

    2022

  5. [5]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE transactions on robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  6. [6]

    LIO-SAM: Tightly-coupled lidar inertial odometry via smoothing and mapping,

    T. Shan, B. Englot, D. Meyers, W. Wang, C. Ratti, and D. Rus, “LIO-SAM: Tightly-coupled lidar inertial odometry via smoothing and mapping,” inProc. IEEE/RSJ Int. Conf. Intell. Robots and Syst.IEEE, 2020, pp. 5135–5142

    2020

  7. [7]

    6n-dof pose tracking for tensegrity robots,

    S. Lu, W. R. J. I. au2, K. Wang, X. Huang, J. Booth, R. Kramer- Bottiglio, and K. Bekris, “6n-dof pose tracking for tensegrity robots,” inRobotics Research, A. Billard, T. Asfour, and O. Khatib, Eds. Cham: Springer Nature Switzerland, 2023, pp. 136–152

    2023

  8. [8]

    Sensor Tendons for Soft Robot Shape Estimation,

    W. R. Johnson, A. Agrawala, X. Huang, J. Booth, and R. Kramer- Bottiglio, “Sensor Tendons for Soft Robot Shape Estimation,” in2022 IEEE Sensors. Dallas, TX, USA: IEEE, Oct. 2022, pp. 1–4

    2022

  9. [9]

    Tensegrity robot proprioceptive state estimation with geometric constraints,

    W. Tong, T.-Y . Lin, J. Mi, Y . Jiang, M. Ghaffari, and X. Huang, “Tensegrity robot proprioceptive state estimation with geometric constraints,” 2024. [Online]. Available: https://arxiv.org/abs/2410. 24226

    2024

  10. [10]

    State esti- mation for tensegrity robots,

    K. Caluwaerts, J. Bruce, J. M. Friesen, and V . SunSpiral, “State esti- mation for tensegrity robots,” in2016 IEEE International Conference on Robotics and Automation (ICRA). Stockholm, Sweden: IEEE, May 2016, pp. 1860–1865

    2016

  11. [11]

    DenseTact 2.0: Optical Tactile Sensor for Shape and Force Reconstruction,

    W. K. Do, B. Jurewicz, and M. Kennedy, “DenseTact 2.0: Optical Tactile Sensor for Shape and Force Reconstruction,” in2023 IEEE International Conference on Robotics and Automation (ICRA). London, United Kingdom: IEEE, May 2023, pp. 12 549–12 555. [Online]. Available: https://ieeexplore.ieee.org/document/10161150/

    arXiv 2023

  12. [12]

    GelSlim3.0: High-Resolution Measurement of Shape, Force and Slip in a Compact Tactile- Sensing Finger,

    I. Taylor, S. Dong, and A. Rodriguez, “GelSlim3.0: High-Resolution Measurement of Shape, Force and Slip in a Compact Tactile- Sensing Finger,” 2021, version Number: 1. [Online]. Available: https://arxiv.org/abs/2103.12269

    Pith/arXiv arXiv 2021

  13. [13]

    Soft-bubble grippers for robust and perceptive manipulation,

    N. Kuppuswamy, A. Alspach, A. Uttamchandani, S. Creasey, T. Ikeda, and R. Tedrake, “Soft-bubble grippers for robust and perceptive manipulation,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Las Vegas, NV , USA: IEEE, Oct. 2020, pp. 9917–9924. [Online]. Available: https://ieeexplore.ieee.org/document/9341534/

    arXiv 2020

  14. [14]

    GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force,

    W. Yuan, S. Dong, and E. Adelson, “GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force,” Sensors, vol. 17, no. 12, p. 2762, Nov. 2017. [Online]. Available: https://www.mdpi.com/1424-8220/17/12/2762

    2017

  15. [15]

    Surface Actuation and Sensing of a Tensegrity Structure Using Robotic Skins,

    J. W. Booth, O. Cyr-Choini `ere, J. C. Case, D. Shah, M. C. Yuen, and R. Kramer-Bottiglio, “Surface Actuation and Sensing of a Tensegrity Structure Using Robotic Skins,”Soft Robotics, vol. 8, no. 5, pp. 531– 541, Oct. 2021

    2021

  16. [16]

    Force-sensing tensegrity for investigating physical human-robot interaction in compliant robotic systems,

    A. R. Barkan, A. Padmanabha, S. R. Tiemann, A. Lee, M. P. Kanter, Y . S. Agarwal, and A. M. Agogino, “Force-sensing tensegrity for investigating physical human-robot interaction in compliant robotic systems,” in2021 IEEE International Conference on Robotics and Automation (ICRA), 2021, pp. 3292–3298

    2021

  17. [17]

    Large-area and low-cost force/tactile capacitive sensor for soft robotic applications,

    A. Pagoli, F. Chapelle, J.-A. Corrales-Ramon, Y . Mezouar, and Y . Lapusta, “Large-area and low-cost force/tactile capacitive sensor for soft robotic applications,”Sensors, vol. 22, no. 11, 2022. [Online]. Available: https://www.mdpi.com/1424-8220/22/11/4083

    2022

  18. [18]

    A tunable magnet-based tactile sensor framework,

    E. Harber, E. Schindewolf, V . Webster-Wood, H. Choset, and L. Li, “A tunable magnet-based tactile sensor framework,” in2020 IEEE SENSORS, 2020, pp. 1–4

    2020

  19. [19]

    Magnetic-based soft tactile sensors with deformable continuous force transfer medium for resolving contact locations in robotic grasping and manipulation,

    A. Mohammadi, Y . Xu, Y . Tan, P. Choong, and D. Oetomo, “Magnetic-based soft tactile sensors with deformable continuous force transfer medium for resolving contact locations in robotic grasping and manipulation,”Sensors, vol. 19, no. 22, 2019. [Online]. Available: https://www.mdpi.com/1424-8220/19/22/4925

    2019

  20. [20]

    Walking on tactip toes: A tactile sensing foot for walking robots,

    E. A. Stone, N. F. Lepora, and D. A. Barton, “Walking on tactip toes: A tactile sensing foot for walking robots,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 9869–9875

    2020

  21. [21]

    Development of a tactile sensor based on biologically inspired edge encoding,

    C. Chorley, C. Melhuish, T. Pipe, and J. Rossiter, “Development of a tactile sensor based on biologically inspired edge encoding,” in2009 International Conference on Advanced Robotics. IEEE, 2009, pp. 1–6

    2009

  22. [22]

    The tactip family: Soft optical tactile sensors with 3d-printed biomimetic morphologies,

    B. Ward-Cherrier, N. Pestell, L. Cramphorn, B. Winstone, M. E. Giannaccini, J. Rossiter, and N. F. Lepora, “The tactip family: Soft optical tactile sensors with 3d-printed biomimetic morphologies,”Soft robotics, vol. 5, no. 2, pp. 216–227, 2018

    2018

  23. [23]

    Tactid: high-performance visuo-tactile sensor-based terrain identification for legged robots,

    Z. Song, C. Li, Z. Quan, S. Mu, X. Li, Z. Zhao, W. Jin, C. Wu, W. Ding, and X.-P. Zhang, “Tactid: high-performance visuo-tactile sensor-based terrain identification for legged robots,”IEEE Sensors Journal, vol. 24, no. 16, pp. 26 487–26 495, 2024

    2024

  24. [24]

    Gelslim 4.0: Focusing on touch and reproducibility,

    A. Sipos, W. v. d. Bogert, and N. Fazeli, “Gelslim 4.0: Focusing on touch and reproducibility,”arXiv preprint arXiv:2409.19770, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  26. [26]

    Aggregated residual transformations for deep neural networks,

    S. Xie, R. Girshick, P. Doll ´ar, Z. Tu, and K. He, “Aggregated residual transformations for deep neural networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1492–1500

    2017

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.