At First Contact: Stiffness Estimation Using Vibrational Information for Prosthetic Grasp Modulation

Anway S. Pimpalkar; Ariel Slepyan; Nitish V. Thakor

REVIEW 3 major objections 6 minor 20 references

At First Contact: Stiffness Estimation Using Vibrational Information for Prosthetic Grasp Modulation

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read First-contact vibrations classify object stiffness with up to 98.6% accuracy and regress it within 15 ms

desk verdict New temporal window for vibration-based stiffness estimation, but a missing velocity control means the central claim is not yet nailed down. read the letter →

arxiv 2411.18507 v2 pith:MV6BS6AW submitted 2024-11-27 cs.RO

Anway S. Pimpalkar , Ariel Slepyan , Nitish V. Thakor This is my paper · ORCID

classification cs.RO

keywords stiffnessestimationprosthetichandpiezoelectricsensingtactilegraspmodulationvibrationanalysismachinelearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a prosthetic fingertip can estimate object stiffness from vibration alone in the first 15 milliseconds after initial contact, before the grasp closes. It shows that a piezoelectric sensor combined with machine learning classifies stiffness levels with up to 98.6% accuracy and regresses stiffness on real objects with mean squared error as low as 2.39 Shore A, all in under 1.5 ms. Because the average delay between first and second finger contact in a pinch is 16.65 ms, the estimate arrives early enough to modulate the grasp before full force develops. This opens a new sensing window that traditional force-based methods cannot exploit.

What carries the argument

The load-bearing mechanism is a biomimetic fingertip sensor that places a PZT piezoelectric disk on the back of a PCB to capture vibration and a piezoresistive array on the front to measure force, mimicking the layered placement of Pacinian and Merkel/Meissner receptors in human skin. During a pinch, one finger contacts the object before the other, producing a vibration-only signal before any force is registered, and the paper treats the 15 ms post-contact window as input to support vector machines and a convolutional neural network for stiffness classification and regression.

What would settle it

Test objects matched in Shore A stiffness but differing in surface texture or curvature; if first-contact vibration amplitude or the trained models' predictions shift with texture or curvature at constant stiffness, the claim that vibration encodes stiffness alone would be weakened.

Watch

Extended reading notes

Core claim

The paper's central claim is that the transient vibration created when one finger first touches an object carries reliable information about that object's stiffness, even though no reaction force has yet built up. In pinch grasps the fingers contact the object sequentially, and the authors use the 15 ms window between first and second contact as input. The amplitude of the piezoelectric response grows with stiffness, and both an SVM and an EfficientNetV2 CNN extract this structure from raw 15 ms vibration windows. The CNN reaches 98.6% discrimination accuracy and 2.39 Shore A regression MSE on unseen everyday objects while running in under 1.5 ms, confirming that first-contact vibration can generalize beyond the flat silicone training blocks.

Load-bearing premise

The results assume that first-contact vibration depends primarily on stiffness (Shore A) and not on confounds such as surface roughness, local curvature, mass, or material damping, so models trained on flat silicone blocks generalize to curved, textured real objects.

Editorial extensions

If this is right

Stiffness can be estimated before grasp completion, enabling prosthetic grasp-force modulation and reducing the risk of crushing fragile objects.
The SVM-based contact detector achieved 100% accuracy against ground truth, offering a reliable event trigger for prosthetics control.
Because piezoelectric response amplitude is proportional to stiffness, simple amplitude thresholds could provide low-compute stiffness estimates in embedded systems.
The regression results on apples, oranges, tennis balls, and avocados suggest the method can transfer to robotic grippers equipped with similar fingertip sensors.

Reading between the lines

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

The same first-contact vibration principle might extend to other contact geometries, such as sliding or tapping, but the paper only tests pinches.
The measured 16.65 ms inter-contact interval comes from a fixed palm-up setup; real reach-to-grasp movements may show different asymmetry, changing the usable time window.
A testable extension would vary grasp speed and object mass to see whether the vibration–stiffness relationship is invariant to those factors.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The genuinely new thing here is the temporal window: the 15 ms between first and second finger contact, which prior stiffness methods leave unused because force sensors read nothing until full contact. The authors show that a PZT sensor in that window yields enough signal to classify silicone blocks (97–98.6%) and regress Shore A on eight real objects (MSE down to 2.39). If the result holds, it is a real advance for prosthetic grasp modulation, where early estimates are safety-relevant. I also credit the careful contact-event ground truth: copper strips plus 960 fps video validation is the right way to establish that you are actually measuring what you claim to measure. The sensor integration into a TASKA hand is concrete and reproducible in spirit.

The soft spots are real, though. The biggest one is the one the stress-test flags: no control for closing speed. The PZT amplitude grows with impact impulse, which is a function of both stiffness and approach velocity. The training data uses one fixed hand posture and presumably one motor command; the real-object validation inherits that. So the learned mapping may be 'stiffness given this particular grasp velocity,' not stiffness generally. Since the stated application is user-driven prosthetics with varying speeds, this is a load-bearing gap. A simple experiment varying motor speed across otherwise identical silicone blocks would test it. The second issue is confound control on the real objects: apples, oranges, tennis balls, and avocados differ in curvature, surface texture, mass, and damping. With one trial protocol, those variables alias into the stiffness prediction. Third, the reported accuracies and MSEs come from a single 90-10 split with no error bars, and the number of test trials per object is not stated. That makes it hard to judge stability. Fourth, there is no baseline comparison to existing force-based or post-contact methods; the reader cannot see how much better this is in a head-to-head. Minor: no code or data release, which would strengthen the empirical case.

I do not think the paper is circular or dishonest—the models are trained on held-out stiffness blocks, and the self-citation to [13] is just the sensor lineage. The logic is clear; it is the controls that need work. The central idea is worth serious referee time, but a desk reject would be premature. The right move is to send it to review and require a velocity-sweep experiment, confound controls or at least a discussion of their absence, variance reporting, and a force-based baseline before publication.

Referee Report

3 major / 6 minor

Summary. The manuscript presents a piezoelectric fingertip sensor integrated into a prosthetic hand and uses the first 15 ms of vibration after initial contact to estimate object stiffness. Models are trained on 2,500 pinches of five silicone blocks of known Shore A stiffness and validated on eight real objects; the authors report 97.2% (SVM) and 98.6% (CNN) classification accuracy, regression MSE of 4.17 and 2.39 Shore A, and inference under 1.5 ms, which is faster than the mean inter-contact interval of 16.65 ms. The paper argues that this enables stiffness-based grasp modulation before full contact.

Significance. If the central claim holds, the paper offers a genuinely useful result: exploiting the stochastic finger-contact asymmetry to obtain stiffness information before a grasp is complete, with inference latency compatible with real-time prosthetic control. Strengths include the physical sensor integration, a reasonably sized training dataset, validation of contact timing with high-speed video, and the comparison of two machine-learning pipelines. There is no obvious circularity in the training/evaluation design. The main limitation is that the evidence for a general 'stiffness encoding' is not yet conclusive: the training protocol holds closing speed constant, and the real-object validation is small and statistically underreported.

major comments (3)

[Section III-A and Fig. 6] The training protocol uses a single prosthetic hand in a fixed palm-up pose with no variation in motor command or closing speed, and Fig. 6(b) interprets peak PZT amplitude as the stiffness correlate. Since the peak amplitude of a piezoelectric element during impact scales with the impulse delivered at contact, which depends on both stiffness and approach velocity, the learned mapping may be specific to the trained closure speed rather than a general stiffness encoding. Please add a control experiment that varies closing speed (e.g., several motor commands or approach velocities) and show that the amplitude-stiffness relationship and the SVM/CNN predictions are invariant; without this control, the central generalization claim that vibration 'encodes stiffness' is not established.
[Section III-D and Fig. 7] The real-world generalization result is based on a single 90-10 split and eight objects (two each of apple, orange, tennis ball, and avocado) with no per-object trial counts, no repeated cross-validation, and no confidence intervals; the reported MSE values of 4.17 and 2.39 Shore A therefore have no associated uncertainty. Please report the number of trials per object, run repeated splits or k-fold cross-validation, and provide per-object errors so the reader can assess whether performance is driven by one object class or by surface texture and curvature rather than stiffness.
[Section III-D] The ground-truth stiffness for real objects was measured with a Shore A durometer on curved, nonuniform surfaces such as apples, oranges, and avocados. Shore A durometer readings are geometry- and contact-area-dependent, and for curved substrates or thin skins they do not measure a well-defined material modulus. This affects the validity of the reported MSE values as ground truth. Please report how the durometer measurements were made on curved surfaces, provide an independent stiffness measure if possible, or discuss the uncertainty in the ground-truth values.

minor comments (6)

[Section IV] The statement that 'force data remains absent due to the lack of a reaction force until the second finger establishes contact' is physically unclear, since the first finger does experience a reaction force, and it is not supported by a displayed force trace; please clarify or provide evidence.
[Section IV] The units of the reported regression errors, 'MSE values of 4.17 and 2.39 Shore A', are inconsistent: mean squared error should be in squared Shore A units, or the values should be reported as root mean square error.
[Section III-C] Hyperparameter details for the SVM (C and gamma ranges) and the EfficientNetV2 input representation and architecture are missing; please include these details for reproducibility.
[Section II] The Savitzky-Golay filter parameters (window length and polynomial order) and the exact alignment of the 15 ms analysis window relative to the detected contact are not specified; please report them.
[Fig. 4(b)] The distribution statistics for the finger-contact timing offset (mean 16.65 ms, SD 10.35 ms) lack the number of grasps used to compute them; please add N.
[Fig. 7] The description of 'true values represented by extended color bars' is unclear; conventional ground-truth markers or explicit labels would make the figure easier to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the stiffness result is an empirical, out-of-sample ML finding with held-out real-world objects.

full rationale

The derivation chain is empirical and self-contained. The models are trained on 2,500 pinches of five silicone blocks with durometer-verified Shore A labels and tested on eight held-out real objects (apples, oranges, tennis balls, avocados) whose stiffness was measured with the same durometer standard. No parameter is fitted to the test labels, and no prediction is computed from the training inputs by construction; the 90-10 split and cross-object generalization are genuine out-of-sample evaluations. The stated amplitude-stiffness proportionality is presented as an observed sensor response, not as a derived consequence, and the SVM and CNN are standard regressors rather than ansatze that embed the answer. The only self-citation ([13], a prior sensor paper by overlapping authors) is invoked for 'previous approaches with tactile data' in model selection; it is not load-bearing for the stiffness-estimation claim, which is validated independently by the held-out real-object data. Concerns that closing speed was not varied (so impact impulse may alias with stiffness) are a missing-control or external-validity issue, not a circularity: the models' inputs and labels are still distinct, and the test objects were unseen. Therefore no circular step is identified.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper relies on the assumption that Shore A durometer hardness is the relevant ground-truth stiffness, that first-contact vibration is a function of stiffness rather than confounds, and that silicone-block training generalizes to real objects. These are domain assumptions rather than derivations. The hand-chosen processing parameters (window length, smoothing, model hyperparameters) are listed as free parameters because they affect results but are not justified from first principles.

free parameters (5)

Analysis window length = 15 ms
Chosen by hand as the input duration for SVM and CNN; based on the observed inter-contact interval.
Exponential smoothing alpha = 0.5
Applied to vibration data on the microcontroller before saving; value chosen by hand.
SVM hyperparameters (C, gamma)
Tuned via optimization; exact values not reported in the paper.
CNN training hyperparameters = lr=0.001, 40 epochs, step decay 0.5 every 5 epochs
Standard choices that affect model performance; chosen without systematic justification.
Contact detection threshold = 3 sigma
Used in the thresholding method, later replaced by SVM; a manual threshold.

assumptions (4)

domain assumption Shore A durometer measurement is a valid ground-truth for the stiffness relevant to pinch grasp modulation.
Used to label both training blocks (Section III-A) and test objects (Section III-D).
domain assumption First-contact vibrational signals are determined primarily by object stiffness, not by other object properties.
Underlies the entire transfer from silicone blocks to real objects; not directly tested (Section III-D).
domain assumption The training data (silicone blocks) is representative of the test objects in terms of contact mechanics.
Required for the claimed generalization to apples, oranges, tennis balls, and avocados (Section III-D).
domain assumption The timing asymmetry of finger contacts (mean 16.65 ms) is a consistent property of the prosthetic hand's pinch grasps.
Used to argue the inference time is fast enough; measured on the authors' dataset (Section IV).

how reviews work

0 comments

Cite this review

Pith. "Pith review of At First Contact: Stiffness Estimation Using Vibrational Information for Prosthetic Grasp Modulation." pith.science (2026). https://pith.science/paper/MV6BS6AW

@misc{pith2026241118507,
  author       = {Pith},
  title        = {Pith review of: At First Contact: Stiffness Estimation Using Vibrational Information for Prosthetic Grasp Modulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MV6BS6AW}},
  note         = {Machine review of arXiv:2411.18507}
}

read the original abstract

Stiffness estimation is crucial for delicate object manipulation in robotic and prosthetic hands but remains challenging due to dependence on force and displacement measurement and real-time sensory integration. This study presents a piezoelectric sensing framework for stiffness estimation at first contact during pinch grasps, addressing the limitations of traditional force-based methods. Inspired by human skin, a multimodal tactile sensor that captures vibrational and force data is developed and integrated into a prosthetic hand's fingertip. Machine learning models, including support vector machines and convolutional neural networks, demonstrate that vibrational signals within the critical 15 ms after first contact reliably encode stiffness, achieving classification accuracies up to 98.6% and regression errors as low as 2.39 Shore A on real-world objects of varying stiffness. Inference times of less than 1.5 ms are significantly faster than the average grasp closure time (16.65 ms in our dataset), enabling real-time stiffness estimation before the object is fully grasped. By leveraging the transient asymmetry in grasp dynamics, where one finger contacts the object before the others, this method enables early grasp modulation, enhancing safety and intuitiveness in prosthetic hands while offering broad applications in robotics.

Figures

Figures reproduced from arXiv: 2411.18507 by the authors.

**Figure 5.** Contact events validation using 960 fps videography: (a) frame before [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗

**Figure 6.** (a) Representative piezoelectric sensor responses for varying stiffness [PITH_FULL_IMAGE:figures/full_fig_p003_6.png] view at source ↗

Figures from the paper (1 more)

**Figure 7.** Figure 7: Stiffness regression on real-world objects with varying stiffness levels [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[13]

Flexible multilayer tactile sensor on a soft robotic fingertip,

S. Sankar, A. Slepyan, M. M. Iskarous, W.-Y . Cheng, R. DeBrabander, J. Zhang, A. Gupta, and N. V . Thakor, “Flexible multilayer tactile sensor on a soft robotic fingertip,” in 2022 IEEE Sensors . IEEE, Oct. 2022, p. 1–4

work page 2022
[1]

Coding and use of tactile signals from the fingertips in object manipulation tasks,

R. S. Johansson and J. R. Flanagan, “Coding and use of tactile signals from the fingertips in object manipulation tasks,” Nature Reviews Neuroscience, vol. 10, no. 5, p. 345–359, Apr. 2009

work page 2009
[2]

The role of stiffness in versatile robotic grasping,

C. J. Stabile, D. J. Levine, G. M. Iyer, C. Majidi, and K. T. Turner, “The role of stiffness in versatile robotic grasping,” IEEE Robotics and Automation Letters, vol. 7, no. 2, p. 4733–4740, Apr. 2022

work page 2022
[3]

Variable stiffness robotic hand for stable grasp and flexible handling,

S. Mahboubi, S. Davis, and S. Nefti-Meziani, “Variable stiffness robotic hand for stable grasp and flexible handling,” IEEE Access , vol. 6, p. 68195–68209, 2018

work page 2018
[4]

Grasped object stiffness detection for adaptive pid sliding mode position control of a prosthetic hand,

R. Andrecioli and E. D. Engeberg, “Grasped object stiffness detection for adaptive pid sliding mode position control of a prosthetic hand,” in 2012 4th IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics (BioRob). IEEE, Jun. 2012, p. 526–531

work page 2012
[5]

Current-sensor-based con- tact stiffness detection for prosthetic hands,

H. Deng, X. Xu, W. Zhuo, and Y . Zhang, “Current-sensor-based con- tact stiffness detection for prosthetic hands,” IEEE Access , vol. 8, p. 29456–29466, 2020

work page 2020
[6]

Stiffness-estimation-based grasping force fuzzy control for underactuated prosthetic hands,

Y . Zhang, X. Xu, R. Xia, and H. Deng, “Stiffness-estimation-based grasping force fuzzy control for underactuated prosthetic hands,” IEEE/ASME Transactions on Mechatronics , vol. 28, no. 1, p. 140–151, Feb. 2023

work page 2023
[7]

An introductory study of common grasps used by adults during perfor- mance of activities of daily living,

M. Vergara, J. Sancho-Bru, V . Gracia-Ib ´a˜nez, and A. P ´erez-Gonz´alez, “An introductory study of common grasps used by adults during perfor- mance of activities of daily living,” Journal of Hand Therapy , vol. 27, no. 3, p. 225–234, Jul. 2014

work page 2014

Show all 20 references

[8]

Trends and challenges in robot manipulation,

A. Billard and D. Kragic, “Trends and challenges in robot manipulation,” Science, vol. 364, no. 6446, Jun. 2019

2019
[9]

The sensory neurons of touch,

V . Abraira and D. Ginty, “The sensory neurons of touch,” Neuron, vol. 79, no. 4, p. 618–639, Aug. 2013

2013
[10]

A survey of thresholding tech- niques,

P. Sahoo, S. Soltani, and A. Wong, “A survey of thresholding tech- niques,” Computer Vision, Graphics, and Image Processing , vol. 41, no. 2, p. 233–260, Feb. 1988

1988
[11]

Cnn-based methods for object recognition with high-resolution tactile sensors,

J. M. Gandarias, A. J. Garcia-Cerezo, and J. M. Gomez-de Gabriel, “Cnn-based methods for object recognition with high-resolution tactile sensors,” IEEE Sensors Journal , vol. 19, no. 16, p. 6872–6882, Aug. 2019

2019
[12]

Smart tactile sensing systems based on embedded cnn implementations,

M. Alameh, Y . Abbass, A. Ibrahim, and M. Valle, “Smart tactile sensing systems based on embedded cnn implementations,” Micromachines, vol. 11, no. 1, p. 103, Jan. 2020

2020
[14]

Efficientnetv2: Smaller models and faster training,

M. Tan and Q. V . Le, “Efficientnetv2: Smaller models and faster training,” arXiv, 2021

2021
[15]

The organization of digit contact timing during grasping,

L. F. Schettino, A. Pallottie, C. Borland, S. Nessa, A. Nawroj, and Y .-C. Yu, “The organization of digit contact timing during grasping,” Experimental Brain Research , vol. 227, no. 4, p. 477–486, Apr. 2013. 5

2013
[16]

Hand grasping synergies as biometrics,

V . Patel, P. Thukral, M. K. Burns, I. Florescu, R. Chandramouli, and R. Vinjamuri, “Hand grasping synergies as biometrics,” Frontiers in Bioengineering and Biotechnology , vol. 5, May 2017

2017
[17]

Comprehensive review of artificial neural network applications to pattern recognition,

O. I. Abiodun, M. U. Kiru, A. Jantan, A. E. Omolara, K. V . Dada, A. M. Umar, O. U. Linus, H. Arshad, A. A. Kazaure, and U. Gana, “Comprehensive review of artificial neural network applications to pattern recognition,” IEEE Access, vol. 7, p. 158820–158846, 2019

2019
[18]

Mlino bench: A comprehensive benchmarking tool for evaluating ml models on edge devices,

V .-E. Baciu, J. Stiens, and B. da Silva, “Mlino bench: A comprehensive benchmarking tool for evaluating ml models on edge devices,” Journal of Systems Architecture, vol. 155, p. 103262, Oct. 2024

2024
[19]

Current rates of prosthetic usage in upper-limb amputees – have innovations had an impact on device acceptance?

S. Salminger, H. Stino, L. H. Pichler, C. Gstoettner, A. Sturma, J. A. Mayer, M. Szivak, and O. C. Aszmann, “Current rates of prosthetic usage in upper-limb amputees – have innovations had an impact on device acceptance?” Disability and Rehabilitation , vol. 44, no. 14, p. 370...

2020
[20]

The latest research progress on bionic artificial hands: A systematic review,

K. Guo, J. Lu, Y . Wu, X. Hu, and H. Yang, “The latest research progress on bionic artificial hands: A systematic review,” Micromachines, vol. 15, no. 7, p. 891, Jul. 2024

2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [13]

Flexible multilayer tactile sensor on a soft robotic fingertip,

S. Sankar, A. Slepyan, M. M. Iskarous, W.-Y . Cheng, R. DeBrabander, J. Zhang, A. Gupta, and N. V . Thakor, “Flexible multilayer tactile sensor on a soft robotic fingertip,” in 2022 IEEE Sensors . IEEE, Oct. 2022, p. 1–4

work page 2022

[2] [1]

Coding and use of tactile signals from the fingertips in object manipulation tasks,

R. S. Johansson and J. R. Flanagan, “Coding and use of tactile signals from the fingertips in object manipulation tasks,” Nature Reviews Neuroscience, vol. 10, no. 5, p. 345–359, Apr. 2009

work page 2009

[3] [2]

The role of stiffness in versatile robotic grasping,

C. J. Stabile, D. J. Levine, G. M. Iyer, C. Majidi, and K. T. Turner, “The role of stiffness in versatile robotic grasping,” IEEE Robotics and Automation Letters, vol. 7, no. 2, p. 4733–4740, Apr. 2022

work page 2022

[4] [3]

Variable stiffness robotic hand for stable grasp and flexible handling,

S. Mahboubi, S. Davis, and S. Nefti-Meziani, “Variable stiffness robotic hand for stable grasp and flexible handling,” IEEE Access , vol. 6, p. 68195–68209, 2018

work page 2018

[5] [4]

Grasped object stiffness detection for adaptive pid sliding mode position control of a prosthetic hand,

R. Andrecioli and E. D. Engeberg, “Grasped object stiffness detection for adaptive pid sliding mode position control of a prosthetic hand,” in 2012 4th IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics (BioRob). IEEE, Jun. 2012, p. 526–531

work page 2012

[6] [5]

Current-sensor-based con- tact stiffness detection for prosthetic hands,

H. Deng, X. Xu, W. Zhuo, and Y . Zhang, “Current-sensor-based con- tact stiffness detection for prosthetic hands,” IEEE Access , vol. 8, p. 29456–29466, 2020

work page 2020

[7] [6]

Stiffness-estimation-based grasping force fuzzy control for underactuated prosthetic hands,

Y . Zhang, X. Xu, R. Xia, and H. Deng, “Stiffness-estimation-based grasping force fuzzy control for underactuated prosthetic hands,” IEEE/ASME Transactions on Mechatronics , vol. 28, no. 1, p. 140–151, Feb. 2023

work page 2023

[8] [7]

An introductory study of common grasps used by adults during perfor- mance of activities of daily living,

M. Vergara, J. Sancho-Bru, V . Gracia-Ib ´a˜nez, and A. P ´erez-Gonz´alez, “An introductory study of common grasps used by adults during perfor- mance of activities of daily living,” Journal of Hand Therapy , vol. 27, no. 3, p. 225–234, Jul. 2014

work page 2014