Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Robot Tactile Gesture Recognition Based on Full-body Modular E-skin

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Equivariant GNN reads robot skin touch at 91% accuracy.

desk verdict A genuine full-body e-skin system with an equivariant GNN recognizer that runs in real time; the central accuracy claim lacks significance testing and the robustness claims are not experimentally supported. read the letter →

arxiv 2506.18256 v1 pith:B5YIJ3SM submitted 2025-06-23 cs.RO

Shuo Jiang , Boce Hu , Linfeng Zhao , Lawson L.S. Wong This is my paper
classification cs.RO
keywords tactilegesturerecognitionelectronicskinequivariantgraphneuralnetworkSE(3)equivariancehuman-robotinteractionmodulare-skinsensinggeometricdeeplearning
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 develops a modular electronic skin for a robot arm and a recognizer that classifies human touch gestures as commands. The central claim is that an equivariant graph neural network which incorporates the real-time 3D pose of each sensing point along with its pressure reading can recognize poke, double-pat, grab, and stroke gestures accurately and quickly enough for real-time human-robot interaction. On the authors' own 581-gesture dataset the recognizer reaches 91.1 percent test accuracy, beating a CNN-LSTM baseline at 80.4 percent and traditional machine-learning baselines. The practical upshot is that a robot can be commanded purely through touch, with gestures mapped to actions such as pausing, proceeding, or closing the gripper.

What carries the argument

The central mechanism is the Equivariant Graph Convolutional Layer (EGCL), an SE(3)-equivariant message-passing layer for physical systems. The paper combines it with a dynamic K-nearest-neighbors edge construction based on kinematic distance, so the graph structure updates as the robot moves, while edges inside each skin patch stay static to cut computation. Node features are pressure plus real-time pose, which lets the network use deformation rather than only pressure images to tell gestures apart.

What would settle it

Collect a tactile gesture dataset while the robot is moving, or compare against a shear-sensitive skin variant: if the equivariant GNN's margin over the CNN-LSTM baseline shrinks or disappears when shear or motion is present, then the pose-plus-pressure assumption is the limiting factor.

Watch

Extended reading notes

Core claim

The paper's discovery is that treating tactile gesture recognition as an SE(3)-equivariant graph problem makes the classifier insensitive to skin patch shape, sensor density, and robot-pose-dependent deformation. Each sensor is a graph node carrying scalar pressure plus its real-time pose from forward kinematics; edges are built by kinematic (Euclidean) distance, with static edges precomputed within patches and dynamic KNN edges added only around triggered sensors. The equivariant graph convolutional layers preserve the spatial symmetry of the interaction, so the model generalizes from fewer samples than non-equivariant baselines. The result is a full-body e-skin system for a UR5 robot, with 2112 sensing points, that recognizes four gestures at 50 Hz and runs each prediction step in about 3.8 ms.

Load-bearing premise

The recognizer assumes that normal pressure readings plus 3D sensor positions carry enough information to tell all four gestures apart, because the skin measures no shear forces and the local x- and y-axes are set arbitrarily.

Editorial extensions

If this is right

  • A robot can be controlled entirely by touch, with poke mapped to pause or resume, double-pat to advance to the next waypoint, and grab to close or release the gripper.
  • Because the recognizer uses graph geometry rather than a fixed grid, the same approach can be mounted on differently shaped skins without retraining for patch shape or density.
  • The single-step runtime of about 3.8 ms is fast enough for real-time 50 Hz tactile command on the robot.
  • Equivariance lets the model learn from fewer labeled gestures by exploiting the SE(3) symmetry of touch interactions.
  • Static edges within skin patches plus dynamic KNN edges only around triggered sensors cut computation enough for full-body coverage.

Reading between the lines

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

  • A natural next test is a dataset collected while the robot moves during gestures; the current data keeps the robot stationary, so the dynamic graph construction is not yet stress-tested under real motion.
  • Adding shear or tangential force channels would likely make stroke and directional gestures easier to tell apart, since the current local x- and y-axes are arbitrary without shear sensing.
  • The architecture depends only on node poses and kinematic distances, so it could be reused as a standard interface for any full-body e-skin layout, regardless of patch geometry or sensor density.
  • The threshold-based auto-labeling with a minimum window size may mislabel gesture boundaries, so a more fine-grained labeling procedure could push the reported accuracies upward.
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.

Referee Report

4 major / 5 minor

Summary. This paper presents a modular robot e-skin with 2112 piezoresistive sensing points mounted on a UR5 robot arm, together with a tactile gesture recognizer based on an E(n)-equivariant graph neural network. Sensor nodes carry pressure readings and real-time positions from forward kinematics; edges are constructed by KNN on kinematic distance, with static in-patch edges precomputed and dynamic edges for activated sensors. The recognizer classifies poke, double-pat, grab, and stroke, and the authors report a validation accuracy of 96.1% and a test accuracy of 91.1%, outperforming CNN-LSTM, SVM, KNN, and MLP baselines on their dataset, with a one-step runtime of 3.81 ms. A human-robot collaboration demonstration maps gestures to robot actions.

Significance. The hardware and system integration are a concrete contribution: the modular skin supports irregular patch shapes and the GNN architecture is a sensible match to non-uniform taxel layouts. If the reported gains are reproducible and statistically sound, the work would demonstrate a practical full-body tactile control interface. The paper's strengths include real-time runtime evaluation, a real robot demonstration, and use of a principled equivariant layer. However, the central empirical claim is currently not supported by the statistics or by the experimental scope of the robustness assertions.

major comments (4)
  1. [Table II, Section V.B] The claim that the GNN 'significantly outperforms' baselines is not supported by the reported results. The test set contains only 95 samples; with test accuracies of 91.1% and 80.4%, the 95% Wilson confidence intervals are approximately [84.3%, 96.3%] and [71.7%, 87.9%], which overlap. The paper reports no error bars, no repeated seeds, and no paired significance test such as McNemar's test. Additionally, Section V.A and Fig. 11 show that the pooling method was selected using test-set accuracy, which can bias the reported 91.1% upward. Please provide confidence intervals or repeated cross-validation, a valid paired test, and a test set used only for final evaluation.
  2. [Section IV.D, IV.A.3] The robustness claim regarding robot motion is not tested by the data. Section IV.D states that during data collection 'the robot remains stationary in each pose,' so all training and test graphs come from static configurations. Section IV.A.3 argues that deformation patterns help distinguish gestures, and the Introduction claims the method is 'unaffected by geometric changes due to the robot's movements,' but no experiment involves motion during a gesture. Either collect data with the robot moving, or explicitly limit the claim to stationary poses with arbitrary arm configuration.
  3. [Section IV.C.1, Fig. 3] The sensor provides only normal pressure, and the local x- and y-axes are set arbitrarily. This raises a correctness risk for gestures such as stroke, where tangential motion may be the discriminative signal. The paper should show a per-gesture confusion matrix and, ideally, an ablation or additional sensor modality to demonstrate that normal pressure plus 3D sensor positions are sufficient to distinguish all four gestures. Without this, the generalization of the reported accuracy to natural touching behavior is uncertain.
  4. [Section IV.A.1, IV.A.2, V] The claims that the method is insensitive to skin size, shape, and density are not demonstrated empirically. The experiments use one fixed skin layout on one robot arm; there is no evaluation across patches of different density, shape, or coverage. Please either provide such an evaluation or revise the claims to state that these properties are architectural expectations rather than validated results.
minor comments (5)
  1. [Section IV.B] The phrase 'limited be the recognizer's speed' should read 'limited by the recognizer's speed'.
  2. [Table I] Comparing accuracy numbers across different robots, gesture sets, and sensor layouts is not a meaningful quantitative comparison; please mark the table as qualitative or add a column with dataset and protocol characteristics.
  3. [Fig. 3 caption] The caption says 'both cases' but the figure has two rows; please specify which ambiguity corresponds to the top row and which to the bottom row.
  4. [Section IV.C.3] The text 'knearest neighbors' should be 'k nearest neighbors'.
  5. [Reference [23] paragraph] The sentence 'the solutions has limitation' should be 'the solution has limitations'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy comparison is empirical and self-contained; the sole self-citation is motivational, not load-bearing.

full rationale

No circular step is present. The equivariant graph convolutional layer is imported from [25], an external publication with a proof of SE(3)-equivariance, and the paper's contribution is the application of that layer to tactile gesture recognition on a custom full-body e-skin. The only self-citation, [26], appears in Section IV.C.3 as motivational support that equivariant models 'demonstrated efficiency in leveraging scarce data in robot applications'; it is not used to justify the architecture, the dataset, or the reported accuracy numbers. The central claim is an empirical comparison in Table II: the proposed GNN is trained on 581 labeled gestures and evaluated on a held-out test set of 95 gestures, and this is a standard evaluation procedure rather than a fitted parameter renamed as a prediction. The geometric and kinematic distance edge features are defined from robot kinematics and sensor geometry, not derived from the gesture labels, so the classification does not reduce to a self-definitional input. Section V.A's use of test-set accuracy to justify the choice of max pooling is a methodological concern about optimistic bias (and the lack of a significance test is a statistical concern), but it is not circularity: the chosen pooling method does not make the reported 91.1% test accuracy equal to a training objective by construction. The arbitrary x- and y-axes for sensor frames, noted in Section IV.C.1, are also a modeling assumption about sensor modalities, not a circular derivation. Overall, the paper's derivation chain is self-contained and no claim is forced by self-citation or by definition.

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

The central result is empirical and rests on a private dataset plus several unreported thresholds and architecture details. The shear-free assumption is the most consequential domain assumption because it is load-bearing for stroke and grab recognition and is not validated experimentally.

free parameters (5)
  • KNN neighbor count K = 8
    Selected based on validation accuracy (Fig. 11); lower K chosen for speed.
  • FIFO time window = 2 seconds
    Assumes all tactile gestures are completed within 2 seconds (Section IV.C.3).
  • Activation threshold for dynamic connections = not reported
    Used to select triggered sensors for real-time KNN (Section IV.C.4); value not given.
  • Auto-labeling threshold and minimum window size = not reported
    Used in threshold-based automatic labeling of gesture segments (Section IV.D).
  • GNN architecture hyperparameters = not reported
    Hidden dimensions, number of layers, learning rate are not specified in the paper.
assumptions (5)
  • domain assumption The robot skin bends but does not stretch, so geometric distances between sensors remain constant.
    Stated in Section IV.A.5 to justify focusing on kinematic distance.
  • domain assumption Tactile gestures are typically completed within 2 seconds.
    Used to set the FIFO buffer window in Section IV.C.3.
  • domain assumption Normal pressure and 3D positions are sufficient to recognize gestures; x/y axes can be arbitrary.
    No shear force is measured (Section IV.C.1); this is a structural input to the recognizer.
  • standard math The EGCL layer is SE(3)-equivariant as proven in [25].
    Relied upon in Section III.B to justify the architecture.
  • ad hoc to paper A pressure threshold separates gesture from no-gesture within manually marked ranges.
    Used for automatic labeling in Section IV.D; threshold values are not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robot Tactile Gesture Recognition Based on Full-body Modular E-skin." pith.science (2026). https://pith.science/paper/B5YIJ3SM

@misc{pith2026250618256,
  author       = {Pith},
  title        = {Pith review of: Robot Tactile Gesture Recognition Based on Full-body Modular E-skin},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5YIJ3SM}},
  note         = {Machine review of arXiv:2506.18256}
}
read the original abstract

With the development of robot electronic skin technology, various tactile sensors, enhanced by AI, are unlocking a new dimension of perception for robots. In this work, we explore how robots equipped with electronic skin can recognize tactile gestures and interpret them as human commands. We developed a modular robot E-skin, composed of multiple irregularly shaped skin patches, which can be assembled to cover the robot's body while capturing real-time pressure and pose data from thousands of sensing points. To process this information, we propose an equivariant graph neural network-based recognizer that efficiently and accurately classifies diverse tactile gestures, including poke, grab, stroke, and double-pat. By mapping the recognized gestures to predefined robot actions, we enable intuitive human-robot interaction purely through tactile input.

Figures

Figures reproduced from arXiv: 2506.18256 by the authors.

Figure 1
Figure 1. Our modular robot skin system enables customized robot [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Interactive tactile gestures investigated in this work: (a) poke; [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. (a) irregular skin patch; (b) closer view at patch mounting. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Example of geometric distance and kinematic distance. IV. TACTILE GESTURE RECOGNITION In this work, we primarily address the problem of how to perceive and recognize tactile gestures applied to a robot equipped with full-body electronic skin. We will introduce our robo…
Figure 6
Figure 6. Figure 6: (a) Skin perception system; (b) flexible skin patch fabricating; (c) Connector design. [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 8
Figure 8. Figure 8: (a) Dynamic edge connection by real-time KNN. The [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: During data collection, the robot remains stationary in some [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 10
Figure 10. Figure 10: Threshold-based automatic data labeling. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Classification accuracy: (left) on validation set; (right) on [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 12
Figure 12. Figure 12: (a) Human-robot collaboration task; Mapping tactile gestures to robot actions: (b) double-pat: to the next waypoint; (c) grab: [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    The role of affective touch in human- robot interaction: Human intent and expectations in touching the haptic creature,

    S. Yohanan and K. E. MacLean, “The role of affective touch in human- robot interaction: Human intent and expectations in touching the haptic creature,”International Journal of Social Robotics, vol. 4, pp. 163– 180, 2012

    work page 2012

  2. [2]

    Friends and pets as companions: Strategies for coping with loneliness among homeless youth,

    L. Rew, “Friends and pets as companions: Strategies for coping with loneliness among homeless youth,”Journal of child and adolescent psychiatric nursing, vol. 13, no. 3, pp. 125–132, 2000

    work page 2000

  3. [3]

    Cardiovascular effects of human-pet dog interactions,

    J. K. V ormbrock and J. M. Grossberg, “Cardiovascular effects of human-pet dog interactions,”Journal of behavioral medicine, vol. 11, pp. 509–517, 1988

    work page 1988

  4. [4]

    An event-triggered low-cost tactile perception system for social robot’s whole body interaction,

    S. Lin, J. Su, S. Song, and J. Zhang, “An event-triggered low-cost tactile perception system for social robot’s whole body interaction,” IEEE Access, vol. 9, pp. 80 986–80 995, 2021

    work page 2021

  5. [5]

    Deep learning classification of touch gestures using distributed normal and shear force,

    H. Choi, D. Brouwer, M. A. Lin, K. T. Yoshida, C. Rognon, B. Stephens-Fripp, A. M. Okamura, and M. R. Cutkosky, “Deep learning classification of touch gestures using distributed normal and shear force,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 3659–3665

    work page 2022

  6. [6]

    Large-area conformable sensor for proximity, light touch, and pressure-based gesture recognition,

    M. S. Sarwar and K. Yamane, “Large-area conformable sensor for proximity, light touch, and pressure-based gesture recognition,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 2700–2707

    work page 2021

  7. [7]

    Automatic recognition of touch gestures in the corpus of social touch,

    M. M. Jung, M. Poel, R. Poppe, and D. K. Heylen, “Automatic recognition of touch gestures in the corpus of social touch,”Journal on multimodal user interfaces, vol. 11, pp. 81–96, 2017

    work page 2017

  8. [8]

    Bioinspired co-design of tactile sensor and deep learning algorithm for human–robot interaction,

    D. Kong, G. Yang, G. Pang, Z. Ye, H. Lv, Z. Yu, F. Wang, X. V . Wang, K. Xu, and H. Yang, “Bioinspired co-design of tactile sensor and deep learning algorithm for human–robot interaction,”Advanced Intelligent Systems, vol. 4, no. 6, p. 2200050, 2022

    work page 2022

Show all 27 references
  1. [9]

    Categories of touch: Classifying human touch using a soft tactile sensor,

    J. Sun, E. Billing, F. Seoane, B. Zhou, D. H ¨ogberg, and P. Hemeren, “Categories of touch: Classifying human touch using a soft tactile sensor,” inThe robotic sense of touch: From sensing to understanding, workshop at the IEEE International Conference on Robotics and Automati...

    2017

  2. [10]

    Detecting and classifying human touches in a social robot through acoustic sensing and machine learning,

    F. Alonso-Mart ´ın, J. J. Gamboa-Montero, J. C. Castillo, ´A. Castro- Gonz´alez, and M. ´A. Salichs, “Detecting and classifying human touches in a social robot through acoustic sensing and machine learning,”Sensors, vol. 17, no. 5, p. 1138, 2017

    2017

  3. [11]

    Enable natural tactile interaction for robot dog based on large-format distributed flexible pressure sensors,

    L. Zhan, Y . Cao, Q. Chen, H. Guo, J. Gao, Y . Luo, S. Guo, G. Zhou, and J. Gong, “Enable natural tactile interaction for robot dog based on large-format distributed flexible pressure sensors,”arXiv preprint arXiv:2303.07595, 2023

    2023 arXiv

  4. [12]

    Organization and understanding of a tactile information dataset tacact for physical human-robot interaction,

    P. Wang, J. Liu, F. Hou, D. Chen, Z. Xia, and S. Guo, “Organization and understanding of a tactile information dataset tacact for physical human-robot interaction,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 7328– 7333

    2021

  5. [13]

    Robust recognition of tactile gestures for intuitive robot programming and control,

    D. Kubus, A. Muxfeldt, K. Kissener, J. Haus, and J. Steil, “Robust recognition of tactile gestures for intuitive robot programming and control,” in2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 1643–1650

    2017

  6. [14]

    Textile pressure mapping sensor for emotional touch detection in human-robot inter- action,

    B. Zhou, C. A. Velez Altamirano, H. Cruz Zurian, S. R. Atefi, E. Billing, F. Seoane Martinez, and P. Lukowicz, “Textile pressure mapping sensor for emotional touch detection in human-robot inter- action,”Sensors, vol. 17, no. 11, p. 2585, 2017

    2017

  7. [15]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veli ˇckovi´c, “Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,”arXiv preprint arXiv:2104.13478, 2021

    2021 arXiv

  8. [16]

    Pressure distribution classification and segmentation of human hands in contact with the robot body,

    A. Albini and G. Cannata, “Pressure distribution classification and segmentation of human hands in contact with the robot body,”The International Journal of Robotics Research, vol. 39, no. 6, pp. 668– 687, 2020

    2020

  9. [17]

    Tactile images generation from contacts involving adjacent robot links,

    ——, “Tactile images generation from contacts involving adjacent robot links,” in2018 27th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN). IEEE, 2018, pp. 306–312

    2018

  10. [18]

    Tactilegcn: A graph convolutional network for predicting grasp stability with tactile sensors,

    A. Garcia-Garcia, B. S. Zapata-Impata, S. Orts-Escolano, P. Gil, and J. Garcia-Rodriguez, “Tactilegcn: A graph convolutional network for predicting grasp stability with tactile sensors,” in2019 International Joint Conference on Neural Networks (IJCNN). IEEE, 2019, pp. 1–8

    2019

  11. [19]

    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. Sug- ano, 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, p...

    2022

  12. [20]

    Tacgnn: Learning tactile-based in-hand manipulation with a blind robot using hierarchical graph neural network,

    L. Yang, B. Huang, Q. Li, Y .-Y . Tsai, W. W. Lee, C. Song, and J. Pan, “Tacgnn: Learning tactile-based in-hand manipulation with a blind robot using hierarchical graph neural network,”IEEE Robotics and Automation Letters, vol. 8, no. 6, pp. 3605–3612, 2023

    2023

  13. [21]

    Graph neural networks for interpretable tactile sensing,

    W. Fan, H. Bo, Y . Lin, Y . Xing, W. Liu, N. Lepora, and D. Zhang, “Graph neural networks for interpretable tactile sensing,” in2022 27th International Conference on Automation and Computing (ICAC). IEEE, 2022, pp. 1–6

    2022

  14. [22]

    Tac- tile object property recognition using geometrical graph edge features and multi-thread graph convolutional network,

    S. Kulkarni, S. Funabashi, A. Schmitz, T. Ogata, and S. Sugano, “Tac- tile object property recognition using geometrical graph edge features and multi-thread graph convolutional network,”IEEE Robotics and Automation Letters, 2024

    2024

  15. [23]

    Intrinsic sense of touch for intuitive physical human-robot interaction,

    M. Iskandar, A. Albu-Sch ¨affer, and A. Dietrich, “Intrinsic sense of touch for intuitive physical human-robot interaction,”Science Robotics, vol. 9, no. 93, p. eadn4008, 2024. [Online]. Available: https://www.science.org/doi/abs/10.1126/scirobotics.adn4008

    2024 doi

  16. [24]

    Sample efficient grasp learning using equivariant models,

    X. Zhu, D. Wang, O. Biza, G. Su, R. Walters, and R. Platt, “Sample efficient grasp learning using equivariant models,”arXiv preprint arXiv:2202.09468, 2022

    2022 arXiv

  17. [25]

    E (n) equivariant graph neural networks,

    V . G. Satorras, E. Hoogeboom, and M. Welling, “E (n) equivariant graph neural networks,” inInternational conference on machine learn- ing. PMLR, 2021, pp. 9323–9332

    2021

  18. [26]

    Orbitgrasp: SE(3)-equivariant grasp learning,

    B. Hu, X. Zhu, D. Wang, Z. Dong, H. Huang, C. Wang, R. Walters, and R. Platt, “Orbitgrasp: SE(3)-equivariant grasp learning,” in8th Annual Conference on Robot Learning, 2024

    2024

  19. [27]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Van- derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,”Journal of Machine Lea...

    2011

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.