Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Neural network modelling of kinematic and dynamic features for signature verification

T0 review · 4 major / 5 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 A simple neural network can estimate a robot's joint motions and torques from signature coordinates alone, and those estimated features verify signatures as well as or better than the robot's own measurements.

desk verdict A genuinely useful paper: real UR5e torque data, a public MLP estimator, and a true cross-database test; the main caveat is that the torque labels are unvalidated current-derived proxies, so the dynamic-feature claim rests on an assumption the paper never checks. read the letter →

arxiv 2411.17506 v1 pith:5RCL4ANW submitted 2024-11-26 cs.LG

Moises Diaz , Miguel A. Ferrer , Jose Juan Quintana , Adam Wolniakowski , Roman Trochimczuk , Konstantsin Miatliuk , Giovanna Castellano , Gennaro Vessio This is my paper · ORCID
classification cs.LG
keywords onlinesignatureverificationkinematicfeaturesdynamictorqueestimationUR5eroboticarmmultilayerperceptrontimewarpingbiometricgeneralization
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

The paper asks whether an automatic signature verifier needs a physical robot to capture arm kinematics and dynamics, or whether a cheap neural network can infer those hidden quantities from the 2D pen trajectory. The authors record a UR5e robot signing 16,500 signatures from MCYT330, collecting six joint angles, six angular velocities, and six joint torques. They then train a multilayer perceptron to map each signature point, together with its local neighborhood, to those eighteen values. The central claim is that the MLP's estimates are accurate enough that a DTW-based verifier using them matches or outperforms one using the robot's direct measurements, and generalizes to five other signature databases the robot never saw.

What carries the argument

The enabling object is a two-stage pipeline. First, a UR5e robotic arm replays signatures using forward and inverse kinematics, with joint torques computed from motor currents via tau_i = r_i * K_I,i * I_i. Second, a multilayer perceptron with one ReLU-hidden layer of twelve units, dropout, and three sigmoid output heads (six units each for angles, velocities, torques) is trained on the robot's recorded sequences; each input is a sliding window of eleven (x,y) points. The DTW-based verifier with Sakoe-Chiba band and two-stage score normalization then consumes these estimated features as function-based signals.

What would settle it

Mount a force-torque sensor on the UR5e's end effector while it replays signatures, compare the sensor-measured torques against the current-derived torques at the same joints, and retrain the MLP on the sensor values; if the verifier's equal error rate does not improve or match, then the current-derived torque space is not a faithful dynamic measure and the paper's headline dynamic feature is unsupported.

Watch

Extended reading notes

Core claim

The paper's central discovery is that angular positions, angular velocities, and force torques of a signing robotic arm can be estimated from plain (x,y) trajectory coordinates by a small MLP with a sliding window, and that these estimates carry enough signer-specific information for online signature verification. The MLP beats recurrent baselines (RNN, LSTM, GRU) on angular velocity and torque estimation, and in the verification experiments the estimated angular velocities and torques achieve lower equal error rates than the features recorded directly from the UR5e robot. The model, trained only on MCYT330, transfers to BiosecurID, Visual, Blind, OnOffSigDevanagari-75, and OnOffSigBengali-75, indicating that the inferred dynamic features are not tied to the robot's particular signing corpus.

Load-bearing premise

The torque labels used to train and judge the network are computed from motor currents by a linear formula supplied by the robot manufacturer, with no independent verification against a force-torque sensor or a physical dynamic model.

Editorial extensions

If this is right

  • If the estimates hold up, signature verification systems can adopt torque-like dynamic features without needing any robotic hardware, since the MLP generates them from standard digitizer coordinates.
  • The estimated angular velocities and torques generalize across capture devices and populations, so a single model trained once can serve multiple databases and writing styles.
  • The fact that estimated features beat measured robot features in verification suggests that the MLP acts as a regularizer, smoothing away robot-specific noise and overfitting.
  • Because the model outputs a full 18-dimensional kinematic-dynamic state per point, the same features could be reused for tasks beyond verification, such as handwriting analysis or robotic trajectory assessment.

Reading between the lines

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

  • The paper compares datasets but does not test whether the MLP's estimated torques correlate with physiological joint loads in humans; a plausible extension is to validate the estimated dynamics against motion-capture and electromyography data from human signers.
  • The sliding-window horizon and network size were chosen by preliminary trials, and the authors do not report sensitivity to these hyperparameters; a systematic ablation would clarify how much of the gain comes from the window versus the architecture.
  • Because the robot held the pen perpendicular to the writing surface, the model may be biased toward that grip; a multi-orientation training set could make the estimated features invariant to pen tilt and improve robustness on natural signing postures.
  • The consistent superiority of estimated torques over robot-measured torques hints that current-derived torque labels contain noise that the MLP cannot reproduce, and that the verifier benefits from this denoising; testing on force-torque-sensor ground truth would separate label noise from genuine dynamic information.
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. The paper proposes a pipeline in which a UR5e robot executes 16,500 MCYT signatures and records joint angular positions, angular velocities, and current-derived force torques, and then an MLP with a sliding-window input is trained to predict these 18 features from (x,y) signature coordinates. The estimated features are used in a DTW-based signature verifier and tested on MCYT (DS1) and five additional databases (BiosecurID, Visual, Blind, OnOffSigDevanagari-75, OnOffSigBengali-75). The main claims are that the MLP estimates the robot features accurately and that the estimated features support signature verification as well as or better than the robot's own measurements, including on databases never seen by the robot.

Significance. If the claims hold, the paper offers a low-cost way to obtain arm-joint kinematic and torque-like dynamic features for online signature verification, with a genuinely cross-database out-of-sample transfer experiment (DS2-DS6) and a public release of the trained model. The comparison against RNN, LSTM, and GRU baselines is also a useful contribution. However, the significance is conditional on validating the torque labels, correcting the theta/omega labeling confusion, and making the EER results numerically accessible; as written, the dynamic-feature novelty rests on unvalidated motor-current torque estimates.

major comments (4)
  1. [Section 3.1] The torque labels are the only ground truth for the dynamic-feature head, but they are not validated. The paper computes tau_i = r_i * K_I,i * I_i from motor currents and explicitly states that the UR5e has no built-in force-torque sensors; no comparison to an external force-torque sensor, an independent inverse-dynamics model, or any physical consistency check is reported. Because the controller parameters were modified (kgain=2000, tlookahead=0.03 s), the current-to-torque relation may not reflect true joint torques. This is load-bearing for the paper's dynamic-feature claim: if the current-derived torques are inaccurate proxies, the MLP learns an unvalidated proxy, and the torque-based results in Table 1, Figures 2-4, and Section 4.5 inherit that uncertainty. Please validate the torque labels against a sensor or an independent dynamic model, or substantially soften the dynamic-feature claims.
  2. [Table 1] The row labels in Table 1 are swapped relative to the definitions used in the rest of the paper. The introduction and Section 3.2 define theta as angular position and omega as angular velocity, and Section 4.2 uses theta_r for position and omega_r for velocity; however, Table 1 lists 'Angular velocities, theta(t)' and 'Angular positions, omega(t)'. This reverses the theta/omega rows and makes the result narrative in Section 4.4 ('MLP significantly outperforms the RNN, LSTM, and GRU models in estimating angular velocities and force torques') inaccurate or at least unverifiable. Please correct the labels and re-check the claims that depend on them.
  3. [Section 4.5] The verification results are reported only as DET curves. Section 4.3 promises that 'the EER is reported as the average +/- standard deviation,' but no table contains EER values for DS1-DS6, for UR5e versus MLP, or for random versus skilled forgeries. As a result, the claims in Section 4.5 (e.g., 'omega and tau delivered results comparable to DS1' for DS2, and 'angular velocities and force torques showing improved EERs') cannot be checked quantitatively. Please include a table with mean +/- std EER for each database, feature type (theta, omega, tau), source (UR5e vs MLP), and forgery scenario.
  4. [Sections 4.3 and 4.5] The DS1 verification comparison is not an independent test of generalization. The MLP was trained on DS1 (Section 4.3), and the four-fold procedure still allows signatures from the same writers to appear in the training set when estimating features for the held-out quarter. The genuinely out-of-sample evidence is the DS2-DS6 transfer reported in Section 4.5; the paper should state this explicitly and avoid presenting the DS1 comparison as evidence of generalization. If the authors want DS1 to serve as an in-database test, a writer-disjoint split should be used.
minor comments (5)
  1. [Abstract and Section 1] The dataset name is inconsistent: the abstract says MCYT300, while the contributions in Section 1 say MCYT330; please use one consistent name with the correct corpus identifier.
  2. [Section 3.2] The sentence introducing the estimation task repeats the theta/omega swap: it says 'estimating the angular velocities theta(t), angular positions omega(t)', whereas the paper's own definitions make theta the angular position and omega the angular velocity.
  3. [Figure 2] The subcaption for (p) reads 'tau4 t)' and is missing the closing parenthesis; also, Section 4.2 states that omega6 is excluded from verification, but Figure 2 displays omega6, so the figure should note that this channel is shown for completeness but excluded from the verifier.
  4. [Section 3.2] The sliding-window horizon and MLP hyperparameters are described only as chosen based on 'preliminary trials'; a short sensitivity analysis would strengthen the claim that the simple MLP is robust and would let readers assess the effect of the window length.
  5. [References] Reference [5] is incomplete (it lacks full author names), and reference [16] is an arXiv preprint rather than the published conference version; please update the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MLP mapping is learned from robot targets, and the cross-database experiments provide independent evidence.

full rationale

The paper's central derivation is an empirical supervised mapping, not an analytic one: an MLP is trained on input (x,y) trajectories to regress robot-recorded angular positions, angular velocities, and current-derived torques. Because the target values are obtained from a separate physical/control pipeline in Section 3.1, and the input features are only the signature coordinates, the estimation task is not definitionally equivalent to its output; the mapping could fail and is evaluated by MAE/MSE in Table 1. The DS1 comparison of MLP-estimated versus UR5e features is in-sample in the sense of using the same database, but the paper uses a four-fold split with early stopping, and the genuine generalization evidence is the transfer to DS2-DS6, which were never processed by the robot. The self-citations to [7] and [19] concern feature normalization and prior robot writing demonstrations; neither is used as a uniqueness theorem or as a substitute for the present experiments, so they are not load-bearing. The concern that torque labels from motor currents (tau_i = r_i * K_I,i * I_i) are unvalidated against a force-torque sensor is a measurement-validity and correctness issue rather than circularity, because the labels do not depend on the MLP outputs or on the verification result. Overall, the claimed generalization is independently testable and is not forced by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on trusting current-derived torque labels, assuming one robot represents human signing, and assuming a learnable coordinate-to-feature map. No new physical entities are introduced.

free parameters (2)
  • sliding window horizon = 5 preceding and 5 succeeding points
    Chosen based on effectiveness in preliminary trials (Section 3.2), not derived from theory.
  • MLP hyperparameters = 12 hidden units, dropout 0.3, learning rate 0.01, patience 1
    Set by common design choices (Section 3.2, 4.3) and not tuned on a held-out test set.
assumptions (4)
  • domain assumption Torque estimated from motor currents is a valid proxy for true joint torque
    Section 3.1 defines tau_i = r_i * K_I,i * I_i with no force-torque sensor validation; the paper treats these values as ground truth labels.
  • domain assumption Pen held perpendicular to the writing surface captures signature dynamics
    Section 3.1 assumes the pen is perpendicular to the surface, ignoring human pen tilt and azimuth variations.
  • domain assumption A single UR5e robot configuration represents all signers
    Section 2.3 and the conclusions note that individual arm and forearm anatomy is not modeled; the entire feature space is robot-specific.
  • domain assumption The (x,y)-to-feature mapping is learnable and transferable across databases
    Section 3.2 and 4.5 assume that a 22-dimensional local coordinate window determines 18 robot outputs across different digitizers, scripts, and writing conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural network modelling of kinematic and dynamic features for signature verification." pith.science (2026). https://pith.science/paper/5RCL4ANW

@misc{pith2026241117506,
  author       = {Pith},
  title        = {Pith review of: Neural network modelling of kinematic and dynamic features for signature verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5RCL4ANW}},
  note         = {Machine review of arXiv:2411.17506}
}
read the original abstract

Online signature parameters, which are based on human characteristics, broaden the applicability of an automatic signature verifier. Although kinematic and dynamic features have previously been suggested, accurately measuring features such as arm and forearm torques remains challenging. We present two approaches for estimating angular velocities, angular positions, and force torques. The first approach involves using a physical UR5e robotic arm to reproduce a signature while capturing those parameters over time. The second method, a cost effective approach, uses a neural network to estimate the same parameters. Our findings demonstrate that a simple neural network model can extract effective parameters for signature verification. Training the neural network with the MCYT300 dataset and cross validating with other databases, namely, BiosecurID, Visual, Blind, OnOffSigDevanagari 75 and OnOffSigBengali 75 confirm the models generalization capability.

Figures

Figures reproduced from arXiv: 2411.17506 by the authors.

Figure 2
Figure 2. Visual representation of the sequences of angular positions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Performance results across different databases, trained with DS1 data, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

  1. [1]

    Bhavani, S., & Bharathi, R. (2024). A multi-dimensional review on hand- written signature verification: strengths and gaps. Multimedia Tools and Applications, 83, 2853–2894

    work page 2024

  2. [2]

    J., Naser, A., & Lotfi, A

    Bird, J. J., Naser, A., & Lotfi, A. (2023). Writer-independent signature verification; evaluation of robotic and generative adversarial attacks. In- formation Sciences, 633, 170–181

    work page 2023

  3. [3]

    Chandra, S., Dillenbourg, P., & Paiva, A. (2020). Children teach hand - writing to a social robot with different learning competencies. International Journal of Social Robotics, 12, 721–748

    work page 2020

  4. [4]

    Corke, P. (2023). Representing position and orientation. In Robotics, Vision and Control: Fundamental Algorithms in Python (pp. 23–86)

    work page 2023

  5. [5]

    Introduction to robotics : mechanics & control / John J

    Craig, (1986). Introduction to robotics : mechanics & control / John J. Craig.. Reading, Mass.: Addison-Wesley Pub. Co.,. Includes bibliographies and index

    work page 1986

  6. [6]

    Deng, Z., Huang, L., & Wang, C. (2023). Enhanced in-air signature verification via hand skeleton tracking to defeat robot-level replays. In 39th Annual Computer Security Applications Conf. (pp. 451–462)

    work page 2023

  7. [7]

    A., & Quintana, J

    Diaz, M., Ferrer, M. A., & Quintana, J. J. (2019). Anthropomorphic features for on -line signatures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41, 2807–2819

    work page 2019

  8. [8]

    Diaz, M. et al. (2021). Sequence-based dynamic handwriting analysis for Parkinson’s disease detection with one-dimensional convolutions and BiGRUs. Expert Systems with Applications, 168, 114405

    work page 2021

Show all 29 references
  1. [9]

    A., Diaz, M., Tolosana, R., & Plamondon, R

    Faundez-Zanuy, M., Fierrez, J., Ferrer, M. A., Diaz, M., Tolosana, R., & Plamondon, R. (2020). Handwriting biometrics: Applications and future trends in e-security and e-health. Cognitive Computation, 12, 940–953

    2020

  2. [10]

    Ferrer, M. A. et al. (2017). Static and dynamic synthesis of bengali and devanagari signatures. IEEE transactions on cybernetics, 48, 2896–2907

    2017

  3. [11]

    Fischer, A., Diaz, M., Plamondon, R., & Ferrer, M. A. (2015). Robust score normalization for DTW-based on-line signature verification. In 13th Int. Conf. on document analysis and recognition (ICDAR) (pp. 241–245)

    2015

  4. [12]

    Galbally, J. et al. (2015). On-line signature recognition through the combination of real dynamic data and synthetically generated static data. Pattern Recogn., 48, 2921–2934

    2015

  5. [13]

    Giorgi, G., Saracino, A., & Martinelli, F. (2021). Using recurrent neural networks for continuous authentication through gait analysis. Pattern Recognition Letters, 147, 157–163

    2021

  6. [14]

    M., Alwais, S., Abdi, H., & Nahavandi, S

    Kebria, P. M., Alwais, S., Abdi, H., & Nahavandi, S. (2016). Kinematic and dynamic modelling of UR5 manipulator. In IEEE Int. Conf. on Sys- tems, Man, and Cybernetics (SMC) (pp. 004229–004234)

    2016

  7. [15]

    Kholmatov, A., & Yanikoglu, B. (2009). SUSIG: an on-line signature database, associated protocols and benchmark results. Pattern Analysis and Applications, 12, 227–236

    2009

  8. [16]

    P., & Ba, J

    Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

    2014 arXiv

  9. [17]

    Lei, H., & Govindaraju, V. (2005). A comparative study on the consistency of features in on-line signature verification. Pattern Recognition Letters, 26, 2483–2489

    2005

  10. [18]

    Lin, Y.-F. et al. (2020). Forecasting fluctuations in the financial index using a recurrent neural network based on price features. IEEE Tran. on Emerging Topics in Computational Intell., 5, 780–791

    2020

  11. [19]

    Miatliuk, K. et al. (2019). Universal robot employment to mimic human writing. In 20th Int. Carpathian Control Conf. (ICCC) (pp. 1–5). IEEE

    2019

  12. [20]

    Nalwa, V. S. (1997). Automatic on-line signature verification. Proceed- ings of the IEEE, 85, 215–239

    1997

  13. [21]

    Ortega-Garcia, J. et al. (2003). MCYT baseline corpus: a bimodal bio - metric database. IEE Proceedings-Vision, Image and Signal Processing, 150, 395–401

    2003

  14. [22]

    E., & Ugur, E

    Seker, M., Tekden, A. E., & Ugur, E. (2019). Deep effect trajectory pre- diction in robot manipulation. Robotics & Autono. Syst., 119, 173–184

    2019

  15. [23]

    Singhal, M., & Shinghal, K. (2024). Secure deep multimodal biometric authentication using online signature and face features fusion. Multimedia Tools and Applications, 83, 30981–31000

    2024

  16. [24]

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdi - nov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15, 1929–1958

    2014

  17. [25]

    Tan, H., He, L., Huang, Z.-C., & Zhan, H. (2024). Online signature ver- ification based on dynamic features from gene expression programming. Multimedia Tools and Applications, 83, 15195–15221

    2024

  18. [26]

    Tolosana, R. et al. (2021). ICDAR 2021 competition on on-line signature verification. In Document Analysis and Recognition– ICDAR 2021: 16th Int. Conf., Proceedings, Part IV (pp. 723–737). Springer. 11

    2021

  19. [27]

    Tresset, P., & Leymarie, F. F. (2013). Portrait drawing by paul the robot. Computers & Graphics, 37, 348–363

    2013

  20. [28]

    Woz´niak, M., Wieczorek, M., Siłka, J., & Połap, D. (2020). Body pose prediction based on motion sensor data and recurrent neural network. IEEE Trans. on Indust. Informatics, 17, 2101–2111

    2020

  21. [29]

    Xu, D. et al. (2019). Survey on multi-output learning. IEEE transactions on neural networks and learning systems , 31, 2409– 2429

    2019

Pith tools

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