{"id":"51ef5d48-4b6e-4812-b539-3142efe55815","arxiv_id":"2411.17506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An MLP trained on UR5e robot signatures estimates kinematic and dynamic joint features from signature coordinates and supports competitive signature verification across multiple databases.","lead":"Researchers trained a small neural network to estimate a robot arm's joint angles, speeds, and force torques from the 2D coordinates of a handwritten signature. The estimated features match or beat the robot's own measurements in automatic signature verification and generalize to five other signature databases.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated motor-current torque labels in Section 3.1 undermine the dynamic-feature claim; if the current-to-torque conversion is inaccurate, the paper's headline dynamic contribution is unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: torque labels are computed from motor currents without independent validation. I agree that this is the most consequential issue. The paper's advertised contribution is not just kinematic estimation but the introduction of force torques as a dynamic feature space; if the torque labels are inaccurate, the MLP's low MAE and the downstream verification performance for torque features are only evidence of fitting to an arbitrary current-derived signal, not to physically meaningful dynamics. This does not require rejecting the paper, because the concern is empirically addressable: a sensor-based or model-based validation study would either confirm the current-derived labels or show how much they deviate. The paper already has a conditional verdict, and this stress-test identifies a concrete validation step that should be part of the acceptance condition. I also note the systematic θ/ω notation swaps in Section 3.2 and Table 1, but these appear to be typographical inconsistencies rather than a distinct load-bearing flaw: the numerical results and text can be read consistently once the variable-name convention is corrected. The unvalidated torque labels, by contrast, directly affect the truth of the central claim and deserve the primary focus.","tokens_in":9499,"tokens_out":7016,"duration_ms":73441,"concrete_test":"Replay a subset of MCYT330 signatures on the UR5e while mounting a six-axis force/torque sensor at the end effector or using the robot controller's independently estimated joint torques, and compare these measurements against the current-derived torques tau_i = r_i * K_I,i * I_i from Section 3.1. Report per-joint correlation, mean absolute error, and normalized RMS error across the trajectory. If the discrepancy exceeds a pre-specified tolerance, such as 10% RMS after removing offsets, or shows systematic bias correlated with acceleration, the dynamic ground truth is unsupported and the paper's central dynamic-feature claim requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the introduction of force torques as a new dynamic feature space, and the MLP is trained and evaluated against torque labels computed by Section 3.1 as tau_i = r_i * K_I,i * I_i. These labels are derived from motor currents with manufacturer-calibrated coefficients, and the paper explicitly states that the UR5e has no built-in force-torque sensors. No independent validation against a force-torque sensor, a calibrated dynamic model, or any external reference is provided. If the current-derived torques differ systematically from the true joint torques—for example, due to unmodeled friction, gravity, inertial coupling, controller gain effects, or the modified kgain and tlookahead settings—then the MLP is trained to reproduce an unvalidated proxy rather than the physical dynamic features the paper claims. Consequently, the feature-estimation results in Table 1, the visual agreement in Figure 2, and the verification comparisons in Figures 3 and 4 for torque features all inherit this uncertainty. The conclusion that force torques generalize across DS2-DS6 (Section 4.5) rests entirely on these labels, and the paper offers no evidence that the labels are physically accurate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9749,"tokens_out":5699,"duration_ms":53430,"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":[{"comment":"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.","section":"Section 3.1"},{"comment":"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.","section":"Table 1"},{"comment":"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.","section":"Section 4.5"},{"comment":"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.","section":"Sections 4.3 and 4.5"}],"minor_comments":[{"comment":"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.","section":"Abstract and Section 1"},{"comment":"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.","section":"Section 3.2"},{"comment":"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.","section":"Figure 2"},{"comment":"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.","section":"Section 3.2"},{"comment":"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.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the cross-database transfer experiment is a real strength. My main concern is the unvalidated torque labels, because the dynamic-feature claim is the paper's headline novelty; if the authors cannot provide at least one independent validation, the dynamic-feature conclusions should be substantially hedged. The theta/omega label swap in Table 1 and the absence of tabulated EERs are fixable but need to be corrected before the claims can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something new: a UR5e robot signs 16,500 MCYT300 signatures while recording angular positions, velocities, and motor-current-based torques, and a simple MLP learns to estimate all 18 channels from (x,y) coordinates. The trained model and data are public. The cross-database test on BiosecurID, Visual, Blind, and two Devanagari/Bengali sets is a genuine out-of-sample check, and the verification results there are the paper's strongest evidence. I think the central claim—the estimated features generalize across databases—holds, at least for angular velocity and torque features.\n\nWhat it does well: the MLP is appropriately simple, the comparison against RNN/LSTM/GRU is fair, and the authors acknowledge the single-robot limitation both in Section 2.3 and in the conclusions. Releasing the model and data is a real contribution.\n\nSoft spots, in rough order of weight:\n\n1. The torque labels are the headline novelty, but they are never validated. Section 3.1 computes tau from motor currents using manufacturer coefficients, and the paper explicitly says the UR5e has no force-torque sensor. No comparison against an external sensor, a calibrated dynamic model, or any sanity check on friction or gravity effects is provided. If those torque traces are not physically accurate, the dynamic-feature contribution is unsupported. This is not fatal—the verification gains could still hold if the estimator learns a useful proxy—but the paper should either validate the currents or soften the physical claim.\n\n2. There is a systematic label swap: Table 1 and Section 3.2 write “angular velocities, theta(t)” and “angular positions, omega(t)”, reversing standard notation. The numbers are probably fine, but as written it confuses which quantity the MAE belongs to.\n\n3. EERs are described but only shown in DET figures; no numeric EER table is given for Figures 3 and 4, despite the text saying “EER is reported as average +/- std.” That is a reproducibility gap.\n\n4. The DS1 verification comparison is in-sample in spirit: four-fold cross-validation on signatures from the same database used for training. The genuinely independent evidence is DS2–DS6, which the paper correctly emphasizes.\n\nMinor: the GitHub repo exists but the public artifacts are incomplete, and “MCYT300”/“MCYT330” are used inconsistently across the paper.\n\nWho this is for: researchers working on online signature verification, especially feature engineering or robotic replay. It deserves peer review with a request for major revision: validate or hedge the torque labels, fix the notation, and put EERs in a table. I would not desk-reject this.","headline":"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.","tokens_in":10272,"tokens_out":2479,"would_cite":true,"duration_ms":23475,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["online signature verification","kinematic features","dynamic features","torque estimation","UR5e robotic arm","multilayer perceptron","dynamic time warping","biometric generalization"],"falsifier":"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.","tokens_in":9348,"feed_emoji":"✍️","tokens_out":1846,"duration_ms":35200,"temperature":0.7,"pith_summary":"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.","feed_headline":"A small MLP extracts arm torques from signature strokes, beating the robot that measured…","feed_subtitle":"Estimated joint angles, velocities, and torques verify signatures as well as or better than UR5e hardware, across five unseen databases.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines the anthropomorphic kinematic features (angular positions) that this paper extends to velocities and torques, and supplies the feature-normalization and verification methodology followed here.","marker":"[7]"},{"why":"Provides the MCYT330 corpus (DS1) whose 16,500 signatures were replayed by the UR5e robot to generate the training labels.","marker":"[21]"},{"why":"Establishes that the UR5e robot can accurately replicate human signature trajectories and velocities, which is the premise for using the robot's measurements as ground truth.","marker":"[19]"},{"why":"Supplies the robust score normalization scheme and DTW implementation used in the automatic signature verifier.","marker":"[11]"},{"why":"Provides the BiosecurID (DS2) online signature data used to test the model's cross-database generalization.","marker":"[12]"},{"why":"Provides the SUSIG Visual (DS3) and Blind (DS4) sub-corpora, two additional unseen databases for generalization testing.","marker":"[15]"},{"why":"Provides the OnOffSigDevanagari-75 (DS5) and OnOffSigBengali-75 (DS6) databases with non-Latin scripts, testing the model across different writing systems.","marker":"[10]"}],"fun_headline_variants":["Small MLP predicts signing arm's joint angles and torques from pen coordinates","Neural net estimates arm dynamics from signatures, generalizes across datasets","MLP estimates signing arm torques from pen strokes, beating robot measurements","NN infers joint angles and torques from signatures, outperforms robot features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Small MLP predicts signing arm's joint angles and torques from pen coordinates","Neural net estimates arm dynamics from signatures, generalizes across datasets","MLP estimates signing arm torques from pen strokes, beating robot measurements","NN infers joint angles and torques from signatures, outperforms robot features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3453,"prompt_tokens":859,"completion_tokens":2594,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":2514}},"tokens_in":475,"tokens_out":2594,"duration_ms":15654,"temperature":1.0,"reasoning_tokens":2514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:01:17.300057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A., & Quintana, J","cited_arxiv_id":null,"evidence_quote":"Defines the anthropomorphic kinematic features (angular positions) that this paper extends to velocities and torques, and supplies the feature-normalization and verification methodology followed here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MCYT330 corpus (DS1) whose 16,500 signatures were replayed by the UR5e robot to generate the training labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that the UR5e robot can accurately replicate human signature trajectories and velocities, which is the premise for using the robot's measurements as ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the robust score normalization scheme and DTW implementation used in the automatic signature verifier."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BiosecurID (DS2) online signature data used to test the model's cross-database generalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SUSIG Visual (DS3) and Blind (DS4) sub-corpora, two additional unseen databases for generalization testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OnOffSigDevanagari-75 (DS5) and OnOffSigBengali-75 (DS6) databases with non-Latin scripts, testing the model across different writing systems."}],"review_version":1}