{"id":"a722fb34-28aa-4682-b118-1d5a3364c489","arxiv_id":"2411.13335","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single-step, full-array tactile calibration method enables 3D force estimation and real-time closed-loop interaction force control on a robotic hand.","lead":"The authors built a single calibration routine that maps readings from many small tactile sensors on a robot hand into an estimate of the force applied anywhere on the hand, and they used that estimate to drive a closed-loop force controller at 100 Hz. On the Allegro hand with uSkin sensors, the best estimate reached 0.12±0.08 N error on a test indenter, with larger errors on softer objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cross-geometry/material generalization claim rests on Table III object trials, but those trials use non-collocated ground truth through deformable objects while the estimator sits inside the control loop, so the reported errors do not cleanly measure the learned mapping H's generalization.","rationale":"In good faith, the paper does what it says at the level of a system demonstration: it calibrates full tactile arrays on a robotic hand with a single data-collection session, evaluates five estimators with cross-validation, and closes a 100 Hz force loop using the estimate. The spherical-indenter result is a clean, held-out-geometry validation and supports the feasibility of the approach. The reader's veredict of CONDITIONAL is appropriate because the broader claim that the learned mapping generalizes to arbitrary object geometries and material properties is not yet cleanly demonstrated. My stress-test converges on the same weakest assumption but sharpens it: Table III's object rows cannot serve as evidence either for or against that generalization, because the ground-truth force is non-collocated and the estimator is inside the feedback loop. The authors themselves flag the non-collocation limitation in the Discussion, which is a point in their favor, but the limitation is not merely a nuisance: it prevents separating estimator error from object internal dynamics and controller compensation. Without that separation, the observed M4 degradation on Softball and Abrasive sponge and the apparent M3λ robustness are confounded. A rigid-replica experiment is the minimal check that isolates the material-compliance variable while keeping geometry fixed. If the errors persist on rigid replicas, the current object results are generalizable; if they shrink, the paper's central contribution should be stated as 'estimation and control for rigid, unseen geometries,' with deformable-object performance left as an open issue. This does not require rejecting the paper; it requires an additional validation before the strong cross-material claim can be accepted.","tokens_in":12481,"tokens_out":6518,"duration_ms":80435,"concrete_test":"Repeat closed-loop experiments (2)-(5) with rigid 3D-printed replicas of the same four YCB objects (identical external geometry, no material compliance), keeping the same training data, controller, and command sequence, with the object mounted directly on the FT sensor so the ground-truth force is effectively collocated with the contact. If the M4/M3λ estimation and tracking errors on the rigid replicas match Table III, the soft-object results are not dominated by the non-collocation confound; if errors drop substantially (e.g., M4 on the Softball replica returns to roughly 0.04-0.14 N), the current object trials do not support the claimed generalization to deformable materials and the abstract should be re-qualified to rigid/unseen-geometry cases until a collocated deformable-object test is available.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that a single calibration on five rigid PLA push-plate indenters produces full-array force estimates accurate enough for closed-loop interaction force control on unseen geometries and materials. The strongest clean evidence is the held-out spherical indenter: in Table III (1), M3λ achieves ê=0.12±0.08 N and M4 achieves 0.04±0.03 N, and this is the basis of the abstract's 'up to 0.12±0.08 [N]' phrasing. The evidence for the broader 'varying geometries' and 'materials' claim is Table III (2)-(5), where YCB objects are taped between the fingertip and the reference force-torque sensor. In these trials the reference measurement is not the force applied at the fingertip-object contact: the object deforms, so its internal dynamics and the non-collocated mounting contaminate the ground truth. Section V-C acknowledges this ('the estimated finger forces and measured sensor forces were not collocated') and the Discussion concedes that the non-collocated measurements 'likely degraded our average estimation performance.' Therefore the differing errors for M4 and M3λ on Softball and Abrasive sponge cannot be attributed to the estimator alone; they may reflect the measurement setup rather than H's generalization. In addition, because H's output is the feedback signal in Eq. (8), the integral controller can make the estimated force track fd even if H is biased, so low tracking error does not establish low estimation bias; the true tracking error is not a substitute for a collocated estimation benchmark. The online evaluation in Table II on the held-out spherical plate gives average errors of 0.21±0.10 N (M4) and 0.31±0.14 N (M3λ), so the closed-loop numbers reflect favorable dynamics and controller compensation, not typical estimator accuracy. Thus the central generalization claim is under-supported by the current object trials, and the abstract's error margin is a best-case rigid-geometry number rather than a representative cross-object accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a data-driven calibration method for estimating 3D interaction forces from raw tactile array activations on a robotic hand (Allegro Hand with Xela uSkin sensors). The authors compare five models—M1 to M5—ranging from linear and quadratic regressions to fully connected and convolutional neural networks. Models are trained on data collected from five PLA push-plate indenters with different geometries, evaluated offline with 5-fold cross-validation, tested online on a held-out spherical indenter, and then used inside a closed-loop integral admittance force controller (Eq. 8) for interaction force tracking against the held-out sphere and four YCB objects. The paper reports that the best models achieve low estimation errors in the spherical test (e.g., M3λ with 0.12±0.08 N) and that closed-loop control reduces true tracking error by about 74% compared with open-loop control. The central claim is that a single calibration step on a limited set of rigid indenters yields full-array force estimates accurate enough for closed-loop force control across varying geometries and materials.","tokens_in":12779,"tokens_out":2692,"duration_ms":33616,"significance":"If the central claim holds, the paper makes a useful contribution: it directly addresses practical barriers in tactile-based interaction control, namely cumbersome element-wise calibration and slow inference, by providing an on-hand full-array calibration method with fast prediction (sub-millisecond) and an explicit integration into an interaction force control loop. The authors also provide code, data, and videos, which supports reproducibility. The offline cross-validation over indenter geometries, the held-out spherical-plate online test with an external ATI force-torque sensor, and the comparison against an open-loop controller are appropriate validation steps. However, the evidence for generalization across materials and deformable objects is weaker than the abstract implies: the closed-loop trials on YCB objects use non-collocated reference measurements, and the controller feedback path can mask estimation bias. The discussion candidly acknowledges some of these limitations, which is to the authors' credit, but the load-bearing generalization claim needs stronger, cleaner evidence.","major_comments":[{"comment":"The closed-loop experiments (2)–(5) on YCB objects cannot cleanly support the generalization claim because the ground-truth force-torque sensor is not collocated with the fingertip-object contact: each deformable object is taped between the finger and the sensor, so the measured force includes the object's internal dynamics and the mounting interface. The paper itself acknowledges in V-C that the estimated finger forces and measured sensor forces were not collocated and in the Discussion that this 'likely degraded our average estimation performance.' Therefore, the differing errors for M4 and M3λ on Softball (0.25±0.16 vs. 0.13±0.07 N) and Abrasive sponge (0.25±0.29 vs. 0.18±0.11 N) cannot be attributed to the learned mapping H alone; they may reflect the measurement setup. A cleaner test would be to obtain collocated ground truth at the fingertip-object contact for the deformable objects, or to validate H open-loop on the same objects with a collocated reference and only then assess closed-loop behavior.","section":"V-C and Table III"},{"comment":"Using the estimated force as the feedback signal in the integral controller can make the estimated force track the desired force even when H is biased, because the integral action acts on the error between fd and f-hat. Thus a low tracking error etrack does not establish low estimation bias, and the true tracking error ~e is a joint property of H and the controller. For example, in the Softball and Abrasive sponge trials, M4 has etrack of 0.12±0.05 and 0.13±0.10 N while ~e is 0.22±0.14 and 0.25±0.25 N, suggesting that the controller is driving the estimate toward the reference despite a larger true error. The paper should either report open-loop estimation accuracy on the YCB objects with collocated sensing or explicitly analyze how controller gain and estimator bias interact before concluding that M4's performance indicates weak generalization.","section":"Eq. (8) and Table III"},{"comment":"The abstract's headline error of 'up to 0.12±0.08 [N]' is taken from a single favorable condition, the closed-loop spherical-indenter trial with M3λ (Table III, row 1). Across the YCB-object trials, the same model's absolute errors are 0.13–0.18 N and M4's errors reach 0.25±0.16 N and 0.25±0.29 N. The phrase 'up to' is technically defensible but the overall narrative in the abstract and introduction implies that a single calibration generalizes well across varying geometries and materials. The reported evidence only supports strong generalization across rigid PLA indenter geometries (offline CV and the online spherical test); material and deformable-object generalization is currently confounded by non-collocated measurement. The authors should qualify the claim in the abstract or provide additional empirical support for material generalization.","section":"Abstract and Section V-C"},{"comment":"The paper notes in the Discussion that 'neural network performance for this type of task depends more strongly on training set composition than linear models do.' This is consistent with Table III, where M4 outperforms M3λ on the spherical indenter but degrades on the Softball and Abrasive sponge. However, because those YCB-object trials are non-collocated, the paper cannot conclude that M3λ has 'superior generalization capabilities' for out-of-distribution objects. The model ranking in closed-loop YCB trials may be substantially affected by the measurement setup. Please provide a collocated open-loop comparison on the same objects to separate estimator generalization from controller and measurement effects.","section":"Section VI and Table III"}],"minor_comments":[{"comment":"There are several typos and formatting issues, including 'vairants' and 'Fingertac -' in the references section, and the caption of Fig. 4 contains a stray 's. that H(x) = f-hat.' Please proofread.","section":"Abstract and Introduction"},{"comment":"The description of data collection says 'For each plate geometries, except the spherical one which is kept for online validation, four sequences of 100 seconds each are collected,' which yields 40,000 samples per plate, yet the text later states 200,000 data points per array. Clarify whether the 200,000 figure includes all plates and arrays, and specify the exact total per array.","section":"Section IV-A"},{"comment":"The first row block is labeled 'Flat Curved Array' which appears to be a typo; it should be 'Curved Array.' Also, the table would be easier to read with separate rows for each model rather than grouping the two array types under a single heading.","section":"Table I"},{"comment":"The online evaluation reports results after retraining M4 and M3λ on the complete dataset, including the spherical plate. This is reasonable for testing the online pipeline, but it means the online numbers are not directly comparable to the cross-validated offline numbers in Table I. Please state this explicitly.","section":"Section V-B"},{"comment":"The definitions of er and e-hat are clear, but note that the relative error in Eq. (9) excludes points with |f_i| < 0.5 N only for the denominator; it would be helpful to state whether such points are also excluded from the numerator and from e-hat computation in the closed-loop results.","section":"Eq. (9) and Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and addresses a relevant problem, but the central generalization claim currently rests on evidence that is partly confounded by non-collocated ground truth in the YCB-object trials. The authors' own acknowledgments support this concern. I would encourage the editor to invite a major revision that either adds collocated validation on deformable objects or substantially softens the generalization claims in the abstract and conclusions. The code and data release are a plus and should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline: this is a solid empirical paper that delivers on its main pipeline claim—single-step, full-array tactile force calibration on a real hand, with a held-out geometry validation and closed-loop interaction force control at 100 Hz. The material-generalization claim is weaker than the abstract suggests, but the core result holds.\n\nWhat's actually new and good: they avoid element-by-element calibration, which is the practical bottleneck for uSkin-type arrays. They compare five models (linear, quadratic, raw-input linear with damping, MLP, CNN) under 5-fold cross-validation over indenter geometries, then validate online on a spherical indenter not used in training. The curved-array results show neural models beating rotation-prior models, which is a useful empirical finding. The closed-loop experiments show a large reduction in true tracking error compared to open-loop (about 74% average for M3λ), and they report against a calibrated ATI Mini45. Code and dataset are public. That is real, reproducible work.\n\nThe soft spots are in the generalization claim. The abstract's 'up to 0.12±0.08 N' is the M3λ result on the rigid spherical indenter, the best case. On the YCB objects—especially the softball and abrasive sponge—M4 degrades to 0.25 N errors, and the paper attributes this to distribution shift. But the stress-test note is right: those object trials use non-collocated ground truth (object taped to the FT sensor), so the errors do not cleanly measure the estimator's generalization to new materials. On top of that, the closed-loop integral term can make the estimated force track the desired force even if the estimate is biased, so low tracking error isn't evidence of low estimation bias. The paper acknowledges the non-collocation confound in the Discussion, but the abstract overclaims. Also, hyperparameters (λ=33, NN architecture) were tuned on the validation folds with no nested CV, so the offline numbers are likely optimistic.\n\nNone of this breaks the central pipeline claim; it does mean the paper should be read as demonstrating geometry generalization, not material generalization. The fix is straightforward: soften the abstract, report the model-selection procedure, and ideally add a collocated deformable-object experiment or explicitly scope the claims.\n\nThis deserves a serious referee—I'd send it to a robotics venue (ICRA/IROS/RAL) with requests for revision. It's a useful contribution for anyone working with uSkin-type arrays or tactile force feedback.","headline":"Solid pipeline paper: single-step full-array calibration with 100 Hz closed-loop force control works on held-out geometry, but the material-generalization claim is stretched by non-collocated object trials.","tokens_in":13450,"tokens_out":3243,"would_cite":true,"duration_ms":35310,"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":"This paper shows that a single calibration step on push-plate indenters lets a robotic hand estimate interaction forces from raw tactile array data well enough to close a 100 Hz force control loop, achieving as low as 0.12±0.08 N error.","keywords":["tactile sensing","force estimation","robotic hand","interaction control","admittance control","uSkin sensor","calibration","closed-loop control"],"falsifier":"Press a new object with known stiffness far outside the five training plates (for example a very soft foam or a highly curved rigid surface) against the finger while a calibrated reference sensor measures the true force, and check whether the closed-loop controller's estimation error stays below about 0.25 N; the paper's own Table III already shows errors reaching 0.25±0.29 N on an abrasive sponge, so a dedicated out-of-distribution test would determine the limit of the claim.","tokens_in":12256,"feed_emoji":"🤖","tokens_out":4681,"duration_ms":46310,"temperature":0.7,"pith_summary":"This paper aims to make tactile sensor arrays on a robotic hand directly usable for closed-loop interaction force control, without per-sensor pre-calibration. It claims that a single data-collection step, pressing the hand's tactile arrays against a few 3D-printed push-plates, is enough to train a mapping from raw sensor activations to 3D contact force, and that this mapping holds across curved fingertips and flat phalanxes as well as objects not in the training set. The payoff is practical: a force estimate that runs at 100 Hz and can be fed back into an admittance controller, letting the finger apply a desired force trajectory with errors as low as 0.12±0.08 N. If true, this removes a major bottleneck in using rich tactile feedback for dexterous manipulation.","feed_headline":"Tactile force feedback nails 0.12 N on robot fingers","feed_subtitle":"One-step calibration turns skin readings into 100 Hz force control, cutting tracking error 74%.","key_machinery":"The load-bearing object is the estimator H(x; θ), a parametric map from a whole array's raw tactile activations to a 3D force vector, trained with ordinary least squares or Adam on data collected by pressing each sensor array against 3D-printed push-plates while an external force-torque sensor records ground truth. Two versions carry the results: M3λ, a damped linear model on the raw concatenated taxel vector, and M4, a one-hidden-layer MLP with 16 ReLU units. The controller is an integral force feedback law (Eq. 8) built on a position-controlled Allegro Hand, with the estimator's output f̂ replacing a physical force sensor in the loop; the integral term plus anti-windup and torque saturation make the contact force converge to the desired value.","core_discovery":"The central claim is that uncalibrated uSkin tactile arrays mounted on a full robotic hand can be calibrated as a whole, in one pass, by training a parametric model H that maps the concatenated raw taxel activations x to the resultant contact force f. The paper evaluates five such models, from a rotation-assisted linear model to a convolutional network, and shows on a 5-fold cross-validation over push-plate geometries that the nonlinear MLP (M4) and the damped linear model (M3λ) give the best accuracy. The decisive demonstration is closed-loop: the estimated force is used as the feedback signal in an integral force controller (Eq. 8) running at 100 Hz, and against four deformable YCB objects plus an unseen spherical indenter, the M3λ-based controller reduces true force tracking error by about 74% compared to the open-loop version, with average absolute errors around 0.12–0.17 N. The paper also reports that the linear model generalizes better to out-of-distribution objects than the neural network, an observation it ties to training-set composition.","pith_inferences":["Because the calibration relies on a small set of push-plate indenters, the method could likely be extended to other magnetic or capacitive arrays by re-using the same training pipeline, provided the taxel layout is known.","The observed out-of-distribution advantage of the linear model hints that the raw taxel sum z (Eq. 1) already encodes most of the force information, and that nonlinear models mainly help on in-distribution shapes; this could be tested by ablating M4's hidden layer on a broader set of objects.","If the 100 Hz rate and 0.12 N accuracy hold under dynamic motion (sliding, fast contact changes), the approach could enable force-controlled manipulation tasks like assembly or fragile-object handling; the paper only tests quasi-static presses, so dynamic validation is a natural next step.","Combining the linear model's generalization with a small online adaptation step (e.g., a correction on the residual) might close the gap on highly deformable objects, though the paper itself does not explore this."],"forward_implications":["A single calibration session can produce force feedback for all 18 arrays of an Allegro Hand, since both selected models run in about 0.05 ms per prediction, well within the 100 Hz control rate.","Force estimates are validated against an external sensor, so the reported 0.12±0.08 N error is a statement about real applied force, not just about the model's internal consistency.","When the controller is closed with the estimated force, actual tracking error drops roughly 74% versus open-loop, showing the estimation is accurate enough to be used as feedback, not just monitoring.","The linear model M3λ is more robust than the neural network on objects outside the training distribution, suggesting simpler models may be preferable for unseen materials and geometries."],"supporting_citations":[{"why":"Supplies the uSkin modular, distributed 3-axis tactile sensor system that the whole hand is equipped with.","marker":"[19]"},{"why":"Provides the curved fingertip uSkin array design and the earlier element-wise calibration approach that this paper extends to full-array, single-step calibration.","marker":"[20]"},{"why":"Documents the time-consuming shear-force data collection that limited earlier calibration tests to a single chip, motivating the proposed data-efficient method.","marker":"[21]"},{"why":"Gives the prior Fingertac calibration accuracy numbers (0.21, 0.16, 0.44 N for X, Y, Z) that serve as a baseline for the claimed generalization improvement.","marker":"[22]"},{"why":"Offers a benchmark of different models for force magnitude and direction estimation from a resistive tactile array, a comparative reference for the proposed approach.","marker":"[24]"},{"why":"Supplies the YCB object set used for the closed-loop interaction control experiments on deformable objects.","marker":"[37]"},{"why":"Provides the base control law on the Allegro Hand without force or tactile feedback, which is extended here with an integral force correction term.","marker":"[41]"},{"why":"Justifies the integral-only force feedback design, argued to be more robust for a robot in contact with a rigid surface.","marker":"[42]"}],"fun_headline_variants":["Whole-hand tactile calibration cuts force error to 0.12 N","One-step skin model drives 100Hz finger control","Hand-scale tactile calibration enables 74% error cut","Raw taxels to force: fast calibration for dexterous control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned mapping from tactile signals to force, trained on five push-plate geometries, is assumed to transfer to unseen objects with different stiffness and curvature without retraining.","fun_headline_variants_meta":{"raw":{"variants":["Whole-hand tactile calibration cuts force error to 0.12 N","One-step skin model drives 100Hz finger control","Hand-scale tactile calibration enables 74% error cut","Raw taxels to force: fast calibration for dexterous control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1260,"prompt_tokens":928,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":544,"tokens_out":332,"duration_ms":4479,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:32:02.506125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Press a new object with known stiffness far outside the five training plates (for example a very soft foam or a highly curved rigid surface) against the finger while a calibrated reference sensor measures the true force, and check whether the closed-loop controller's estimation error stays below about 0.25 N; the paper's own Table III already shows errors reaching 0.25±0.29 N on an abrasive sponge, so a dedicated out-of-distribution test would determine the limit of the claim.","supporting_citations":[{"cited_title":"A modular, distributed, soft, 3- axis sensor system for robot hands,","cited_arxiv_id":null,"evidence_quote":"Supplies the uSkin modular, distributed 3-axis tactile sensor system that the whole hand is equipped with."},{"cited_title":"Covering a robot fingertip with uskin: A soft electronic skin with distributed 3-axis force sensitive elements for robot hands,","cited_arxiv_id":null,"evidence_quote":"Provides the curved fingertip uSkin array design and the earlier element-wise calibration approach that this paper extends to full-array, single-step calibration."},{"cited_title":"A new silicone structure for uskin—a soft, distributed, digital 3-axis skin sensor and its integration on the humanoid robot icub,","cited_arxiv_id":null,"evidence_quote":"Documents the time-consuming shear-force data collection that limited earlier calibration tests to a single chip, motivating the proposed data-efficient method."},{"cited_title":"Fingertac - an interchangeable and wearable tactile sensor for the fingertips of human and robot hands,","cited_arxiv_id":null,"evidence_quote":"Gives the prior Fingertac calibration accuracy numbers (0.21, 0.16, 0.44 N for X, Y, Z) that serve as a baseline for the claimed generalization improvement."},{"cited_title":"Force estimation and slip detec- tion/classification for grip control using a biomimetic tactile sensor,","cited_arxiv_id":null,"evidence_quote":"Offers a benchmark of different models for force magnitude and direction estimation from a resistive tactile array, a comparative reference for the proposed approach."},{"cited_title":"The ycb object and model set: Towards common benchmarks for manipulation research,","cited_arxiv_id":null,"evidence_quote":"Supplies the YCB object set used for the closed-loop interaction control experiments on deformable objects."},{"cited_title":"Kitech- hand: A highly dexterous and modularized robotic hand,","cited_arxiv_id":null,"evidence_quote":"Provides the base control law on the Allegro Hand without force or tactile feedback, which is extended here with an integral force correction term."},{"cited_title":"Integral force control with robustness enhancement,","cited_arxiv_id":null,"evidence_quote":"Justifies the integral-only force feedback design, argued to be more robust for a robot in contact with a rigid surface."}],"review_version":1}