{"id":"3e4be65c-6bd9-423f-ac25-59a0532accb5","arxiv_id":"1908.00452","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A time-delay neural network trained on simulated vehicle data estimates per-wheel tire-road friction from lateral force and slip angle, beating a model-based nonlinear least-squares baseline in simulation.","lead":"This paper trains a time-delay neural network to estimate the tire-road friction coefficient at each wheel from lateral force and slip angle signals. The authors compare it with a nonlinear least-squares model-based estimator in simulations and report lower error and lower compute time, but no real-world validation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The per-wheel friction claim depends on the EKF force estimator, which fixes μ=0.8; no sensitivity analysis is provided, so a force-estimate bias at off-nominal friction would propagate into the TDNN inputs and could invalidate the claimed per-wheel accuracy.","rationale":"The reader's verdict is CONDITIONAL with high confidence, and I agree. The paper is a competent, honest simulation study: the TDNN architecture, training setup, and comparison to NLS are described in enough detail to reproduce; the error histograms, R-values, and RMS tables are presented; and the conclusion explicitly flags the lack of field tests. My stress test did not reveal a fatal internal inconsistency in the reported results. The load-bearing weak point is the same one the reader identified: the EKF force estimator is built on a fixed μ=0.8 Dugoff model and no sensitivity analysis is given. This matters because the TDNN does not observe μ directly; it observes Fy/Fz and α, and if those are biased by the filter's μ assumption, the network's output is biased in a way that the in-distribution simulation metrics cannot reveal. A concrete sensitivity test on the EKF force estimates across the true μ range would settle whether the per-wheel claim survives off-nominal friction. No code or data artifacts are released, which limits independent checking but is not by itself a correctness flaw. I therefore keep the CONDITIONAL verdict rather than upgrading or downgrading it.","tokens_in":13950,"tokens_out":10088,"duration_ms":109391,"concrete_test":"Run the full estimation chain on identical constant-steer and ramp-steer maneuvers with the true road friction held at μ=0.3, 0.5, 0.7, 1.0, and 1.2, while keeping the EKF's internal μ fixed at 0.8. Compare the EKF-estimated per-wheel Fy/Fz and slip angle with the simulator's ground-truth wheel forces, and quantify steady-state bias per wheel. Then feed these estimates into the trained TDNN and compute the RMS error in μ against ground truth, separating the per-wheel results as in Tables 7-8. If the force-estimate bias remains at sensor-noise level for all μ and the TDNN RMS stays in the 0.03-0.07 range, the fixed-μ concern is resolved; if bias grows with |μ-0.8| and TDNN error tracks it, the per-wheel claim is conditional on that assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The estimation chain in Section 4 (Fig. 3) is only as good as its first stage: the EKF tire-force estimator uses a Dugoff model with a constant μ=0.8, and the paper asserts robustness to TRFC variations without showing a sensitivity analysis. The TDNN inputs are the per-wheel ratio Fy/Fz and slip angle α; any systematic bias in the estimated forces when the true friction departs from 0.8 is fed directly into the network. Inertial acceleration can correct the total vehicle force, but it does not automatically correct the front/rear or left/right distribution of that force, which is precisely what the per-wheel claim needs. Because training and evaluation are carried out in the same simulated environment, a consistent filter bias could be learned by the TDNN and would not show up in the reported R≈0.94 or RMS values. The manuscript also does not state unambiguously whether the 200,000 training samples pass through this EKF or are raw simulated tire-model outputs, so the train/deploy distribution match is itself unverified. The conclusion's own sentence that 'an actual benchmark or field test is needed' correctly limits the claim, but the abstract's unqualified per-wheel statement is stronger than the fixed-μ assumption supports without analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a hierarchical estimator for the tire-road friction coefficient (TRFC) at each wheel. A cascade of extended Kalman filters estimates per-wheel lateral and vertical forces using a vehicle model whose internal Dugoff tire model fixes the friction coefficient at μ=0.8; the slip angle is computed kinematically; and a time-delay neural network (TDNN) with a 50-sample observation window maps (F_y/F_z, α) to a per-wheel estimate of μ. The TDNN is trained on about 200,000 simulated samples with μ ranging from 0.3 to 1.2 and evaluated in three simulated maneuvers: a ramp steer on constant μ=0.8, a constant steer with step changes in μ for all wheels (1.0→0.6), and a constant steer with different μ for left and right front wheels (0.9/0.8→0.8/0.7). The TDNN is compared with nonlinear least-squares (NLS) fits using Dugoff and Brush tire models; the reported RMS errors are slightly lower for the TDNN (e.g., 0.0346 vs 0.0421 in case 1), and the average computation time per iteration is 0.594 ms vs 3.379 ms for the NLS.","tokens_in":14207,"tokens_out":7812,"duration_ms":70796,"significance":"If the claims hold, the approach would offer a real-time, per-wheel friction estimate without an explicit tire model in the network itself, which is of practical interest for active-safety and vehicle-control systems. The paper's strengths are the use of a realistic 32-state vehicle simulation with sensor noise, a comparison against two NLS model-based baselines, and a clear demonstration of the computational advantage of the TDNN. The central novelty—using a TDNN for per-wheel TRFC estimation from lateral force information—is clearly stated. However, the significance is tempered because the evaluation is entirely in the same simulated environment used for training, the training-data pipeline is not fully specified, and the fixed-μ EKF force estimator's robustness is asserted without sensitivity analysis.","major_comments":[{"comment":"The claim that the EKF force estimator is robust to TRFC variations is not supported in this paper. The EKF uses a Dugoff model with constant μ=0.8, and the TDNN's inputs are the estimated F_y/F_z and slip angle α. If the true friction departs from 0.8, a systematic bias in the per-wheel force estimates—especially in the distribution of lateral force between front/rear and left/right wheels, which inertial acceleration does not constrain—propagates directly into the network inputs. Please provide a sensitivity analysis that varies the true μ while keeping the EKF's μ fixed, quantifies the force-estimation error, and shows that the TDNN's per-wheel μ estimates remain accurate; alternatively, close the loop by updating the EKF's μ.","section":"Section 4, Fig. 3 and paragraph 'The aforementioned works...'"},{"comment":"It is not stated whether the 200,000 training samples are raw outputs of the simulated tire model (the Pacejka model in the vehicle simulator) or the EKF-estimated forces and kinematically computed slip angles that are actually fed to the network in deployment (Fig. 3). If training uses clean model outputs, the training and deployment input distributions differ (the deployed inputs are filtered EKF estimates with estimator bias), and the reported R≈0.94 and mse values do not demonstrate the accuracy of the deployed system. Please specify the pipeline precisely and, if necessary, train on the actual deployed input features.","section":"Section 4, paragraph 'In the neural network data collecting stage...'"},{"comment":"All test maneuvers use friction values (0.6-1.0) inside the training range (0.3-1.2) and the same simulation environment as the training data. The results are therefore interpolation within the fitted envelope and do not provide evidence of generalization to out-of-range conditions or to a different vehicle or tire model. Add a hold-out test with a friction value not in the training grid (e.g., μ=0.5) or with a mismatched tire model in the simulator, and report the resulting RMS errors. The conclusion's own statement that a field test is needed should be reflected in the abstract's unqualified per-wheel claim.","section":"Section 5, Tables 7-8 and Figs. 11-14"},{"comment":"The RMS values are computed from a single simulation run for each scenario. Because the measurements are corrupted with Gaussian sensor noise, the reported small differences between TDNN and NLS (e.g., 0.0346 vs 0.0421 in Table 7) could be within run-to-run variability. Please report the mean and standard deviation (or at least the range) of the RMS errors over several Monte Carlo runs for each maneuver to assess the significance of the TDNN's advantage.","section":"Section 5, Tables 7-8"}],"minor_comments":[{"comment":"The phrase 'avoiding the use of standard mathematical tire models' overstates the architecture, since the EKF force estimator uses a Dugoff tire model with μ=0.8; please qualify the claim to say the TDNN itself is model-free.","section":"Abstract and Section 4"},{"comment":"The displayed summation has a typo: 'F (x)−Fk=1' should be 'F(x)-F_k' with the subscript on F_k.","section":"Equation (6)"},{"comment":"The text first describes a 'constant left turn maneuver' and then refers to a 'right-hand steer maneuver'; please clarify which steering direction is simulated and ensure Fig. 13 matches the text.","section":"Section 5.3"},{"comment":"Please state whether overlapping sliding windows were used to form the N=50 delayed input vectors from the 200,000 samples and report how many training windows resulted.","section":"Section 4, data collection"},{"comment":"The correlation coefficients are presented without a statement of the data source or the number of samples; add this information for reproducibility.","section":"Table 2"},{"comment":"The histograms do not show the mean or standard deviation of the computation-time distributions; include the reported means (0.594 ms vs 3.379 ms) in the figure or caption for clarity.","section":"Fig. 15"}],"recommendation":"major_revision","confidential_remarks":"The paper's main weaknesses are the unverified fixed-μ EKF assumption and the ambiguous training-data pipeline; both are fixable with additional analysis and clarification. The novelty of TDNN for this application is modest but sufficient for the journal's scope. I do not see concerns about citation practices; the related literature is adequately represented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: competent simulation study of TDNN for per-wheel tire-road friction estimation; the per-wheel claim is plausible but not yet supported outside the simulator because the EKF force estimator fixes μ=0.8 and no sensitivity analysis is given.\n\nThe actual new bit is narrow but real: a time-delay neural network fed with per-wheel Fy/Fz and slip angle, trained to output friction coefficient, with an observation window sized to capture friction-change transients. The authors compare against nonlinear least squares using two tire models and show lower RMS errors and roughly six times lower computation time. That is a legitimate incremental contribution, and the simulation setup is more realistic than many NN papers: full multibody vehicle model, sensor noise per commercial IMU specs, three maneuvers including split-friction left/right.\n\nCredit where due: they are honest about the lateral-excitation requirement, they acknowledge the NLS model discrepancy explicitly, and the conclusion admits field testing is needed. The citation pattern looks fine; the relevant prior work on NLS and NN-based estimation is cited.\n\nThe soft spots are real but not disqualifying. The load-bearing one is the force estimator. The EKF uses a Dugoff model with fixed μ=0.8; the paper asserts robustness to TRFC variations based on earlier work, but does not show any sensitivity analysis for off-nominal friction. The TDNN inputs are Fy/Fz and α, so any bias in the estimated forces at e.g. μ=0.4 feeds straight into the network. The stress-test worry that inertial acceleration corrects total force but not its axle/wheel distribution is exactly right and is not addressed. Also, the text does not make unambiguous whether the 200,000 training samples are the raw simulated tire outputs or the EKF-filtered estimates; that matters because the deployed network sees EKF outputs. If the training data skip the EKF, the train/test distribution mismatch is another unverified link.\n\nOther issues are minor: single simulation runs, no variance across noise seeds; RMS differences between TDNN and NLS (0.0346 vs 0.0421) are small, so the 'lower RMS' claim is modest; no code or data released. But none of this undermines the paper's internal validity; the authors never overclaim beyond simulation.\n\nWho this is for: people working on practical friction estimation for chassis control, especially those wanting a fast inference alternative to recursive least squares. It deserves serious peer review. My recommendation: send it to review, but request (a) a sensitivity study of the EKF force estimates across the μ range, (b) clarification of the training-data pipeline, and (c) reporting of multiple runs.","headline":"Per-wheel TDNN friction estimation is a solid simulation study, but the fixed-μ EKF force estimator and in-distribution evaluation keep it from supporting the abstract's general claim.","tokens_in":14738,"tokens_out":2224,"would_cite":false,"duration_ms":22434,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-14T15:55:26.947467+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}