{"id":"e443e26f-a3ed-4e70-8b25-e4c887d2c02a","arxiv_id":"2505.19560","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LF-GNSS integrates a learned measurement-noise model and innovation correction into an extended Kalman filter, using a DOP-based satellite feature and a hard-example-mining loss, and reports improved urban GNSS positioning accuracy.","lead":"This paper presents LF-GNSS, a system that combines a deep learning network with a Kalman filter to improve satellite positioning in cities. The network learns to adjust the filter's noise settings and correct its measurement errors, and the authors report better accuracy than several existing GNSS tools on public and private datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth reference quality is the load-bearing assumption; a comparison of L1/L5-band and RTK/INS references against the same raw pseudoranges is needed to confirm whether the reported RMSE improvements are real.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: ground-truth reference trajectories are used as exact in both public and private evaluations, so the RMSE differences become unreliable if the references carry meter-level errors. I agree this is the single most load-bearing issue. In my reading, this is not a circularity or soundness defect in the derivation of LF-GNSS; the architecture is plausible, the DPC feature is a reasonable addition, and the ablation table is directionally consistent with the paper's claims. The problem is empirical: the central claim is an accuracy claim, and every measurement against which it is tested depends on an unvalidated reference. The paper discloses that RTK/INS TC solutions serve as ground truth in Sec. IV-C and that public dataset references are used, but it provides no accuracy estimate for those references in urban canyons and elevated bridges, precisely where RTK/INS can degrade. There is also no statistical testing or error bars, so the reported gaps in Tables I and II are single-run point estimates. The second issue raised by the reader, backpropagation through the EKF without discussion, is real but less load-bearing because the tables could still be true even if training is unstable; a conditional accept can require the authors to document the backward pass. A concrete test that re-derives the RMSE with an independent reference would settle the concern. Verdict remains CONDITIONAL, with the condition strengthened: the authors should quantify reference accuracy and re-run the evaluation on an independent trajectory reference (or at least validate reference drift on segments with known ground-control points or stationary periods).","tokens_in":15144,"tokens_out":1625,"duration_ms":15106,"concrete_test":"Re-evaluate the private ALGO sequences with an independent kinematic trajectory reference (post-processed dual-frequency RTK/INS with fixed ambiguities, or a visual-inertial SLAM solution with loop closures) and recompute each method's RMSE against the original RTK/INS TC reference. If LF-GNSS's improvement over TDL-GNSS on the same raw pseudoranges persists under the independent reference, the concern is settled; if the margin shrinks to near zero or changes sign, it confirms reference-induced bias.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that LF-GNSS outperforms CSSRLIB, goGPS, RTKLIB, and TDL-GNSS in urban RMSE. The weakest load-bearing assumption is the accuracy of the ground-truth trajectories: in the private experiments (Sec. IV-C), RTK/INS tightly coupled solutions are treated as exact references, and in the public datasets (Sec. IV-B), the provided reference trajectories are likewise taken as exact. The paper states this use explicitly but never quantifies the accuracy of those references in the tested segments. Because all comparisons, the hard-example-mining loss (Sec. III-E), and the reported improvement percentages (79.55%–84.99% vs traditional, 53.68% vs DL, Sec. IV-C) are computed against these references, any meter-level reference error directly biases the measured RMSE differences. The paper also lacks error bars or statistical tests, so the table cells are point estimates against a single reference run. A second fragile link is the backpropagation through the EKF update (Eqs. 35–40), which is not described and could be numerically unstable; however, the ground-truth reference issue is more load-bearing because it affects every reported number in Tables I and II and is entirely absent from the paper's limitations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LF-GNSS, a deep-learning/Kalman-filter fusion framework for GNSS pseudorange positioning in urban environments. A multi-head-attention network takes per-satellite features (SNR, elevation, azimuth, pseudorange residual, and a new leave-one-out DOP contribution feature) and outputs a per-satellite measurement noise covariance R and an innovation compensation vector vc, which are injected into the EKF measurement update. Training uses a dynamic hard-example-mining loss on final ENU position error. The framework is evaluated against CSSRLIB, goGPS, RTKLIB, and TDL-GNSS on three public dataset families and a private vehicle dataset, with ablations of the DPC feature and the hard-example-mining loss on the private dataset only.","tokens_in":15374,"tokens_out":11364,"duration_ms":98893,"significance":"If the reported results hold, the framework would be a valuable sensor-free approach to urban GNSS positioning, with a simple and plausible integration of learned measurement noise and innovation compensation into a standard EKF, plus a novel per-satellite geometry feature. The open-source intent and the use of multiple public datasets are positive features. However, the current manuscript contains a mathematical inconsistency in the proposed hard-example-mining loss, and the evaluation has important gaps: single-run RMSEs without statistical tests, unquantified ground-truth reference quality, and ablations limited to the private dataset. These issues prevent the central claim of robust superiority from being accepted as stated.","major_comments":[{"comment":"The proposed hard-example-mining weight is inconsistent with its stated purpose. For the sample with Lbase = max(Lbase) in the batch, w = (1 - 1)^{γdynamic} = 0, so the hardest sample contributes zero to L_dhem. Moreover, for all samples w ∈ [0,1], and since γdynamic decreases as Lbase increases, w is actually larger for smaller Lbase; that is, easier samples receive the larger relative weights. Thus Eq. (45) implements a downweighting of large-error samples, not 'prioritizing challenging satellite signals.' The ablation in Table II therefore cannot be interpreted as evidence for hard-example mining as described. Please correct the formula (e.g., a focal-style factor that increases with error) or reframe the contribution as a robust loss, and re-run the ablation.","section":"Section III-E, Eqs. (43)-(45)"},{"comment":"The evaluation treats the RTK/INS tightly coupled solutions (private dataset) and the provided reference trajectories (public datasets) as exact ground truth, but nowhere quantifies their accuracy over the tested segments. Because the loss (Eqs. 41-46), all RMSE values in Tables I and II, and the claimed improvement percentages in Section IV-C are computed against these references, any meter-level reference error would directly bias the comparisons. Please report reference-quality metrics (e.g., percentage of fixed RTK epochs, baseline lengths, or an independent cross-check) for at least the private dataset, and discuss how reference error affects the reported margins.","section":"Section IV-C and IV-B"},{"comment":"The training procedure requires gradients of the final-position loss with respect to the network outputs R_k and vc_k through the EKF measurement update, including the inversion S_k^{-1} and the Joseph-form covariance update. The manuscript does not describe how this backpropagation is implemented, whether automatic differentiation is used, or how numerical stability is guaranteed for learned R_k. This is essential for reproducibility and for assessing gradient health; please add the implementation details and any stability measures.","section":"Section III-D, Eqs. (35)-(40)"},{"comment":"All results are single-run RMSE point estimates with no standard deviations, confidence intervals, or significance tests, even though LF-GNSS training is stochastic. Furthermore, the ablations for DPC and HEM are reported only on the private dataset (Table II), so the contributions are not validated on the public datasets. Please provide multi-seed statistics and, if possible, ablations on at least one public split.","section":"Tables I and II, Section IV-D"},{"comment":"Several hyperparameters that determine the filter and loss behavior are not specified: the process noise covariance Q_k, the initial state covariance sigmas σ_p0, σ_v0, σ_cb0, σ_cd0, σ_ISB0, and the loss coefficients α, γ, λ. Without these values, the experiments cannot be reproduced and the sensitivity of the method to these settings cannot be assessed. Please report all hyperparameter values and, ideally, a sensitivity analysis.","section":"Section IV-A and Eqs. (31), (34), (43)-(45)"}],"minor_comments":[{"comment":"The word 'EXPRIMENT' should be 'EXPERIMENT'.","section":"Table II caption"},{"comment":"DPC is defined via a generic 'DOP'; please specify whether HDOP, VDOP, PDOP, or GDOP is used in the leave-one-out difference, since the values and their interpretation differ.","section":"Equation (24)"},{"comment":"The output layer is described as mapping to a '2-dimensional output' but Rdiag and vc are both of dimension N; this should read 'two output heads' or similar.","section":"Section III-C1"},{"comment":"The input dimension is stated as d=8, while Section III-B lists only five features; please clarify what the remaining three feature channels are.","section":"Section III-C2"},{"comment":"The mapping from 'KLT-[3, 1, 2] and Whampoa' to 'IPNL-Train, IPNL-A, IPNL-B, and IPNL-C' is ambiguous; an explicit one-to-one table would improve reproducibility.","section":"Section IV-B"},{"comment":"The sentence about 79.55%-84.99% and 53.68% improvements should state that these percentages correspond to the ALGO-C 3D RMSE values; as written, 'final statistical analysis' could be read as an aggregate over all private test sets.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. The hard-example-mining formula issue is a clear, checkable technical error that should be fixed; the ground-truth and statistical concerns require additional experiments. The large margins over baselines on several public datasets suggest the core framework may be viable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible incremental contribution to learning-based GNSS positioning, with two genuinely new pieces (the DPC feature and the DHEM loss). The reported gains on urban datasets are large, but the evaluation has several soft spots, and the DHEM loss equation as written appears to do the opposite of what the paper claims.\n\nWhat's new and good: DPC is a sensible leave-one-out DOP contribution feature that captures geometry per satellite. Coupling an attention network that outputs R and vc to an EKF is a natural extension of TDL-GNSS and LEAR-EKF, and the paper compares against several baselines including ablation on the private dataset. They also promise code and data, which is nice.\n\nSoft spots: First, Eqs. (43)-(44): w = (1 - sqrt(Lbase/max))^gamma_dynamic. For the sample with the largest error in a batch, w = 0, so the hardest sample contributes zero loss. That contradicts the stated goal of focusing on hard examples. Either the formula is mis-stated or the implementation differs; this needs correction. Second, all results are single-run RMSE values with no error bars or significance tests. Third, train/test splits are from the same dataset families (IPNL, GREAT, PLANET, and the private ALGO), so cross-dataset generalization is untested. Fourth, the ground truth references (RTK/INS TC in private, provided trajectories in public) are taken as exact, but their accuracy is never quantified; this is the load-bearing assumption for every reported improvement. Fifth, key implementation details are missing: process noise Q, initial covariances, and the backward pass through the EKF update. The claimed 79.55%-84.99% and 53.68% improvements are roughly consistent with Table II if you pick the right baselines, but the text doesn't say which, which is sloppy.\n\nThis paper is for researchers working on deep learning for GNSS positioning and differentiable filtering. The core idea is worth pursuing, and the DPC feature could be useful independently. But the paper needs major revision before its claims can be accepted.\n\nRecommendation: send to peer review. The framework is novel enough and the questions are real. A good reviewer should ask for a corrected or clarified loss, error bars, cross-dataset evaluation, and quantification of reference accuracy.","headline":"A plausible learning-plus-EKF GNSS fusion with two genuinely new twists, but the hard-example loss as written zeros out the hardest samples and the evaluation lacks error bars and cross-dataset validation.","tokens_in":15927,"tokens_out":6160,"would_cite":false,"duration_ms":63251,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A learning-filtering deep fusion framework claims more accurate urban GNSS positioning by letting a neural network set the Kalman filter's noise and innovation corrections from per-satellite signal features.","keywords":["GNSS positioning","urban canyon","Extended Kalman Filter","deep learning","hard example mining","DOP contribution","NLOS and multipath mitigation","pseudorange-only positioning"],"falsifier":"Re-run the private route comparison with ground truth from survey-grade total-station marks or a post-processed open-sky RTK/INS solution, and check whether LF-GNSS's 3D RMSE margins over TDL-GNSS and RTKLIB (1.91 m versus 2.70 m on ALGO-B, for example) survive; if the margin collapses to noise, the central accuracy claim is falsified.","tokens_in":14899,"feed_emoji":"📡","tokens_out":11833,"duration_ms":100153,"temperature":0.7,"pith_summary":"LF-GNSS is a satellite-positioning framework that couples a deep network to an extended Kalman filter. The network reads per-satellite signal features and outputs the filter's measurement-noise covariance and an innovation-compensation vector, so bad satellite signals are downweighted and biased pseudoranges are corrected in urban canyons. The paper claims this produces better accuracy than classical open-source GNSS engines and than an earlier deep-learning least-squares approach, across three public benchmark suites and a 1.5-hour road test. If the claim holds, it offers a sensor-free way to make GNSS more dependable for autonomous driving and city navigation, without requiring an IMU or camera.","feed_headline":"Neural network tunes Kalman filter for sharper urban GNSS fixes","feed_subtitle":"Dynamic noise and innovation correction trims urban GNSS positioning error.","key_machinery":"The load-bearing mechanism is the learned EKF update. For each epoch, a network with multi-head self-attention and fully connected layers (dimensions 64, 128, 64) consumes per-satellite features — signal-to-noise ratio, elevation, azimuth, pseudorange residual, and a new DOP-contribution feature — and produces (1) the diagonal observation-noise covariance matrix $R$, kept positive via Softplus, and (2) the innovation compensation vector $v_c$. The EKF uses $R$ inside the innovation covariance and updates the state with $v_k + v_c$, using the Joseph-stabilized covariance update. The DOP-contribution feature is computed by removing each satellite from the design matrix and recomputing DOP, so the network sees each satellite's geometric importance. The dynamic hard-example-mining loss reweights training epochs by Euclidean ENU error, focusing learning on hard positioning cases.","core_discovery":"The central claim is that the tight fusion of learning and filtering works: a multi-head attention network that outputs a per-epoch diagonal measurement-noise matrix $R$ and a compensated innovation vector $v_c$, fed into a standard EKF, yields more accurate and stable pseudorange-only positioning than hand-tuned filters or earlier learning-based estimators. The paper reports 2D and 3D RMSE improvements of 79.55%–84.99% over CSSRLIB, goGPS, and RTKLIB, and a 53.68% improvement over TDL-GNSS on the private route, with the full framework reaching 3D RMSE of 1.91 m on ALGO-B while IE-PPP reaches 2.55 m. Ablations show that both the DOP-contribution feature and the dynamic hard-example-mining loss contribute to the improvement.","pith_inferences":["The framework's features are all derivable from pseudorange measurements and ephemeris, so the same learned $R$ and $v_c$ mechanism could plausibly be attached to other estimators, such as factor-graph optimization or an unscented Kalman filter, not just an EKF.","The paper does not report cross-city generalization; a natural testable extension is to train on one city's urban dataset and evaluate on another, using the released code, to see whether the learned noise and innovation models transfer or overfit local canyon geometry.","On open-sky segments, where all compared methods should agree, LF-GNSS should not be worse than the classical engines; a diagnostic that checks per-segment error would reveal whether the urban improvement comes with an open-sky cost."],"forward_implications":["The network adds only a small per-epoch forward pass, so LF-GNSS can run as a real-time pseudorange-only positioning engine on vehicles without IMU or camera.","Because the learned outputs are noise covariances and innovation offsets rather than positions, the same model can handle changing satellite geometry and multiple constellations (GPS, BeiDou, Galileo, GLONASS) through the inter-system-bias state.","On the private route, the full framework reaches 2D RMSE of 1.83 m on ALGO-A; removing the DOP feature raises it to 2.45 m and removing hard-example mining to 2.53 m, indicating both components carry part of the gain.","Against carrier-phase PPP, LF-GNSS keeps 3D RMSE at 1.91 m on ALGO-B while IE-PPP gives 2.55 m, suggesting better stability in deep urban scenes.","The planned open-source release of code and a collection of urban satellite-positioning datasets gives other groups the means to reproduce the comparisons."],"supporting_citations":[{"why":"Supplies the KLT and Whampoa public urban-canyon datasets used for the IPNL training and test splits.","marker":"[33]"},{"why":"Supplies the Campus public multi-sensor urban dataset series used for the GREAT training and test splits.","marker":"[34]"},{"why":"Supplies the SmartPNT-POS long-duration urban dataset split into PLANET-Train, A, B, and C.","marker":"[35]"},{"why":"Defines the TDL-GNSS baseline, the earlier deep-learning plus least-squares framework that LF-GNSS must beat.","marker":"[23]"},{"why":"Provides the LEAR-EKF precedent of learning adaptive robustness for GNSS filtering, motivating the learned noise formulation.","marker":"[26]"},{"why":"Shows a graph convolutional network coupled with a Kalman filter for GNSS corrections, one of the learning-filtering fusions LF-GNSS extends.","marker":"[24]"},{"why":"Couples a graph neural network with a Bayesian Kalman filter for smartphone positioning, the closest prior learning-filtering architecture.","marker":"[25]"},{"why":"Supplies the LightGBM-based NLOS identification work whose signal features inform the four baseline features in LF-GNSS.","marker":"[18]"}],"fun_headline_variants":["AI-tuned Kalman filter cuts urban GNSS error by 80%","Hard example mining boosts GNSS accuracy in urban canyons","Deep fusion network sharpens satellite positioning in cities","Learning-filtering framework trims multipath errors for GNSS","DOP-weighted AI filter improves urban GNSS fixes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation treats the reference trajectories as exact ground truth — RTK/INS tightly coupled solutions on the private route and the provided references on the public datasets — so if those references carry meter-level errors in urban canyons, the reported RMSE differences between methods become unreliable.","fun_headline_variants_meta":{"raw":{"variants":["AI-tuned Kalman filter cuts urban GNSS error by 80%","Hard example mining boosts GNSS accuracy in urban canyons","Deep fusion network sharpens satellite positioning in cities","Learning-filtering framework trims multipath errors for GNSS","DOP-weighted AI filter improves urban GNSS fixes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00115,"raw_usage":{"total_tokens":4790,"prompt_tokens":989,"completion_tokens":3801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":3717}},"tokens_in":605,"tokens_out":3801,"duration_ms":26203,"temperature":1.0,"reasoning_tokens":3717,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:11:16.608516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the private route comparison with ground truth from survey-grade total-station marks or a post-processed open-sky RTK/INS solution, and check whether LF-GNSS's 3D RMSE margins over TDL-GNSS and RTKLIB (1.91 m versus 2.70 m on ALGO-B, for example) survive; if the margin collapses to noise, the central accuracy claim is falsified.","supporting_citations":[{"cited_title":"Fisheye camera aided gnss nlos detection and learning-based pseudorange bias correction for intelligent vehicles in urban canyons,","cited_arxiv_id":null,"evidence_quote":"Supplies the KLT and Whampoa public urban-canyon datasets used for the IPNL training and test splits."},{"cited_title":"GREAT Dataset: A vehicle-mounted multi- sensor raw observation dataset in complex urban environment,","cited_arxiv_id":null,"evidence_quote":"Supplies the Campus public multi-sensor urban dataset series used for the GREAT training and test splits."},{"cited_title":"A large-scale diverse gnss/sins dataset: construction, publication, and application,","cited_arxiv_id":null,"evidence_quote":"Supplies the SmartPNT-POS long-duration urban dataset split into PLANET-Train, A, B, and C."},{"cited_title":"pyrtklib: An open-source package for tightly coupled deep learning and GNSS integration for positioning in urban canyons","cited_arxiv_id":"2409.12996","evidence_quote":"Defines the TDL-GNSS baseline, the earlier deep-learning plus least-squares framework that LF-GNSS must beat."},{"cited_title":"Learning-enhanced adaptive robust gnss navigation in challenging environments,","cited_arxiv_id":null,"evidence_quote":"Provides the LEAR-EKF precedent of learning adaptive robustness for GNSS filtering, motivating the learned noise formulation."},{"cited_title":"Learning gnss positioning corrections for smartphones using graph convolution neural networks,","cited_arxiv_id":null,"evidence_quote":"Shows a graph convolutional network coupled with a Kalman filter for GNSS corrections, one of the learning-filtering fusions LF-GNSS extends."},{"cited_title":"Tightly coupled graph neural network and kalman filter for smartphone positioning,","cited_arxiv_id":null,"evidence_quote":"Couples a graph neural network with a Bayesian Kalman filter for smartphone positioning, the closest prior learning-filtering architecture."},{"cited_title":"A reliable nlos error identification method based on lightgbm driven by multiple features of gnss signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the LightGBM-based NLOS identification work whose signal features inform the four baseline features in LF-GNSS."}],"review_version":1}