{"id":"d4088579-fb4e-4d7c-a54d-5cf8090de3df","arxiv_id":"2506.03537","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Rao-Blackwellized particle filter that estimates position and velocity separately, with NLOS rejection, improves GNSS positioning accuracy in urban environments without integer ambiguity resolution.","lead":"This paper describes a GNSS positioning method that combines a particle filter for position with a Kalman filter for velocity, avoiding the need to resolve integer ambiguities in carrier-phase measurements. Tests on urban driving data show the method stays accurate more often than a standard particle filter or conventional RTK-GNSS when satellite signals are obstructed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 14-point accuracy advantage may hinge on an undisclosed, position-dependent NLOS-rejection threshold η in Eq. (15); without its value and sensitivity, the central claim is not established.","rationale":"The central claim is that the proposed RBPF, combining per-particle KF velocity estimation with pseudorange-residual-based NLOS Doppler rejection, achieves significantly better urban positioning than a conventional PF and RTKLIB. For this claim to hold, the gate in Eq. (15) must robustly separate LOS from NLOS at the particle positions used inside the filter, and its setting must not be a tuned artifact of the one reported trajectory. This is the least secure condition in the paper. The reader identified the threshold η as the weakest assumption; I agree and sharpen the point: the residual in Eq. (7) conflates the particle position error with multipath, so the gate is not an invariant NLOS classifier. A particle far from the true position can have all residuals exceed η regardless of satellite health, while an NLOS satellite can pass if its pseudorange delay compensates the geometry error. Thus the gate's behavior, and hence the velocity updates and particle transitions that distinguish the method from the conventional PF, depends on the distribution of particles as much as on the environment. The single trajectory and absence of repeated trials make this especially serious: the 14-point CDF improvement reported in Fig. 7 could be due to threshold overfitting. Secondary issues include the under-derived AFV-based KF innovation and missing error bars, but the threshold fragility is the most load-bearing because it directly controls which measurements enter the KF and can change the qualitative behavior of the algorithm, not just the tuning of a nuisance parameter. Since the reader's CONDITIONAL verdict already requires threshold disclosure, sensitivity analysis, and additional experiments, my read does not change the verdict.","tokens_in":10145,"tokens_out":8693,"duration_ms":111400,"concrete_test":"Run the proposed RBPF and the conventional PF on the same recorded Nagoya dataset with η swept over {∞, 10, 5, 2, 1, 0.5, 0.1} m, and also with the actual value if recovered from the authors. Plot the within-0.3 m percentage (the Fig. 7 metric) and the full CDF for each η. If the RBPF advantage over the conventional PF is maintained across at least two decades of η and repeats on a second urban route, the concern is resolved; if the reported 68.5% is an isolated peak near one η, the central claim fails to generalize.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy gain is attributed to per-particle NLOS Doppler rejection via the gate in Eq. (15): |d(ρ_k, x_i^n)| > η, using the DD pseudorange residual in Eq. (7). The value of η is not reported—only \"determined experimentally\"—and no sensitivity analysis is given. The concern is stronger than nondisclosure: d(ρ_k, x_i^n) at an arbitrary particle position is not a pure NLOS indicator. It equals the DD multipath error plus the projection of the particle's position error onto the satellite line of sight. A fixed η therefore gates different satellites differently depending on how far the particle is from the true position: a LOS satellite can be rejected for a particle with large position error, and an NLOS satellite can pass when its multipath roughly cancels the particle's geometric residual. Since the claimed 68.5% versus 54.4% within 0.3 m (Fig. 7) comes from a single urban trajectory, the 14-point difference could be produced by an η tuned to that specific dataset. Without the threshold value, a sensitivity sweep, and at least one independent route, the improvement over the conventional PF—which uses the same AFV likelihood—is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Rao-Blackwellized particle filter (RBPF) for GNSS carrier-phase positioning that avoids integer ambiguity resolution. In the proposed method, each particle carries a 3D position estimated by a particle filter and a 3D velocity estimated by a per-particle Kalman filter. The velocity update uses Doppler observations after rejecting NLOS satellites with a threshold on double-differenced pseudorange residuals evaluated at each particle position. The position likelihood is computed from the ambiguity function value (AFV) of double-differenced carrier phase, following the authors' earlier PF work. A single-vehicle urban experiment compares the proposed method with a conventional PF and RTKLIB. The reported results show that the proposed method attains 68.5% of position errors within 0.3 m versus 54.4% for the conventional PF and 52.7% for RTKLIB, and also reports improved velocity accuracy and robustness to reduced particle count.","tokens_in":10389,"tokens_out":5102,"duration_ms":55897,"significance":"If the reported results hold, the paper offers a useful extension of AFV-based particle filtering to urban GNSS positioning, adding a Rao-Blackwellized velocity state and an NLOS rejection mechanism that improves particle diversity and robustness. The empirical demonstration on real data, the per-particle NLOS gating idea, and the observation that accuracy is maintained with fewer particles are valuable contributions. However, the central accuracy claim rests on a single trajectory, an undisclosed rejection threshold, no ablation isolating the proposed components, and no repeated-run statistics. The paper also does not release code or data, which limits reproducibility. The significance is therefore conditional on addressing these validation gaps.","major_comments":[{"comment":"The NLOS rejection gate in Eq. (15) is a load-bearing component of the proposed method, but the threshold eta is only described as 'determined experimentally'. The paper does not report its value, its units, or any sensitivity analysis. Moreover, the residual d(rho_k, x_i^n) at an arbitrary particle position is not a clean NLOS indicator: it contains the DD multipath error plus the projection of the particle position error onto the satellite line of sight. A fixed threshold can therefore reject LOS observations for particles that are far from the true position and can accept NLOS observations when the multipath error approximately cancels the geometric residual. Since the reported 68.5% versus 54.4% improvement in Fig. 7 comes from a single urban trajectory, please report the threshold value, provide a sensitivity sweep over eta, show that the gate separates NLOS from LOS observations in this dataset, and validate on at least one independent route.","section":null},{"comment":"The use of the AFV as an innovation term in the Kalman filter measurement update is not derived. Equation (11) is written as a standard KF update with innovation y_t - h(x_t^n) - C xhat_t|t-1, but the text then states that the AFV is applied directly to the nonlinear-state innovation y_t - h(x_t^n). The AFV is not a position or velocity measurement, and the observation model in Eq. (6) does not contain an AFV term. The covariance R_t for this pseudo-measurement is not specified. Because the AFV-based correction is part of the proposed velocity estimation, please derive the pseudo-measurement model, define the associated covariance, and justify the Kalman gain expression, or remove the AFV correction from the KF update.","section":null},{"comment":"The experimental evaluation is based on a single trajectory, and the PF-based methods are stochastic. The paper reports no multiple independent runs, no error bars, and no statistical comparison. The difference between 68.5% and 54.4% of estimates within 0.3 m could be within the run-to-run variation caused by particle resampling. Please report results over multiple random seeds (e.g., mean and spread of the CDF), state whether common random numbers are used, and give the number of epochs as well as the temporal correlation structure of the errors.","section":null},{"comment":"The comparison between the proposed RBPF and the conventional PF in Figs. 4-5 and 7 confounds three changes at once: the Rao-Blackwellized velocity state, the per-particle Kalman filter, and the NLOS rejection gate. The claimed improvement in velocity and position accuracy is attributed to 'Rao-Blackwellization and NLOS rejection', but no ablation is provided. Please include at least an RBPF variant without NLOS rejection and a conventional PF with the same NLOS rejection to isolate the contribution of each component.","section":null}],"minor_comments":[{"comment":"The sentence after Eq. (14) says that M_t denotes the covariance matrix of the updated state following the measurement update, but M_t as defined in Eq. (14) is the innovation covariance in a standard Kalman filter. Please correct this wording.","section":null},{"comment":"The symbols A_t^n, A_t^l, \\bar{A}_t^l, and \\bar{A}_t^n are used in the time-update equations but are not all defined in the text. Please define each matrix and clarify the relationship between A_t and \\bar{A}_t.","section":null},{"comment":"The index term 'Partilce Filter' contains a typo and should read 'Particle Filter'.","section":null},{"comment":"The values of the observation standard deviation sigma_Phi in Eq. (9) and the KF covariance matrices Q, R, and the initial covariance are not reported. Since these are additional tuning parameters, please state their values in Section IV.","section":null},{"comment":"The paper reports percentages within 0.3 m but does not state the total number of epochs used to compute the CDFs in Fig. 7. Please add the epoch count and the duration of the dataset.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonably clear empirical extension of the authors' prior AFV-PF work, but the current validation is too thin for a journal-level claim of robust improvement: one trajectory, one stochastic run, and an undisclosed threshold on the key NLOS gate. I would encourage the editor to require the sensitivity/ablation and repeated-run statistics described in the major comments. If the authors can provide those, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. It applies Rao-Blackwellized particle filtering to GNSS-only carrier-phase positioning, with each particle running a Kalman filter for velocity and a pseudorange-residual gate to drop NLOS Doppler measurements before the velocity update. That combination is genuinely new—I checked the cited literature and the authors are right that RBPF hasn't been used for GNSS-only positioning before. The paper also gives a fair baseline comparison: conventional PF and a stock RTKLIB kinematic solution, on real urban data with a high-grade ground truth.\n\nThe good: the velocity estimation result is believable, especially at the two underpasses where the conventional PF loses velocity entirely and the RBPF keeps estimating via the position-change innovation. The position CDF shows a 68.5% within-0.3 m rate against 54.4% and 52.7% for the baselines; the particle-count sweep is a nice robustness check and shows the method degrades more gracefully.\n\nThe soft spots are all around the experimental support. The NLOS rejection threshold η in Eq. (15) is never reported, only \"determined experimentally,\" and there is no sensitivity analysis. That matters more than usual because the gate is not a clean NLOS indicator: the residual also contains the particle's position error projected onto the line of sight, so a fixed threshold can reject LOS satellites for a badly misplaced particle and let NLOS satellites through when multipath happens to cancel the geometric term. On a single urban route, a tuned η could explain a good chunk of the 14-point gap. There are no error bars or repeated runs, and the abstract's \"centimeter-accurate\" overstates the actual 0.3 m accuracy level achieved. The KF/AFV measurement update is also under-derived: treating the AFV as an innovation is plausible but needs a clearer statistical justification than the paragraph in Sec. III-C.\n\nThe concerns are addressable, not fatal. The method is plausible, the empirical evidence is suggestive, and the idea is worth serious referee time. I'd want the authors to report η, run a sensitivity sweep, and add at least one more independent route before I'd trust the accuracy claim; the architecture itself is solid enough to engage with. I would not cite it as a definitive result yet, but I'd definitely bring it to a reading group interested in urban GNSS state estimation.\n\nRecommendation: send to peer review. It deserves a serious referee. The revision should require the threshold disclosure and a sensitivity analysis.","headline":"A genuinely new RBPF formulation for GNSS-only urban positioning, but the headline accuracy gain rests on an undisclosed, position-dependent NLOS threshold and a single trajectory.","tokens_in":10909,"tokens_out":2720,"would_cite":false,"duration_ms":29249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An urban GNSS filter keeps 68.5% of position fixes within 0.3 m without ever resolving integer ambiguities.","keywords":["GNSS positioning","Rao-Blackwellized particle filter","integer ambiguity resolution","ambiguity function","NLOS multipath rejection","urban positioning","Doppler velocity estimation","carrier phase positioning"],"falsifier":"A direct check: sweep the rejection threshold from zero to infinity on the same logged urban trajectories while keeping everything else fixed; if no threshold reproduces the reported gap over the conventional particle filter, or if the best threshold changes sharply between routes, the residual gate is not carrying the reported gain.","tokens_in":9925,"feed_emoji":"🛰️","tokens_out":9836,"duration_ms":95388,"temperature":0.7,"pith_summary":"Conventional centimeter-grade GNSS positioning (RTK) depends on resolving integer ambiguities in carrier-phase measurements, and in cities those ambiguities constantly break under multipath and signal blockage. This paper tries to remove that fragile step: a particle filter scores candidate positions by the fractional-wavelength residual of the carrier phase, which needs no ambiguity resolution, and the paper's contribution is to make that filter survive urban conditions. The fix is Rao-Blackwellization: position stays a particle-filter state, while each particle's velocity becomes a linear state tracked by its own Kalman filter. Doppler observations from satellites flagged as non-line-of-sight by large pseudorange residuals at the particle's position are rejected before the velocity update, which improves velocity accuracy and lets different particles keep different velocity hypotheses. In the reported urban vehicle test, 68.5% of position estimates were within 0.3 m of ground truth, against 54.4% for the conventional particle filter and 52.7% for conventional RTK-GNSS.","feed_headline":"New GNSS filter beats RTK in cities, skips fragile ambiguity step","feed_subtitle":"A particle filter with per-particle Kalman velocities rejects NLOS multipath; 68.5% of fixes land within 0.3 m.","key_machinery":"The central object is a Rao-Blackwellized (marginalized) particle filter for GNSS: the state is factored into a nonlinear three-dimensional position, sampled by particles, and a linear three-dimensional velocity, estimated per particle by a Kalman filter. The position likelihood is computed from the ambiguity function value (AFV), the fractional-wavelength residual of each double-differenced carrier phase at a particle's position, which is zero at the true position and has multiple wavelength-spaced peaks that the product over satellites suppresses. The velocity Kalman filter is the workhorse: its measurement update uses Doppler observations after rejecting non-line-of-sight satellites with Eq. (15), and its time update uses the particle's own displacement, so velocity survives gaps in GNSS observations. The paper credits this per-particle Kalman velocity, not the likelihood, with preserving particle diversity and carrying the accuracy gain.","core_discovery":"At its center, the paper claims that the practical bottleneck for ambiguity-free carrier-phase positioning in urban environments is the state transition, not the measurement model. The measurement model already has a sharp likelihood based on the ambiguity function value of double-differenced carrier phase, but that sharp peak is useless if the particle cloud never moves near the true position. The proposed method therefore splits the state into a nonlinear position component, sampled by particles, and a linear three-dimensional velocity component, estimated for every particle by a Kalman filter conditioned on that particle's position. The Kalman update uses Doppler velocities built from only those satellites whose pseudorange residual at the particle's position is below a rejection threshold, so each particle can use a different satellite set. Because the Kalman time update also incorporates the observed displacement of the particle, velocity estimates continue even through underpasses where Doppler observations vanish. The paper reports that this design places more particles near the true position and yields better position accuracy than both the previous particle-filter method and conventional RTK-GNSS on a real urban route.","pith_inferences":["An inference the authors leave implicit: the NLOS rejection threshold could be estimated per epoch from the distribution of pseudorange residuals instead of being fixed experimentally, which would make the method less sensitive to environment changes.","Because each particle keeps its own velocity, the same architecture could ingest raw Doppler measurements directly in the Kalman filter (the authors list this as future work), removing the intermediate least-squares velocity solution and helping when fewer than four satellites are visible.","The AFV likelihood's wavelength-spaced side peaks mean performance likely depends on satellite geometry; a stress test with fewer satellites or a single-constellation urban canyon would show where the method starts to fail.","If the thresholding works as reported, a 3D city model could make it sharper: pseudorange residuals computed against a map-constrained particle position should separate NLOS rays more cleanly than residuals from an unconstrained position."],"forward_implications":["Urban positioning no longer requires integer ambiguity resolution or its re-initialization after cycle slips; the ambiguity-free likelihood remains the source of position information.","Velocity estimation continues through underpasses where Doppler observations disappear, because the Kalman time update combines the particle's own displacement with the previous velocity.","The reported accuracy gain is a jump to 68.5% of fixes within 0.3 m, from 54.4% for the previous particle filter and 52.7% for conventional RTK-GNSS.","Accuracy is maintained at 1000 particles, so the method can run with lower computational cost than the conventional particle filter.","Per-particle velocity estimates preserve particle diversity, so the filter is less likely to collapse when a single shared velocity would push all particles away from the true position."],"supporting_citations":[{"why":"The previous particle-filter method that computes position likelihood from carrier-phase residuals without integer ambiguity resolution; it is the baseline the paper extends.","marker":"[3]"},{"why":"Source of the ambiguity function value used in Eq. (8) to score particle positions from the fractional carrier phase.","marker":"[9]"},{"why":"Introduces Rao-Blackwellized particle filtering, the factorization that lets the paper add a Kalman-filtered velocity state.","marker":"[13]"},{"why":"Marginalized particle filtering for mixed linear/nonlinear state-space models, the basis for treating velocity as the linear state.","marker":"[14]"},{"why":"Open-source RTK-GNSS software used as the conventional baseline in the experimental comparison.","marker":"[26]"}],"fun_headline_variants":["Rao-Blackwellized GNSS filter hits cm accuracy in cities","Ambiguity-free GNSS positioning improved by per-particle Kalman velocity","Urban GNSS filter rejects NLOS, beats RTK with particle velocity updates","Particle filter skips integer fixing, uses Kalman velocities for urban accuracy","GNSS carrier-phase positioning in cities: no ambiguity, just better state transitions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single experimentally tuned cutoff cleanly separates corrupted non-line-of-sight satellite measurements from clean ones by the size of the distance residual at each candidate position, and the paper does not report the cutoff value or how much the accuracy gain depends on it.","fun_headline_variants_meta":{"raw":{"variants":["Rao-Blackwellized GNSS filter hits cm accuracy in cities","Ambiguity-free GNSS positioning improved by per-particle Kalman velocity","Urban GNSS filter rejects NLOS, beats RTK with particle velocity updates","Particle filter skips integer fixing, uses Kalman velocities for urban accuracy","GNSS carrier-phase positioning in cities: no ambiguity, just better state transitions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":1192,"prompt_tokens":1029,"completion_tokens":163,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":74}},"tokens_in":645,"tokens_out":163,"duration_ms":3072,"temperature":1.0,"reasoning_tokens":74,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:59:48.684755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: sweep the rejection threshold from zero to infinity on the same logged urban trajectories while keeping everything else fixed; if no threshold reproduces the reported gap over the conventional particle filter, or if the best threshold changes sharply between routes, the residual gate is not carrying the reported gain.","supporting_citations":[{"cited_title":"Multiple Update Particle Filter: Position Estimation by Com- bining GNSS Pseudorange and Carrier Phase Observations,","cited_arxiv_id":null,"evidence_quote":"The previous particle-filter method that computes position likelihood from carrier-phase residuals without integer ambiguity resolution; it is the baseline the paper extends."},{"cited_title":"Miniature interferometer terminals for earth surveying: ambiguity and multipath with global positioning system,","cited_arxiv_id":null,"evidence_quote":"Source of the ambiguity function value used in Eq. (8) to score particle positions from the fractional carrier phase."},{"cited_title":"Particle filters for state estimation of jump Markov linear systems,","cited_arxiv_id":null,"evidence_quote":"Introduces Rao-Blackwellized particle filtering, the factorization that lets the paper add a Kalman-filtered velocity state."},{"cited_title":"Marginalized particle filters for mixed linear/nonlinear state-space models,","cited_arxiv_id":null,"evidence_quote":"Marginalized particle filtering for mixed linear/nonlinear state-space models, the basis for treating velocity as the linear state."},{"cited_title":"Development of the low-cost RTK-GPS receiver with an open source program package RTKLIB,","cited_arxiv_id":null,"evidence_quote":"Open-source RTK-GNSS software used as the conventional baseline in the experimental comparison."}],"review_version":1}