{"id":"bbd8f229-bad0-4a5b-b486-c6ac256d7457","arxiv_id":"2508.19736","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Three real-world 5G testbeds using OpenAirInterface achieve 1-2 m positioning accuracy in 90% of cases via filtered UL-TDoA with PSO, plus CIR-based fingerprinting for NLoS cases.","lead":"OpenAirInterface-based 5G testbeds in a factory, an outdoor terrace, and an Airbus hall show that uplink TDoA positioning with filtering and a particle swarm optimizer can reach 1-2 meter accuracy in most tested conditions. The authors also release the measured channel impulse response datasets to support further work on data-driven positioning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 1–2 m / 90% claim is not supported by Table 2: the PSO/TDoA pipeline exceeds 2 m at GEO-5G per-RU (2.02), handheld (2.36), and Airbus (8.94); only ML fingerprinting reaches <2 m at Airbus.","rationale":"The paper makes a genuine experimental contribution: an open OAI LMF implementation, three distinct testbeds, and a public CIR dataset. The reader's weakest_assumption about the strongest-peak ToA (Eq. 1) is real and contributes to the Airbus TDoA failure (CE90 = 8.94 m). However, I find a more direct, load-bearing problem in the paper's own Table 2: even the proposed PSO/TDoA pipeline does not consistently achieve 1-2 m at the 90th percentile outside two favorable configurations (GEO-5G with 8 non-collinear antennas, Stellantis filtered). The per-RU reference at GEO-5G is 2.02 m and the handheld mobile scenario is 2.36 m, so the '1-2 m in 90% of cases' claim is not supported by the data as stated. Additionally, the paper does not report any train/test separation for the filter thresholds or the CNN model, which is a serious missing support for the sub-meter Airbus fingerprinting result. These issues do not destroy the value of the work, but they require conditionally revising the central claim, adding held-out evaluations, and reporting per-configuration statistical details. The reader's CONDITIONAL verdict is therefore appropriate, though for a broader set of reasons than the ToA assumption alone.","tokens_in":10965,"tokens_out":9507,"duration_ms":103196,"concrete_test":"Run a strict held-out evaluation: split each testbed's labeled measurements by time or by spatial grid into calibration/training and test sets; fit the ToA/TDoA filter thresholds (μτ, στ) and the CNN (including masking threshold γ) only on the training split; report CE90 on the test split for the PSO/TDoA pipeline and for the fingerprinting method. If the test-split CE90 for any testbed exceeds the reported value by more than 20% or the aggregate TDoA CE90 crosses 2 m, the central claim fails as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim ('1-2 meter positioning accuracy in 90% of cases in different testbeds') is not supported by the paper's own results. Table 2 reports CE90 (90th percentile error) for the proposed PSO/TDoA pipeline as: GEO-5G 8 non-collinear 1.96 m, per-RU reference 2.02 m, handheld mobile 2.36 m; Stellantis filtered 1.99 m; Airbus 16-antenna TDoA 8.94 m. Thus among TDoA configurations, only the 8-non-collinear GEO-5G static case and the filtered Stellantis case meet '<2 m'; the per-RU reference and mobile UE exceed 2 m, and Airbus fails badly. The sub-meter Airbus result (0.74 m) comes from a separate CNN fingerprinting method, not from the LMF PSO pipeline. Moreover, Section 5.1's ToA filter thresholds and Section 5.3.2's NLoS mask threshold γ are derived empirically from CIR datasets, and no train/test separation is reported for either the filter calibration or the CNN training/evaluation; if these are fit on the same data used to compute CE90, the reported accuracies, including 0.74 m, are optimistic. The abstract therefore overstates both the magnitude and the generality of the result: the PSO/TDoA pipeline does not achieve 1-2 m in 90% of cases across the three testbeds as claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports experimental UL-TDoA positioning results from three 5G testbeds (EURECOM GEO-5G outdoor, STELLANTIS indoor, Airbus factory hall) built on OpenAirInterface with an LMF. It proposes a processing pipeline comprising statistical ToA filtering, geometry-based TDoA filtering, a PSO position estimator, and an optional CIR-based CNN fingerprinting method with NLoS antenna masking. The authors release the datasets publicly. The central claim, stated in the abstract and introduction, is that 1–2 m positioning accuracy is achieved in 90% of cases across the testbeds.","tokens_in":11418,"tokens_out":6110,"duration_ms":70831,"significance":"If the claims hold, this is a valuable experimental contribution: it demonstrates an open-source, 3GPP-compliant LMF positioning pipeline on real O-RAN hardware, validates a geometric TDoA bound, releases a public CIR dataset, and shows that a CNN fingerprinting method can outperform TDoA in a dense factory environment. The PSO estimator is self-contained and independent of the reported error metrics, and the geometric bound in Theorem 5.1 is correct. However, the headline accuracy claim is not supported by Table 2, and the calibration of the ToA filter and NLoS mask uses the same data on which performance is evaluated, which may make the reported numbers optimistic.","major_comments":[{"comment":"The abstract states that results demonstrate '1–2 meter positioning accuracy in 90% of cases in different testbeds', but Table 2's CE90 values for the proposed PSO/TDoA pipeline are: GEO-5G 8 non-collinear antennas 1.96 m, per-RU reference 2.02 m, handheld mobile UE 2.36 m, Stellantis filtered 1.99 m, and Airbus 16-antenna TDoA 8.94 m. Only two of these configurations are below 2 m at CE90; the per-RU and handheld configurations exceed 2 m, and the Airbus TDoA result fails by a large margin. The sub-meter Airbus result (0.74 m) is from the CNN fingerprinting method, not the LMF TDoA pipeline. The conclusion (§7) itself is more modest ('approximately 2m accuracy in 90% of cases under favorable conditions'). The abstract and introduction should be revised to state which method and which configurations meet the 1–2 m target, and the 90%-of-cases claim should be made configuration-specific.","section":"Abstract, §6, Table 2"},{"comment":"The ToA filter uses empirical mean μτ and standard deviation στ of the max-peak delay distribution, but the manuscript does not state whether this distribution is estimated on a dataset disjoint from the CIRs whose ToAs are filtered and then scored in Table 2. If the same data are used to fit the filter bounds and to compute CE90, the reported accuracy is optimistically biased. The authors should either report filter parameters calibrated on an independent training set and evaluate on a held-out test set, or quantify the sensitivity of the results to the choice of the ±1σ threshold. This is load-bearing for all filtered TDoA results, including the claimed 1.99 m at Stellantis.","section":"§5.1, Eq. (3)"},{"comment":"The NLoS masking threshold γ=0.4 is selected 'empirically based on analysis of a labeled dataset' (Fig. 6b) and then applied in both training and inference. No train/validation/test partition is described for this threshold selection, and no training details for the CNN (loss function, optimizer, epochs, split ratio, repeated runs) are provided. The reported Airbus fingerprinting result, 0.74 m CE90, is therefore not established as a genuine test-set result. The authors should report a clear data partition, show that γ is chosen only on the training/validation portion, and give test-set metrics with confidence intervals.","section":"§5.3.2, Eq. (15), Table 2"},{"comment":"ToA is defined as the delay of the strongest multipath component, i.e., the peak of the CIR magnitude. In NLoS conditions this peak is a reflected path, not the first arrival. The Airbus TDoA result (CE90 = 8.94 m) is consistent with this limitation. The manuscript should explicitly state that the TDoA pipeline is expected to be reliable only when the strongest peak is a valid proxy for the true arrival time, and the '1–2 m' claim should be scoped accordingly. Alternatively, a first-peak detector or an LoS/NLoS classification step before TDoA estimation would be needed to support a broader claim.","section":"§3, Eq. (1); §4.3, Table 2"}],"minor_comments":[{"comment":"Eq. (11) uses u_p^(t) on the right-hand side, but the iteration index should be i, not t. Also, the PSO hyperparameters are partially given (w=0.9, c1=0.5, c2=0.9), but the number of particles, number of iterations, and moving-average window size are not reported; these are needed for reproducibility.","section":"§5.3.1, Eqs. (10)-(11)"},{"comment":"The notation for the speed of light is inconsistent: Eq. (6) uses v, while Eq. (7) uses c. Unify the notation and define the symbol explicitly.","section":"§5.1, Eq. (7)"},{"comment":"Table 2 reports MAE and CE90 without sample sizes, confidence intervals, or the number of test points. The reader cannot assess the statistical significance of differences between configurations (e.g., 1.96 m vs 2.02 m). Please add these details.","section":"§6, Table 2"},{"comment":"The conversion from RTK geographic coordinates to the local Cartesian frame is described as 'a linear transformation', but the transformation parameters (translation, rotation, scaling) and how they were estimated are not given. This is important because ground truth accuracy underpins all reported errors.","section":"§4.1"},{"comment":"The description of the CNN input states the CIR matrix is in C^{M×C} and uses complex notation, but the layers in Table 1 appear to treat it as real. Clarify whether the absolute value/real part is used as input, and describe the preprocessing consistently.","section":"§5.3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper gives the 5G positioning community something it actually lacks: a complete open LMF pipeline on OpenAirInterface, run on three real testbeds, with CIR data and ground truth released publicly. The per-RU TDoA reference trick for mitigating inter-RU clock drift is a sensible engineering fix, and the Airbus story is honest — they show TDoA falling apart under NLoS and then demonstrate a CIR fingerprinting CNN with NLoS masking that gets CE90 down to 0.74 m. That's a real result if it holds up.\n\nBut the abstract oversells. Table 2 shows the PSO/TDoA pipeline only hits sub-2 m CE90 in two configurations: the 8-antenna GEO-5G static case (1.96 m) and the filtered Stellantis case (1.99 m). Per-RU reference is 2.02 m, handheld 2.36 m, and Airbus TDoA is 8.94 m. The 1–2 m / 90% claim works only if you cherry-pick the fingerprinting number and ignore the TDoA failures. The conclusion says 'under favorable conditions,' which is closer to the truth.\n\nThe bigger issue is circularity. The ToA filter in Eq. 3 uses the empirical mean and standard deviation of the max-peak delay distribution from the same dataset it filters; the NLoS masking threshold gamma=0.4 is chosen from a labeled dataset; and there is no train/test split reported for either the filter calibration or the CNN. If those thresholds are fit to the same CIRs used for evaluation, the accuracy numbers, including the 0.74 m, are optimistic. The PSO estimator itself is self-contained, so the geometric TDoA results aren't as tainted, but the preprocessing steps are.\n\nMinor concerns: no error bars, no sample sizes, and no comparison to classical estimators like Chan's method. The strongest-path ToA assumption is a known weak point, and the authors don't dwell on it, though Airbus shows the consequence.\n\nBottom line: this is a workshop-level experimental paper with a strong dataset component. It deserves a serious referee, not a desk reject, but the referee should push for an abstract rewrite and a proper train/test separation. I'd cite it for the dataset and the LMF implementation, and I'd bring it to a reading group to discuss where data-dependent filtering gets dangerous.","headline":"Useful testbed paper with a real dataset, but the abstract's 1-2 m / 90% claim overstates what the TDoA pipeline achieves and the filters look data-fit.","tokens_in":11898,"tokens_out":3923,"would_cite":true,"duration_ms":39458,"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":"This paper reports that a fully open-source 5G positioning pipeline, built with OpenAirInterface and uplink TDoA, achieves 1–2 meter accuracy in 90% of cases across three real-world testbeds after adding ToA/TDoA filtering and particle-swar","keywords":["5G positioning","UL-TDoA","OpenAirInterface","Location Management Function","Particle Swarm Optimization","Channel Impulse Response","fingerprinting","O-RAN"],"falsifier":"Run the same filtered LMF pipeline in a factory hall with known ground truth while deliberately placing a strong metal reflector so that the reflected CIR peak is the strongest; if more than 10% of test points exceed 2 m error, the generalised 1–2 m/90% claim does not hold in such NLoS-dominated settings.","tokens_in":10910,"feed_emoji":"📡","tokens_out":5138,"duration_ms":54471,"temperature":0.7,"pith_summary":"This paper tries to establish that high-precision 5G positioning is practical with open-source network components, not just in simulation. It reports results from three real testbeds—an outdoor rooftop deployment, an indoor automotive lab, and an aircraft factory hall—using uplink time-difference-of-arrival with a newly integrated Location Management Function in the OpenAirInterface stack. To make TDoA work despite clock drift and multipath, the authors add filters that discard physically impossible measurements and a particle-swarm optimizer that estimates position. They claim 1–2 meter accuracy in 90% of test points under favorable conditions, and show a channel-impulse-response fingerprinting method that outperforms TDoA in the hardest NLoS factory environment. A public dataset of measured CIRs, timestamps, and ground-truth positions accompanies the paper.","feed_headline":"1–2 meter 5G positioning achieved on open-source testbeds","feed_subtitle":"ToA/TDoA filtering plus particle-swarm optimization puts 90% of fixes inside 2 m on live networks.","key_machinery":"The load-bearing objects are: the strongest-peak ToA read off the channel impulse response (Eq. 1); the per-RU TDoA difference (Eq. 5) that cancels inter-RU clock drift; a TDoA bound from the triangle inequality (Theorem 5.1) used as a physical-feasibility filter; and a particle-swarm optimizer that minimizes the TDoA loss (Eqs. 8–11). For the data-driven branch, the mechanism is a TDoA-aligned, peak-normalized CIR matrix (Eq. 14) with a binary LoS/NLoS mask (Eq. 15) fed into a CNN that regresses 2D position.","core_discovery":"The central claim is that a complete positioning pipeline implemented in the OpenAirInterface Location Management Function—empirical ToA filtering, geometry-bounded TDoA filtering, per-RU TDoA references, and a particle-swarm optimizer—makes UL-TDoA reliable enough to reach 1–2 m error in 90% of cases on three distinct live 5G networks. The paper also demonstrates that where TDoA degrades in dense multipath, an O-RAN-inspired fingerprinting framework using TDoA-aligned, peak-normalized CIRs with a learned LoS/NLoS mask achieves sub-meter accuracy, and that this data-driven method degrades gracefully when fewer antennas are available.","pith_inferences":["The strongest-peak ToA definition ties the pipeline's success to environments where the direct path is also the strongest path; extending the filters to detect first-arrival peaks (e.g., through super-resolution or learned timing features) would likely close the gap in the hardest NLoS factory cases.","The per-RU reference strategy effectively treats each radio unit as a self-synchronized cluster, suggesting a natural extension where inter-RU clock offsets are estimated jointly with the UE position, potentially removing the need for precision time protocol altogether.","The LoS/NLoS masking threshold (gamma = 0.4) is tuned to this dataset; a self-calibrating threshold based on training CIR statistics could make the fingerprinting method transferable across sites without manual adjustment.","The MQTT-based CIR offloading used here is a stand-in for the standard E2 interface; validating the same fingerprinting model over E2 with a near-real-time RIC would test whether the approach holds under standard O-RAN latency and signaling constraints."],"forward_implications":["Operators can replicate 1–2 m UL-TDoA positioning using open-source network functions and commercial O-RAN radio units, without proprietary positioning modules.","The per-RU TDoA reference strategy mitigates clock drift in multi-RU deployments, so large outdoor coverage does not require tight cross-RU synchronization.","ToA/TDoA filtering based on empirical delay-spread bounds and the triangle-inequality limit removes physically impossible measurements, improving robustness in dense multipath.","CIR-based fingerprinting with NLoS masking can outperform TDoA accuracy in factory halls and degrades gracefully as antenna count drops.","The public dataset of CIRs with ground-truth labels enables other groups to benchmark and train data-driven positioning methods against a common reference."],"supporting_citations":[{"why":"Prior work that introduced the open-source NRPPa and LMF implementation in OpenAirInterface, which this paper extends with filtering and PSO.","marker":"[3]"},{"why":"Earlier comparison study that justifies choosing PSO over gradient and least-squares TDoA methods.","marker":"[4]"},{"why":"3GPP NRPPa specification defining positioning signaling between gNB and LMF, used to retrieve ToAs.","marker":"[1]"},{"why":"3GPP specification defining the LMF and its service-based interface, which the pipeline implements.","marker":"[2]"},{"why":"Classic hyperbolic TDoA estimator used as a baseline comparison for the PSO approach.","marker":"[6]"},{"why":"Standard least-squares TDoA positioning method that PSO is compared against.","marker":"[12]"},{"why":"Super-resolution ToA estimation method cited as an alternative for dense multipath, which the proposed low-complexity filtering avoids.","marker":"[15]"},{"why":"Metric-space reference supplying the triangle inequality used to bound feasible TDoA values in Theorem 5.1.","marker":"[14]"},{"why":"GPS textbook describing RTK carrier-phase positioning, the centimeter-level ground-truth method used for mobile trajectories.","marker":"[18]"}],"fun_headline_variants":["OpenAirInterface 5G testbeds hit 1–2 meter positioning 90% of the time","Particle swarm boosts open-source 5G positioning to 1–2 m accuracy","5G UL-TDoA meets 1–2 m error on three live testbeds with OpenAirInterface","Open-source 5G positioning overcomes multipath for 90% sub-2m fixes","CIR fingerprinting beats TDoA in dense multipath on 5G testbeds"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ToA is taken as the delay of the strongest multipath peak, which is a valid timing reference only when the direct path is also the strongest path; in dense NLoS environments that condition fails, as the paper's own Airbus TDoA results show.","fun_headline_variants_meta":{"raw":{"variants":["OpenAirInterface 5G testbeds hit 1–2 meter positioning 90% of the time","Particle swarm boosts open-source 5G positioning to 1–2 m accuracy","5G UL-TDoA meets 1–2 m error on three live testbeds with OpenAirInterface","Open-source 5G positioning overcomes multipath for 90% sub-2m fixes","CIR fingerprinting beats TDoA in dense multipath on 5G testbeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1238,"prompt_tokens":795,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":317}},"tokens_in":539,"tokens_out":443,"duration_ms":5369,"temperature":1.0,"reasoning_tokens":317,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:29:29.723243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same filtered LMF pipeline in a factory hall with known ground truth while deliberately placing a strong metal reflector so that the reflected CIR peak is the strongest; if more than 10% of test points exceed 2 m error, the generalised 1–2 m/90% claim does not hold in such NLoS-dominated settings.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier comparison study that justifies choosing PSO over gradient and least-squares TDoA methods."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"3GPP NRPPa specification defining positioning signaling between gNB and LMF, used to retrieve ToAs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"3GPP specification defining the LMF and its service-based interface, which the pipeline implements."},{"cited_title":"and Ho, K.C","cited_arxiv_id":null,"evidence_quote":"Classic hyperbolic TDoA estimator used as a baseline comparison for the PSO approach."},{"cited_title":"Prediction of number of cases expected and estimation of the final size of coronavirus epidemic in India using the logistic model and genetic algorithm","cited_arxiv_id":"2003.12017","evidence_quote":"Standard least-squares TDoA positioning method that PSO is compared against."},{"cited_title":"Pahlavan","cited_arxiv_id":null,"evidence_quote":"Super-resolution ToA estimation method cited as an alternative for dense multipath, which the proposed low-complexity filtering avoids."},{"cited_title":"Khamsi and William A","cited_arxiv_id":null,"evidence_quote":"Metric-space reference supplying the triangle inequality used to bound feasible TDoA values in Theorem 5.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GPS textbook describing RTK carrier-phase positioning, the centimeter-level ground-truth method used for mobile trajectories."}],"review_version":1}