Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

UWB TDoA Error Correction using Transformers: Patching and Positional Encoding Strategies

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A transformer encoder that reads raw UWB channel responses from all anchors can correct TDoA position errors directly, reporting 0.39 m mean absolute error in a near-all-NLOS industrial environment—a 73.6 % improvement over the TDoA…

desk verdict A genuinely new TDoA error-correction method with a promising real-world result, but the headline 0.39 m is an overfit-to-the-evaluation-trajectory best-of-252 and should not be taken at face value. read the letter →

arxiv 2507.03523 v1 pith:F4TAY5SD submitted 2025-07-04 eess.SP cs.LG

Dieter Coppens , Adnan Shahid , Eli De Poorter This is my paper · ORCID
classification eess.SPcs.LG
keywords UWBTDoAtransformererrorcorrectionchannelimpulseresponsepositionalencodingNLOSindoorlocalization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the usual remedy for non-line-of-sight UWB localization—dropping 'bad' links—is both geometrically harmful and infeasible when nearly every link is NLOS. It proposes instead a transformer encoder that ingests raw channel impulse responses from all anchors in one pass and outputs a corrected 3D position, learning to exploit multipath information rather than discard it. In real-world industrial measurements, the best configuration reaches 0.39 m mean absolute error (73.6 % better than the uncorrected TDoA baseline), and 0.19 m on the subset where prior link-exclusion methods could operate. The paper also introduces anchor-coordinate 'spatial encoding' and a per-CIR patching scheme that keeps the approach scalable to larger anchor arrays.

What carries the argument

The load-bearing object is an encoder-only transformer whose self-attention processes tokens derived from the channel impulse responses. Patching turns the $N\times150$ CIR matrix into a sequence of tokens: multi-CIR patches combine samples from all anchors at the same time offset, while per-CIR patches isolate each anchor's CIR (here each full CIR is one token at $L_{\mathrm{patch}}=150$). The paper's novel positional encoding, 'spatial encoding', maps each anchor's normalized $(x,y,z)$ coordinate through logarithmically spaced sinusoids, giving the model the physical origin of each token; this is what lets time-ordered inputs (sorted by arrival time) work without zero-padding. A combined spatial+time encoding adds inter-arrival time differences but gives no clear gain. The machinery matters because it replaces the explicit geometry of hyperboloid intersection with a learned, attention-based aggregation of all received signals, and it allows one forward pass instead of one per anchor pair.

What would settle it

Re-run the full 252-configuration sweep under a nested protocol—hold out the evaluation trajectory, select configurations on the training/validation split, then evaluate on the held-out trajectory, ideally with several random seeds per configuration. If the best selected configuration no longer beats the 1.48 m TDoA baseline by the claimed margin (or its MAE rises substantially above 0.39 m), the headline improvement would not transfer.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that TDoA position error can be corrected by a transformer encoder operating directly on the raw CIRs of every anchor that hears the tag, including NLOS links. The model takes the initial TDoA estimate together with the CIR set and learns the mapping to the true position, so no anchor pair is discarded and no separate per-pair correction is needed. The authors identify per-CIR patching with $L_{\mathrm{patch}}=150$ (one token per complete anchor CIR), fixed ordering, and their novel 3D anchor-coordinate sinusoidal positional encoding as the strongest configuration, and they report that spatial encoding preserves performance under time-based ordering, which removes zero-padding and keeps complexity constant as the environment scales. They further claim that this position-correction approach outperforms prior DDoA correction with a CNN and the unsupervised DEC+k-means bad-link exclusion method on the same testbed.

Load-bearing premise

The reported accuracy assumes that picking the best of 252 hyperparameter configurations on the single evaluation trajectory does not materially overfit that trajectory; the paper reports no repeated-seed runs and no nested validation split, so the true deployed improvement could be smaller than 0.39 m even if the method works.

Editorial extensions

If this is right

  • TDoA positioning can remain available in regions where almost all links are NLOS, because no link has to be classified as unusable and removed.
  • Deploying more anchors need not blow up compute if the model uses time-based per-CIR patching: only the anchors that hear the tag are tokenized, so sequence length stays near the number of visible anchors rather than the total installed.
  • A single correction pass over all CIRs replaces per-anchor-pair correction, reducing operations below the CNN DDoA-correction baseline for most patch widths.
  • The best reported configuration cuts mean absolute error from 0.45 m to 0.19 m on the comparable subset and from 1.48 m to 0.39 m on the full difficult dataset, with CEP99 of 2.35 m versus 9.22 m for raw TDoA.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because spatial encoding ties tokens to anchor coordinates rather than sequence position, a natural extension is that the model should tolerate changes in which anchors are present or their input order; feeding randomly permuted anchor orders at inference would test whether the encoding alone carries the geometric information.
  • The same 'correct, don't discard' pattern could transfer to other hyperbolic positioning systems (for example cellular or acoustic TDoA) where NLOS-dominated geometries arise and raw channel information is available; the paper does not test those modalities.
  • The time-based variant's complexity staying constant in installed-anchor count suggests a path to very large deployments where a tag is heard by only a handful of the installed anchors; the paper's 15-anchor testbed does not itself demonstrate that scale.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a transformer-encoder architecture that corrects UWB TDoA position estimates using raw channel impulse responses (CIRs) from all anchors, rather than excluding NLOS links. It introduces two CIR ordering schemes (fixed and time-based), two patching strategies (multi-CIR and per-CIR), and three positional encodings (learnable, spatial, and spatial-plus-time-difference). The method is evaluated on real UWB measurements collected in an industrial IIoT lab, with separate training and evaluation trajectories. The authors report a best-case MAE of 0.39 m versus a 1.48 m TDoA baseline (a 73.6% improvement) and compare against prior link-exclusion and DDoA-correction approaches. A complexity analysis and Pareto-front study are also included.

Significance. If the quantitative claims are supported, this is a useful contribution to UWB localization: it is, to my knowledge, the first direct TDoA position-correction method that consumes raw CIRs from all anchors including bad links, and it is evaluated on real measurements with a separate evaluation trajectory. The paper also provides a systematic ablation of patching, ordering, and positional-encoding choices, and a complexity comparison that is helpful for practitioners. The main strengths are the real-world dataset, the external evaluation trajectory, and the breadth of the parameter study. However, the headline accuracy is currently a best-of-family statistic selected on the same evaluation trajectory, and the comparison with prior work uses a filtered common ground, so the strongest quantitative conclusions need additional support.

major comments (3)
  1. [Section VI, Table III, Figure 7a] The headline result of 0.39 m is the minimum MAE over a 252-configuration sweep evaluated on the single held-out evaluation trajectory, with no repeated-seed runs and no independent validation split used for model selection. Within the best-performing family (per-CIR patching, fixed ordering, spatial encoding), the median MAE is 0.55 m, so the reported 0.39 m is a favorable draw from a broad distribution rather than an estimate of the expected performance of a deployed model. This directly affects the abstract's 'up to 0.39 m' and '73.6%' claims. Please report mean and standard deviation over repeated training seeds, select the final configuration on a validation split (or use nested cross-validation), and report the median and an interval for the selected model.
  2. [Section VIII.A, Table VII] The comparison against [9] and [12] is performed on a filtered 'common ground' that excludes the difficult region, changing the TDoA baseline from 1.48 m on the full evaluation trajectory to 0.45 m on the filtered subset. The text states that the filter excludes 'all samples with x-values between -21 m and 14 m', which is inconsistent with the preceding statement that [9] excluded x-values between -21 m and -15 m. Since the asterisked baseline numbers come from different data subsets, the table does not currently support the conclusion's claim of a 42.4% improvement over [9]. Please report results on the identical full evaluation trajectory and separately on the filtered subset, or state explicitly that the 42.4% figure is not a like-for-like comparison.
  3. [Section VI.B.1] The variability of the spatial-encoding results is large: the same configuration family that yields the best single MAE of 0.39 m has a median of 0.55 m (and the combined encoding has a median of 0.56 m). This variability is presented only as boxplots aggregated over many configurations, with no per-configuration variance across training runs. The paper should quantify run-to-run variance for the selected configuration, otherwise the reader cannot distinguish a robust architectural advantage from optimization noise.
minor comments (5)
  1. [Section VI] The sentence 'This section will evaluate the performance of our proposed approaches.' appears twice, once immediately after the section heading and once at the start of the first paragraph.
  2. [Section V.D.2] The text refers to 'The concatenation in Equation (17)', but the spatial encoding is defined in Equation (15); the equation numbering should be corrected.
  3. [Table VII] The entry '0.47 m 0.56 m / 1.22 m' uses an unexplained slash; please clarify which value corresponds to CEP75, CEP90, and CEP95, and whether '/' indicates a missing value or a range.
  4. [Section IV] The code and dataset are described as 'being prepared for public release'; for a reproducibility-oriented journal, please provide the actual repository link or a detailed data description at revision time.
  5. [Table I and Section II.A.3] Reference [11] is categorized as a cellular DDoA-selection method with hyperbola intersections, but the cited paper is a TOA error-mitigation method; please correct the categorization or the reference.

Circularity Check

0 steps flagged · score 2.0 of 10

No structural circularity: the central accuracy claim is an empirical best-of-family result; self-citations are not load-bearing, and the 0.39 m headline suffers from evaluation-set selection bias rather than equation-level circularity.

full rationale

The derivation chain is not circular: Eq. (7) defines the TDoA baseline, Eq. (10) defines the learned correction pcorr = F(pTDoA, {CIR_n}), and the model is trained with MSE against MOCAP ground truth on one trajectory and evaluated on a separate trajectory. That is a genuine external check, and no prediction is defined in terms of its own target. The transformer, patching, and positional-encoding choices are architectural design decisions, not equations that presuppose the reported accuracy. Self-citations [8], [9], and [12] are related-work context and comparison baselines from the same group, but they are not used to prove the central claim and are published measurements from separate studies, so they do not create load-bearing circularity. The main caveat is statistical rather than circular: the abstract's 0.39 m is the minimum MAE over 252 configurations measured on the single evaluation trajectory, meaning the same data were used to select the best configuration. This makes the headline optimistic and is a selection-bias / statistical-validity concern, not a circular derivation, because the model weights are still trained on the training trajectory and the paper phrases the claim as 'up to 0.39 m.' Score 2 reflects the minor self-citation and the selection-bias burden, not a structural circularity.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central empirical claim rests on a supervised training setup rather than a parameter-free derivation. The model weights are learned from data, and the reported best result depends on hyperparameters selected via a 252-configuration sweep. No new physical entities are introduced.

free parameters (6)
  • d_model = 64 for the best per-CIR configuration
    Embedding dimension selected via the 252-configuration parameter sweep; affects accuracy and complexity.
  • L_patch = 150 for the best per-CIR configuration
    Full 150-sample CIR embedded as one token; chosen from the sweep and argued to improve hierarchical processing.
  • Positional encoding variant = Spatial encoding for the best configuration
    Learned, spatial, and spatial-plus-time variants were compared; spatial encoding gave the best full-evaluation result.
  • CIR ordering = Fixed order for the best full-evaluation result; time-based order for some Pareto configurations
    Fixed versus time-based ordering was compared; choice changes whether zero padding is needed and affects accuracy.
  • Transformer hyperparameters = 4 layers, 8 heads, FFN dimension 256, dropout 0.15, Adam lr 0.001, batch size 64
    Fixed across all sweep experiments, not tuned per configuration, but still chosen by the authors.
  • CIR window = 150 samples (50 before, 100 after the first path)
    Hand-selected based on the assumption that this captures relevant multipath in indoor industrial environments.
assumptions (6)
  • domain assumption A learnable mapping from raw CIRs and the initial TDoA position to a corrected position exists and is approximable by a transformer encoder.
    The whole method rests on the existence of the function F in Eq. 10; no theoretical guarantee is provided.
  • domain assumption The MOCAP ground truth is accurate enough to serve as a supervised training label.
    Section IV states millimeter-range uncertainty from an 18-camera Qualisys system, and labels are used to train the correction model.
  • domain assumption The training trajectory (systematic straight-line paths) is representative enough for the random evaluation trajectory.
    Section IV claims the split tests interpolation to areas between training paths, but no distribution-shift analysis is provided.
  • ad hoc to paper The 150-sample CIR window (50 samples before, 100 after the first path) captures the multipath information needed for correction.
    Section V.A selects this window based on assumed indoor ranges; it is not validated across environments.
  • domain assumption The DW1000 NLOS estimate used to characterize the environment is reliable.
    Section VIII uses the DW1000 user manual NLOS metric to argue that most positions have few or no LOS links.
  • standard math Standard supervised training with MSE loss, Adam, and early stopping produces a model that reduces position error.
    Section VI specifies the training setup but provides no formal convergence or optimality argument.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UWB TDoA Error Correction using Transformers: Patching and Positional Encoding Strategies." pith.science (2026). https://pith.science/paper/F4TAY5SD

@misc{pith2026250703523,
  author       = {Pith},
  title        = {Pith review of: UWB TDoA Error Correction using Transformers: Patching and Positional Encoding Strategies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4TAY5SD}},
  note         = {Machine review of arXiv:2507.03523}
}
read the original abstract

Despite their high accuracy, UWB-based localization systems suffer inaccuracies when deployed in industrial locations with many obstacles due to multipath effects and non-line-of-sight (NLOS) conditions. In such environments, current error mitigation approaches for time difference of arrival (TDoA) localization typically exclude NLOS links. However, this exclusion approach leads to geometric dilution of precision problems and this approach is infeasible when the majority of links are NLOS. To address these limitations, we propose a transformer-based TDoA position correction method that uses raw channel impulse responses (CIRs) from all available anchor nodes to compute position corrections. We introduce different CIR ordering, patching and positional encoding strategies for the transformer, and analyze each proposed technique's scalability and performance gains. Based on experiments on real-world UWB measurements, our approach can provide accuracies of up to 0.39 m in a complex environment consisting of (almost) only NLOS signals, which is an improvement of 73.6 % compared to the TDoA baseline.

Figures

Figures reproduced from arXiv: 2507.03523 by the authors.

Figure 1
Figure 1. Comparison of TDoA and TWR localization techniques [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the different UWB localization improvement approaches, showing different techniques and how our work [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) The IIoT lab environment and (b) a comparison of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Schematic overview of the TDoA correction method. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the two proposed patching strategies: (top) multi-CIR patching and (bottom) per-CIR patching. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Results of the parameter sweep for Multi-CIR patching. Each boxplot shows the distribution of MAE across multiple [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: MAE results from the parameter sweep for Per-CIR patching, comparing different positional encoding methods. Each [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Results of the parameter sweep for Per-CIR patching, analyzing the impact of model dimension and patch width. Each [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: The total complexity as a function of the patch width (with [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Comparison of computational complexity versus [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. OpenAlex reports about 3 citations worldwide. Full citation record

  1. Geometry-Aware Infrastructure-Anchored Denoiser for UWB Sensing and Work-Zone Reconstruction

    cs.LG 2026-07 conditional novelty 4.5 of 10

    Geometry-aware UWB denoising with latent anchor-layout estimation and deterministic distance projection improves real outdoor work-zone polygon IoU by 15.5% over PoseMLP while lowering range MSE.

Reference graph

Works this paper leans on

29 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [9]

    Error Mitigation for TDoA UWB Indoor Localization using Unsupervised Machine Learning

    P. B. Duong, B. V . Herbruggen, A. Broering, A. Shahid, and E. D. Poorter, “Error mitigation for tdoa uwb indoor localization using unsupervised machine learning,” 2024. [Online]. Available: https://arxiv.org/abs/2404.06824

    work page Pith review arXiv 2024

  2. [12]

    Anchor pair selection for error correction in time difference of arrival (tdoa) ultra wideband (uwb) positioning systems,

    B. Van Herbruggen, J. Fontaine, and E. De Poorter, “Anchor pair selection for error correction in time difference of arrival (tdoa) ultra wideband (uwb) positioning systems,” in 2021 International Conference on Indoor Positioning and Indoor Navigation (IPIN) . IEEE, 2021, pp. 1–8

    work page 2021

  3. [1]

    A survey of enabling technologies for network localization, tracking, and navigation,

    C. Laoudias, A. Moreira, S. Kim, S. Lee, L. Wirola, and C. Fischione, “A survey of enabling technologies for network localization, tracking, and navigation,” IEEE Communications Surveys & Tutorials , vol. 20, no. 4, pp. 3607–3644, 2018

    work page 2018

  4. [2]

    Precision positioning for smart logistics using ultra-wideband technology-based indoor naviga- tion: A review,

    M. Elsanhoury, P. M ¨akel¨a, J. Koljonen, P. V ¨alisuo, A. Shamsuzzoha, T. Mantere, M. Elmusrati, and H. Kuusniemi, “Precision positioning for smart logistics using ultra-wideband technology-based indoor naviga- tion: A review,” IEEE Access, vol. 10, pp. 44 413–44 445, 2022

    work page 2022

  5. [3]

    An overview of uwb standards and organizations (ieee 802.15.4, fira, apple): Interoperability aspects and future research directions,

    D. Coppens, A. Shahid, S. Lemey, B. Van Herbruggen, C. Marshall, and E. De Poorter, “An overview of uwb standards and organizations (ieee 802.15.4, fira, apple): Interoperability aspects and future research directions,” IEEE Access, vol. 10, pp. 70 219–70 241, 2022

    2022

  6. [4]

    High-accuracy ranging and localization with ultrawideband commu- nications for energy-constrained devices,

    L. Flueratoru, S. Wehrli, M. Magno, E. S. Lohan, and D. Niculescu, “High-accuracy ranging and localization with ultrawideband commu- nications for energy-constrained devices,” IEEE Internet of Things Journal, vol. 9, no. 10, pp. 7463–7480, 2022

    work page 2022

  7. [5]

    A probabilistic learning approach to uwb ranging error mitigation,

    C. Mao, K. Lin, T. Yu, and Y . Shen, “A probabilistic learning approach to uwb ranging error mitigation,” in 2018 IEEE Global Communications Conference (GLOBECOM). IEEE, 2018, pp. 1–6

    work page 2018

  8. [6]

    Edge inference for uwb ranging error correction using autoencoders,

    J. Fontaine, M. Ridolfi, B. Van Herbruggen, A. Shahid, and E. De Poorter, “Edge inference for uwb ranging error correction using autoencoders,” IEEE Access, vol. 8, pp. 139 143–139 155, 2020

    work page 2020

Show all 29 references
  1. [7]

    A variational learning approach for concurrent distance estimation and environmental identification,

    Y . Li, S. Mazuelas, and Y . Shen, “A variational learning approach for concurrent distance estimation and environmental identification,” IEEE Transactions on Wireless Communications , 2023

    2023

  2. [8]

    Re- moving the need for ground truth uwb data collection: Self-supervised ranging error correction using deep reinforcement learning,

    D. Coppens, B. van Herbruggen, A. Shahid, and E. de Poorter, “Re- moving the need for ground truth uwb data collection: Self-supervised ranging error correction using deep reinforcement learning,” IEEE Transactions on Machine Learning in Communications and Networking, vol. 2, ...

    2024

  3. [10]

    How to select the best sensors for tdoa and tdoa/aoa localization?

    Y . Zhao, Z. Li, B. Hao, P. Wan, and L. Wang, “How to select the best sensors for tdoa and tdoa/aoa localization?” China Communications, vol. 16, no. 2, pp. 134–145, 2019

    2019

  4. [11]

    A nonline-of-sight error- mitigation method for toa measurements,

    C. Ma, R. Klukas, and G. Lachapelle, “A nonline-of-sight error- mitigation method for toa measurements,” IEEE Transactions on Ve- hicular Technology, vol. 56, no. 2, pp. 641–651, 2007

    2007

  5. [13]

    Anomaly detection based on generalized gaussian distribution approach for ultra-wideband (uwb) indoor positioning system,

    F. Che, Q. Z. Ahmed, F. A. Khan, and P. I. Lazaridis, “Anomaly detection based on generalized gaussian distribution approach for ultra-wideband (uwb) indoor positioning system,” in2021 26th International Conference on Automation and Computing (ICAC) , 2021, pp. 1–5

    2021

  6. [14]

    Non-line-of-sight identification based on unsupervised machine learning in ultra wideband systems,

    J. Fan and A. S. Awan, “Non-line-of-sight identification based on unsupervised machine learning in ultra wideband systems,”IEEE Access, vol. 7, pp. 32 464–32 471, 2019

    2019

  7. [15]

    Identification of nlos and multi-path conditions in uwb localization using machine learning methods,

    C. L. Sang, B. Steinhagen, J. D. Homburg, M. Adams, M. Hesse, and U. R ¨uckert, “Identification of nlos and multi-path conditions in uwb localization using machine learning methods,” Applied Sciences , vol. 10, no. 11, 2020. [Online]. Available: https://www.mdpi.com/ 2076-3417...

    2020

  8. [16]

    Uwb nlos/los classification using deep learning method,

    C. Jiang, J. Shen, S. Chen, Y . Chen, D. Liu, and Y . Bo, “Uwb nlos/los classification using deep learning method,” IEEE Communications Let- ters, vol. 24, no. 10, pp. 2226–2230, 2020

    2020

  9. [17]

    Uwb los/nlos identification in multiple indoor environments using deep learning methods,

    Q. Liu, Z. Yin, Y . Zhao, Z. Wu, and M. Wu, “Uwb los/nlos identification in multiple indoor environments using deep learning methods,” Physical Communication, vol. 52, p. 101695, 2022

    2022

  10. [18]

    Transfer learning for uwb error correction and (n) los classification in multiple environments,

    J. Fontaine, F. Che, A. Shahid, B. Van Herbruggen, Q. Z. Ahmed, W. B. Abbas, and E. De Poorter, “Transfer learning for uwb error correction and (n) los classification in multiple environments,” IEEE Internet of Things Journal, 2023

    2023

  11. [19]

    Transformer- based nlos detection in uwb localization systems,

    S. Tomovi ´c, K. Bregar, T. Javornik, and I. Radusinovi ´c, “Transformer- based nlos detection in uwb localization systems,” in 2022 30th Telecom- munications Forum (TELFOR). IEEE, 2022, pp. 1–4

    2022

  12. [20]

    Uwb nlos identifica- tion and mitigation based on bidirectional encoder representations from transformer (bert) deep learning,

    H. Yang, Y . Wang, C. Seow, M. Sun, and D. Plets, “Uwb nlos identifica- tion and mitigation based on bidirectional encoder representations from transformer (bert) deep learning,” in 2024 14th International Conference on Indoor Positioning and Indoor Navigation (IPIN) , 2024, pp. 1–6

    2024

  13. [21]

    A machine learning approach to ranging error mitigation for uwb localization,

    H. Wymeersch, S. Maran `o, W. M. Gifford, and M. Z. Win, “A machine learning approach to ranging error mitigation for uwb localization,” IEEE transactions on communications , vol. 60, no. 6, pp. 1719–1728, 2012

    2012

  14. [22]

    A semi-supervised learning approach for ranging error mitigation based on uwb waveform,

    Y . Li, S. Mazuelas, and Y . Shen, “A semi-supervised learning approach for ranging error mitigation based on uwb waveform,” in MILCOM 2021-2021 IEEE Military Communications Conference (MILCOM) . IEEE, 2021, pp. 533–537

    2021

  15. [23]

    Self-supervised deep location and ranging error correction for uwb localization,

    B. Yang, J. Li, Z. Shao, and H. Zhang, “Self-supervised deep location and ranging error correction for uwb localization,” IEEE Sensors Jour- nal, vol. 23, no. 9, pp. 9549–9559, 2023

    2023

  16. [24]

    Deep-learning-based wi-fi indoor positioning system using continuous csi of trajectories,

    Z. Zhang, M. Lee, and S. Choi, “Deep-learning-based wi-fi indoor positioning system using continuous csi of trajectories,” Sensors, vol. 21, no. 17, p. 5776, 2021

    2021

  17. [25]

    5g positioning advancements with ai/ml,

    M. Alawieh and G. Kontes, “5g positioning advancements with ai/ml,” arXiv preprint arXiv:2401.02427 , 2023

    2023 arXiv

  18. [26]

    Available: https://idlab.ugent.be/resources/industrial-iot-lab

    [Online]. Available: https://idlab.ugent.be/resources/industrial-iot-lab

  19. [27]

    Wi-pos: A low-cost, open source uwb hardware platform with long range sub-ghz backbone,

    B. Van Herbruggen, B. Jooris, J. Rossey et al. , “Wi-pos: A low-cost, open source uwb hardware platform with long range sub-ghz backbone,” Sensors, vol. 19, no. 7, p. 1548, 2019

    2019

  20. [28]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023. [Online]. Available: https://arxiv.org/abs/1706.03762

    2023 arXiv

  21. [29]

    Ltd., DW1000 User Manual Version 2.18

    D. Ltd., DW1000 User Manual Version 2.18

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.