Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Uncertainty-Driven Radar-Inertial Fusion for Instantaneous 3D Ego-Velocity Estimation

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Radar network + IMU fusion beats existing ego-velocity methods

desk verdict The raw-radar CVNN is a real contribution, but the paper's own Table II undercuts the uncertainty-driven headline and the EKF derivation is not reproducible. read the letter →

arxiv 2506.14294 v1 pith:HC55YAY3 submitted 2025-06-17 cs.RO cs.AIeess.SP

Prashant Kumar Rai , Elham Kowsari , Nataliya Strokina , Reza Ghabcheloo This is my paper · ORCID
classification cs.ROcs.AIeess.SP
keywords ego-velocityestimation4Dradarcomplex-valuedneuralnetworkradar-inertialfusionextendedKalmanfilteruncertaintyquantificationColodataset
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

This paper claims that a single radar scan, read as a complex-valued data cube, contains enough information to estimate the sensor's instantaneous three-dimensional velocity, and that fusing that estimate with an inertial measurement unit lowers estimation error below existing radar-based methods. The authors build a complex-valued convolutional network, CV-RDCNet, that outputs both a velocity vector and a covariance that is meant to tell a downstream extended Kalman filter how much to trust each radar reading. On four held-out ColoRadar sequences, the fused system reports lower mean-squared and mean-absolute velocity errors than the two public baselines it is compared with, RIO and 4DEgo. The paper also reports a result its authors did not expect: feeding the network's covariance into the filter did not beat using only the mean velocity, which they explain by the Gaussian assumption in the EKF.

What carries the argument

The central mechanism is the CV-RDCNet (Complex Value - Range Doppler Channel Network), a complex-valued convolutional network that consumes the $256 \times 16 \times 192$ complex radar cube (range, Doppler, angle bins) and produces a predicted 3D ego-velocity mean $\hat{y} \in \mathbb{R}^3$ and a covariance $\Sigma \in \mathbb{R}^{3 \times 3}$ built from six parameters via a Cholesky decomposition. The feature extractor uses complex residual blocks with channel and spatial attention and a strided convolution, and the two heads are trained with a negative log-likelihood loss plus a diagonal regularizer. The other half of the mechanism is an extended Kalman filter with a 13-dimensional state composed of an orientation quaternion, velocity, and gyroscope and accelerometer biases; its measurement model is the radar velocity equation $V_W^R = V_I + R_I^W [\omega_I]_\times P_I^R$, and its measurement covariance is either fixed or set from the network's predicted $\Sigma$, which lets the filter down-weight radar updates when the network reports high uncertainty.

What would settle it

Recompute the fusion on the four ColoRadar test sequences with a 13-dimensional error-state EKF whose Jacobians match the state (measurement matrix $H$ of size $3 \times 13$) and with the exact noise parameters from the paper; if the MSE and MAE in Table II are not reproduced, or if the filter cannot run because the quaternion error representation is underspecified, the paper's central claim would not hold.

Watch

Extended reading notes

Core claim

The central claim is that processing minimally processed complex radar tensors with a complex-valued network, trained with a negative log-likelihood loss to output both a mean velocity $\hat{y} \in \mathbb{R}^3$ and a covariance $\Sigma \in \mathbb{R}^{3\times3}$, produces an instantaneous ego-velocity estimate that, when fused with IMU data in an extended Kalman filter, outperforms both instantaneous Doppler-based methods and scan-matching/heatmap-registration methods on the ColoRadar dataset. Concretely, the paper reports that CV-RDCNet + EKF reduces MSE and MAE for the $V_x$, $V_y$, $V_z$ components on the Outdoor0, Longboard2, Longboard5, and EdgarArmy5 sequences relative to RIO and 4DEgo. A secondary finding is that using the network-predicted covariance $\Sigma$ as the filter's measurement noise did not improve over using the mean velocity alone, which the authors attribute to the EKF's Gaussian noise assumption and imperfect calibration of the predicted uncertainty.

Load-bearing premise

The results stand on the extended Kalman filter being implemented exactly as the authors intended, including a mathematically consistent orientation-error representation and the unstated process noise, measurement noise, and initial covariance matrices; if that implementation is wrong, the fused numbers in Table II cannot be reproduced.

Editorial extensions

If this is right

  • If the claim holds, a robot can estimate its full 3D translational velocity from a single radar frame, without feature tracking, scan matching, or RANSAC outlier rejection.
  • The approach is designed for GNSS-denied, low-visibility, all-weather conditions where cameras and LiDAR degrade, because complex-valued radar data remain informative.
  • The network's predicted covariance provides a per-scan confidence that can, in principle, make the EKF robust to radar outliers and multipath reflections, though the paper's results show calibration is needed to realize the benefit.
  • Because propagation runs at IMU rate while radar corrections arrive at 5 FPS, the system can output ego-velocity at high frequency with bounded drift over short periods.
  • The performance gain over 4DEgo, an end-to-end registration method, suggests that instantaneous single-scan estimation can be more accurate than multi-scan registration when given complex-valued raw data.

Reading between the lines

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

  • The mean-only fusion outperforming the uncertainty-aware variant suggests the covariance head is not well calibrated; a natural extension is to calibrate it with a scalar temperature or to use it only as a gate for accepting radar updates.
  • Because the method is instantaneous, it could be used to initialize or correct sliding-window radar-inertial odometry without adding latency, a temporal composition the paper does not explore.
  • The reported comparison covers only four sequences from one dataset; rerunning the same pretrained network on more ColoRadar splits or other 4D radar datasets would reveal how much of the margin is environment-specific.
  • The EKF's Jacobian dimensions as printed (a 3×3 orientation block and a 3×12 measurement matrix for a 13-state vector) imply an unspecified quaternion error representation; an implementation with a standard three-dimensional error state would be needed to reproduce the table.
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 / 6 minor

Summary. The paper proposes an ego-velocity estimation pipeline for autonomous navigation that combines a complex-valued neural network (CV-RDCNet) processing raw ADC radar data with an Extended Kalman Filter (EKF) fusing radar velocity predictions with IMU measurements. The network outputs both a velocity mean and a covariance representing aleatoric uncertainty. The authors evaluate the method on the ColoRadar dataset, comparing the raw network, a mean-only EKF fusion variant, and an uncertainty-aware EKF variant against two public baselines (RIO and 4DEgo). The abstract claims the uncertainty-aware filter improves robustness and accuracy and achieves significantly lower error than existing methods.

Significance. The raw-ADC, complex-valued network architecture for instantaneous 3D ego-velocity estimation is a timely and potentially useful contribution, and the evaluation on a public dataset against public baselines is a strength. The paper is also commendably honest in Section V in reporting that the uncertainty-aware variant does not beat the mean-only variant. However, the headline claim of 'uncertainty-driven' fusion improving accuracy is directly contradicted by the paper's own Table II, and the EKF formulation contains a state-dimension inconsistency. If the central claim were revised to focus on the raw-radar CVNN and the mean-only EKF, the work could be of interest; as written, the uncertainty-fusion contribution is unsupported.

major comments (3)
  1. [Abstract and Section V, Table II] The abstract claims the filter 'leverages the network-predicted uncertainty to refine... improving the overall robustness and accuracy,' and the third contribution states uncertainty is used to 'mitigat[e] IMU drift and bias.' However, Table II shows that the mean-only variant CV-RDCNet + EKF (vm) has lower or equal MSE/MAE than the uncertainty-aware variant CV-RDCNet + EKF (vm, σm) in 20 of 24 component-level cells (18 strictly lower, 2 ties), while the uncertainty-aware variant is strictly better in only 4 cells. The paper's own Section V states the mean-only variant 'slightly outperforms' the uncertainty-aware variant. This is an internal contradiction with the central contribution; the reported data do not support the uncertainty-driven mechanism.
  2. [Section III-B, Eqs. (12), (16), (18)] The state vector in Eq. (12) has 13 elements (4D quaternion q_W^I, 3D velocity, 3D gyroscope bias, 3D accelerometer bias). Yet the Jacobian F in Eq. (16) is a 4x4 block matrix of 3x3 blocks, i.e., 12x12, and the measurement Jacobian H in Eq. (18) has four 3x3 blocks, i.e., 3x12. No orientation-error representation (such as a 3D error quaternion or a multiplicative error state) is introduced to reduce the quaternion to three degrees of freedom. As written, the Jacobian dimensions do not match the state vector, so the EKF prediction and update steps are undefined and the filter cannot be reproduced from the paper.
  3. [Section V and Algorithm 1] The experimental comparison reports single-run MSE and MAE on four test sequences without error bars, multiple random seeds, or statistical significance tests. The abstract's phrase 'significantly lower error' is therefore not supported by the evidence as presented. Additionally, the EKF requires process noise Q, measurement noise R, and initial covariance P0, none of which are specified in Algorithm 1 or in the text; similarly, the regularization coefficient λ1 in Eq. (9) and the stability constant ε in Eq. (8) are not given values. These omissions make the fusion results non-reproducible and prevent a reader from separating the reported differences from noise.
minor comments (6)
  1. [Throughout] The dataset name is spelled inconsistently: 'Coloradar' and 'ColoRadar' are both used; the latter is the official name.
  2. [Section IV-B] The text says 'A data sett of approximately 25,000 instances' — 'sett' should be 'set.'
  3. [Table I and Table II] Table I lists 'Edger Army' while Table II and the text use 'EdgarArmy5'; these should be unified.
  4. [Section V, bullet list] The bullet 'The predicted covariance accounts for epistemic uncertainty' is inconsistent with the paper's earlier definition of the predicted covariance as aleatoric uncertainty (Section III-A3 and [9]). Please correct this.
  5. [Eq. (16)] The symbols \bar{R}_W^I and \bar{a}_I are used in Eq. (16) but never defined in the text; please define them (presumably the estimated rotation and acceleration at the linearization point).
  6. [Figure 4] The caption and figure do not identify which curve corresponds to CV-RDCNet + EKF (vm), CV-RDCNet + EKF (vm, σm), and CV-RDCNet (vm); add a legend or explicit labels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports supervised network training and external benchmark comparisons, and the self-cited 4DEgo baseline is used only for comparison, not as load-bearing derivation.

full rationale

The paper's central derivation chain is an empirical one: a complex-valued neural network is trained on ColoRadar radar scans with ground-truth ego-velocities, and its output is fused with IMU data in an EKF. The predicted ego-velocity is evaluated on held-out sequences against external baselines (RIO, 4DEgo) and an unfused network variant. Nothing in the derivation defines the predicted velocity in terms of the fitted parameters of the evaluation, and no fitted input is renamed as a prediction. The only self-citation of note is reference [6] (4DEgo), by overlapping authors, which is used as a baseline and as the source of the twist-extraction method for ground-truth labels; neither use forces the reported outcome. The paper honestly reports that the uncertainty-aware EKF does not outperform the mean-only EKF in most component-level cells, which undermines a headline claim but is a correctness concern, not circularity. The EKF state-dimension inconsistencies and unstated noise parameters affect reproducibility but do not make the derivation equivalent to its inputs. Overall, the result is self-contained as an empirical benchmark, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result is a trained neural network plus an EKF. It depends on learned weights, unreported regularization and filter noise parameters, and the assumption that ColoRadar ground truth and complextorch operations are correct. No new physical entities are introduced.

free parameters (4)
  • CV-RDCNet weights theta = Trained on approximately 25,000 ColoRadar instances.
    The network predictions depend entirely on these learned weights; no weights, training seed, or final checkpoints are released.
  • Covariance regularization coefficient lambda_1 = Not reported in the paper.
    Equation 9 introduces lambda_1 to penalize small covariance diagonals; the chosen value affects both velocity accuracy and the predicted uncertainty used in fusion.
  • EKF noise covariances Q, R, and initial P0 = Not reported in the paper.
    The EKF fusion behavior depends on these matrices; the paper does not state their values or how the fixed-R variant was set.
  • Numerical stability constant epsilon and Cholesky covariance construction = Not reported in the paper.
    Equation 8 adds epsilon to stabilize the covariance; its value and the exact Cholesky parameterization affect the NLL loss and the uncertainty estimates.
assumptions (4)
  • domain assumption ColoRadar ground-truth poses are accurate enough to serve as velocity labels.
    Section IV-A2 derives linear velocity labels from the dataset poses at 10 FPS; any pose error propagates into training and evaluation.
  • domain assumption A single radar scan contains sufficient Doppler information to regress 3D ego-velocity.
    The learning approach assumes the raw complex radar data carries enough velocity-related signal for the network to recover 3D linear velocity.
  • domain assumption The EKF Gaussian noise model is adequate for fusing radar velocity and IMU data.
    The paper itself notes in Section V that the Gaussian assumption is violated by the network-predicted uncertainty, and this is used to explain why the uncertainty-aware fusion performed worse.
  • standard math Complex-valued convolution, batch normalization, ReLU, and attention operations as implemented in complextorch are correct.
    Section III-A relies on these operations; no formal verification or numerical checks are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Driven Radar-Inertial Fusion for Instantaneous 3D Ego-Velocity Estimation." pith.science (2026). https://pith.science/paper/HC55YAY3

@misc{pith2026250614294,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Driven Radar-Inertial Fusion for Instantaneous 3D Ego-Velocity Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HC55YAY3}},
  note         = {Machine review of arXiv:2506.14294}
}
read the original abstract

We present a method for estimating ego-velocity in autonomous navigation by integrating high-resolution imaging radar with an inertial measurement unit. The proposed approach addresses the limitations of traditional radar-based ego-motion estimation techniques by employing a neural network to process complex-valued raw radar data and estimate instantaneous linear ego-velocity along with its associated uncertainty. This uncertainty-aware velocity estimate is then integrated with inertial measurement unit data using an Extended Kalman Filter. The filter leverages the network-predicted uncertainty to refine the inertial sensor's noise and bias parameters, improving the overall robustness and accuracy of the ego-motion estimation. We evaluated the proposed method on the publicly available ColoRadar dataset. Our approach achieves significantly lower error compared to the closest publicly available method and also outperforms both instantaneous and scan matching-based techniques.

Figures

Figures reproduced from arXiv: 2506.14294 by the authors.

Figure 1
Figure 1. TI millimeter-wave (AWR2243) cascade radar signal processing pipeline illustrating the conversion of raw ADC data [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the proposed CV-RDCNet: a complex-valued convolutional network with attention mechanisms and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Radar-Inertial Sensor Fusion Framework: (a) EKF model for estimating ego-velocity using radar-derived velocity from [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of estimated ego-velocity across different [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: CV-RDCNet predicted mean and sigma vs ground truth [PITH_FULL_IMAGE:figures/full_fig_p007_5.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. Full citation record

  1. H-RINS: Hierarchical Tightly-coupled Radar-Inertial State Estimation via Smoothing and Mapping

    cs.RO 2026-03 conditional novelty 5.5 of 10

    Hierarchical dual factor graphs with continuous bias-and-covariance injection from a persistent full-state backend suppress radar-inertial drift while delivering real-time odometry.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Direct egomotion estimation,

    C. Silva and J. Santos-Victor, “Direct egomotion estimation,” in Proceed- ings of 13th International Conference on Pattern Recognition , vol. 1, 1996, pp. 702–706 vol.1

    work page 1996

  2. [2]

    High performance automotive radar signal processing on ti’s tda3x platform,

    P. Swami, A. Jain, P. Goswami, K. Chitnis, A. Dubey, and P. Chaudhari, “High performance automotive radar signal processing on ti’s tda3x platform,” in IEEE Radar Conference (RadarConf) , 2017

    work page 2017

  3. [3]

    Deep learning-based robust positioning for all-weather autonomous driving,

    Y . Almalioglu, M. Turan, N. Trigoni, and A. Markham, “Deep learning-based robust positioning for all-weather autonomous driving,” Nature Machine Intelligence , vol. 4, no. 9, pp. 749–760, Sep. 2022. [Online]. Available: https://doi.org/10.1038/s42256-022-00520-5

    work page doi:10.1038/s42256-022-00520-5 2022

  4. [4]

    Instantaneous ego-motion estimation using doppler radar,

    D. Kellner, M. Barjenbruch, J. Klappstein, J. Dickmann, and K. Di- etmayer, “Instantaneous ego-motion estimation using doppler radar,” in 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013) , 2013, pp. 869–874

    work page 2013

  5. [5]

    A lidar odometry for outdoor mobile robots using ndt based scan matching in gps-denied environments,

    B. Zhou, Z. Tang, K. Qian, F. Fang, and X. Ma, “A lidar odometry for outdoor mobile robots using ndt based scan matching in gps-denied environments,” in IEEE International Conference on CYBER Technology in Automation, Control, and Intelligent Systems (CYBER) , 2017

    work page 2017

  6. [6]

    4dego: ego-velocity esti- mation from high-resolution radar data,

    P. K. Rai, N. Strokina, and R. Ghabcheloo, “4dego: ego-velocity esti- mation from high-resolution radar data,” Frontiers in Signal Processing , vol. 3, 2023

    work page 2023

  7. [7]

    Real-time loop closure in 2d lidar slam,

    W. Hess, D. Kohler, H. Rapp, and D. Andor, “Real-time loop closure in 2d lidar slam,” in IEEE International Conference on Robotics and Automation (ICRA) , 2016

    work page 2016

  8. [8]

    The performance analysis of complex- valued neural network in radio signal recognition,

    J. Xu, C. Wu, S. Ying, and H. Li, “The performance analysis of complex- valued neural network in radio signal recognition,” IEEE Access, vol. 10, pp. 48 708–48 718, 2022

    work page 2022

Show all 41 references
  1. [9]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” ArXiv, vol. abs/1703.04977, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:71134

    2017 arXiv

  2. [10]

    Radars for autonomous driving: A review of deep learning methods and challenges,

    A. Srivastav and S. Mandal, “Radars for autonomous driving: A review of deep learning methods and challenges,” IEEE Access , vol. 11, pp. 97 147–97 168, 2023

    2023

  3. [11]

    Radar odometry for au- tonomous ground vehicles: A survey of methods and datasets,

    N. J. Abu-Alrub and N. A. Rawashdeh, “Radar odometry for au- tonomous ground vehicles: A survey of methods and datasets,” IEEE Transactions on Intelligent V ehicles, vol. 9, no. 3, pp. 4275–4291, 2024

    2024

  4. [12]

    Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,

    T. Shan, B. Englot, D. Meyers, W. Wang, C. Ratti, and R. Daniela, “Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2020

    2020

  5. [13]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 40, no. 3, pp. 611–625, 2018

    2018

  6. [14]

    Lidar-level localization with radar? the cfear approach to accurate, fast, and robust large-scale radar odometry in diverse environments,

    D. Adolfsson, M. Magnusson, A. Alhashimi, A. J. Lilienthal, and H. Andreasson, “Lidar-level localization with radar? the cfear approach to accurate, fast, and robust large-scale radar odometry in diverse environments,” IEEE Transactions on Robotics, vol. 39, no. 2, pp. 1476– 1...

    2023

  7. [15]

    Scan denoising and normal distribution transform for accurate radar odometry and positioning,

    R. Zhang, Y . Zhang, D. Fu, and K. Liu, “Scan denoising and normal distribution transform for accurate radar odometry and positioning,” IEEE Robotics and Automation Letters , vol. 8, no. 3, pp. 1199–1206, 2023

    2023

  8. [16]

    Under the radar: Learning to predict robust keypoints for odometry estimation and metric localisation in radar,

    D. Barnes and I. Posner, “Under the radar: Learning to predict robust keypoints for odometry estimation and metric localisation in radar,” in 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 9484–9490

    2020

  9. [17]

    Tbv radar slam – trust but verify loop candidates,

    D. Adolfsson, M. Karlsson, V . Kubelka, M. Magnusson, and H. An- dreasson, “Tbv radar slam – trust but verify loop candidates,” IEEE Robotics and Automation Letters , vol. 8, no. 6, pp. 3613–3620, 2023

    2023

  10. [18]

    Real-time pose graph slam based on radar,

    M. Holder, S. Hellwig, and H. Winner, “Real-time pose graph slam based on radar,” in IEEE Intelligent V ehicles Symposium (IV) , 2019

    2019

  11. [19]

    Advances in automotive radar: A framework on compu- tationally efficient high-resolution frequency estimation,

    F. Engels, P. Heidenreich, A. M. Zoubir, F. K. Jondral, and M. Win- termantel, “Advances in automotive radar: A framework on compu- tationally efficient high-resolution frequency estimation,” IEEE Signal Processing Magazine, 2017

    2017

  12. [20]

    Scan registration for autonomous mining vehicles using 3d-ndt,

    M. Magnusson, A. Lilienthal, and T. Duckett, “Scan registration for autonomous mining vehicles using 3d-ndt,” Journal of Field Robotics , 2007

    2007

  13. [21]

    4drvo-net: Deep 4d radar–visual odometry using multi-modal and multi-scale adaptive fusion,

    G. Zhuoins, S. Lu, L. Xiong, H. Zhouins, L. Zheng, and M. Zhou, “4drvo-net: Deep 4d radar–visual odometry using multi-modal and multi-scale adaptive fusion,” IEEE Transactions on Intelligent V ehicles , pp. 1–15, 2023

    2023

  14. [22]

    Self-supervised 4-d radar odometry for autonomous vehicles,

    H. Zhou, S. Lu, and G. Zhuo, “Self-supervised 4-d radar odometry for autonomous vehicles,” in 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) , 2023, pp. 764–769

    2023

  15. [23]

    Coloradar: The direct 3d millimeter wave radar dataset,

    A. Kramer, K. Harlow, C. Williams, and C. Heckman, “Coloradar: The direct 3d millimeter wave radar dataset,” The International Journal of Robotics Research, 2022

    2022

  16. [24]

    Instantaneous ego-motion estimation using multiple doppler radars,

    D. Kellner, M. Barjenbruch, J. Klappstein, J. Dickmann, and K. Di- etmayer, “Instantaneous ego-motion estimation using multiple doppler radars,” in 2014 IEEE International Conference on Robotics and Au- tomation (ICRA) , 2014, pp. 1592–1597

    2014

  17. [25]

    An ekf based approach to radar inertial odometry,

    C. Doer and G. F. Trommer, “An ekf based approach to radar inertial odometry,” in 2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI) , 2020, pp. 152– 159

    2020

  18. [26]

    Radar-inertial ego-velocity estimation for visually degraded environments,

    A. Kramer, C. Stahoviak, A. Santamaria, A.-a. Agha-mohammadi, and C. Heckman, “Radar-inertial ego-velocity estimation for visually degraded environments,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) , 05 2020, pp. 5739–5746

    2020

  19. [27]

    Less is more: Physical-enhanced radar-inertial odometry,

    Q. Huang, Y . Liang, Z. Qiao, S. Shen, and H. Yin, “Less is more: Physical-enhanced radar-inertial odometry,” 2024. [Online]. Available: https://arxiv.org/abs/2402.02200

    2024 arXiv

  20. [28]

    Dero: Dead reckoning based on radar odometry with accelerometers aided for robot localization,

    H. V . Do, Y . H. Kim, J. H. Lee, M. H. Lee, and J. W. Song, “Dero: Dead reckoning based on radar odometry with accelerometers aided for robot localization,” 2024. [Online]. Available: https://arxiv.org/abs/2403.05136

    2024 arXiv

  21. [29]

    Complex-valued channel attention and application in ego-velocity estimation with automotive radar,

    H. Cho, S. Choi, Y .-R. Cho, and J. Kim, “Complex-valued channel attention and application in ego-velocity estimation with automotive radar,” IEEE Access , vol. PP, pp. 1–1, 01 2021

    2021

  22. [30]

    Complex-valued neural networks: A comprehensive survey,

    C. Lee, H. Hasegawa, and S. Gao, “Complex-valued neural networks: A comprehensive survey,”IEEE/CAA Journal of Automatica Sinica , vol. 9, no. 8, pp. 1406–1426, 2022

    2022

  23. [31]

    Complex-valued neural networks for millimeter wave fmcw- radar angle estimations,

    K. Kaiser, J. Daugalas, J. L ´opez-Randulfe, A. Knoll, R. Weigel, and F. Lurz, “Complex-valued neural networks for millimeter wave fmcw- radar angle estimations,” in 2022 19th European Radar Conference (EuRAD), 2022, pp. 145–148

    2022

  24. [32]

    Complex- valued convolutional neural networks for enhanced radar signal de- noising and interference mitigation,

    A. Fuchs, J. Rock, M. Toth, P. Meissner, and F. Pernkopf, “Complex- valued convolutional neural networks for enhanced radar signal de- noising and interference mitigation,” in 2021 IEEE Radar Conference (RadarConf21), 2021, pp. 1–6

    2021

  25. [33]

    Complex-valued channel attention and application in ego-velocity estimation with automotive radar,

    H.-W. Cho, S. Choi, Y .-R. Cho, and J. Kim, “Complex-valued channel attention and application in ego-velocity estimation with automotive radar,” IEEE Access , vol. 9, pp. 17 717–17 727, 2021

    2021

  26. [34]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” arXiv preprint arXiv:1512.03385 , 2015

    2015 arXiv

  27. [35]

    Leveraging heteroscedastic uncertainty in learning complex spectral mapping for single-channel speech enhancement,

    K.-L. Chen, D. D. E. Wong, K. Tan, B. Xu, A. Kumar, and V . K. Ithapu, “Leveraging heteroscedastic uncertainty in learning complex spectral mapping for single-channel speech enhancement,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Proce...

    2023

  28. [36]

    Aleatory or epistemic? does it matter?

    A. D. Kiureghian and O. Ditlevsen, “Aleatory or epistemic? does it matter?” Structural Safety , vol. 31, no. 2, pp. 105–112, 2009

    2009

  29. [37]

    Estimating uncertainty with gaussian log-likelihood loss,

    A. Kumar, E. Marks, W. Mou, C. Feng, and X. Liu, “Estimating uncertainty with gaussian log-likelihood loss,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , 2019, pp. 773–782

    2019

  30. [38]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K ¨opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, PyTorch: An Imperative Style, High- ...

    2019

  31. [39]

    Adam: A method for stochastic optimization,

    D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” International Conference on Learning Representations , 12 2014

    2014

  32. [40]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks,

    M. Seitzer, A. Tavakoli, D. Antic, and G. Martius, “On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks,” 2022. [Online]. Available: https://arxiv.org/abs/2203.09168

    2022 arXiv

  33. [41]

    Drio: Robust radar-inertial odometry in dynamic environments,

    H. Chen, Y . Liu, and Y . Cheng, “Drio: Robust radar-inertial odometry in dynamic environments,” IEEE Robotics and Automation Letters , vol. 8, no. 9, pp. 5918–5925, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.