Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Keeping every radar Doppler cell at a soft weight cuts ego-motion pose error by roughly a third to nearly half versus CFAR point clouds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 15:05 UTC pith:RRE3ED3B

load-bearing objection Clean systems paper: dense analytic soft weights beat CFAR point-cloud RIO under a shared ESKF, with real transfer and embedded timing—limited mainly by thin ColoRadar sampling and residual sidelobe bias in U. the 4 major comments →

arxiv 2607.26980 v1 pith:RRE3ED3B submitted 2026-07-29 cs.RO eess.SP

Dense Soft Weighting for Radar Ego-Velocity Estimation

Atar Babgei , Chenyu Zhao , Michael Breza , Julie A. McCann This is my paper
classification cs.RO eess.SP
keywords FMCW radarDoppler radarego-velocity estimationradar-inertial odometrysensor fusionembedded sensingDense Soft WeightingCFAR
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Conventional single-chip radar ego-velocity pipelines first threshold the dense range-Doppler spectrum with CFAR, throwing away weak cells before any velocity fit. This paper argues that those discarded cells still carry usable Doppler geometry, especially off boresight, and that an analytic continuous weight built from beam-domain power and concentration lets them contribute without learning or per-platform retuning. The same weights feed a closed-form robust least-squares velocity and a measurement-derived covariance that plug straight into a shared inertial filter. On two public datasets and a self-collected 3D handheld rig spanning different chips and apertures, the method lowers mean absolute translational pose error by 31–45% against the strongest matched point-cloud baseline while staying real-time on embedded hardware. A sympathetic reader cares because radar is often the only reliable motion sensor in darkness, fog, or dust, and a training-free front-end that transfers across chips lowers the barrier to deploying radar-inertial odometry.

Core claim

Dense Soft Weighting shows that mapping every range-Doppler cell to a continuous confidence from peak power and peak-to-median concentration, then solving one robust weighted least-squares fit with a closed-form three-term covariance, yields more accurate ego-velocity and fused trajectories than CFAR-to-point-cloud pipelines under an identical inertial back-end, without platform-specific training.

What carries the argument

Dense Soft Weighting: the per-cell weight w_i = u_i v_i, where u_i is normalized square-root peak power and v_i is a fixed-parameter sigmoid of the log peak-to-median beam ratio; this weight multiplies every line-of-sight Doppler constraint in a single Cauchy-reweighted least-squares solve and directly supplies the residual term of the velocity covariance.

Load-bearing premise

A single fixed soft gate on peak-to-median concentration turns sub-threshold cells into net-positive velocity geometry across scenes, rather than letting sidelobe or multipath bias pull the direction vectors and spoil the fit.

What would settle it

Re-run the identical shared-filter comparison on sequences with heavy multipath or strong moving clutter: if mean translational APE no longer improves over the strongest CFAR baseline, or if lateral velocity bias grows, the claim that soft dense weights systematically help fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Weak off-boresight Doppler cells can be kept at low confidence instead of discarded, improving lateral velocity conditioning without RANSAC.
  • A training-free analytic front-end with fixed hyperparameters can transfer across single-chip radar layouts and chirp configs.
  • Closed-form measurement-derived velocity covariance lets dense radar plug into standard inertial filters rather than private pose regressors.
  • Real-time embedded latency on the full dense cube makes on-chip or near-sensor deployment of the front-end plausible.
  • Under an identical back-end the accuracy gap is attributable to the front-end representation, isolating detector-first information loss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Sidelobe-aware direction-of-arrival (Capon/MUSIC-style) on the same weighted cells could remove the residual lateral bias the paper already flags when angle-FFT peaks pull toward boresight.
  • The same soft-weight field may serve as a drop-in measurement model for tightly coupled per-cell filter updates, not only per-scan velocity.
  • Longer-horizon smoothers that consume the reported covariance could convert the per-frame gains into lower drift over multi-minute trajectories.
  • If the concentration gate remains stable at higher speeds or on aerial platforms, dense soft weighting becomes a default analytic baseline before any learned radar odometry is trained.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Dense Soft Weighting (DSW), an analytic single-chip FMCW radar front-end that assigns every range–Doppler cell a continuous confidence from beam-domain peak power and peak-to-median concentration, then estimates 3D ego-velocity by robust weighted least squares with a closed-form, measurement-derived covariance for a loosely coupled ESKF. Unlike CFAR→point-cloud pipelines, DSW retains sub-threshold cells at graded weight and avoids platform-specific learning. Under a shared inertial back-end (identical Q, temporal offset, and χ² gate), the authors report 31–45% lower mean translational APE versus the strongest CFAR baselines on ColoRadar and a self-collected AWR6843AOP set, competitive pose accuracy against learning-based methods on Radarize without training, fixed hyperparameters across chips, and real-time latency on a Jetson Orin NX.

Significance. If the controlled gains hold, this is a practically useful contribution to radar-inertial odometry: it isolates the front-end effect with a clean shared-ESKF harness (Fig. 5, §4.3), supplies analytic velocity and covariance without per-platform training, and demonstrates cross-chip transfer plus embedded feasibility. Strengths include the fixed-hyperparameter protocol, ablations showing both weight factors matter and hard gating diverges (Table 4), per-frame Δv improvements before fusion (Fig. 6, Table 2), and honest discussion of residual lateral bias under angle-FFT sidelobes (§5.7). The work is incremental relative to sparse RIO and learned dense methods, but the representation-level claim—soft dense Doppler evidence can beat detector-first pipelines under identical fusion—is well motivated and deployable.

major comments (4)
  1. [§5.7, Eqs. (7)–(8)] §5.7 and Eqs. (7)–(8): the paper itself states that under sidelobe leakage, angle-FFT peaks pull û_i toward boresight, biasing v̂_y through the geometry matrix U rather than the residuals, so soft weighting cannot remove the error. This is the main limit on the claim that dense soft evidence systematically improves velocity conditioning. Please quantify this failure mode (e.g., lateral-motion subsets, bias vs. RMSE breakdown, or frames where DSW underperforms Capon/on-chip clouds) and tighten abstract/conclusion language so “retains useful sub-threshold cues” is scoped to regimes where the FFT/parabolic LOS model remains adequate, or add a stronger DoA baseline as promised future work with at least one quantitative teaser.
  2. [Table 2, §5.4, Abstract] Table 2 / §5.4: the headline ColoRadar 45% mean APEt reduction (2.12 m → 1.17 m) rests on only three sequences. That is thin for a primary quantitative claim even with consistent per-sequence wins. Either expand the ColoRadar evaluation to a larger advertised subset with the same protocol, or report uncertainty (bootstrap/sequence-level spread) and present the percentage as indicative rather than the lead abstract number. The self-collected set (three sequences shown) has the same issue for the 31% figure.
  3. [§5.5, Abstract] §5.5 and Fig. 7(h): over the full Radarize held-out split, DSW mean translational APE is 1.11 m vs. 0.81 m for the Radarize network (and much better than milliEgo). The abstract and §5.5 framing (“closely matches or exceeds learned dense odometry”) overstates parity on the native learned benchmark. Please lead with the full-split numbers, keep sequence-level wins as secondary, and clarify that the main advantage claimed is training-free transfer and embedded cost rather than beating the in-distribution learned dense baseline on mean APE.
  4. [§3.6, Eqs. (11)–(14), Table 4] §3.6, Eqs. (11)–(14): the closed-form Σ_v is central to “measurement-derived covariance for a standard inertial back-end,” but there is no calibration check (NEES, innovation consistency, or empirical residual covariance vs. predicted Σ). Ablation Table 4 shows only a small APEt change for scalar vs. 3-component Σ. Either validate consistency of the reported covariances on held-out frames or soften claims that the three-term model is a validated uncertainty product rather than a convenient positive-definite floor for the shared ESKF.
minor comments (6)
  1. [§4.1] Table 1 / §4.1: ColoRadar text says “336.8 s and 300.34 m” for three sequences; confirm units and totals (300.34 m path length seems intended).
  2. [§3.3] Eq. (2) sign convention d_i = −û_i^T v_R is fine but should be stated once relative to radar radial-velocity polarity used in the datasets.
  3. [Fig. 2, Fig. 7] Fig. 2 and Fig. 7: ensure colorblind-safe palettes and legible axis labels in print; some trajectory panels are dense.
  4. [§4.3] §4.3 Reproducibility: listing CFAR/RANSAC settings is good; releasing code or a reference configuration file would substantially strengthen reproducibility of the shared-ESKF claim.
  5. [§2.3] Related work: UNRIO is cited as arXiv:2604.13584 with a 2026 date alongside this manuscript’s 2026 arXiv stamp—double-check citation metadata before camera-ready.
  6. [§5.4, Abstract] Minor typos: “arpg_lab_run0” spacing in captions; “1 mself-collected” missing space in §5.4; abstract “31-45%” vs body “31–45%” en-dash consistency.

Circularity Check

0 steps flagged

No significant circularity: empirical front-end evaluated against external trajectories under a fixed shared back-end.

full rationale

Dense Soft Weighting is an analytic design (per-cell power × soft peak-to-median gate, robust WLS, closed-form covariance) whose hyperparameters (τ=200, κ=0.5, c=2) are fixed a priori across datasets rather than fitted to the reported APEt metric. The Doppler measurement model (Eqs. 2–3) is the standard LOS radial-velocity constraint; the WLS solution and covariance assembly (Eqs. 8–14) follow from that model and do not encode the trajectory-error claim. Gains are measured against independent ground truth (lidar-SLAM, MoCap, T265 pseudo-GT) with an identical ESKF (same Q, t_d, χ² gate) isolating the front-end. Ablations (Table 4) and cross-chip transfer further show the result is not forced by construction. Citations to prior RIO work supply baselines and filter structure, not a self-proving uniqueness chain. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard FMCW Doppler geometry plus a small set of hand-chosen soft-gate and robust-loss constants, not on new physics. Load-bearing modeling choices are the radial-velocity cell model, the peak/median confidence proxy as a stand-in for detection quality, white residual covariance assembly, and the fairness of isolating front-ends via one ESKF. No new particles or forces; the ‘entity’ is the weighting procedure itself.

free parameters (5)
  • concentration gate center τ = 200
    Log peak-to-median ratio at which the sigmoid gate equals 1/2; fixed once for all datasets and never fit per sequence, but still a hand-chosen operating point that controls how many weak cells survive.
  • concentration gate slope κ = 0.5
    Sigmoid steepness in Eq. (6); ablation shows sensitivity (0.25/1.0 worse than 0.5) so the default is an empirical design choice.
  • Cauchy IRLS cutoff c = 2
    Single-pass robust loss cutoff on normalized residuals; fixed at 2 for both apertures.
  • per-platform covariance scalar floor / Doppler floor σ_dop = per-platform scalar (unspecified numeric)
    Section 4.3 notes a per-platform scalar floor added to covariance; Doppler-bin term is configuration-derived but still sets a tunable uncertainty floor that affects ESKF trust.
  • baseline CFAR and RANSAC thresholds = P_FA=1e-2; inlier 0.15 m/s
    P_FA=1e-2, guard/training cells, 0.15 m/s inlier threshold, 100 hypotheses—chosen for baselines and can move the relative gap versus DSW.
axioms (5)
  • domain assumption Each range-Doppler cell obeys d_i = −û_iᵀ v_R + ε_i with a single shared rigid-body sensor velocity (static-dominant scene plus outlier residual model).
    Section 3.3 stacked measurement model; standard since Kellner et al., but violated by moving reflectors and multipath except insofar as Cauchy weights suppress them.
  • ad hoc to paper Peak beam power and cross-beam median ratio are sufficient statistics for continuous cell confidence comparable to what CFAR inspects.
    Equations (4)–(6); motivated as a soft CFAR analogue but not derived from a detection-theoretic optimality proof.
  • domain assumption Velocity covariance is adequately captured by residual WLS term plus diagonal angle Jacobian term plus isotropic Doppler quantization floor (Eqs. 11–14), with approximately white residuals.
    Section 3.6; paper later admits neglected sidelobe correlation may under-estimate uncertainty (§5.7).
  • domain assumption Holding the loose-coupled ESKF, process noise Q, temporal offset state, and χ² gate fixed isolates front-end quality in trajectory metrics.
    Sections 3.7 and 4.3; standard systems-evaluation assumption, reasonable but ignores possible front-end–filter mismatch (e.g., optimistic Σ).
  • standard math Closed-form weighted least squares and elementary sigmoid/parabolic interpolation are correctly applied.
    Equations (8)–(10); routine linear algebra and interpolation.
invented entities (1)
  • Dense Soft Weighting confidence field w_i = u_i v_i no independent evidence
    purpose: Replace binary CFAR keep/discard with a continuous per-cell weight so sub-threshold Doppler cells can enter ego-velocity WLS at reduced leverage.
    Core proposed object of the paper; defined from normalized sqrt peak power and sigmoid log peak-to-median. Independent evidence is only the empirical odometry gains on the evaluated sets, not an external physical constant.

pith-pipeline@v1.2.0-daily-grok45 · 24328 in / 4143 out tokens · 77421 ms · 2026-07-30T15:05:12.029934+00:00 · methodology

0 comments
read the original abstract

Sensing ego-velocity estimation is fundamental to state estimation in visually degraded environments, where camera- and LiDAR-based pipelines can become unreliable. Millimetre-wave radar is well suited to these conditions because it provides direct Doppler velocity sensing and remains robust to poor illumination, textureless scenes, and airborne particulates. However, conventional radar ego-velocity pipelines typically apply constant false alarm rate (CFAR) thresholding to convert dense radar spectra into sparse point clouds, prematurely discarding sub-threshold returns that may still retain useful Doppler motion cues. We present Dense Soft Weighting, an analytic radar front-end that maps every range-Doppler cell to a continuous confidence metric rather than enforcing a binary detection threshold. Ego-velocity is then estimated using a deterministic robust weighted least-squares formulation, while the same weighted measurements provide a closed-form, measurement-derived velocity covariance for integration with a shared inertial back-end. The method requires no platform-specific training data or learning-based uncertainty model, supporting transfer across single-chip radar configurations. Across two public datasets and one self-collected dataset, Dense Soft Weighting reduces mean absolute pose error by 31-45% relative to the strongest CFAR point-cloud baseline under an identical inertial back-end, while running in real time on embedded hardware.

Figures

Figures reproduced from arXiv: 2607.26980 by Atar Babgei, Chenyu Zhao, Julie A. McCann, Michael Breza.

Figure 1
Figure 1. Figure 1: Motivating example for the dense radar ego-velocity [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Soft weighting versus binary thresholding. Left: (a) beam-domain power at one [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System overview of a dense radar front-end. A standard FMCW chain produces a dense range-Doppler-angle represen [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensor acquisition rig comprising a TI AWR6843AOP-EVM mmWave FMCW radar, DCA1000 data-capture card, NVIDIA Jetson Orin NX single-board computer, and Xsens MTi-320-3A IMU. Retro-reflective markers provide indoor motion-capture ground truth. (RGB-D camera not used in this evaluation.) the sparse point-cloud baselines run on identical recordings. The dataset provides two ground-truth sources: a lidar-SLAM tra… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-method comparison harness (left to right); [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-frame body-frame 𝑣𝑥 and 𝑣𝑦 over time on one ColoRadar sequence arpg_lab_run1. The proposed system matches the ground-truth velocity most closely, attaining the lowest per-frame error among the compared methods. Point-cloud baselines. We compare the dense front-end with sparse, point-cloud-based ego-velocity pipelines that follow the RANSAC family [3–5]: 3-point RANSAC rejects gross Doppler out￾liers, a… view at source ↗
Figure 7
Figure 7. Figure 7: Estimated trajectories across the three evaluation datasets: ColoRadar (top, XY view), Radarize (middle), and the [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 6 linked inside Pith

  1. [1]

    Barfoot, Ayoung Kim, and Christoffer Heckman

    Kyle Harlow, Hyesu Jang, Timothy D. Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmWave radar applications in robotics.IEEE Transactions on Robotics, 40:4544–4560, 2024

    2024

  2. [2]

    Instantaneous ego-motion estimation using Doppler radar

    Dominik Kellner, Michael Barjenbruch, Jens Klappstein, Jürgen Dickmann, and Klaus Dietmayer. Instantaneous ego-motion estimation using Doppler radar. In IEEE International Conference on Intelligent Transportation Systems (ITSC), pages 869–874, Piscataway, NJ, USA, 2013. IEEE

    2013

  3. [3]

    Christopher Doer and Gert F. Trommer. An EKF based approach to radar inertial odometry. InIEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 152–159, Piscataway, NJ, USA, 2020. IEEE

    2020

  4. [4]

    Christopher Doer and Gert F. Trommer. x-RIO: Radar inertial odometry with multiple radar sensors and yaw aiding.Gyroscopy and Navigation, 12(4):329–339, 2021

    2021

  5. [5]

    EKF- based radar-inertial odometry with online temporal calibration.IEEE Robotics and Automation Letters, 10(7):7230–7237, 2025

    Changseung Kim, Geunsik Bae, Woojae Shin, Sen Wang, and Hyondong Oh. EKF- based radar-inertial odometry with online temporal calibration.IEEE Robotics and Automation Letters, 10(7):7230–7237, 2025. Also available as arXiv:2502.00661

    Pith/arXiv arXiv 2025

  6. [6]

    Digital beamforming enhanced radar odometry

    Jingqi Jiang, Shida Xu, Kaicheng Zhang, Jiyuan Wei, Jingyang Wang, and Sen Wang. Digital beamforming enhanced radar odometry. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4601–4607, Piscataway, NJ, USA, 2025. IEEE. Dense Soft Weighting for Radar Ego-Velocity Estimation

    2025

  7. [7]

    High resolution point clouds from mmWave radar

    Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghai, Jeff Bilmes, Swarun Kumar, and Anthony Rowe. High resolution point clouds from mmWave radar. InIEEE International Conference on Robotics and Automation (ICRA), pages 4135–4142, Piscataway, NJ, USA, 2023. IEEE

    2023

  8. [8]

    RadarSFD: Single-frame diffusion with pretrained priors for radar point clouds

    Bin Zhao and Nakul Garg. RadarSFD: Single-frame diffusion with pretrained priors for radar point clouds. InIEEE International Conference on Robotics and Automation (ICRA), Piscataway, NJ, USA, 2026. IEEE. Also available as arXiv:2509.18068

    Pith/arXiv arXiv 2026

  9. [9]

    CREVE: An acceleration-based constraint approach for robust radar ego-velocity estimation

    Hoang Viet Do, Bo Sung Ko, Yong Hun Kim, and Jin Woo Song. CREVE: An acceleration-based constraint approach for robust radar ego-velocity estimation. arXiv preprint arXiv:2409.16847, 2024. arXiv-only; v2 updated April 2025

    Pith/arXiv arXiv 2024

  10. [10]

    VGC-RIO: A tightly integrated radar-inertial odometry with spatial weighted Doppler velocity and local geometric constrained RCS histograms

    Jianguang Xiang, Xiaofeng He, Zizhuo Chen, Lilian Zhang, Xincan Luo, and Jun Mao. VGC-RIO: A tightly integrated radar-inertial odometry with spatial weighted Doppler velocity and local geometric constrained RCS histograms. IEEE Robotics and Automation Letters, 10(11):11642–11649, 2025

    2025

  11. [11]

    Saputra, Peijun Zhao, Yasin Almalioglu, Pedro P

    Chris Xiaoxuan Lu, Muhamad Risqi U. Saputra, Peijun Zhao, Yasin Almalioglu, Pedro P. B. de Gusmao, Changhao Chen, Ke Sun, Niki Trigoni, and Andrew Markham. milliEgo: Single-chip mmWave radar aided egomotion estimation via deep sensor fusion. InProceedings of the 18th ACM Conference on Embedded Networked Sensor Systems (SenSys), pages 109–122, New York, NY...

    2020

  12. [12]

    BatMobility: Towards flying without seeing for autonomous drones

    Emerson Sie, Zikun Liu, and Deepak Vasisht. BatMobility: Towards flying without seeing for autonomous drones. InProceedings of the 29th Annual International Conference on Mobile Computing and Networking (MobiCom), pages 1–16, New York, NY, USA, 2023. ACM

    2023

  13. [13]

    Radarize: Enhancing radar SLAM with generalizable doppler-based odometry

    Emerson Sie, Xinyu Wu, Heyu Guo, and Deepak Vasisht. Radarize: Enhancing radar SLAM with generalizable doppler-based odometry. InProceedings of the 22nd ACM International Conference on Mobile Systems, Applications and Services (MobiSys), pages 331–344, New York, NY, USA, 2024. ACM

    2024

  14. [14]

    UNRIO: Uncertainty-aware velocity learning for radar-inertial odometry

    Jui-Te Huang, Tinashu Huang, Anthony Rowe, and Michael Kaess. UNRIO: Uncertainty-aware velocity learning for radar-inertial odometry. arXiv preprint arXiv:2604.13584, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    Cedric Le Gentil, Leonardo Brizi, Daniil Lisus, Xinyuan Qiao, Giorgio Grisetti, and Timothy D. Barfoot. DRO: Doppler-aware direct radar odometry with gyroscope. InRobotics: Science and Systems (RSS), Los Angeles, CA, USA, 2025. Robotics: Science and Systems Foundation

    2025

  16. [16]

    Christopher Doer and Gert F. Trommer. Radar inertial odometry with online calibration. InEuropean Navigation Conference (ENC), pages 1–10, Piscataway, NJ, USA, 2020. IEEE

    2020

  17. [17]

    Christopher Doer and Gert F. Trommer. Yaw aided radar inertial odometry using Manhattan World assumptions. In28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1–9, Piscataway, NJ, USA, 2021. IEEE

    2021

  18. [18]

    Multi-radar inertial odometry for 3D state estimation using mmWave imaging radar

    Jui-Te Huang, Ruoyang Xu, Akshay Hinduja, and Michael Kaess. Multi-radar inertial odometry for 3D state estimation using mmWave imaging radar. InIEEE International Conference on Robotics and Automation (ICRA), pages 12006–12012, Yokohama, Japan, 2024. IEEE

    2024

  19. [19]

    RIV-SLAM: Radar-inertial- velocity optimization based graph SLAM

    Dong Wang, Stefan May, and Andreas Nüchter. RIV-SLAM: Radar-inertial- velocity optimization based graph SLAM. InIEEE 20th International Conference on Automation Science and Engineering (CASE), pages 774–781, Piscataway, NJ, USA, 2024. IEEE

    2024

  20. [20]

    4DEgo: Ego- velocity estimation from high-resolution radar data.Frontiers in Signal Processing, 3:1198205, 2023

    Prashant Kumar Rai, Natalia Strokina, and Reza Ghabcheloo. 4DEgo: Ego- velocity estimation from high-resolution radar data.Frontiers in Signal Processing, 3:1198205, 2023

    2023

  21. [21]

    S3E: Self-supervised state estimation for radar-inertial system

    Shengpeng Wang, Yulong Xie, Qing Liao, and Wei Wang. S3E: Self-supervised state estimation for radar-inertial system. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 26686–26695, Piscataway, NJ, USA, 2025. IEEE. Also available as arXiv:2509.25984

    arXiv 2025

  22. [22]

    Impact of temporal delay on radar-inertial odometry

    Vlaho-Josip Štironja, Luka Petrović, Juraj Peršić, Ivan Marković, and Ivan Petro- vić. Impact of temporal delay on radar-inertial odometry. arXiv preprint arXiv:2503.02509, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Tightly-coupled EKF-based radar-inertial odometry

    Jan Michalczyk, Roland Jung, and Stephan Weiss. Tightly-coupled EKF-based radar-inertial odometry. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12336–12343, Piscataway, NJ, USA, 2022. IEEE

    2022

  24. [24]

    Ground-optimized 4D radar- inertial odometry via continuous velocity integration using Gaussian process

    Wooseong Yang, Hyesu Jang, and Ayoung Kim. Ground-optimized 4D radar- inertial odometry via continuous velocity integration using Gaussian process. InIEEE International Conference on Robotics and Automation (ICRA), pages 3815– 3821, Piscataway, NJ, USA, 2025. IEEE

    2025

  25. [25]

    Less is more: Physical-enhanced radar-inertial odometry

    Qiucan Huang, Yuchen Liang, Zhijian Qiao, Shaojie Shen, and Huan Yin. Less is more: Physical-enhanced radar-inertial odometry. InIEEE International Con- ference on Robotics and Automation (ICRA), pages 15966–15972, Piscataway, NJ, USA, 2024. IEEE

    2024

  26. [26]

    Quaternion kinematics for the error-state Kalman filter

    Joan Solà. Quaternion kinematics for the error-state Kalman filter. arXiv preprint arXiv:1711.02508, 2017. Standard ESKF reference

    Pith/arXiv arXiv 2017

  27. [27]

    ColoRadar: The direct 3D millimeter wave radar dataset.The International Journal of Robotics Research, 41(4):351–360, 2022

    Andrew Kramer, Kyle Harlow, Thomas Williams, and Christoffer Heckman. ColoRadar: The direct 3D millimeter wave radar dataset.The International Journal of Robotics Research, 41(4):351–360, 2022

    2022

  28. [28]

    evo: Python package for the evaluation of odometry and SLAM

    Michael Grupp. evo: Python package for the evaluation of odometry and SLAM. https://github.com/MichaelGrupp/evo, 2017

    2017