Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

EKF-Based Radar-Inertial Odometry with Online Temporal Calibration

T0 review · 2 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims that the time offset between an IMU and a radar can be estimated online from the Doppler ego-velocity of a single radar scan, and that using that estimate reduces trajectory error substantially.

desk verdict Solid, well-validated RIO contribution with a real but fixable flaw in the time-offset Jacobian derivation. read the letter →

arxiv 2502.00661 v2 pith:P3FTTFME submitted 2025-02-02 cs.RO

Changseung Kim , Geunsik Bae , Woojae Shin , Sen Wang , Hyondong Oh This is my paper · ORCID
classification cs.RO
keywords radar-inertialodometrytemporalcalibrationtimeoffsetestimationextendedKalmanfilterDopplerego-velocityFMCW4Dradarsensorfusiononline
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 tries to establish that a radar-inertial odometry system can estimate the time offset between its inertial measurement unit (IMU) and radar online, using only the radar's ego-velocity estimated from Doppler shifts in a single scan, and that correcting for this offset substantially improves trajectory accuracy. It embeds the time offset as a filter state in an error-state extended Kalman filter (EKF) and writes the radar ego-velocity measurement model so that the predicted velocity is evaluated at the IMU time shifted by the estimated offset. On its own seven-sequence handheld dataset, the offset converges to about -0.113 seconds and the filter reduces average absolute pose error by 56% in translation and 75% in rotation compared with the same filter assuming a zero offset. The practical point is that radar sensors carry much larger delays than cameras or LiDARs, often exceeding one radar frame, so temporal calibration cannot be ignored in radar-inertial fusion.

What carries the argument

The carrying object is the radar ego-velocity measurement model, Eq. (13), which expresses the radar's velocity from a single scan in terms of the IMU attitude, velocity, gyroscope bias, and pre-calibrated extrinsics. The paper's move is to evaluate that model at $t' = t + t_d$ and to derive the Jacobian $H_{t_d}$, Eq. (16), by differentiating through the attitude and velocity states with respect to the time offset. This Jacobian is what turns a Doppler-derived velocity residual into an online correction for the IMU-radar delay; without it, the offset would stay unobservable and the update would only correct pose. The filter also models $t_d$ as a random walk with its own process noise so that a slowly varying delay can be tracked.

What would settle it

Run the filter on a synthetic trajectory with pure rotation and a known radar delay, and compute the ego-velocity residual Jacobian $H_{t_d}$ by numerical differentiation of Eq. (13) with respect to the time offset while holding the IMU states fixed. If the filter using the published Eq. (16) produces an offset estimate that disagrees with the numerical Jacobian or fails to converge under rotation-only motion, the Jacobian is incomplete and the online calibration is biased for rotation-dominated motion.

Watch

Extended reading notes

Core claim

The paper augments an error-state EKF's state with the scalar time offset $t_d$ and rewrites the standard radar ego-velocity measurement model so that the predicted ego-velocity is evaluated at the shifted IMU time $t' = t + t_d$. The residual between the Doppler-derived ego-velocity and this shifted prediction then carries information about $t_d$ whenever the platform accelerates or rotates. A chain-rule Jacobian $H_{t_d}$ maps that residual into a correction for the offset, and $t_d$ itself is propagated as a random walk. The claim is that this single-scan ego-velocity residual is sufficient for real-time temporal calibration, with no feature matching, no scan matching, and no hardware trigger, and that using the estimated offset produces more accurate odometry than the same filter with a zero offset across self-collected, ICINS, and ColoRadar sequences.

Load-bearing premise

The load-bearing premise is that the Jacobian in Eq. (16) completely and correctly captures how a shift in the IMU time changes the predicted radar ego-velocity; if the derivative of the gyroscope measurement with respect to time is missing and matters, the filter's online estimate of the time offset will be biased in exactly the rotation-dominated motions where the offset matters most.

Editorial extensions

If this is right

  • A robot can drop hardware synchronization triggers for radar-IMU fusion and still keep both measurements on a common time stream, as long as the platform moves enough to excite ego-velocity changes.
  • Any existing RIO filter that already uses radar ego-velocity in its update can adopt this temporal calibration with a small state and Jacobian addition, because the underlying measurement is the same.
  • The estimated offset in the self-collected and ColoRadar datasets, both using the same TI radar, is consistently around -0.11 seconds, suggesting the delay is dominated by the radar's signal-processing pipeline rather than the environment.
  • On the self-collected dataset the improvement is uniform: average APE translation drops 56%, APE rotation 75%, RPE translation 50%, and RPE rotation 57%, relative to the same filter with a zero offset.
  • On lower-rotation public sequences the gains are smaller but still present, with a 33% average reduction in RPE translation on ColoRadar.

Reading between the lines

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

  • A concrete check of the central mechanism would be to include the time-derivative of the gyroscope measurement in the Jacobian; rotation-only tests would then reveal whether the published chain rule omits a term that matters.
  • The near-identical offset estimates across datasets using the same radar model suggest the delay is a sensor-family property, so initializing the filter near that value could speed convergence in practice.
  • The same ego-velocity residual could drive online temporal calibration in optimization-based or scan-matching RIO, not just EKF, since the measurement model does not depend on point-cloud density.
  • The convergence behavior implies $t_d$ is only weakly observable during stationary or low-motion segments; an explicit observability analysis would map when the filter is learning the offset versus coasting on its random-walk prior.
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

2 major / 4 minor

Summary. This paper proposes an error-state extended Kalman filter for radar-inertial odometry with online estimation of the time offset between the IMU and the radar. The filter augments the state with a scalar time offset td and constructs the radar ego-velocity measurement model at the IMU time, t' = t + td, using the standard single-scan Doppler ego-velocity. The authors derive the measurement Jacobian, including a time-offset Jacobian H_td obtained by chain rule, and evaluate the method in Gazebo simulation, on a self-collected handheld dataset with motion-capture ground truth, and on the public ICINS and ColoRadar datasets. Compared with the same filter without temporal calibration, the method reports an average 56% reduction in APE translation and 75% in APE rotation on the self-collected dataset, and confirms on ICINS that artificially inserted delays are recovered to within about 15 ms. The implementation is open source.

Significance. The practical significance is high if the equations are corrected. The idea of using single-scan radar ego-velocity for online temporal calibration is novel and removes the need for hardware triggers. The validation is unusually thorough for a letter: 15 real sequences, 100 repeated trials for RANSAC averaging, artificial-delay experiments on a hardware-synchronized dataset, and a fixed-offset sweep that independently brackets the estimated offset. The open-source release makes the claims checkable. The main caveat is that the mathematical core of the calibration, the measurement Jacobian in Eqs. (15)-(16), contains sign and chain-rule issues that are load-bearing for the online update; these must be fixed before the results can be accepted as a faithful description of the implemented method.

major comments (2)
  1. [§IV-D, Eq. (16)] The time-offset Jacobian H_td is incomplete. Equation (13) contains the explicit term R_R^I (I_omega_m(t') - b_g(t'))_x I_p_R, and since t' = t + td, the total derivative of the predicted ego-velocity with respect to td includes R_R^I (d I_omega_m/dt')_x I_p_R. Equation (16) differentiates only through G_theta_I and G_v_I and omits this term. Section IV-E explicitly identifies I_omega_m as a factor affected by temporal misalignment, so this is not a harmless simplification. In rotation-dominated trajectories with nonzero angular acceleration, the omitted term contributes to the sensitivity of the radar ego-velocity to td and therefore to the Kalman gain for the time-offset update. As printed, the online update is incomplete and should be redone with the full chain rule.
  2. [§IV-D, Eq. (15)] The measurement Jacobian in Eq. (15) is inconsistent with the error-state convention stated in Eq. (3). With x = x_hat + x_tilde, the model in Eq. (13) implies H_bg = -R_R^I [I_p_R]_x, not +R_R^I [I_p_R]_x; a positive sign drives the gyro-bias update in the opposite direction. Similarly, for the attitude block H_q, the stated right-perturbation error q = q_hat (x) q_tilde leads to H_q = -R_R^I [G_R_I^T G_v_I]_x, not the printed positive expression. If an unconventional error-state convention was used, it must be defined explicitly in Eq. (3). As printed, the sign errors make the filter equations unsuitable for direct implementation and need to be corrected or justified.
minor comments (4)
  1. [§IV-D, Eq. (14)] The residual is defined as r = h(\tilde{x}) + n_r, but in an EKF the residual is z - h(\hat{x}) and the first-order approximation is r ≈ H \tilde{x} + n_r. As written, the equation conflates the residual with the noise-perturbed measurement function and should be clarified.
  2. [§IV-D, Eq. (16)] The notation \tilde{t}' and \tilde{t}_d is confusing: t' and td are scalar quantities, not error-state vectors. The chain rule should be written as a derivative with respect to td evaluated at the current estimate, not as partial derivatives with respect to tilde variables.
  3. [Table I] The paper states that the lowest error values are highlighted in red and the second-lowest in blue, but these colors will not be visible in grayscale or to color-blind readers. Adding numeric labels such as bold or asterisks would make the table readable in all formats.
  4. [§IV-C, Eq. (8)] The covariance propagation in Eq. (8) does not include coupling between the time-offset error and the attitude/velocity errors, even though a change in td shifts the IMU measurements used in the propagation interval. A brief explanation of why this coupling is neglected, or an inclusion of the cross terms, would improve the consistency of the derivation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the time offset is estimated from live radar ego-velocity residuals and checked against known artificial delays, not fitted to the reported trajectory errors.

full rationale

The claimed central derivation is not circular. The time offset t_d is included in the filter state in Eq. (2)/(4) and enters the radar ego-velocity prediction only through evaluating the model at t' = t + t_d in Eqs. (13)-(14). The Jacobian H_td in Eq. (16) is obtained by differentiating that prediction with respect to the state and t_d; it is not a restatement of the quantity being evaluated. The estimated offset is validated against controlled artificial delays in simulation (-0.15 s) and in hardware-triggered ICINS data (average estimation error ~0.015 s), so it is an estimate checked against independent ground truth rather than a parameter fitted to the reported APE/RPE values. The fixed-offset sweep in Table I is a sensitivity analysis, not a source of filter parameters; the online estimator is initialized to 0.0 s rather than to the sweep's best offset. The only possible self-citation is Ref. [23], whose S. Wang may coincide with a coauthor; it is used merely as an external LiDAR-inertial time-offset comparison value and is not load-bearing. The printed H_td omits the gyroscope's own time-derivative term from Eq. (13), and Eq. (15) appears to have a sign issue in H_bg; these are correctness concerns in the Jacobian, but they do not make the estimate circular, because the residual in Eq. (14) still measures an independent Doppler ego-velocity discrepancy rather than being defined by the construction of the Jacobian.

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

The method introduces no new physical entities, forces, or conserved quantities. It adds the time offset td as an extra filter state, which is a modeling choice rather than an invented physical entity. The central claim rests on standard EKF assumptions, a prior radar ego-velocity model, hand-chosen noise covariances, and a manual extrinsic calibration, plus the completeness of the displayed Jacobian derivation.

free parameters (4)
  • Time-offset random-walk noise variance
    Section IV-E models td as a random walk. Section V-B3 shows that increasing this noise accelerates convergence but reduces stability, and the paper does not state the exact value used for the reported results.
  • Measurement noise covariance R for radar ego-velocity update
    Used in Eq. (17). The paper states all parameters are kept identical to baseline EKF-RIO [11], so R is inherited rather than fitted here, but it is a hand-set parameter that affects all reported errors.
  • IMU process noise covariance and bias random-walk parameters
    Defined in Section IV-C through Q. These are chosen parameters, inherited from the baseline for fair comparison, and they influence the filter behavior and final accuracy.
  • Manual extrinsic calibration R_R^I and I p_R
    Section V-A states the extrinsic calibration is performed manually. Eq. (13) assumes these are correct and constant. Errors in these values would directly bias the ego-velocity prediction and the time-offset estimate.
assumptions (4)
  • domain assumption Radar ego-velocity measurement model Eq. (13) is valid, with rigid mounting and constant pre-calibrated extrinsics.
    Invoked in Section IV-D to construct the residual and Jacobians. The paper assumes the IMU and radar are rigidly connected and that the manual extrinsic calibration is accurate.
  • domain assumption IMU measurements are corrupted by slowly varying biases plus zero-mean Gaussian noise, and the time offset evolves as a random walk.
    Section IV-C and IV-E use this model for propagation and for the added td state. The random-walk model is adopted because offsets vary across sensor models.
  • domain assumption The time offset is observable from ego-velocity discrepancies during the motion in each test sequence.
    Section IV-E and Fig. 4(b) note that td is difficult to estimate when the platform is stationary or mildly moving. The reported sequences contain enough rotational and translational variation to make the offset observable.
  • standard math Error-state EKF linearization is valid, meaning first-order Taylor expansion around the current estimate is accurate.
    Section IV-B and IV-D use the ES-EKF framework. The chain rule in Eq. (16) and the Kalman update rely on this standard small-error assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EKF-Based Radar-Inertial Odometry with Online Temporal Calibration." pith.science (2026). https://pith.science/paper/P3FTTFME

@misc{pith2026250200661,
  author       = {Pith},
  title        = {Pith review of: EKF-Based Radar-Inertial Odometry with Online Temporal Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3FTTFME}},
  note         = {Machine review of arXiv:2502.00661}
}
read the original abstract

Accurate time synchronization between heterogeneous sensors is crucial for ensuring robust state estimation in multi-sensor fusion systems. Sensor delays often cause discrepancies between the actual time when the event was captured and the time of sensor measurement, leading to temporal misalignment (time offset) between sensor measurement streams. In this paper, we propose an extended Kalman filter (EKF)-based radar-inertial odometry (RIO) framework that estimates the time offset online. The radar ego-velocity measurement model, derived from a single radar scan, is formulated to incorporate the time offset into the update. By leveraging temporal calibration, the proposed RIO enables accurate propagation and measurement updates based on a common time stream. Experiments on both simulated and real-world datasets demonstrate the accurate time offset estimation of the proposed method and its impact on RIO performance, validating the importance of sensor time synchronization. Our implementation of the EKF-RIO with online temporal calibration is available at https://github.com/spearwin/EKF-RIO-TC.

Figures

Figures reproduced from arXiv: 2502.00661 by the authors.

Figure 1
Figure 1. Illustration of the temporal misalignment between IMU and radar [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The handheld platform configuration, including the radar, IMU, and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison of estimated trajectories with the ground truth. The [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Simulation results: (a) Trajectory comparison, and (b) top: difference [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Time offset estimation with 3-sigma boundaries for different initial [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Time offset estimation in ICINS 1: (a) estimation results for different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Dense Soft Weighting for Radar Ego-Velocity Estimation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Dense soft weighting of all range-Doppler cells yields training-free radar ego-velocity and covariance that cuts fused pose error 31–45% versus CFAR point-cloud baselines under a shared ESKF.

  2. Learning Point Correspondences In Radar 3D Point Clouds For Radar-Inertial Odometry

    cs.RO 2025-06 conditional novelty 6.0 of 10

    A self-supervised transformer and PointNet matcher for sparse, noisy consumer radar point clouds improves radar-inertial odometry position accuracy by 14 to 19 percent on average in real flights.

Reference graph

Works this paper leans on

23 extracted references · 21 canonical work pages · cited by 2 Pith papers

  1. [1]

    OpenVINS: A Research Platform for Visual-Inertial Estimation,

    P. Geneva, K. Eckenhoff, W. Lee, Y . Yang, and G. Huang, “OpenVINS: A Research Platform for Visual-Inertial Estimation,” in IEEE Interna- tional Conference on Robotics and Automation , 2020, pp. 4666–4672

    work page 2020

  2. [2]

    ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. M. Montiel, and J. D. Tard´os, “ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM,” IEEE Transactions on Robotics , vol. 37, no. 6, pp. 1874–1890, 2021

    work page 2021

  3. [3]

    LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping,

    T. Shan, B. Englot, D. Meyers, W. Wang, C. Ratti, and D. Rus, “LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 2020, pp. 5135–5142

    work page 2020

  4. [4]

    FAST-LIO2: Fast Direct LiDAR-Inertial Odometry,

    W. Xu, Y . Cai, D. He, J. Lin, and F. Zhang, “FAST-LIO2: Fast Direct LiDAR-Inertial Odometry,” IEEE Transactions on Robotics , vol. 38, no. 4, pp. 2053–2073, 2022

    2022

  5. [5]

    Laser–visual–inertial odometry and mapping with high robustness and low drift,

    J. Zhang and S. Singh, “Laser–visual–inertial odometry and mapping with high robustness and low drift,” Journal of Field Robotics , vol. 35, no. 8, pp. 1242–1264, 2018

    work page 2018

  6. [6]

    Degradation Resilient LiDAR- Radar-Inertial Odometry,

    M. Nissov, N. Khedekar, and K. Alexis, “Degradation Resilient LiDAR- Radar-Inertial Odometry,” inIEEE International Conference on Robotics and Automation , 2024, pp. 8587–8594

    work page 2024

  7. [7]

    A New Wave in Robotics: Survey on Recent MmWave Radar Applications in Robotics,

    K. Harlow, H. Jang, T. D. Barfoot, A. Kim, and C. Heckman, “A New Wave in Robotics: Survey on Recent MmWave Radar Applications in Robotics,” IEEE Transactions on Robotics , vol. 40, pp. 4544–4560, 2024

    work page 2024

  8. [8]

    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 IEEE Conference on Intelligent Transportation Systems , 2013, pp. 869–874

    work page 2013

Show all 23 references
  1. [9]

    4D mmWave Radar for Autonomous Driving Perception: A Comprehensive Survey,

    L. Fan, J. Wang, Y . Chang, Y . Li, Y . Wang, and D. Cao, “4D mmWave Radar for Autonomous Driving Perception: A Comprehensive Survey,” IEEE Transactions on Intelligent V ehicles, vol. 9, no. 4, pp. 4606–4620, 2024

    2024

  2. [10]

    Do we need scan-matching in radar odometry?

    V . Kubelka, E. Fritz, and M. Magnusson, “Do we need scan-matching in radar odometry?” in IEEE International Conference on Robotics and Automation, 2024, pp. 13710–13716

    2024

  3. [11]

    An EKF Based Approach to Radar Inertial Odometry,

    C. Doer and G. F. Trommer, “An EKF Based Approach to Radar Inertial Odometry,” in IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems , 2020, pp. 152–159

    2020

  4. [12]

    Radar Inertial Odometry With Online Calibration,

    C. Doer and G. F. Trommer, “Radar Inertial Odometry With Online Calibration,” in European Navigation Conference , 2020, pp. 1–10

    2020

  5. [13]

    Yaw aided Radar Inertial Odometry using Manhattan World Assumptions,

    C. Doer and G. F. Trommer, “Yaw aided Radar Inertial Odometry using Manhattan World Assumptions,” in International Conference on Integrated Navigation Systems , 2021, pp. 1–9

    2021

  6. [14]

    Tightly-Coupled EKF-Based Radar-Inertial Odometry,

    J. Michalczyk, R. Jung, and S. Weiss, “Tightly-Coupled EKF-Based Radar-Inertial Odometry,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 2022, pp. 12336–12343

    2022

  7. [15]

    Multi-State Tightly- Coupled EKF-Based Radar-Inertial Odometry With Persistent Land- marks,

    J. Michalczyk, R. Jung, C. Brommer, and S. Weiss, “Multi-State Tightly- Coupled EKF-Based Radar-Inertial Odometry With Persistent Land- marks,” in IEEE International Conference on Robotics and Automation , 2023, pp. 4011–4017

    2023

  8. [16]

    4D iRIOM: 4D Imaging Radar Inertial Odometry and Mapping,

    Y . Zhuang, B. Wang, J. Huai, and M. Li, “4D iRIOM: 4D Imaging Radar Inertial Odometry and Mapping,” IEEE Robotics and Automation Letters, vol. 8, no. 6, pp. 3246–3253, 2023

    2023

  9. [17]

    Online Temporal Calibration for Monocular Visual- Inertial Systems,

    T. Qin and S. Shen, “Online Temporal Calibration for Monocular Visual- Inertial Systems,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 2018, pp. 3662–3669

    2018

  10. [18]

    Online temporal calibration for camera–IMU systems: Theory and algorithms,

    M. Li and A. I. Mourikis, “Online temporal calibration for camera–IMU systems: Theory and algorithms,” The International Journal of Robotics Research, vol. 33, no. 7, pp. 947–964, 2014

    2014

  11. [19]

    Efficient Multi-sensor Aided Inertial Navigation with Online Calibration,

    W. Lee, Y . Yang, and G. Huang, “Efficient Multi-sensor Aided Inertial Navigation with Online Calibration,” in IEEE International Conference on Robotics and Automation , 2021, pp. 5706–5712

    2021

  12. [20]

    Quaternion kinematics for the error-state Kalman filter,

    J. Sola, “Quaternion kinematics for the error-state Kalman filter,” arXiv preprint arXiv:1711.02508, 2017

    2017 arXiv

  13. [21]

    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, vol. 41, no. 4, pp. 351–360, 2022

    2022

  14. [22]

    evo: Python package for the evaluation of odometry and SLAM,

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

    2017

  15. [23]

    Two-step LiDAR/- Camera/IMU spatial and temporal calibration based on continuous-time trajectory estimation,

    S. Li, X. Li, S. Chen, Y . Zhou, and S. Wang, “Two-step LiDAR/- Camera/IMU spatial and temporal calibration based on continuous-time trajectory estimation,” IEEE Transactions on Industrial Electronics , vol. 71, no. 3, pp. 3182–3191, 2024

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.