{"id":"c503a626-e6f6-4da2-ba9f-74311a76a1cd","arxiv_id":"2504.20339","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Direct radar odometry that uses all radar intensity information, continuous-time motion and Doppler distortion correction, and an optional Doppler-based velocity constraint outperforms point-based radar odometry on driving and off-road benchmarks.","lead":"A new radar odometry method registers full radar intensity images directly against a local map instead of extracting sparse features, and corrects for motion and Doppler distortion in a continuous-time optimization. Aided by a gyroscope, it reports state-of-the-art accuracy on public driving benchmarks and works in featureless tunnels when the radar supports Doppler velocity readings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Doppler velocity objective's row-as-instant model may introduce an environment-dependent bias that the lateral low-pass filter cannot track, exactly where the Tunnel/Skyway claim depends on it.","rationale":"The reader identified the same weakest assumption: the Doppler velocity objective relies on the row-as-instant model and an environment-dependent residual bias handled only by an online low-pass filter. My stress-test pass converges on this and sharpens it: this assumption is load-bearing specifically for the Tunnel/Skyway claims, because without Od the method fails in those environments (DRO-G errors of 7.12% and 45.76% in Table I). The paper itself flags the modeling simplification in Section V-A4 and the environment dependence in Appendix B, and the ablation excludes Skyway (Section V-C4), so no evidence is provided that the bias correction converges or is small where it matters most. This is a correctness risk, not a framing issue, and it is testable with the authors' platform and public code, which the paper provides. The concern is specific: a faster-varying or non-lateral component of the row-as-instant bias would corrupt Od and degrade the exactly the feature-deprived robustness that distinguishes DRO-GD from DRO-G. I therefore recommend CONDITIONAL: the claim is plausible and well-supported in structured environments, but the feature-deprived generalization is not yet established without a direct test of the bias dynamics. No change to the reader's verdict is needed; the paper should be accepted only conditional on addressing this concern.","tokens_in":22235,"tokens_out":1568,"duration_ms":17714,"concrete_test":"Run DRO-GD on held-out Tunnel and Skyway sequences from the authors' platform, logging the lateral-bias filter output and per-scan velocity against RTK-GNSS ground truth. Then recompute the odometry (a) with the lateral-bias correction disabled and (b) with the row-as-instant model in Eq. (10) replaced by per-range-bin azimuth/timestamp interpolation over the continuous scan. If translation error changes by more than 0.1 percentage points, or velocity RMSE by more than 0.02 m/s, the row-as-instant assumption materially drives the claimed feature-deprived performance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that DRO-GD enables odometry in feature-deprived Tunnel and Skyway scenarios rests on the Doppler-based velocity objective Od (Eq. 13), since Oψ alone fails there (Table I: DRO-G Tunnel 7.12%, Skyway 45.76%). The authors themselves identify a residual velocity bias 'partly explained by the disparity between the actual measurement process and the simple measurement model that assumes the information contained in a row of the radar data corresponds to a single azimuth/timestamp' (Section V-A4), and show the bias is environment-dependent (Appendix B). Only the lateral component is corrected online by a low-pass filter, with no stated time constant or convergence guarantee. If the bias varies faster than the filter tracks, or has a non-lateral component, Od will corrupt the velocity estimate in precisely the degenerate geometries (Tunnel, Skyway) where Oψ cannot compensate. The ablation omits Skyway (Section V-C4) and reports velocity RMSE, not bias residual, so no evidence is given that the correction converges or that the bias is small in the regimes that support the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRO, a direct SE(2) radar odometry method for spinning FMCW radars. Instead of extracting features or point clouds, DRO registers the full radar intensity image against an on-the-fly local map, using a locally continuous trajectory model that corrects both motion distortion and Doppler-induced range shifts. A second objective function (Od, Eq. 13) exploits triangular frequency modulation by GP-infilling alternating up- and down-chirp azimuths and cross-correlating the two images after a velocity-dependent Doppler shift. Orientation is obtained either from a constant angular velocity model or from gyroscope preintegration. The method is evaluated on a new automotive dataset (Suburbs, Highway, Tunnel, Skyway), on the Boreas and MulRan public benchmarks, and on an off-road dataset, with ablations and timing results. The paper reports state-of-the-art performance, including 0.26% average translation error on the Boreas leaderboard when aided by a gyroscope, and claims that the Doppler constraint enables odometry in feature-deprived tunnels and skyways.","tokens_in":22502,"tokens_out":6952,"duration_ms":72051,"significance":"If the reported results hold, this is a significant contribution to radar odometry. The direct, feature-free formulation with continuous-time motion and Doppler distortion correction is a principled departure from point-cloud-based methods, and the Doppler velocity objective (Eqs. 5-13) is derived clearly and without assuming that consecutive radar beams observe identical objects. The paper also provides substantial experimental evidence: more than 250 km of on-road data, public benchmarks, off-road validation, ablation studies, sensitivity analysis, and a real-time GPU implementation with public code. These strengths are real: the work addresses a known weakness of radar odometry in geometrically degenerate environments, and the evaluation is broader than in most related papers. However, several load-bearing points need attention before the claims can be fully accepted, in particular the consistency of the headline 0.18% number, the correctness of the printed gyroscope preintegration formula, and the validation of the online Doppler-bias correction in exactly the Tunnel/Skyway regimes where the Doppler objective carries the solution.","major_comments":[{"comment":"The abstract states that when using data with the Doppler-enabling frequency modulation, the translation error is reduced to 0.18% in similar environments. This number does not appear in Table I for any DRO-GD configuration: the DRO-GD entries are 0.19% (Suburbs), 0.24% (Highway), 0.34% (Tunnel), and 0.40% (Skyway). The only 0.18% entry in Table I is DRO-G in Suburbs, which does not use the Doppler objective Od. Please reconcile the abstract with the reported results, or state explicitly which sequences and configuration produce the 0.18% value.","section":"Abstract and Table I"},{"comment":"The gyroscope preintegration formula appears incorrect as written. For a constant angular accel? Actually linearly interpolated angular rate ω(s) between ω_i and ω_{i+1}, the exact integral from t_i to t is θ_i + ω_i (t-t_i) + (ω_{i+1}-ω_i)(t-t_i)^2/(2(t_{i+1}-t_i)). The printed expression θ(t_i) + (ω_{i+1}+ω_i)(t-t_i)/(2(t_{i+1}-t_i)) + ω_i(t-t_i) is dimensionally inconsistent: the first term has units of angular rate, not angle, and it does not evaluate to the trapezoidal integral at t=t_{i+1}. Since orientation integration is a central component of the gyro-aided results, please correct Eq. (15) and confirm that the implementation matches the corrected formula.","section":"Eq. (15), Section IV-D2"},{"comment":"The claim that DRO-GD enables odometry in feature-deprived Tunnel and Skyway scenarios rests on the Doppler objective Od, because DRO-G alone fails there (Table I: 7.12% and 45.76%). The paper acknowledges a residual Doppler velocity bias 'partly explained by the disparity between the actual measurement process and the simple measurement model that assumes the information contained in a row of the radar data corresponds to a single azimuth/timestamp' and shows in Appendix B that the bias is environment-dependent. Only the lateral component is corrected, by a low-pass filter with no stated time constant or convergence guarantee, and the ablation study in Table II excludes Skyway and reports velocity RMSE rather than bias residual. Please provide evidence that the bias correction converges and tracks the environment-dependent bias in Tunnel/Skyway (e.g., bias residual against ground-truth velocity, sensitivity to the filter time constant, or a bound on the resulting error in Eq. (13)), or discuss why the remaining bias cannot corrupt the registration that supports the headline claim.","section":"Section V-A4 and Appendix B"},{"comment":"The Doppler objective assumes that, after GP infilling, the up-chirp and down-chirp images observe exactly the same geometry. This is a modeling simplification: the infilled rows are constructed from measurements taken at different azimuths and timestamps, and the mismatch will grow with radar rotation rate, range, and vehicle speed. The paper acknowledges a related limitation in Section V-A4 but does not analyze how the interpolation error affects Od or the estimated velocity. A simulation or synthetic-shift experiment varying the infill neighborhood (U x V), rotation rate, and velocity would make the validity domain of this assumption explicit.","section":"Section IV-C"},{"comment":"The MulRan evaluation applies post-hoc corrections: a 50 ms IMU timestamp offset removed from all sequences and a 0.172 deg radar heading adjustment. The text reports that without the heading correction DRO-G gives 1.46% and 0.38 deg/100m, while with the correction the table shows 1.45% and 0.38 deg/100m. It is not stated whether these corrections were selected on the test sequences themselves or on held-out calibration data. Please state the selection protocol explicitly; if the corrections were tuned on the evaluation sequences, the comparison to CFEAR is not fully apples-to-apples and the interpretation of the MulRan results should be softened accordingly.","section":"Section V-D2, MulRan"}],"minor_comments":[{"comment":"The caption says the best radar-based method is in bold, but no bold entries appear in the table as rendered; please add the emphasis or remove the statement.","section":"Table I caption"},{"comment":"The ablation study states that Skyway sequences were left out due to their extreme challenge, yet Fig. 8 (sensitivity to gamma) includes a Skyway curve. Please clarify whether Skyway was included in the sensitivity analysis or adjust the text.","section":"Section V-C4 and Fig. 8"},{"comment":"The robust weighting formula rho = (|psi - Gamma| - 1)^6 gives weight 1 for exact agreement, zero near absolute difference 1, and large weights again for differences greater than 1 (which can occur if intensities are not strictly normalized). Please clarify the intended behavior and, if intensities are normalized to [0,1], state this explicitly or correct the formula.","section":"Eq. (18)"},{"comment":"The 'No local map' ablation shows a lower velocity RMSE (0.107 m/s) than the full DRO-GD (0.119 m/s), which is counterintuitive; a one-sentence explanation would help the reader interpret the ablation.","section":"Table II"},{"comment":"The low-pass filter used for lateral velocity bias estimation is described qualitatively; please report its time constant or update rule and the velocity threshold value used in the release.","section":"Section V-A4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong candidate if the authors fix the abstract/table inconsistency, correct Eq. (15), and provide the requested validation of the Doppler-bias correction in the Tunnel/Skyway regime. The Eq. (15) issue is the kind of error that can be corrected locally, but it is load-bearing for the gyro-aided results, so I would not accept before seeing the correction. The abstract's 0.18% claim should be checked carefully by the editor, as it currently overstates what Table I reports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth your time. This is the first direct radar odometry method that treats Doppler and motion distortion in a principled continuous-time way, and the broad claims hold up. On Boreas the gyro-aided version roughly halves the translation error of the best prior radar odometry (0.26% vs 0.51%), and the Doppler constraint genuinely enables odometry in tunnels and skyways where geometry-only registration fails. The GP-based infill that decouples up- and down-chirps without assuming consecutive-beam correspondence is a real idea, not a repackaging of prior work.\n\nWhat's done well: the derivations are clean (Eqs. 5–13), the objective functions are well-motivated cross-correlations, and the experiments are extensive: 250+ km across Boreas, MulRan, and their own automotive/off-road sets, with public code and real-time performance. The ablation is fair, and the paper is honest about its limitations, including the gyro-bias dependence and the environment-dependent Doppler velocity bias.\n\nThe soft spots are real but manageable. The abstract's 0.18% figure doesn't appear in Table I; either it needs a pointer to the right table or a correction. MulRan required post-hoc corrections (50 ms IMU timestamp offset, 0.172° heading adjustment); they're disclosed, but it weakens the comparison. More importantly, the ablation omits Skyway, the hardest environment, and reports velocity RMSE rather than bias residual. Since the Tunnel/Skyway claim leans entirely on the Doppler objective, and the authors themselves document that the row-as-instant model creates an environment-dependent velocity bias that only a low-pass filter tracks laterally, we never see direct evidence that the bias is small or that the filter converges in the regimes that matter. I don't think this sinks the paper—the method works on the data—but it deserves a closer look from the referees.\n\nThe unreleased automotive and off-road datasets are a nuisance but not a dealbreaker, given the public benchmarks.\n\nBottom line: this is a solid, genuinely new contribution that deserves a serious referee. I'd support sending it to peer review with requests for clarification on the 0.18% figure and a proper analysis of the Doppler bias correction in degenerate environments.","headline":"Genuine advance in direct radar odometry with a clean Doppler-aware formulation; the main claims hold, but the headline tunnel/skyway result rests on a Doppler bias correction that needs more evidence.","tokens_in":23032,"tokens_out":2490,"would_cite":true,"duration_ms":24203,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Radar odometry can run directly on the full intensity image, with Doppler and motion distortion corrected inside the optimization, and this beats point-based radar odometry.","keywords":["radar odometry","direct registration","Doppler distortion","motion distortion","continuous-time trajectory","FMCW radar","Gaussian process infill","gyroscope preintegration"],"falsifier":"Drive a car with centimetre-level ground truth suddenly from an open road into a long featureless tunnel using Doppler-only estimation. If the estimated sideways velocity jumps or drifts faster than the paper's low-pass filter can follow it, and the odometry error grows sharply, the single-row/single-instant radar model is the weak link.","tokens_in":22049,"feed_emoji":"📡","tokens_out":8076,"duration_ms":81710,"temperature":0.7,"pith_summary":"This paper proposes a new way to estimate a robot's motion from a spinning millimetre-wave radar. Instead of extracting sparse points or features, the method registers the full radar intensity image directly against a local map, while correcting Doppler and motion distortion inside a continuous-time trajectory estimate. With a triangular frequency modulation that makes radial Doppler velocities observable, an additional Doppler objective lets the system estimate velocity even in geometrically featureless tunnels. The reported evidence is over 250 km of on-road data plus off-road runs, with an average relative translation error of 0.26% on the Boreas leaderboard when a gyroscope is used, falling to 0.18% when the Doppler constraint is active. If correct, the direct use of all radar information is not only simpler than point-based radar odometry but more accurate in exactly the geometries where points run out.","feed_headline":"Doppler-aware direct radar odometry beats point-based baselines","feed_subtitle":"All radar intensity, no extracted points: DRO hits 0.26% error on Boreas and runs in featureless tunnels.","key_machinery":"The load-bearing object is a pair of objective functions optimized by gradient ascent. The first, $\\mathcal{O}_\\psi$, is the summed product of measured intensity $\\psi_{nm}$ and local-map intensity queried at the motion- and Doppler-corrected position of each range cell. The second, $\\mathcal{O}_d$, applies the range shift $\\Delta r_n = \\beta\\,[\\cos\\alpha_n,\\ \\sin\\alpha_n]^\\top R^\\top v$ to the Gaussian-process-infilled up-chirp image and correlates it with the down-chirp image, so that maximizing it recovers the body-centric velocity. The continuous trajectory connects the two objectives: orientation comes from gyroscope preintegration (or a constant angular rate), and translation comes from a constant body-centric velocity model integrated over the scan. The Gaussian process infill for interleaved chirp directions is what makes the Doppler comparison valid without assuming consecutive beams see the same object.","core_discovery":"The central claim is that radar odometry can be formulated as a continuous, gradient-based optimization over the full radar intensity image, with both motion distortion and Doppler-induced range shifts corrected inside the cost rather than in a pre-processing step. The paper introduces two cross-correlation objectives: direct scan-to-local-map intensity registration, and, for triangular-chirp radars, a Doppler velocity constraint that compares Gaussian-process-infilled up-chirp and down-chirp images shifted by the range shift $\\Delta r_n = \\beta\\,[\\cos\\alpha_n,\\ \\sin\\alpha_n]^\\top R^\\top v$. This makes DRO, in the authors' words, the first direct method that accounts in a principled way for both distortions, and it is what lets the system estimate velocity purely from Doppler information in environments where geometry gives almost no constraint, such as a featureless tunnel. The supporting evidence is a set of benchmarks over more than 250 km of driving plus off-road runs, with headline numbers of 0.26% relative translation error on Boreas and 0.18% when the Doppler constraint is available.","pith_inferences":["Editorial inference: the Gaussian-process infill trick that makes interleaved up-chirp and down-chirp rows comparable is a general mechanism; any sensor with two interleaved measurement patterns could use it to recover a velocity-dependent shift without explicit data association.","Editorial inference: the observation that radar-based orientation constraints degrade rotation estimates relative to pure gyroscope integration suggests that, for low-resolution spinning radars, orientation is best left to the inertial sensor while radar constrains translation and velocity only.","Editorial inference: a testable extension is applying the direct Doppler objective to 3D FMCW radars with dense beams; if the single-azimuth-per-row assumption is the main source of the residual velocity bias, per-beam timing in 3D data should reduce that bias.","Editorial inference: the environment-dependent lateral velocity bias, only partially corrected by a low-pass filter, is the most likely bottleneck; replacing it with a model-based or consistency-based correction could push the reported error below the 0.18% level."],"forward_implications":["On commercial sawtooth-pattern radars, direct registration alone reaches 0.26% average relative translation error on the Boreas odometry leaderboard, ahead of point-extraction methods.","When the radar emits triangular chirps, the Doppler objective keeps relative translation error at about 0.34% in a featureless tunnel where continuous-time point-based radar registration fails at roughly 35%.","Orientation from gyroscope integration stays accurate even where radar returns are weak, and the paper reports that adding radar constraints to inertial methods can actually worsen rotation estimates.","The full Doppler-aided pipeline runs in about 89 ms per 250 ms radar frame on a laptop GPU, so the extra processing does not prevent real-time use.","Because the method uses all intensity information, it avoids feature extraction, data association, and hand-tuned discretization of the search space."],"supporting_citations":[{"why":"Supplies the per-azimuth Doppler velocity extraction baseline (DG) and documents the lateral velocity bias that DRO addresses with its online correction.","marker":"[49]"},{"why":"Provides the continuous-time ICP radar teach-and-repeat baseline CT-R that DRO is compared against in automotive and off-road tests.","marker":"[15]"},{"why":"Defines the Boreas dataset and its odometry leaderboard, the primary benchmark where DRO-G reports 0.26% translation error.","marker":"[16]"},{"why":"CFEAR is the point-based radar odometry baseline that DRO outperforms on Boreas and MulRan.","marker":"[3]"},{"why":"STEAM-RIO++ is the radar-inertial point-based baseline on the Boreas leaderboard that DRO-G exceeds in translation and rotation accuracy.","marker":"[17]"},{"why":"Supplies the Gaussian process regression machinery used to infill missing up- and down-chirp azimuth rows.","marker":"[55]"},{"why":"Provides the MulRan dataset, which benchmarks the method with a lower-quality IMU and shows the orientation estimate remains competitive.","marker":"[40]"}],"fun_headline_variants":["Radar odometry without feature extraction: 0.26% error","Doppler-aided radar odometry navigates featureless tunnels","Direct intensity radar odometry with motion and Doppler correction","DRO: radar odometry from raw intensity, no points needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each row of radar data corresponds to one beam direction at one instant, and that the interpolated up-chirp and down-chirp views show the same scene; because a real spinning radar sweeps continuously, this mismatch shows up as a velocity bias that the online correction may not track quickly enough.","fun_headline_variants_meta":{"raw":{"variants":["Radar odometry without feature extraction: 0.26% error","Doppler-aided radar odometry navigates featureless tunnels","Direct intensity radar odometry with motion and Doppler correction","DRO: radar odometry from raw intensity, no points needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000541,"raw_usage":{"total_tokens":2656,"prompt_tokens":1072,"completion_tokens":1584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":1512}},"tokens_in":688,"tokens_out":1584,"duration_ms":12194,"temperature":1.0,"reasoning_tokens":1512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:31:27.980363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Drive a car with centimetre-level ground truth suddenly from an open road into a long featureless tunnel using Doppler-only estimation. If the estimated sideways velocity jumps or drifts faster than the paper's low-pass filter can follow it, and the odometry error grows sharply, the single-row/single-instant radar model is the weak link.","supporting_citations":[{"cited_title":"Are We Ready for Radar to Replace Lidar in All-weather Mapping and Localization? IEEE Robotics and Automation Letters (RA-L), 7(4):10328–10335, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-time ICP radar teach-and-repeat baseline CT-R that DRO is compared against in automotive and off-road tests."},{"cited_title":"Schoellig, and Timothy D","cited_arxiv_id":null,"evidence_quote":"STEAM-RIO++ is the radar-inertial point-based baseline on the Boreas leaderboard that DRO-G exceeds in translation and rotation accuracy."},{"cited_title":"Gaussian Processes for Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process regression machinery used to infill missing up- and down-chirp azimuth rows."}],"review_version":1}