{"id":"6317076d-b355-4e78-ba36-4dd8024fa798","arxiv_id":"2507.04311","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Modeling the residual distortion of each LiDAR point as a per-point covariance and using it in matching and residual weighting reduces odometry error under intense vibration.","lead":"This paper adds a per-point uncertainty estimate to LiDAR data after vibration correction, and uses it to make LiDAR-inertial odometry more accurate on rough terrain. It matters for field robots that move fast over bumps, where scan distortion usually degrades localization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed gain may depend on the unvalidated scalar gamma: Eq. (17) sets sigma = gamma * dt * k, and the paper only asserts gamma-insensitivity without showing a sweep or calibrating sigma against measured undistortion error.","rationale":"The reader's weakest_assumption is exactly the hand-chosen scalar gamma and the absence of a sensitivity sweep. I agree that this is the most load-bearing concern: the entire contribution reduces to using the post-undistortion covariance, and if its scale is not robust, the claimed benefit is not established. I also extend the concern to the functional form (linear vs sqrt(delta_t)) and to the circularity in estimating k_v from the filter state, but these reinforce rather than replace the core issue. The reader's CONDITIONAL verdict is appropriate: the method is coherent and the experiments show consistent improvements, but reproducibility and sensitivity are not documented. My proposed tests would either validate the gamma robustness or show that the gain is fragile. I found no grounds to reject the paper outright, and no need to move the verdict.","tokens_in":12429,"tokens_out":4857,"duration_ms":62044,"concrete_test":"On the released vibration-platform and 01-circle sequences, rerun the full pipeline with gamma in {0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1.0} while holding all other settings fixed, and report mean/RMSE APE for each value. If the improvement over Fast-LIO persists across a 10x range (e.g., 0.02 to 0.2) and does not invert at extreme values, the gamma concern is mitigated. Additionally, on the vibration platform, which provides ground-truth poses, compute the empirical per-point undistortion error as a function of delta_t and compare it to the predicted sigma = gamma * delta_t * k_omega; fitting sigma = a * delta_t^b will reveal whether the linear-in-delta_t model or a sqrt-delta_t random-walk model is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that uncertainty-aware matching and residual weighting improve LIO under intense vibration. The entire uncertainty model is scaled by one hand-set hyperparameter gamma in Eq. (17): sigma_r = gamma * dt * k_omega and sigma_T = gamma * dt * k_v, with gamma fixed at 0.1. The paper states 'We observe that the proposed method is not sensitive to gamma' but provides no sensitivity sweep, no calibration against ground-truth undistortion error, and no analysis of how gamma should vary with IMU rate, LiDAR model, or terrain. If gamma is too small, the uncertainty is negligible and the method degenerates to Fast-LIO; if too large, the Mahalanobis matching and residual weighting over-trust far-away points, potentially degrading accuracy. A second, compounding issue is that the linear vibration intensity k_v uses the filter's own estimated velocity (Eq. 15), so the uncertainty feedback depends on the state being estimated; a drift in the state changes k_v, which in turn changes the weights, and this loop is not analyzed. Finally, the model assumes uncertainty grows linearly with scan time offset delta_t, but if the dominant error is random-walk IMU noise, the standard deviation should grow as sqrt(delta_t), not delta_t. A gamma sweep alone cannot fix a misspecified functional form. Without calibration evidence or a robustness study, the reported improvements on the vibration platform and public sequences could be a result of favorable gamma tuning rather than the uncertainty model itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a vibration-aware lidar-inertial odometry (LIO) system that assigns a point-wise \"post-undistortion uncertainty\" to each LiDAR point after motion undistortion. The uncertainty is modeled from angular and linear vibration intensities, computed as the mean absolute deviation of gyroscope measurements and of the filter-estimated velocity, scaled by a hand-chosen hyperparameter gamma and the time offset within the scan. This uncertainty is propagated through the point-to-plane observation model and used both to guide scan-to-map matching via Mahalanobis distances and to weight residuals in an iterated Kalman filter. The method is evaluated on a 3DoF vibration platform, several public datasets (NCD, M2DGR, Botanic Garden), and four self-recorded uneven-terrain sequences, reporting improved mean APE over Fast-LIO, Point-LIO, and other baselines in most tested scenarios.","tokens_in":12793,"tokens_out":5967,"duration_ms":64231,"significance":"If the uncertainty model is valid, the proposed method is a practical enhancement to Fast-LIO-style estimators in high-vibration conditions, with the advantage of real-time performance and a released dataset. The core idea of modeling post-undistortion uncertainty rather than seeking perfectly accurate undistortion is novel and well motivated, and the error-propagation chain from rotation/translation perturbations to per-point covariance (Eqs. 5-14) is clearly presented under the stated Gaussian and small-angle assumptions. The experiments cover a useful spread of platforms and environments, and the ablation study demonstrates the contribution of the uncertainty-aware components. However, the link between the heuristic vibration-intensity definition and the resulting covariance scale is not derived or externally validated, which currently limits the strength of the empirical claims.","major_comments":[{"comment":"The entire uncertainty model is scaled by a single hand-chosen hyperparameter gamma (set to 0.1), and the paper only asserts that \"the proposed method is not sensitive to gamma\" without showing a sensitivity sweep or calibrating sigma against measured undistortion errors. Since the claimed benefit of uncertainty-aware matching and residual weighting depends directly on the absolute scale of sigma, please provide experiments varying gamma (e.g., 0.01, 0.02, 0.05, 0.1, 0.2, 0.5) on at least the vibration platform and the 01-circle sequence, or alternatively calibrate gamma from ground-truth undistortion error in simulation.","section":"Sec. III.B.3, Eq. (17)"},{"comment":"The linear vibration intensity k_v is computed from the filter's own estimated velocity GvI, creating a feedback path between the state being estimated and the uncertainty used to estimate that state. The paper does not analyze this loop; a drift or bias in the velocity estimate would change k_v, hence change the point weights, and this interaction could either amplify or mask errors. Please provide either an analysis of the loop's stability or an experiment that uses a vibration-intensity estimate independent of the state estimate (e.g., from raw IMU integration alone) to break the feedback.","section":"Sec. III.B.3, Eqs. (15)-(16)"},{"comment":"The assumed linear growth of sigma with scan time offset delta_t is not derived from the error-propagation model of Sec. III.B.2. If the dominant source of undistortion error is integrated random-walk IMU noise, the standard deviation of the position error should grow as sqrt(delta_t), not delta_t; a scalar gamma cannot fix this misspecification because it rescales all points uniformly and cannot change the relative weighting between early and late points in a scan. Please justify the linear functional form from the IMU noise model or provide empirical evidence from simulation that the undistortion error magnitude scales linearly with the time offset.","section":"Sec. III.B.3, Eq. (17)"},{"comment":"The abstract and conclusion claim that the method \"achieves better performance than other methods when LiDAR undergoes intense vibration,\" but on the Botanic Garden 1008-13 sequence the proposed method (mean/RMSE APE 0.064/0.075 m) is worse than Fast-LIO (0.054/0.059 m). Please qualify the claim to reflect that the improvement holds on most, but not all, tested high-vibration sequences, and discuss possible reasons for this counterexample (e.g., sensor characteristics, vibration frequency content, or the effect of a sparse 16-beam scanner on Mahalanobis matching).","section":"Abstract and Table III"}],"minor_comments":[{"comment":"There is a typo \"Equatoin (1)\" and the phrase \"from Equation (1)\" is ambiguous because Eq. (1) defines the state vector, not the velocity estimates; please specify that GvI(i) are obtained from forward propagation of the current state estimate.","section":"Sec. III.B.3"},{"comment":"In Table II, \"Descirption\" is a typo, and the standard deviations of the proposed method are larger than those of Fast-LIO in most rows (e.g., translation error for 1 Hz z-axis: 1.72 +/- 0.81 vs. 2.19 +/- 0.28); please comment on whether the reported improvements are statistically significant given the overlap.","section":"Table II"},{"comment":"The sequence names in Table III are \"1018-00\" and \"1018-13,\" while the text refers to \"1008-00\" and \"1008-13\"; please unify the notation.","section":"Table III and Sec. IV.B"},{"comment":"The ablation setting \"w/o uncertainty\" with GM enabled is unclear: if the post-undistortion uncertainty is not modeled, what covariance is used in the Mahalanobis-distance matching of Sec. III.C.2? Please clarify how the baseline with only guided matching is implemented.","section":"Sec. IV.E, Table V"},{"comment":"The axis labels in Fig. 9 appear duplicated (\"Y[m]\" and \"X[m]\" on both subplots); please adjust the layout so each axis is labeled once.","section":"Fig. 9"},{"comment":"The sentence \"uncertainty-guided matching (needing 2-3 iterations per scan and taking a total of 22 ms on average)\" is ambiguous; please state whether 22 ms is the total matching time per scan, per iteration, or averaged over all iterations.","section":"Sec. IV.D"},{"comment":"The dataset link is a Google Drive folder; a permanent DOI or a repository identifier would be more stable and citable for a journal publication.","section":"Footnote 1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising and practical idea, and the experimental evidence is mostly supportive, but the three major comments about the unvalidated gamma scale, the feedback path through k_v, and the un-justified linear dependence on delta_t are load-bearing for the core contribution: the point-wise uncertainty model. The missing gamma sensitivity sweep is particularly conspicuous because the paper explicitly claims insensitivity without providing data. The public-dataset counterexample (Botanic Garden 1008-13) also undermines the unqualified abstract claim. These issues are addressable within the manuscript's scope, so I recommend major revision rather than rejection, but the revision must include either new experiments/analysis or a substantive re-derivation of the uncertainty scaling."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea here is genuinely new: instead of trying to make undistortion more accurate under intense vibration, the authors estimate how wrong each point might be after undistortion and feed that uncertainty into matching and Kalman residual weighting. That is a sensible reframing, and the combination of vibration-derived per-point covariance (rotational, translational, plus measurement noise) with an iterated Kalman filter is, to my knowledge, not in the cited literature. The paper does well at the mechanics: the error-propagation derivation (Eq. 5-14) is coherent under the stated small-angle and Gaussian assumptions, and the experiments are consistent and reasonably broad — vibration platform, public datasets, and their own rough-terrain recordings with cm-level RTK truth. The ablations actually separate the contributions of the uncertainty model and the guided matching, which is better than most papers at this level. The one public sequence where they place second (Botanic Garden 1008-13) is reported honestly.\n\nThe soft spots are real but not fatal. The biggest is gamma: Eq. 17 sets sigma = gamma * dt * k, with gamma fixed at 0.1 and only a one-line assertion that the method is insensitive to it. No sensitivity sweep, no calibration against actual undistortion error. The stress-test concern about linear vs sqrt(dt) scaling is worth taking seriously — if the dominant error were random-walk IMU noise, sigma should grow as sqrt(dt), and a fixed gamma cannot fix a misspecified functional form. That said, the experiments still show improvements on multiple independent datasets, which would be a strange accident if the whole effect came from a lucky gamma. I also note the vibration intensity k_v uses the filter's own velocity estimate, creating a feedback path that is not analyzed; in practice this is likely minor because odometry drift is small over a 0.1 s scan. The published results are good enough that I would not block on these concerns, but a revision should show a gamma sweep and ideally compare the dt-linear model against a dt-sqrt alternative.\n\nIs it a new result? Yes — the point-wise post-undistortion covariance propagated from vibration intensity and used to re-rank matches and weight residuals is a combination I have not seen. It is not a paradigm shift, but it is a practical, reusable technique. Who is this for? Anyone doing LIO on ground robots or handheld devices with modest IMUs; the code would be the real deliverable, and the paper does not ship it (only the dataset link is mentioned). I would be comfortable sending this to a serious referee; I would probably ask for the gamma robustness data and a note on the dt dependence, and I would note that the public dataset gains are small but consistent.\n\nRecommendation: accept for peer review, with revision expected.","headline":"A useful, engineering-solid LIO paper that assigns post-undistortion uncertainty from vibration intensity and leverages it in matching and Kalman weighting; the main weakness is a hand-set gamma that is only asserted to be insensitive, not demonstrated.","tokens_in":13259,"tokens_out":688,"would_cite":true,"duration_ms":9422,"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":"This paper introduces point-wise post-undistortion uncertainty, a per-point covariance that guides matching and Kalman updates, and shows it improves odometry on vibrating robots.","keywords":["LiDAR-inertial odometry","vibration","point cloud undistortion","uncertainty estimation","iterated Kalman filter","scan-to-map matching","ground robots","rough terrain"],"falsifier":"Run a sensitivity sweep over gamma across different IMU sampling rates (for example, 100 Hz versus 200 Hz) and different vibration amplitudes: if the optimal gamma changes materially across conditions, or if a single gamma fails to improve or degrades odometry relative to Fast-LIO on some intense-vibration sequence, the core claim of a robust uncertainty scale is undermined. Concretely, running the method with gamma from 0.01 to 1.0 on the 04 all-challenges sequence and checking whether the mean absolute pose error stays below Fast-LIO's 0.062 m would directly test the insensitivity assertion.","tokens_in":12245,"feed_emoji":"🤖","tokens_out":5592,"duration_ms":52814,"temperature":0.7,"pith_summary":"This paper claims that for LiDAR-inertial odometry running on vibrating ground robots, perfect undistortion of LiDAR scans is unnecessary as long as each point's remaining error can be quantified. It introduces point-wise post-undistortion uncertainty, a per-point covariance derived from measured angular and linear vibration intensity, and shows how to feed that covariance into an iterated Kalman filter through uncertainty-guided point-to-map matching and uncertainty-weighted residuals. If the claim holds, robots on rough terrain can get more accurate trajectories without upgrading IMU hardware, since the estimator learns to trust each scan point proportionally to its estimated reliability. The paper supports the claim with vibration-platform tests, public dataset experiments, and own rough-terrain recordings where the method outperforms Fast-LIO and other baselines under intense vibration.","feed_headline":"Vibration-aware LiDAR odometry beats existing methods on rough terrain","feed_subtitle":"By tagging every point with a vibration-derived covariance, the estimator picks better map matches and trims drift.","key_machinery":"The central object is the point-wise post-undistortion uncertainty, a 3x3 covariance attached to each undistorted LiDAR point after IMU-based motion compensation. It is built from three components: (i) rotational uncertainty from angular vibration, expressed as $\\lfloor p \\rfloor_\\times \\mathrm{diag}\\{\\sigma_{rx}^2, \\sigma_{ry}^2, \\sigma_{rz}^2\\} \\lfloor p \\rfloor_\\times^T$, where $\\lfloor p \\rfloor_\\times$ is the skew-symmetric matrix of the point's coordinates; (ii) translational uncertainty $\\mathrm{diag}\\{\\sigma_{Tx}^2, \\sigma_{Ty}^2, \\sigma_{Tz}^2\\}$; and (iii) the LiDAR's own measurement noise, modeled from range and bearing noise and rotated into the global frame. The vibration-dependent scales are set by mean absolute deviation of angular and linear velocity, scaled by the point's time offset from scan start. This covariance is used in two places: to pick the nearest planes by Mahalanobis distance rather than Euclidean distance, and to weigh the point-to-plane residual $R_j$ in the iterated Kalman filter update.","core_discovery":"The central discovery is that undistortion error from intense vibration can be modeled per point as a Gaussian covariance, computed before matching rather than after, and that this covariance contains usable signal for state estimation. The paper defines the post-undistortion uncertainty as the sum of a rotational covariance $\\lfloor p \\rfloor_\\times \\Sigma_r \\lfloor p \\rfloor_\\times^T$, a translational covariance $\\Sigma_T$, and the LiDAR measurement noise rotated by the undistortion rotation: $L\\Sigma_p = \\Sigma_{\\mathrm{rot}} + \\Sigma_{\\mathrm{trans}} + R \\Sigma_{\\mathrm{meas}} R^T$. The rotational and translational error scales are tied to the angular and linear vibration intensity via $\\sigma_r = \\gamma \\, \\delta t \\, k_\\omega$ and $\\sigma_T = \\gamma \\, \\delta t \\, k_v$, where $k_\\omega$ and $k_v$ are mean absolute deviations of LiDAR-frame angular and linear velocity over the scan. This per-point covariance then drives a Mahalanobis-distance selection of matching map points and a residual weight $R_j = u^T {}^G\\Sigma_{p_j} u$ inside an iterated Kalman filter. The paper reports that this pipeline yields lower absolute pose error than the Fast-LIO, Point-LIO, LIO-SAM, LOAM, and LIO-Mapping baselines on intense-vibration sequences, and that it runs in about 36 ms per scan.","pith_inferences":["The scalar gamma at 0.1 is a global calibration; if gamma varies with IMU sampling rate or LiDAR model, the method may need re-tuning per platform, and the paper's insensitivity claim is based on limited ablations that only test different vibration-intensity measures, not gamma values.","The same post-undistortion covariance could be propagated into loop-closure or map-update stages, not only the current scan's matching, potentially reducing drift accumulation further beyond what the paper reports.","The vibration intensity estimate via mean absolute deviation is heuristic; a learned or model-based predictor of undistortion error could replace it and might remove the need to hand-set gamma.","The approach could be tested on legged or tracked robots, whose vibration spectra differ from wheeled robots, to see whether the Gaussian covariance model generalizes across platforms."],"forward_implications":["Odometry on rough-terrain robots can improve without hardware changes, since the method only adds per-point covariance computation and a Mahalanobis matching step.","The uncertainty estimate makes the matching step robust to scan distortion: points with large predicted error are less influential in the state update.","The method keeps real-time performance (about 36 ms per scan with roughly 2,500 points), so it can run on embedded computers.","The formulation is a drop-in modification to Fast-LIO-style iterated Kalman filters, meaning existing LIO systems could adopt the uncertainty weighting with limited changes."],"supporting_citations":[{"why":"Supplies the Fast-LIO iterated Kalman filter, forward propagation, and on-manifold state update that the proposed method extends.","marker":"[5]"},{"why":"Supplies the LiDAR measurement noise model from range and bearing noise used in Equation (12) for the measurement covariance.","marker":"[18]"},{"why":"Provides the ikd-tree structure used to gather candidate map points for uncertainty-guided matching.","marker":"[20]"},{"why":"Point-LIO is a baseline compared on public datasets and own recordings for evaluating intense-vibration performance.","marker":"[15]"},{"why":"The NCD dataset provides handheld LiDAR sequences with intense motion used for evaluation.","marker":"[21]"},{"why":"The M2DGR dataset provides high-speed ground robot street sequences used to compare odometry accuracy.","marker":"[22]"},{"why":"The Botanic Garden dataset offers unstructured rough-terrain sequences where the robot experiences significant vibration.","marker":"[23]"}],"fun_headline_variants":["Post-undistortion uncertainty improves LiDAR odometry on shaky ground","Per-point vibration covariance refines LiDAR-inertial pose estimation","Vibration-aware point uncertainty reduces drift in LiDAR odometry","LiDAR odometry that models vibration beats baselines on rough terrains","Uncertainty-guided matching stabilizes LiDAR odometry under vibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire uncertainty model depends on one hand-set scalar gamma = 0.1 that converts a mean-absolute-deviation measure of vibration into per-point covariance; if that scaling is wrong for a given IMU frequency, terrain, or LiDAR model, the claimed accuracy gain could shrink or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Post-undistortion uncertainty improves LiDAR odometry on shaky ground","Per-point vibration covariance refines LiDAR-inertial pose estimation","Vibration-aware point uncertainty reduces drift in LiDAR odometry","LiDAR odometry that models vibration beats baselines on rough terrains","Uncertainty-guided matching stabilizes LiDAR odometry under vibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000881,"raw_usage":{"total_tokens":3836,"prompt_tokens":1004,"completion_tokens":2832,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2741}},"tokens_in":620,"tokens_out":2832,"duration_ms":23322,"temperature":1.0,"reasoning_tokens":2741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:50:56.206807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a sensitivity sweep over gamma across different IMU sampling rates (for example, 100 Hz versus 200 Hz) and different vibration amplitudes: if the optimal gamma changes materially across conditions, or if a single gamma fails to improve or degrades odometry relative to Fast-LIO on some intense-vibration sequence, the core claim of a robust uncertainty scale is undermined. Concretely, running the method with gamma from 0.01 to 1.0 on the 04 all-challenges sequence and checking whether the mean absolute pose error stays below Fast-LIO's 0.062 m would directly test the insensitivity assertion.","supporting_citations":[{"cited_title":"Point-lio: Robust high-bandwidth light detection and ranging inertial odometry,","cited_arxiv_id":null,"evidence_quote":"Point-LIO is a baseline compared on public datasets and own recordings for evaluating intense-vibration performance."},{"cited_title":"The newer college dataset: Handheld lidar, inertial and vision with ground truth,","cited_arxiv_id":null,"evidence_quote":"The NCD dataset provides handheld LiDAR sequences with intense motion used for evaluation."},{"cited_title":"M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots,","cited_arxiv_id":null,"evidence_quote":"The M2DGR dataset provides high-speed ground robot street sequences used to compare odometry accuracy."},{"cited_title":"Botanicgarden: A high-quality dataset for robot navigation in unstructured natural environments,","cited_arxiv_id":null,"evidence_quote":"The Botanic Garden dataset offers unstructured rough-terrain sequences where the robot experiences significant vibration."}],"review_version":1}