{"id":"a3cbf615-9dec-46e4-b664-ba5871623aab","arxiv_id":"2507.08349","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A targetless calibration method uses the measured height of the GINS above the ground to make the vertical extrinsic components observable under planar vehicle motion.","lead":"A new calibration method lets mining trucks align several laser scanners to their GPS-and-inertia navigation system without calibration targets or overlapping views. It fixes a known blind spot: on flat ground, standard calibration cannot determine sensor heights, so the method adds the measured height of the navigation unit as an extra constraint.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulated translation error matches the height-prior standard deviation, so the paper has not shown that joint optimization, rather than the manual height measurement, produces vertical accuracy.","rationale":"The reader's weakest assumption was that the accuracy of the manually measured GINS installation height hG determines the vertical accuracy, and that the simulated 0.031 m error matching the prior's 0.032 m standard deviation suggests the sensor data adds little. My analysis of the full text confirms this is the most load-bearing concern: the height pseudo-measurement is the only factor acting on the unobservable vertical direction, the paper provides no ablation varying hG, and the real-world evaluation metrics (MME/MPV) reward map sharpness and cannot detect a global vertical offset. The paper's own Table II and Eq. (2) make the numerical coincidence explicit. This does not invalidate the method's usefulness as a practical way to inject a known installation height, but it does undermine the stronger claim that the joint optimization 'resolves' the planar-motion observability problem and that the reported translation accuracy is achieved by the sensor data. The appropriate remedy is exactly the sensitivity analysis and error decomposition the reader requested, so the CONDITIONAL verdict stands. I found no additional internal inconsistency or fatal flaw; the math is standard nonlinear least squares, and the simulation ground truth is a positive piece of evidence. The promised code release would also help, but the height-prior sensitivity is the decisive scientific gap.","tokens_in":14798,"tokens_out":5384,"duration_ms":68322,"concrete_test":"Re-run the simulated calibration (Table II setup) with the true hG perturbed by +5 cm and -5 cm, keeping all other parameters and the same ground-truth evaluation. Then decompose the final G_L0 T translation error into x, y, and z components. If the z-error shifts by approximately the perturbation magnitude (≈5 cm) while x and y remain small, the vertical accuracy is imported from the manual height prior rather than learned from the joint optimization; this would require revising the claim that the observation model resolves the unobservability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the installation-height observation model of Sec. III-B (Eqs. (1)-(2)) and its use in the joint cost of Sec. IV-G (Eq. (19)). The claimed simulated accuracy for the base-LiDAR-to-GINS transform is 0.031 m translation error (Table II, \"Ours\"). The prior covariance in Eq. (2) sets cs = 10^-3, i.e., a standard deviation of sqrt(10^-3) ≈ 0.032 m on the z-translation of VL_G T. The reported translation error is therefore essentially equal to the standard deviation of the manually measured height. This strongly suggests the vertical component of the estimated extrinsic is dominated by the tape-measure prior, not by the LiDAR/GINS data or the joint optimization. The paper reports only total translation error, not a decomposition into x/y/z, and provides no sensitivity analysis for hG. Without such evidence, the claim that the observation model 'resolves' the planar-motion vertical observability failure is not demonstrated: the method may simply inject an external measurement whose uncertainty propagates directly into the result. This is not an internal inconsistency, but it undercuts the headline accuracy and the novelty of the proposed constraint. The real-world evaluation uses map-consistency metrics (MME/MPV) that are also insensitive to a global vertical offset of the map, so it cannot detect this issue either.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a targetless extrinsic calibration method for multiple LiDARs and a GNSS-aided INS on ground vehicles. The method introduces an installation-height observation model that adds a soft constraint on the LiDAR-GINS vertical translation, estimates a virtual LiDAR frame from the flattest local ground, and refines all extrinsics together with the GINS trajectory in a batch optimization. Experiments on a simulated ground-truth dataset and on a real five-LiDAR mining truck are reported; the simulated results show the proposed method outperforming HECalib and LiDAR-align, and the real-world results are evaluated with map-consistency metrics.","tokens_in":14891,"tokens_out":12567,"duration_ms":143327,"significance":"If the result holds, the paper would provide a practical, target-free, non-overlapping-FoV calibration pipeline for heterogeneous LiDARs in flat mining environments, with a simple external height measurement addressing the planar-motion observability problem. The paper has several strengths: the simulated evaluation uses ground truth and includes a no-overlap heterogeneous LiDAR setup; the installation height is a real external measurement rather than a fitted parameter, so the constraint is not circular; the Ours* experiment shows sensitivity to the virtual-frame extrinsic, although its construction is problematic; and the authors state that code is publicly available. The main weakness is that the load-bearing evidence for the vertical-accuracy claim needs strengthening: the simulated vertical error is consistent with the height prior's standard deviation, and the real-world metrics are insensitive to vertical offsets.","major_comments":[{"comment":"The reported simulated translation error for G_L0 T (0.031 m) is essentially equal to the standard deviation sqrt(cs) = sqrt(1e-3) ≈ 0.032 m of the height prior. The paper does not report the x/y/z decomposition of t_err, nor does it provide a sensitivity analysis over hG or cs. Since the vertical component is unobservable under planar motion, the posterior in that direction is determined predominantly by the prior; the current evidence therefore does not demonstrate that the joint optimization, rather than the manually measured height, produces the advertised vertical accuracy. Please report component-wise errors and repeat the simulation with perturbed hG and varied cs.","section":"Sec. III-B/IV-G, Eq. (2), Table II"},{"comment":"The ablation labeled Ours* is not a valid ablation of terrain analysis. The text states that noise is injected into the estimated base-LiDAR-to-VLiDAR extrinsic parameter to 'simulate the impact of not using terrain analysis.' This confounds the effect of the terrain-analysis module with the arbitrarily chosen noise level and does not rerun the pipeline without the terrain-selection step. A proper ablation should execute the whole pipeline with the single global ground or a non-selected local ground, rather than perturbing the output.","section":"Sec. V-C, Table II (Ours*)"},{"comment":"The mean map entropy (MME) and mean plane variance (MPV) metrics are local map-consistency measures that are invariant to a global vertical translation of a point-cloud map. A constant error in the z component of the LiDAR-GINS extrinsic shifts the entire stitched map vertically and does not change local entropy or plane variance. Table I therefore provides no evidence about the vertical accuracy that the height observation model is designed to improve; it can only validate horizontal and rotational consistency. The real-world evaluation of the central claim is thus unsupported.","section":"Sec. V-B/V-D.1, Table I"},{"comment":"The reference solution for the multi-LiDAR extrinsics is described only as 'rigorously verified,' with no verification procedure, independent measurements, or numerical results given. Since all RMSE values in Table III are computed against this reference, its accuracy directly determines the claimed 0.081 m translation RMSE. Please document the construction and validation of the reference, or report alternative consistency metrics that do not depend on an unverified reference solution.","section":"Sec. V-D.2, Table III"},{"comment":"As written, the joint cost function contains no measurement factor for the GINS poses. The GINS data appear only as initial values and in the motion-consistency factor Eq. (22), where the GINS poses are themselves optimization variables. This means the GINS measurements do not directly constrain the solution in the joint optimization, which is inconsistent with the claim that the method 'tightly couples' LiDAR and GINS data and jointly optimizes the GINS trajectory. Please either add explicit GINS measurement factors to Eq. (23) or clarify that the GINS data are used only for initialization; otherwise the description of the optimization should be revised.","section":"Sec. IV-G, Eq. (23)"}],"minor_comments":[{"comment":"There are numerous typos, including 'covirance' in Eq. (2), 'despicted' in Sec. V-A.2, 'algin' in Sec. V-D.1, 'mthod' in Sec. IV-F, 'continously' in Sec. III-D, and 'an method' in Sec. II-B. A thorough copyedit is needed.","section":"Throughout"},{"comment":"The text says that cb = 10^8 indicates significant uncertainty 'in one dimension,' but cb is applied to the first five diagonal entries of Q; this should be clarified to 'in five dimensions.'","section":"Sec. III-B, Eq. (2)"},{"comment":"The notation for the GINS-to-virtual-LiDAR transform is inconsistent: the paper defines G_VL T in Sec. III-A but uses VL_G T in Eq. (19), Eq. (20), and Eq. (24). Please standardize to a single convention.","section":"Sec. III-A and IV-G"},{"comment":"The sentence 'The selection of keyframes helps reduce computational costs and increases the excitation of motion between frames' is misleading: keyframe selection does not increase the motion actually present; it only selects frames with larger relative motion. Please reword.","section":"Sec. III-D"},{"comment":"All reported results are single-run point estimates with no standard deviations or repeated trials. Given the stochasticity of the DIRECT initialization and of the nonlinear optimization, this limits the significance of the comparisons; please add multiple runs or at least state whether the reported numbers are representative.","section":"Tables II and III"}],"recommendation":"major_revision","confidential_remarks":"The core idea is simple and potentially useful in practice, and the simulated ground-truth test is an appropriate starting point. My main concern is that the evidence does not yet establish the vertical-accuracy claim: the simulated result is consistent with the prior standard deviation, the real-world metrics are insensitive to vertical offsets, and the ablation is not a true ablation. I would be willing to reconsider after a revision that adds component-wise errors, sensitivity analysis, a proper ablation, clarification/generalization of the GINS measurement factors, and a described reference for the real-world multi-LiDAR evaluation. The code release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing you should know: the paper's clever height-prior calibration idea is real, but the simulated accuracy it reports for the vertical direction is basically equal to the standard deviation of the tape-measure prior, so the headline claim that the optimization resolves the planar-motion observability failure is not yet backed by the data.\n\nWhat's new and good: the installation-height observation model in Eqs. (1)-(2) is a neat engineering trick. Under near-planar motion, the z-translation between LiDAR and GINS is weakly observable; instead of requiring targets, overlapping views, or per-LiDAR odometry, they use a measured height hG as an explicit pseudo-measurement with a tight covariance in z and loose covariance elsewhere. The ground-aligned virtual LiDAR frame and the joint factor graph that co-optimizes GINS poses and all extrinsics are a sensible combination, and the simulated ground-truth test (Table II) is the right kind of evidence. The Ours* ablation shows terrain analysis does matter. The comparisons with HECalib and LiDAR-align look honest, and the code link is promised.\n\nThe soft spot is exactly the stress-test point. In Table II, the base-LiDAR-to-GINS translation error is 0.031 m, and the prior's standard deviation is sqrt(10^-3)=0.032 m. That's the same number. The paper reports only total translation error, not x/y/z components, and has no sensitivity analysis for hG. So the likely story is that the vertical component inherits the prior's uncertainty, not that the sensor data or joint optimization pins it down. The real-world MME/MPV metrics are also insensitive to a global vertical offset, so they can't detect this. I don't see fatal flaws: the math is standard, the height measurement is genuinely external (no circularity), and the repeated single-run results just need error bars.\n\nAudience: people calibrating multi-LiDAR rigs to a GINS in flat, targetless environments, especially for mining or ground vehicles. The paper deserves a serious referee. The idea is useful, the writing is clear, and the missing evidence is obtainable: repeated trials, z-error decomposition, and a sensitivity sweep on hG. I'd send it to review with a request for those additions.\n\nBest","headline":"Useful height-prior calibration idea, but the headline vertical accuracy matches the prior's uncertainty, so the observability claim is not yet demonstrated.","tokens_in":15623,"tokens_out":3542,"would_cite":false,"duration_ms":37656,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating the GINS installation height as a noisy measurement restores vertical observability for planar-driving ground vehicles, giving centimeter-level multi-LiDAR calibration without targets or overlapping views.","keywords":["targetless extrinsic calibration","multi-LiDAR calibration","LiDAR-GINS calibration","GNSS-aided INS","planar motion observability","installation height observation model","joint optimization","ground vehicles"],"falsifier":"Run the pipeline twice on the same dataset with $h_G$ shifted by $+5$ cm and $-5$ cm while all other data and settings stay fixed, then compare the estimated vertical translation of the base-LiDAR-to-GINS transform. If the estimate moves by almost exactly the shift, the height prior rather than the point-cloud constraints is controlling vertical accuracy, which would show that the reported 0.031 m vertical error is set by the prior and not by the joint optimization over sensor data.","tokens_in":14412,"feed_emoji":"🚚","tokens_out":12710,"duration_ms":126821,"temperature":0.7,"pith_summary":"The paper aims to establish that extrinsic calibration of multiple LiDARs to a GNSS-aided inertial navigation system (GINS) no longer needs artificial targets, overlapping sensor views, or per-LiDAR odometry, even when a ground vehicle's motion is nearly planar. The enabling idea is an observation model that treats the measured installation height of the GINS above the ground, $h_G$, as a noisy constraint inside a joint optimization of extrinsics and GINS poses. If the paper is right, mining trucks and similar vehicles can be calibrated in flat, unstructured sites by driving normally, with reported simulated base-LiDAR-to-GINS accuracy of 0.056 deg and 0.031 m, and real multi-LiDAR RMSE of 0.293 deg and 0.081 m. The load-bearing premise is that the manually measured height and the locally planar ground it refers to are both accurate; if that premise gives way, the vertical components of the extrinsics inherit the error.","feed_headline":"Tape-measured height unlocks multi-LiDAR calibration","feed_subtitle":"Flat-ground trucks can now get accurate sensor alignment without targets or overlapping views.","key_machinery":"The load-bearing object is the ground-aligned virtual LiDAR frame $\\{VL\\}$ together with the installation-height observation model $y = {}^{VL}_G T\\,\\mathrm{Exp}(\\xi)$, where $h_G$ is the measured height of the GINS above the ground and the covariance $Q$ in Eq. (2) makes the height direction the only tightly constrained component ($c_s = 10^{-3}$). This model injects information in the one direction that planar motion never excites, and the residual $\\mathbf{1}_r = \\mathrm{Log}(y\\,{}^{VL}_G T^{-1})$ becomes the first factor of the joint cost in Eq. (23). The other factors are generalized matching factors built from nearest-neighbor point correspondences with covariance-scaled distances, plus a motion-consistency factor that ties base-LiDAR poses to GINS poses through the VLiDAR transform. A terrain-analysis stage selects the flattest local ground segment to define $\\{VL\\}$, and the whole cost is solved with Levenberg-Marquardt, producing refined VLiDAR-GINS and multi-LiDAR extrinsics together with optimized GINS poses.","core_discovery":"The central claim is that a single extra geometric fact—the known height of the GINS frame above the ground—restores observability of the vertical translation direction that planar vehicle motion removes. The paper defines a virtual LiDAR frame $\\{VL\\}$ whose $xy$-plane coincides with the ground, writes the height measurement as $y = {}^{VL}_G T\\,\\mathrm{Exp}(\\xi)$ with a tight height variance $c_s = 10^{-3}$, and folds the residual $\\mathbf{1}_r = \\mathrm{Log}(y\\,{}^{VL}_G T^{-1})$ into the joint cost of Eq. (23) together with generalized matching factors and a motion-consistency factor. Minimizing that cost simultaneously refines the VLiDAR-to-GINS transform, the transforms from the base LiDAR to the other LiDARs, and the GINS trajectory, with the base-LiDAR-to-GINS extrinsic obtained by composition. On simulated data the base-LiDAR-GINS rotation error is 0.056 deg and translation error 0.031 m; on real five-LiDAR truck data the multi-LiDAR RMSE is 0.293 deg and 0.081 m.","pith_inferences":["Editorial extension: the simulated translation error (0.031 m) is nearly identical to the standard deviation of the height prior ($\\sqrt{10^{-3}} \\approx 0.032$ m), suggesting that in the unobservable vertical direction the prior, not the point-cloud geometry, may determine the advertised accuracy.","Editorial extension: a sensitivity analysis over $h_G$ is the natural next check; if a 5 cm bias in the height measurement shifts the estimated vertical extrinsics by roughly 5 cm, users would know the method requires an accurate tape measurement rather than relying on LiDAR data to correct it.","Editorial extension: because the prior is anchored to a locally planar ground plane, extending the method to sloping or undulating terrain likely requires referencing the height to an absolute vertical datum instead of the nearest ground plane.","Editorial extension: the same joint-cost structure could transfer to camera or radar extrinsics by replacing the point-to-distribution matching residual with photometric or range residuals, giving targetless multi-modal calibration for planar-motion vehicles."],"forward_implications":["Near-planar driving no longer forces a degenerate vertical calibration; the height factor supplies the missing constraint, so ordinary mining-vehicle trajectories are sufficient.","A LiDAR fleet with non-overlapping views and mixed sensor types (solid-state and mechanical) can be calibrated in one joint run, removing the need for targets or scene-specific features.","Because GINS poses are optimized alongside the extrinsics, errors from terrain-induced vehicle vibration are absorbed during estimation rather than propagated as calibration bias.","The same pipeline transfers to unstructured outdoor sites, with reported real-world multi-LiDAR RMSE of 0.293 deg and 0.081 m across four transforms on a five-LiDAR truck.","No per-LiDAR laser odometry is required for calibration, which removes a common failure mode for solid-state LiDARs whose odometry can be unstable."],"supporting_citations":[{"why":"Shows that planar vehicle motion leaves vertical translation components weakly observable, the specific failure the installation-height observation model is designed to fix.","marker":"[39]"},{"why":"Provides the global-search LiDAR-GINS initialization used before batch refinement in the proposed pipeline.","marker":"[22]"},{"why":"Defines the generalized matching factor used for covariance-scaled point-cloud correspondences in the LiDAR-GINS and multi-LiDAR costs.","marker":"[37]"},{"why":"Supplies the ground segmentation used by the ground-alignment and terrain-analysis modules.","marker":"[40]"},{"why":"Hand-eye calibration baseline (HECalib) whose LiDAR-GINS estimates the paper compares against.","marker":"[41]"},{"why":"Provides the per-LiDAR laser odometry used by the HECalib baseline to estimate trajectories in the comparative experiments.","marker":"[43]"},{"why":"The DIRECT global optimization algorithm used to solve the initialization residual for the LiDAR-GINS rotational part.","marker":"[38]"}],"fun_headline_variants":["Height prior unlocks targetless multi-LiDAR calibration","Known GINS height fixes flat-ground LiDAR extrinsics","One height measurement calibrates multiple LiDARs","Planar motion? Use GINS height to calibrate LiDARs","Targetless LiDAR calibration via GINS height constraint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the manually measured height of the GINS unit above the ground is accurate to within a few centimeters and that the ground is locally flat at the place where that height was measured; if either condition fails, the vertical component of every estimated transform inherits the error and the LiDAR data alone cannot correct it.","fun_headline_variants_meta":{"raw":{"variants":["Height prior unlocks targetless multi-LiDAR calibration","Known GINS height fixes flat-ground LiDAR extrinsics","One height measurement calibrates multiple LiDARs","Planar motion? Use GINS height to calibrate LiDARs","Targetless LiDAR calibration via GINS height constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1412,"prompt_tokens":1001,"completion_tokens":411,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":330}},"tokens_in":617,"tokens_out":411,"duration_ms":4870,"temperature":1.0,"reasoning_tokens":330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:23:55.063097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline twice on the same dataset with $h_G$ shifted by $+5$ cm and $-5$ cm while all other data and settings stay fixed, then compare the estimated vertical translation of the base-LiDAR-to-GINS transform. If the estimate moves by almost exactly the shift, the height prior rather than the point-cloud constraints is controlling vertical accuracy, which would show that the reported 0.031 m vertical error is set by the prior and not by the joint optimization over sensor data.","supporting_citations":[{"cited_title":"Observability- aware intrinsic and extrinsic calibration of lidar-imu systems,","cited_arxiv_id":null,"evidence_quote":"Shows that planar vehicle motion leaves vertical translation components weakly observable, the specific failure the installation-height observation model is designed to fix."},{"cited_title":"lidar align,","cited_arxiv_id":null,"evidence_quote":"Provides the global-search LiDAR-GINS initialization used before batch refinement in the proposed pipeline."},{"cited_title":"Generalized-icp,","cited_arxiv_id":null,"evidence_quote":"Defines the generalized matching factor used for covariance-scaled point-cloud correspondences in the LiDAR-GINS and multi-LiDAR costs."},{"cited_title":"Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud,","cited_arxiv_id":null,"evidence_quote":"Supplies the ground segmentation used by the ground-alignment and terrain-analysis modules."},{"cited_title":"Hand-eye calibration using dual quaternions,","cited_arxiv_id":null,"evidence_quote":"Hand-eye calibration baseline (HECalib) whose LiDAR-GINS estimates the paper compares against."},{"cited_title":"Direct li- dar odometry: Fast localization with dense point clouds,","cited_arxiv_id":null,"evidence_quote":"Provides the per-LiDAR laser odometry used by the HECalib baseline to estimate trajectories in the comparative experiments."},{"cited_title":"Lipschitzian optimization without the lipschitz constant,","cited_arxiv_id":null,"evidence_quote":"The DIRECT global optimization algorithm used to solve the initialization residual for the LiDAR-GINS rotational part."}],"review_version":1}