{"id":"7ca1f8aa-bd4c-4b03-9c29-49781f4e5b55","arxiv_id":"2505.10018","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LEMON-Mapping merges multi-robot LiDAR maps with spatial bundle adjustment around loops and two-step pose graph optimization for global consistency.","lead":"LEMON-Mapping fuses 3D maps from multiple robots by refining point clouds around loop closures and then smoothing the whole trajectory. It reports large accuracy gains over prior multi-robot map merging tools on public benchmarks, which matters for robot teams that must share one consistent map.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Lemma 3 (Appendix VIII-C) is invalid and the lemma is false; DBA's safety rests entirely on unvalidated Assumption 1, so the paper's central theoretical justification is not established.","rationale":"The reader's conditional verdict is well aligned with the main technical weakness. I independently re-derived the Schur-complement step in Appendix VIII-C and found a concrete counterexample to Lemma 3, so the issue is a false theorem, not merely a missing detail. This matters because Section V-A explicitly presents the covariance ordering as the justification for freezing inner poses 'without assuming inner poses are better measured.' With Lemma 3 removed, the DBA approximation rests on Assumption 1, which is never checked empirically. The practical impact on the headline claim is uncertain: the system may still work well, and the benchmark results are extensive, but a central supporting argument is broken. The other concerns from the reader (no code, no error bars, lenient failure threshold) are secondary; they affect reproducibility and strength of the empirical claim but do not identify a specific false step. A conditional verdict is appropriate: the authors should either repair the lemma if possible, replace it with a direct empirical validation of Assumption 1 and of r_H and r_J, or remove the claim that DBA has a proven covariance guarantee. I therefore leave the reader's verdict unchanged.","tokens_in":20863,"tokens_out":11341,"duration_ms":113111,"concrete_test":"Instrument the DBA module on two representative sequences (e.g., S3E Library and MARS-LVIG Island): after FPGO and loop recall, for every isolated-loop spatial window, compute the Hessian blocks H00, H01, H11 and Jacobians J0, J1 from the BALM2 cost, and evaluate r_H and r_J defined in Eq. (7). If a non-trivial fraction of windows has r_H or r_J not much smaller than 1 (say >0.1), Assumption 1 does not hold in practice and the DBA freezing is unjustified. As a direct check of the impact, rerun the same scenes with joint BA over the active window instead of DBA and compare ATE RMSE and MapEval AWD. If joint BA does not beat DBA, the false lemma has no practical consequence for the headline accuracy claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Lemma 3 (Section V-A, proof in Appendix VIII-C), which is used to justify freezing the inner pose block M_p0 in DBA without invoking Assumption 1. The appendix proof is not just incomplete; the claimed inequality is false. The identity in VIII-C1 is wrong: subtracting diag(0, 1L Σ 1L^T) from L Σ L^T leaves the off-diagonal blocks 0L Σ 1L^T and its transpose, not zero, so the residual is not [0L;0] Σ [0L;0]^T. In addition, the bottom-right block of H^{-1} M H^{-T} is not S^{-1} M_{11} S^{-T}; cross terms involving M_{00} and M_{01} enter and can make the joint covariance smaller than the DBA covariance. A concrete scalar counterexample is H=[[1,0.9],[0.9,1]], M=[[1,0.9999],[0.9999,1]]: DBA gives covariance 1 for the active block, while the joint covariance is about 0.28, contradicting Lemma 3. Therefore DBA's safety rests entirely on unvalidated Assumption 1 (inner poses have significantly smaller noise covariance). If that assumption fails in some loop geometry, the frozen-block approximation can systematically bias the outer poses and the local fused map, and the paper provides no measurement of r_H and r_J (Eq. 7) to show the coupling is negligible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LEMON-Mapping, a multi-session point-cloud merging and optimization framework for LiDAR mapping. The system comprises a loop-processing module with outlier rejection and a false-negative recall mechanism, a spatial bundle-adjustment stage with two variants (diffusion BA for isolated loops and hierarchical BA for clustered loops), and a two-step pose-graph optimization that propagates local corrections globally. The authors evaluate the framework on the S3E, GEODE, MARS-LVIG, and R3LIVE public datasets plus a self-collected dataset, comparing against DCL-SLAM, LAMM, BALM2, and HBA. They report lower ATE and better MapEval metrics, and a scalability study in which 5, 10, and 20 R3LIVE sessions are said to be merged with 100% success. The central empirical claim is that combining loop recall, spatial BA, and a final PGO yields globally consistent, geometrically accurate multi-robot maps.","tokens_in":21213,"tokens_out":3628,"duration_ms":37987,"significance":"If the experimental results hold, LEMON-Mapping is a practically valuable contribution to multi-robot LiDAR mapping: it engages with a real problem (local divergence and blurring in overlapping multi-session maps), is evaluated on multiple public datasets against external baselines, and includes ablations of the proposed components. The scalability experiment, with up to 20 sessions, is a useful stress test that many existing systems do not attempt. However, the theoretical justification for the diffusion BA (DBA) step relies on Lemma 3, whose proof in Appendix VIII-C is algebraically incorrect and whose statement appears false. Since this lemma is used to claim that freezing inner poses is safe even without assuming inner poses are better measured, the theoretical contribution is not currently established. The paper would be materially improved by either repairing or removing this claim and by validating the underlying Assumption 1 empirically.","major_comments":[{"comment":"The proof of Lemma 3 is invalid. In Section VIII-C1, the claimed identity L_ij Σ L_ij^T - diag(0, 1L_ij Σ 1L_ij^T) = [0L_ij; 0] Σ [0L_ij; 0]^T is wrong: subtracting the bottom-right block leaves the off-diagonal blocks 0L_ij Σ 1L_ij^T and 1L_ij Σ 0L_ij^T in the residual, and those are not zero. Consequently inequality (15) does not follow. The later Schur-complement step is also not valid as stated: the bottom-right block of H^{-1} M H^{-T} is not S^{-1} M_11 S^{-T}; cross terms involving M_00 and M_01 enter. In fact the claimed ordering is false: with H = [[1,0.9],[0.9,1]] and M = [[1,0.9999],[0.9999,1]], the joint covariance of the active block is about 0.28 while the DBA covariance is 1, contradicting Σ_DBA^1 ⪯ Σ_joint^11. The lemma therefore cannot be used to justify DBA.","section":"Section V-A and Appendix VIII-C"},{"comment":"With Lemma 3 false, the safety of freezing the inner pose block in DBA rests entirely on Assumption 1, which states that the inner LiDAR poses have significantly smaller measurement noise covariance than the outer poses. The paper does not validate this assumption in any experiment: the two refinement rates r_H and r_J defined in Eq. (7) are never measured, and no loop geometry is reported in which the coupling terms are shown to be negligible. I request either an empirical validation of Assumption 1 (e.g., reporting r_H and r_J across the tested loop closures) or a weakened theoretical statement that explicitly relies on the assumption, together with evidence that the assumption holds in the experiments.","section":"Section V-A, Assumption 1 and Eq. (7)"},{"comment":"The scalability claim of a 100% success rate for 5, 10, and 20 sessions is weakened by the binary success criterion. The definition given, 'correct alignment of each session with all its adjacent sessions that shares sufficient map overlap,' does not state a quantitative threshold, and no trajectory or map-quality metric is reported for these runs. Please report per-session ATE/RMSE or a quantitative alignment-error threshold for the three scalability cases, so that the 'success' claim is falsifiable and comparable with the other experiments.","section":"Section VII-F, Table IX"}],"minor_comments":[{"comment":"There is a typo: 'limitted' should be 'limited.'","section":"Section II-B"},{"comment":"The layout of Table II and the caption do not clearly separate the dataset name from the number of robots; consider a more explicit table structure.","section":"Section VII-A and Table II"},{"comment":"The formatting of Table IV appears corrupted in places (e.g., the spacing between values and the '×' entries); please ensure the table typesets correctly.","section":"Section VII-C, Table IV"},{"comment":"The MapEval metrics are reported for only three sequences of MARS-LVIG; it would be helpful to state explicitly whether the remaining MARS-LVIG sequences were omitted because of missing ground-truth maps or runtime constraints.","section":"Section VII-D, Table V"},{"comment":"The covariance expression in Eq. (9) would benefit from an explicit index of summation over j and from a sentence explaining why the cross terms between different feature clusters vanish or are neglected.","section":"Section V-A, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is credible and well aligned with the journal's scope. The main problem is the false Lemma 3 and its invalid proof; since the lemma is advertised as removing the need for Assumption 1, the authors should either correct the mathematical claim or, more realistically, reframe DBA's safety as an empirically validated assumption and supply the missing measurements of r_H and r_J. The scalability experiment would also be stronger with quantitative error metrics. I see no reason to reject outright, but the current theoretical overreach needs to be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"For the record: this is a solid systems paper for multi-robot LiDAR mapping. The new pieces are real: a loop-recall mechanism that runs after a first pose-graph optimization, a spatial bundle adjustment that optimizes poses from different robots in a geometric window around loops, and a final PGO that propagates those refined constraints globally. The experimental evaluation is broad—public datasets, comparisons to LAMM and DCL-SLAM, and a 20-session scalability test—and the accuracy margins on the successful runs are large enough to be taken seriously. The integration of known components (BALM2, HBA, RING++, g2o, PCM) into a working pipeline is a legitimate contribution; not everything needs to be new under the sun.\n\nThe soft spot is the theory. The proof of Lemma 3 in Appendix VIII-C is simply wrong. The claimed identity in the first step omits the off-diagonal blocks; subtracting diag(0, 1L Σ 1L^T) from L Σ L^T leaves 0L Σ 1L^T and its transpose. And the Schur-complement ordering step is also invalid, because H_11^{-1} ⪯ S^{-1} does not imply that the conjugated sum with the measurement noise is ordered in that direction. The stress-test counterexample is concrete: with H=[[1,0.9],[0.9,1]] and M=[[1,0.9999],[0.9999,1]], joint BA gives the active block a covariance around 0.28, while DBA gives 1. So the lemma is false, and the claim that freezing inner poses is safe without Assumption 1 does not hold.\n\nThis matters because the whole DBA design for isolated loops leans on that lemma. The paper does include Assumption 1 (inner poses have smaller noise covariance), which is probably true in many practical cases, but the authors explicitly try to prove the ordering without it, and that attempt fails. The right fix is either to repair the proof, drop the lemma and state Assumption 1 plainly, or provide direct evidence—e.g., measure r_H and r_J from Eq. 7 on real data—that the coupling is negligible. The rest of the empirical story can stand independently; I would not read this as invalidating the system's engineering claims.\n\nMinor issues: no code or parameters released, no error bars on the reported numbers, the 30m failure threshold is lenient, and a few baselines named in related work (AutoMerge, Disco-SLAM) are not compared. These are all addressable and do not shake my overall read.\n\nWho is this for? Groups working on multi-robot SLAM, map merging, and long-term mapping will get real value from the pipeline and the experiments, even if they skip the appendix. It deserves a serious referee, and I would accept it with the expectation that the theoretical section gets fixed or explicitly weakened.","headline":"A strong engineering paper for multi-robot LiDAR mapping with a genuinely new spatial-BA design, but the proof of its key theoretical lemma is wrong and the safety of the core approximation rests on an unvalidated assumption.","tokens_in":21785,"tokens_out":1961,"would_cite":true,"duration_ms":21500,"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":"LEMON-Mapping claims loop closures should seed map-driven local refinement, not just pose-graph constraints, and reports accurate fusion of up to 20 sessions.","keywords":["multi-robot SLAM","point cloud merging","loop closure","bundle adjustment","pose graph optimization","LiDAR mapping","global consistency","large-scale mapping"],"falsifier":"Compare diffusion BA against full joint BA on a dataset where odometry noise near the loop is artificially inflated so inner and outer poses have comparable uncertainty; if the frozen-inner approximation relies on Assumption 1, the two outputs will diverge and the DBA map will show systematic misalignment, contradicting the claimed covariance ordering.","tokens_in":20663,"feed_emoji":"🗺️","tokens_out":9716,"duration_ms":92552,"temperature":0.7,"pith_summary":"LEMON-Mapping tries to establish that fusing many LiDAR point cloud sessions should be treated as a map-driven registration problem, not just a pose-graph optimization problem. The paper builds a pipeline that filters and recalls loop closures, runs spatial bundle adjustment in local windows around each loop, and then uses two pose graph optimization steps to propagate the local corrections across the whole map. If the claim is right, multi-robot teams exploring large areas can merge many overlapping sessions into a single geometrically accurate map without the divergence and blurring that pure pose-graph methods leave in overlap regions. The reported evidence includes lower absolute trajectory error than two comparison systems on public datasets and a 100% success rate when fusing 5, 10, and 20 sessions from the R3LIVE dataset.","feed_headline":"Merging 20 robot maps at once now succeeds 100% of the time","feed_subtitle":"A loop-driven spatial bundle adjustment fixes local divergence and carries that accuracy to the whole map.","key_machinery":"The load-bearing mechanism is the spatial bundle adjustment: instead of a temporal sliding window, poses are gathered by spatial proximity to each loop closure, so observations of the same plane or line from different robots are optimized together. It comes in two forms—diffusion BA (DBA), which freezes the inner poses near an isolated loop and incrementally adds poses from wider rings, and a spatial hierarchical BA (HBA), which reorders cross-robot poses by principal component analysis so that adjacent poses in the optimization window are spatially close. The rest of the pipeline exists to feed this mechanism and carry its output: a loop processing module rejects false loops and recalls rejected-but-valid ones after the first pose graph, and a final pose graph with sparsified constraints transfers the local alignment to the global map.","core_discovery":"The paper's central claim is that loop closures carry geometric information that traditional multi-robot pose graph optimization wastes: instead of using a loop only as a constraint between two pose nodes, LEMON-Mapping collects every pose inside a spherical window around the loop and jointly optimizes the poses of all robots against the plane features they observe. That spatial bundle adjustment has two variants, one that freezes inner poses and diffuses outward for isolated loops and one that reorders poses by principal component analysis for clusters of loops. A final pose graph, loaded with sparsified bundle-adjustment constraints and odometry edges, then transfers the local alignment to the whole trajectory. The paper reports that this design removes the layered, blurred geometry in overlapping regions, beats two pose-graph-only baselines on the S3E, GEODE, and MARS-LVIG datasets, and merges up to twenty R3LIVE sessions with a 100% success rate.","pith_inferences":["Beyond the paper, the spatial-window idea could be applied to lifelong mapping or collaborative relocalization: any region revisited across sessions becomes a local optimization cell that a global graph stitches together.","The theoretical justification for freezing inner poses during diffusion bundle adjustment is less secure than the paper's Lemma 3 claims; the appendix proof appears to contain a false matrix identity, so the practical safety of the approximation currently rests on Assumption 1 that inner poses are far better measured than outer poses.","A stricter test of the twenty-session scalability result would measure global map quality against ground truth, since the reported success criterion is correct alignment with neighboring overlapping sessions rather than global end-to-end accuracy.","The 2-meter recall threshold and the descriptor-similarity threshold for cross-robot constraints are tunable; datasets with different LiDAR density or motion patterns may require recalibration, making a sensitivity study a natural next step."],"forward_implications":["If the central claim holds, multi-robot teams can merge many overlapping sessions into one geometrically accurate map, because the spatial bundle adjustment repairs overlap regions before the global pose graph distributes the correction.","Local accuracy from loop regions survives global propagation: the final pose graph carries sparsified bundle-adjustment constraints, so the whole map inherits the refinement rather than reverting to pose-only consistency.","The method scales past the usual three-to-five robot limit: the reported 100% success on 5, 10, and 20 sessions indicates the pipeline does not degrade as the number of sessions grows.","Loop recall is a measurable win: recovering rejected-but-valid loops after the first alignment reduces absolute trajectory error in the first pose graph and supplies extra constraints for the later bundle adjustment."],"supporting_citations":[{"why":"Supplies the LiDAR bundle adjustment formulation and closed-form plane-feature residuals that the diffusion BA is built upon.","marker":"[24]"},{"why":"Supplies the hierarchical bundle adjustment scheme that is adapted into the spatial HBA for clustered loops.","marker":"[25]"},{"why":"Provides the RING++ loop detection descriptors used for raw loop closures and for selecting high-similarity cross-robot BA constraints.","marker":"[21]"},{"why":"LAMM is the multi-session map merging baseline that the localization and mapping-quality experiments must beat.","marker":"[29]"},{"why":"DCL-SLAM is the distributed multi-robot SLAM baseline compared in the localization experiments.","marker":"[27]"},{"why":"MapEval supplies the mapping quality metrics (Wasserstein distance, Chamfer distance, consistency scores, plane thickness) used in evaluation.","marker":"[52]"},{"why":"R3LIVE is the dataset source for the scalability experiments with 5, 10, and 20 sessions.","marker":"[51]"},{"why":"MARS-LVIG provides the large-scale aerial datasets used for trajectory and mapping quality evaluation.","marker":"[1]"},{"why":"GEODE provides the heterogeneous LiDAR datasets used in the multi-robot localization comparison.","marker":"[50]"},{"why":"S3E supplies multi-robot campus and tunnel sequences used for localization and ablation studies.","marker":"[44]"}],"fun_headline_variants":["Loop geometry boosts multi-robot map merging to 100% success","100% success merging 20 robot maps with loop-aware optimization","Spatial bundle adjustment: loops fix multi-robot map blur","Multi-robot mapping: loop geometry meets spatial bundle adjustment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 1 in Section V-A, namely that the inner poses frozen during diffusion bundle adjustment have significantly smaller measurement noise than the outer poses being added; the paper's stronger claim that this assumption is unnecessary (Lemma 3) appears to rest on a false matrix identity, so the safety of the approximation in practice depends on that unverified noise-ordering assumption.","fun_headline_variants_meta":{"raw":{"variants":["Loop geometry boosts multi-robot map merging to 100% success","100% success merging 20 robot maps with loop-aware optimization","Spatial bundle adjustment: loops fix multi-robot map blur","Multi-robot mapping: loop geometry meets spatial bundle adjustment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000982,"raw_usage":{"total_tokens":4177,"prompt_tokens":961,"completion_tokens":3216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":3144}},"tokens_in":577,"tokens_out":3216,"duration_ms":23543,"temperature":1.0,"reasoning_tokens":3144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:18:40.932555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare diffusion BA against full joint BA on a dataset where odometry noise near the loop is artificially inflated so inner and outer poses have comparable uncertainty; if the frozen-inner approximation relies on Assumption 1, the two outputs will diverge and the DBA map will show systematic misalignment, contradicting the claimed covariance ordering.","supporting_citations":[{"cited_title":"Ring++: Roto-translation invariant gram for global localization on a sparse scan map,","cited_arxiv_id":null,"evidence_quote":"Provides the RING++ loop detection descriptors used for raw loop closures and for selecting high-similarity cross-robot BA constraints."},{"cited_title":"Large-scale multi-session point-cloud map merging,","cited_arxiv_id":null,"evidence_quote":"LAMM is the multi-session map merging baseline that the localization and mapping-quality experiments must beat."},{"cited_title":"Dcl-slam: A distributed collaborative lidar slam framework for a robotic swarm,","cited_arxiv_id":null,"evidence_quote":"DCL-SLAM is the distributed multi-robot SLAM baseline compared in the localization experiments."},{"cited_title":"Mapeval: Towards unified, robust and efficient slam map evaluation framework,","cited_arxiv_id":null,"evidence_quote":"MapEval supplies the mapping quality metrics (Wasserstein distance, Chamfer distance, consistency scores, plane thickness) used in evaluation."},{"cited_title":"S3e: A multi-robot multimodal dataset for collaborative slam,","cited_arxiv_id":null,"evidence_quote":"S3E supplies multi-robot campus and tunnel sequences used for localization and ablation studies."}],"review_version":1}