Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Improving Map Consistency in Graph-Based LiDAR SLAM Through Information-Aware Odometry and Retroactive Loop Closure

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read A graph-based LiDAR SLAM system can improve map consistency at revisits—fewer duplicated surfaces—using uncertainty-weighted odometry and retroactive loop closure, while matching state-of-the-art trajectory accuracy.

desk verdict Solid LiDAR SLAM systems paper with real new mechanisms, but the headline 'map quality' claim rests on an unvalidated self-consistency metric—send it out, ask for ground-truth validation. read the letter →

arxiv 2607.13516 v1 pith:SKUOOTJR submitted 2026-07-15 cs.RO

classification cs.RO
keywords LiDARSLAMmapconsistencyloopclosureinformationmatrixestimationICPuncertaintyposegraphoptimizationplacerecognitionrevisitevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to prove that a graph-based LiDAR SLAM system can improve the geometric consistency of its maps at revisit locations—fewer duplicated or misaligned surfaces—without sacrificing global trajectory accuracy. It claims this is achieved by replacing fixed, isotropic odometry weights with per-scan information matrices estimated cheaply from the local shape of the ICP objective, and by adding a two-stage loop-closure pipeline: a conservative appearance-based stage that separates place recognition from fine geometric registration, and a retroactive geometry-based stage that uses the optimized pose graph to recover loop closures the first stage missed. The paper supports the claim with experiments on several large-scale outdoor datasets, reporting both standard trajectory errors and a newly proposed revisit-based map-consistency score, where the method is on par with or better than existing systems on trajectory accuracy while consistently scoring better on local map agreement. A sympathetic reader would care because this addresses the practical gap where a low trajectory error does not guarantee a clean map, which matters for robots that rely on maps for navigation and planning.

What carries the argument

The load-bearing object is the estimated ICP information matrix: a block-diagonal Hessian approximation obtained by solving two small linear regressions over sampled perturbations of the converged pose, projected onto the positive-definite cone and used as the odometry constraint weight in pose-graph optimization. The second mechanism is the hierarchical map representation—large local maps (100 m splits) used for appearance-based place recognition and smaller submaps (25 m splits) used for fine registration, linked by Hausdorff-distance (maximum-separation distance) trajectory comparison. The third is the retroactive loop-closure stage: after optimizing the pose graph, local-map trajectory p

What would settle it

A concrete test: pick a revisit region and obtain an independent, survey-grade reference scan of that region. Reconstruct local maps from each method's estimated poses, then rank methods by distance to the reference. If the method with the lowest nearest-neighbor RMS among revisits is not the closest to the reference, the proposed metric rewards self-alignment rather than true map quality, and the stronger reading of the paper's claim fails. A second, simpler check: on a sequence with deliberately large initial drift, verify whether optimized trajectory separation of known missed revisits stay

Watch

Extended reading notes

Core claim

The central claim is that trajectory estimation and local map quality are complementary objectives that can be improved jointly in a graph-based LiDAR SLAM pipeline. Specifically, the paper argues that (i) estimating the information matrix of each ICP odometry constraint by fitting a quadratic local model to perturbed evaluations of the point-to-point ICP objective—rather than using fixed weights or expensive sampling—gives the pose graph principled, geometry-aware constraints; (ii) decoupling place recognition (done on large local maps) from geometric verification (done on smaller submaps) yields reliable loop closures; and (iii) feeding the optimized pose graph back into a retroactive stag

Load-bearing premise

The map-quality claim rests on treating lower nearest-neighbor RMS between local maps built from the estimated poses as 'better map quality'; if this self-consistency proxy is not validated against absolute or independent geometry, the paper's second central claim is not established.

Editorial extensions

If this is right

  • Weighting odometry constraints with geometry-aware information matrices can improve trajectory accuracy on challenging, perceptually aliased sequences without changing the underlying ICP registration.
  • Conservative loop-closure front-ends that trade recall for precision can be combined with a back-end feedback stage that recovers the missed loops, making the system reliable in environments where appearance-based methods fail.
  • Reported SLAM trajectory metrics (ATE/RPE) should be complemented by a revisit-based map-consistency score, since the experiments show comparable trajectory errors with substantially different local map quality.
  • Because the uncertainty-estimation method is stated to be compatible with any ICP objective, it can be transferred to other scan-matching pipelines that already use pose-graph optimization.
  • The retroactive stage's benefit is concentrated on scenes where optimization brings missed revisits into geometric proximity (within 100 m); in well-constrained scenes with little drift, it contributes little, as the ablation shows.

Reading between the lines

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

  • Editorial inference: a natural next validation step is to score the same revisit regions against an independent survey-grade reference map; that would separate 'self-consistent' from 'true' map accuracy and strengthen or refine the paper's second claim.
  • Editorial inference: if the information-matrix estimation is as cheap and generic as described, it could be adopted as a standard drop-in module for other ICP-based odometry front-ends, allowing existing SLAM systems to become uncertainty-aware without architectural changes.
  • Editorial inference: the retroactive stage suggests a general pattern—treat the optimized graph as a geometric prior for a second pass of place recognition. A natural testable extension is to iterate the loop (optimize, search for new geometric revisits, re-optimize) until convergence and check whether map consistency continues to improve or plateaus.
  • Editorial inference: the 100 m Hausdorff threshold controls how much residual drift the retroactive stage can tolerate; adaptively estimating this threshold from the post-optimization trajectory spread would be a testable extension for large-drift sequences.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a graph-based 3D LiDAR SLAM system with three main components: (1) a real-time estimator of ICP local information matrices, obtained by fitting a quadratic model to perturbations of the point-to-point ICP cost and using the block-diagonal Hessian as the odometry constraint weight; (2) a hierarchical loop-closure front-end that uses large local maps for place recognition and smaller submaps for geometric verification; and (3) a retroactive loop-closure module that, after pose-graph optimization, re-examines previously missed revisits by geometric proximity and registers them when their optimized trajectories are close. The authors also propose a revisit-based map-consistency evaluation protocol. The system is evaluated on HeLiPR, MulRan, Apollo, and Newer College against CT-ICP, MULLS, PIN-SLAM, and KISS-SLAM, with ablations for both the information-weighting module and the retroactive loop-closure module. The paper claims, first, global trajectory accuracy on par with or better than state-of-the-art and, second, improved local map quality in revisited areas.

Significance. If the results hold, the paper makes a useful practical contribution. Real-time information-matrix estimation for point-to-point ICP is a genuinely useful capability, since most LiDAR SLAM systems rely on heuristic or isotropic weights. The retroactive loop-closure idea is also well motivated and is a reasonable way to recover missed revisits without inflating false positives in the appearance-based front-end. The experimental evaluation is broad: four datasets, several sensors, multiple baselines run from public implementations, and ablation studies for both principal components. The fact that a single parameter set is used across all datasets strengthens the generality claim for the trajectory results. The trajectory claims are, in my reading, adequately supported by Tables I and II. The map-quality claim, however, is supported only by an unvalidated self-consistency metric that is very close to what the method itself optimizes. This is the main weakness and prevents the paper from being accepted as is. The paper would be significantly strengthened by validating the proposed map-consistency metric against an absolute reference or by presenting independent evidence that lower self

major comments (3)
  1. [Sec. IV-C, Figs. 3 and 4] The second central claim, that the method 'improves the local map quality in revisited areas,' is established only by the evaluation protocol of Sec. IV-C. That protocol identifies revisit cells from the reference trajectory, rebuilds local maps from each method's estimated global poses, and measures nearest-neighbor RMS between those local maps. This is a self-consistency metric: it rewards any method that makes repeated traversals agree, and both pose-graph optimization and the retroactive loop-closure module (Sec. III-E, Eq. (13)-(15)) are directly designed to minimize exactly this kind of geometric disagreement. A method can aggressively align overlapping geometry while biasing the reconstructed map away from absolute truth and still score well. No experiment in the paper validates the metric against ground-truth geometry, a surveyed map, TLS data, or an independent quality assessmen
  2. [Sec. III-B, Eqs. (3)-(7)] The information-matrix estimator is a load-bearing component, but its validity is not directly established. The paper fits a quadratic model to the point-to-point ICP cost and uses the clamped Hessian H+ as the information matrix. This Hessian is the local curvature of a sum of squared Euclidean residuals, not an inverse measurement covariance; the paper states the 'standard interpretation' but does not justify the missing scaling or the relationship between cost curvature and estimation uncertainty when correspondences are recomputed. The perturbation ranges [−0.1, 0.1] m and [−0.01, 0.01] rad are chosen empirically, the block-diagonal truncation in Eq. (6) drops translation-rotation coupling, and the eigenvalue clamp of 1e-3 is ad hoc. The ablation shows that replacing the estimator with a fixed weight hurts ATE, but that only demonstrates that some adaptive weighting helps, not that H
  3. [Sec. III-E and Eq. (15)] The retroactive loop-closure module depends on the assumption that after pose-graph optimization, the trajectories of initially missed revisits come within the Hausdorff threshold tau_H = 100 m (Eq. (15)). This threshold is very loose, and the paper provides no information about how many retroactive constraints are actually recovered, what their precision is, or how sensitive the results are to tau_H. The ablation 'no-feedback' shows only small ATE changes on most sequences, and the map-quality ablation is not statistically assessed. Because this module is the main novelty behind the map-consistency improvement, the evaluation should report, per sequence: number of retroactive loop closures, number verified by geometric registration, the acceptance rate, and a sensitivity study over tau_H. Without this, the reader cannot tell whether the map-consistency differences in Figs. 3-4 are drive
minor comments (5)
  1. [Sec. IV-C, Figs. 3 and 4] The x-axis labels in Fig. 3 are difficult to read ('Bridge TownRoundabout' and a similar concatenation in Fig. 4). Please separate the scenario names and ensure the axis label is visible. Also clarify whether the marker size encodes the standard deviation directly or a scaled value.
  2. [Table II] The MULLS entry on 02-long-sequence reports RPE = 4633.62%. This is likely a formatting or unit issue, and it is inconsistent with the other entries. Please correct or explain.
  3. [Sec. III-C, Eq. (13)] The Hausdorff distance in Eq. (13) is computed between 'trajectories' represented by discrete poses. Please specify whether the comparison is over the raw pose sets, a subsampled set, or an interpolated continuous curve. This affects both the computational cost and the meaning of tau_H = 100 m.
  4. [Sec. III-D] The scan-level pose graph stores every scan as a node. The paper says optimization is infrequent because loop closures are detected at most once per local map, but the retroactive module may add additional constraints. Please clarify how often optimization runs and report the actual runtime or per-scan cost, since the paper emphasizes real-time performance.
  5. [Sec. IV-A] Several threshold parameters (tau_inliers, tau_f, tau_H, perturbation ranges) are stated as fixed, but the sensitivity of the final results to these values is not examined. At least a brief sensitivity discussion would help the reader judge how much tuning is hidden in these choices.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: trajectory and map claims are empirical and benchmarked externally; the map-consistency metric is a self-consistency proxy but not an equation-level reduction to the method's objective.

full rationale

The claimed contributions are experimental, not first-principles derivations. The information matrix in Sec. III-B is estimated by sampling the ICP objective around the converged pose and using the local Hessian as information; it is not fitted to any evaluation target, and trajectory accuracy is assessed against external benchmarks (ATE/RPE on HeLiPR, MulRan, Apollo, Newer College) with public baseline implementations. The retroactive loop closure is a feedback mechanism, not a circular reduction: it uses the optimized graph to propose candidate constraints and re-optimizes, and the ablation shows its removal degrades map-consistency scores. Self-citations ([13], [14], [31]) are used as pipeline components; [31] is explicitly disclaimed as a contribution, [14] is a published place-recognition method, and these are not invoked as a uniqueness theorem or an ansatz to force the paper's conclusions. The closest issue is the map-consistency protocol (Sec. IV-C): it measures nearest-neighbor RMS between local maps built from estimated poses, which is strongly aligned with what the loop-closure optimizer improves. That makes 'improved map quality' dependent on a self-consistency proxy that is not validated against ground-truth geometry; this is a correctness/validation limitation, not an equation-level circularity, because the evaluation is an empirical measurement rather than a derivation from the metric. Overall, no load-bearing circular step; score reflects minor self-citation usage.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The central claims rest on hand-chosen thresholds, an uncalibrated Hessian-to-information mapping, and an unvalidated self-consistency map metric. No new physical or conceptual entities are introduced beyond algorithmic modules.

free parameters (10)
  • perturbation_translation_range = [-0.1, 0.1] m
    Empirically chosen in Sec. III-B; determines the loss-landscape samples used to fit H and therefore the odometry information matrices.
  • perturbation_rotation_range = [-0.01, 0.01] rad
    Empirically chosen in Sec. III-B; affects the rotational Hessian block and the resulting information matrix.
  • eigenvalue_clamp = 1e-3
    All eigenvalues of H below 1e-3 are clamped to 1e-3 during PSD projection; no sensitivity analysis is provided.
  • perturbation_count_M = unspecified
    The number of samples M in Eq. (7) is never given; results depend on it.
  • submap_split_distance_tau_sub = 25 m
    Displacement threshold that defines submap boundaries in Sec. III-C; affects loop-closure granularity.
  • local_map_split_distance_tau_loc = 100 m
    Displacement threshold that defines local maps in Sec. III-C; affects place-recognition context and graph sparsity.
  • Hausdorff_threshold_tau_H = 100 m
    Threshold for accepting submap/local-map trajectory proximity in Eqs. (15) and Section III-E; central to candidate selection.
  • inlier_threshold_tau_inliers = 5
    Minimum feature inlier count for a loop-candidate local-map pair in Sec. III-C.
  • fitness_threshold_tau_f = 0.25
    Minimum registration fitness score for accepting a loop closure into the pose graph in Sec. III-C.
  • voxel_resolutions_nu_sub_nu_loc = 0.5 m
    Voxel grid resolutions for submaps and local maps in Sec. III-C; affect registration accuracy and descriptor content.
assumptions (5)
  • domain assumption The ICP objective is locally well approximated by a quadratic with gradient and Hessian terms (Eq. 3), including under correspondence changes.
    Sec. III-B samples perturbations chosen to induce correspondence changes while assuming a smooth Taylor expansion; no proof of validity is given.
  • standard math The curvature of the ICP loss can be used directly as the information matrix of the odometry constraint.
    Standard nonlinear least-squares interpretation, but for ICP with recomputed correspondences it is an approximation; Sec. III-B.
  • domain assumption The appearance-based place-recognition front-end of Gupta et al. [14] reliably returns conservative loop candidates with approximate relative poses.
    The hierarchical loop-closure module depends entirely on [14], an author-group IJRR paper not summarized in this manuscript.
  • ad hoc to paper Lower nearest-neighbor RMS between local maps reconstructed from estimated poses at revisit locations measures map quality.
    This is the new evaluation protocol in Sec. IV-C; it is not validated against ground-truth or independently rated geometry.
  • ad hoc to paper After pose-graph optimization, the trajectories of missed revisits come within tau_H = 100 m Hausdorff distance, making geometric recovery possible.
    Needed by the retroactive loop-closure module in Sec. III-E; no evidence is given that this holds in all reported environments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Map Consistency in Graph-Based LiDAR SLAM Through Information-Aware Odometry and Retroactive Loop Closure." pith.science (2026). https://pith.science/paper/SKUOOTJR

@misc{pith2026260713516,
  author       = {Pith},
  title        = {Pith review of: Improving Map Consistency in Graph-Based LiDAR SLAM Through Information-Aware Odometry and Retroactive Loop Closure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SKUOOTJR}},
  note         = {Machine review of arXiv:2607.13516}
}
read the original abstract

High-quality maps are fundamental for robotics tasks such as navigation and planning. Although modern graph-based LiDAR SLAM systems achieve good trajectory accuracies, a low trajectory error alone does not guarantee geometrically consistent maps, particularly at revisit locations where missed loop closures and residual drift can produce local misalignments. In this work, we address the problem of jointly improving global trajectory estimation and local map quality in 3D LiDAR SLAM. We first propose a framework to efficiently estimate geometry-dependent information matrices for ICP, enabling principled weighting of odometry constraints in a pose graph. We then introduce a hierarchical loop-closure module that decouples place recognition from geometric registration, together with a retroactive loop-closure module that exploits the optimized pose graph to recover missed loop closures. We also propose an evaluation protocol to measure map consistency at revisit locations. We evaluate our SLAM system on several datasets against state-of-the-art LiDAR SLAM systems. Experimental results demonstrate global trajectory accuracies on par with or better than existing methods while consistently improving local geometric map consistency at revisit locations. These results suggest that coupling uncertainty-aware odometry with geometry-guided loop-closure refinement leads to more accurate trajectories and higher-quality maps.

Figures

Figures reproduced from arXiv: 2607.13516 by the authors.

Figure 1
Figure 1. Trajectory accuracy does not necessarily imply consistent map quality. KISS-SLAM [13], PIN-SLAM [28], and our approach achieve comparable absolute trajectory error (ATE) values, yet have substantial differences in local geometric consistency. Our method better preserves the alignment of repeated observations, reducing duplicated structures in the reconstructed map. The points in all images are colored based on their… view at source ↗
Figure 2
Figure 2. Overview of our LiDAR SLAM approach. The front-end computes LiDAR odometry and information matrix estimates. Registered scans are organized into a hierarchical map representation with large local maps for place recognition and smaller submaps for geometric validation. Odometry and loop-closure constraints are both optimized in a scan-level pose graph. The optimized trajectory is fed back to a retroactive loop-closur… view at source ↗
Figure 3
Figure 3. Map quality evaluation on the HeLiPR dataset averaged over the three recording instances in each scenario across all three LiDARs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Map quality evaluation on the MulRan dataset averaged over [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references

  1. [1]

    Behley and C

    J. Behley and C. Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. InProc. of Robotics: Science and Systems (RSS), 2018

  2. [2]

    Blanco-Claraco

    J.L. Blanco-Claraco. A flexible framework for accurate lidar odometry, map manipulation, and localization.Intl. Journal of Robotics Research (IJRR), 0(0):02783649251316881, 2025

  3. [3]

    Bonnabel, M

    S. Bonnabel, M. Barczyk, and F. Goulette. On the Covariance of ICP- Based Scan-Matching Techniques. InProc. of the IEEE American Control Conf. (ACC), pages 5498–5503, 2016

  4. [4]

    Bosse, P

    M. Bosse, P. Newman, J. Leonard, and S. Teller. Simultaneous Localization and Map Building in Large-Scale Cyclic Environments Using the Atlas Framework.Intl. Journal of Robotics Research (IJRR), 23(12):1113–1139, 2004

  5. [5]

    Brossard, S

    M. Brossard, S. Bonnabel, and A. Barrau. A New Approach to 3D ICP Covariance Estimation.IEEE Robotics and Automation Letters (RA-L), 5(2):744–751, 2020

  6. [6]

    A. Censi. An Accurate Closed-Form Estimate of ICP’s Covariance. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 3167–3172, 2007

  7. [7]

    X. Chen, A. Milioto, E. Palazzolo, P. Gigu `ere, J. Behley, and C. Stach- niss. SuMa++: Efficient LiDAR-based Semantic SLAM. InProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019

  8. [8]

    Dellenbach, J

    P. Dellenbach, J. Deschaud, B. Jacquet, and F. Goulette. CT-ICP Real- Time Elastic LiDAR Odometry with Loop Closure. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022

Show all 34 references
  1. [9]

    J. Deng, Q. Wu, X. Chen, S. Xia, Z. Sun, G. Liu, W. Yu, and L. Pei. NeRF-LOAM: Neural Implicit Representation for Large- Scale Incremental LiDAR Odometry and Mapping. InProc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2023

  2. [10]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets Robotics: The KITTI Dataset.Intl. Journal of Robotics Research (IJRR), 32(11):1231–1237, 2013

  3. [11]

    Grisetti, R

    G. Grisetti, R. K ¨ummerle, C. Stachniss, and W. Burgard. A tutorial on graph-based SLAM.IEEE Trans. on Intelligent Transportation Systems Magazine, 2(4):31–43, 2010

  4. [12]

    Grisetti, R

    G. Grisetti, R. K ¨ummerle, C. Stachniss, U. Frese, and C. Hertzberg. Hierarchical Optimization on Manifolds for Online 2D and 3D Map- ping. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2010

  5. [13]

    Guadagnino, B

    T. Guadagnino, B. Mersch, S. Gupta, I. Vizzo, G. Grisetti, and C. Stachniss. KISS-SLAM: A Simple, Robust, and Accurate 3D LiDAR SLAM System With Enhanced Generalization Capabilities. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5363–5370, 2025

  6. [14]

    Gupta, T

    S. Gupta, T. Guadagnino, B. Mersch, N. Trekel, M. Malladi, and C. Stachniss. Efficiently Closing Loops in LiDAR-Based SLAM Using Point Cloud Density Maps.Intl. Journal of Robotics Research (IJRR), 2026

  7. [15]

    Gupta, T

    S. Gupta, T. Guadagnino, B. Mersch, I. Vizzo, and C. Stachniss. Effectively Detecting Loop Closures using Point Cloud Density Maps. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2024

  8. [16]

    Huang, X

    X. Huang, X. Cheng, Q. Geng, B. Cao, D. Zhou, P. Wang, Y . Lin, and R. Yang. The ApolloScape Dataset for Autonomous Driving. InProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition Workshops, 2018

  9. [17]

    M. Jung, W. Yang, D. Lee, H. Gil, G. Kim, and A. Kim. HeLiPR: Heterogeneous LiDAR Dataset for inter-LiDAR Place Recognition under Spatiotemporal Variations.Intl. Journal of Robotics Research (IJRR), 12(43):1867—-1883, 2024

  10. [18]

    G. Kim, Y . Park, Y . Cho, J. Jeong, and A. Kim. Mulran: Multimodal range dataset for urban place recognition. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2020

  11. [19]

    Kim and A

    G. Kim and A. Kim. Scan Context: Egocentric Spatial Descriptor for Place Recognition within 3D Point Cloud Map. InProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2018

  12. [20]

    H. Kim, J. Choi, T. Sim, G. Kim, and Y . Cho. Narrowing Your FOV With SOLiD: Spatially Organized and Lightweight Global Descriptor for FOV-Constrained LiDAR Place Recognition.IEEE Robotics and Automation Letters (RA-L), 9(11):9645–9652, 2024

  13. [21]

    Landry, F

    D. Landry, F. Pomerleau, and P. Gigu `ere. CELLO-3D: Estimating the Covariance of ICP in the Real World. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 8190–8196, 2019

  14. [22]

    Lu and E

    F. Lu and E. Milios. Globally Consistent Range Scan Alignment for Environment Mapping.Autonomous Robots, 4:333–349, 1997

  15. [23]

    S. Ma, H. Zhang, and M. Toussaint. SVN-ICP: Uncertainty Estimation of ICP-Based LiDAR Odometry Using Stein Variational Newton. IEEE Robotics and Automation Letters (RA-L), 10(11):12181–12188, 2025

  16. [24]

    Magnusson, H

    M. Magnusson, H. Andreasson, A. Nuechter, Achim, and J. Lilienthal. Appearance-Based Loop Detection from 3D Laser Data Using the Normal Distributions Transform. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2009

  17. [25]

    Nieto, T

    J. Nieto, T. Bailey, and E. Nebot. Recursive Scan-Matching SLAM. Journal on Robotics and Autonomous Systems (RAS), 55(1):39–49, 2007

  18. [26]

    Olson and P

    E. Olson and P. Agarwal. Inference on Networks of Mixtures for Robust Robot Mapping.Intl. Journal of Robotics Research (IJRR), 32:826–840, 06 2013

  19. [27]

    Y . Pan, P. Xiao, Y . He, Z. Shao, and Z. Li. MULLS: Versatile LiDAR SLAM Via Multi-Metric Linear Least Square. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2021

  20. [28]

    Y . Pan, X. Zhong, L. Wiesmann, T. Posewsky, J. Behley, and C. Stach- niss. PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency.IEEE Trans. on Robotics (TRO), 40:4045–4064, 2024

  21. [29]

    Ramezani, Y

    M. Ramezani, Y . Wang, M. Camurri, D. Wisth, M. Mattamala, and M. Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. InProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020

  22. [30]

    S ¨underhauf and P

    N. S ¨underhauf and P. Protzel. Towards a Robust Back-End for Pose Graph SLAM. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2012

  23. [31]

    Vizzo, T

    I. Vizzo, T. Guadagnino, B. Mersch, L. Wiesmann, J. Behley, and C. Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way.IEEE Robotics and Automation Letters (RA-L), 8(2):1029–1036, 2023

  24. [32]

    N. Wang, H. Lu, Z. Zheng, Y .H. Liu, and X. Chen. Leveraging Semantic Graphs for Efficient and Robust LiDAR SLAM. InProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2025

  25. [33]

    C. Yuan, J. Lin, Z. Zou, X. Hong, and F. Zhang. STD: Stable Triangle Descriptor for 3D place recognition. InProc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2023

  26. [34]

    Zhang and S

    J. Zhang and S. Singh. LOAM: Lidar Odometry and Mapping in Real-time. InProc. of Robotics: Science and Systems (RSS), 2014

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.