Pith. sign in

REVIEW 5 major objections 5 minor 9 references

Robust 2D lidar-based SLAM in arboreal environments without IMU/GNSS

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A 2D lidar plus modified Hausdorff scan matching can localize and map a robot in orchards with no GNSS, IMU, or handcrafted features.

desk verdict Useful proof-of-concept for 2D-lidar-only orchard SLAM, but the abstract's superiority claim over A-LOAM is contradicted by its own CitrusFarm numbers. read the letter →

arxiv 2505.10847 v1 pith:5QCDOKO5 submitted 2025-05-16 cs.RO cs.SYeess.SY

classification cs.ROcs.SYeess.SY
keywords SLAM2DlidarscanmatchingmodifiedHausdorffdistanceorchardnavigationGNSS-deniedenvironmentsagriculturalroboticsExtendedKalmanFilter
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

The paper claims that a robot can localize and map itself in orchards and forest-like farmland using only a 2D lidar, with no GNSS, no IMU, and no hand-crafted features. The key is a scan-matching loop based on the modified Hausdorff distance, which aligns each lidar scan to an occupancy-grid map without explicit data association. In a controlled field test on a quadruped the method reached about 0.08 m mean positional error and 0.12 degree angular error, and on two of three real orchard datasets it beat A-LOAM on position and angle; on the third (CitrusFarm) A-LOAM had lower position error. If the claim holds, reliable autonomous navigation in GNSS-denied orchards becomes possible with cheaper, simpler sensors.

What carries the argument

The central mechanism is the modified Hausdorff distance (MHD) scan matcher. The MHD is a distance measure between the current lidar scan and the existing map's distance transform, computed with the k best matching coordinates and minimized over translation and rotation to yield the pose estimate. Because it operates on whole scans rather than extracted landmarks, it needs no explicit data association and tolerates outliers from foliage and repetitive tree trunks. The other load-bearing pieces are the horizontal slice of the 3D lidar at z in [0.0,0.2] m, the differential-drive motion model, the EKF for pose refinement, and the recursive Bayesian occupancy-grid update.

What would settle it

Run the same Unitree Go1 with the VLP-16 over a closed grass loop that includes a slope with pitch changes of more than a few degrees, with RTK ground truth; if positional RMSE rises well above the 0.10 m controlled-field value or the loop closure error grows beyond the reported 0.15 m, the planar-slice assumption, and with it the claimed GNSS-free accuracy, fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that accurate and reliable SLAM in complex outdoor agricultural environments does not require GNSS, IMU, or extracted features; a 2D lidar, a horizontal scan slice, and modified Hausdorff distance scan matching are sufficient. The method matches each scan to a growing probabilistic occupancy grid by minimizing the MHD over pose, then fuses the match result with a differential-drive motion model in an EKF and updates the map recursively. On a controlled 16 m loop with RTK ground truth the authors report mean positional error 0.08±0.05 m and mean angular error 0.12±0.14 degrees, versus 0.99 m and 0.83 degrees for A-LOAM. Across public orchard datasets the method was comparable or better than A-LOAM in most metrics, with the explicit exception of CitrusFarm where A-LOAM's richer feature environment gave it lower positional error; in the most degraded case, Pullally, the proposed method kept mean positional error at 0.35 m while A-LOAM exceeded 20 m.

Load-bearing premise

The load-bearing premise is planar motion: the robot is modeled as a differential drive on a flat plane and the lidar scan is a horizontal slice between 0 and 0.2 m height, so significant roll, pitch, or terrain slope can break the 2D scan-to-map correspondence.

Editorial extensions

If this is right

  • Orchard robots can navigate with a single 2D lidar plus wheel or leg odometry, reducing sensor cost and computation.
  • Because matching does not rely on extracted features, repeated trunks and trellises do not cause the association ambiguities that plague landmark-based SLAM in groves.
  • The reported sub-decimeter errors on a legged platform suggest that GNSS-free mapping can meet the precision needs of row-level agricultural operations on moderately uneven ground.
  • The occupancy grid built without GNSS provides a self-consistent map for downstream tasks such as spraying, pruning, and monitoring.

Reading between the lines

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

  • The MHD matcher's independence from extracted landmarks suggests the same objective could be applied to 3D scans or other repetitive, feature-poor environments such as tunnels, though the paper does not test this.
  • The paper notes that adding IMU should improve accuracy; a minimal extension would keep the MHD objective and replace the planar model with a 6-DOF state, testing whether the gain is worth the extra sensor.
  • The CitrusFarm exception indicates the method wins where geometry is sparse and repetitive; a graded test across feature densities would reveal where A-LOAM's feature dependence becomes an advantage.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. This paper presents a 2D-lidar SLAM pipeline for orchard and forest environments that does not use GNSS, IMU, or handcrafted feature extraction. The method projects a horizontal slice of a 3D lidar scan into a 2D polar scan, aligns scans to an occupancy grid by minimizing a modified Hausdorff distance (MHD) over k best matches, and fuses the scan-matching pose with a differential-drive motion model in an EKF. The authors validate the system in a controlled field experiment with a Unitree Go1 quadruped and GNSS-RTK ground truth, and on three orchard datasets (CitrusFarm, Bacchus, Pullally), comparing pose accuracy and loop-closure errors against A-LOAM. The paper claims that the proposed method achieves lower positional and angular errors than A-LOAM and that reliable SLAM is achievable in complex agricultural environments without GNSS, IMU, or handcrafted features.

Significance. The work is practically relevant: a low-cost 2D-lidar solution that avoids explicit data association is attractive for GNSS-denied orchard navigation, and the authors evaluate against external RTK ground truth on multiple public and self-collected datasets, which is a genuine strength. The controlled experiment shows sub-decimeter accuracy and documents local-consistency metrics. However, the central comparative claim is not supported as written: the CitrusFarm results contradict the abstract's blanket statement, and the A-LOAM baseline is run without IMU, which may unfairly handicap it. The planar-motion assumption is also acknowledged by the authors as a source of matching errors on uneven terrain. With accurate qualification of the claims and additional parameter and baseline documentation, the method's feasibility claim would be credible.

major comments (5)
  1. [Abstract and Section III-B.1 / Table III] The abstract's blanket claim that the proposed approach 'achieves lower positional and angular errors' than A-LOAM is contradicted by the paper's own CitrusFarm results in Table III, where the proposed method's mean positional error is 1.48 m versus 0.31 m for A-LOAM, RMS positional error is 1.83 m versus 0.36 m, and loop closure error is 2.56 m versus 0.10 m; Section III-B.1 itself concedes that A-LOAM outperformed. Because the central contribution is a comparative claim, this inconsistency is load-bearing; the abstract, Section III, and Conclusions must be revised to state the comparison accurately, for example by reporting that the proposed method wins on Bacchus and Pullally and loses on CitrusFarm.
  2. [Section III-A and Section III-B] The A-LOAM baseline is reported as running 'using only lidar data (i.e., no IMU)', which is a configuration outside A-LOAM's intended design since A-LOAM is a 3D lidar odometry algorithm that in its reference implementation can consume IMU data. This makes the comparative claim ambiguous: the large A-LOAM errors on Pullally (20.51 m mean positional error) may reflect a handicapped baseline rather than an intrinsic weakness of A-LOAM. The authors should report the exact launch configuration, justify why IMU was disabled for A-LOAM, and discuss how this choice affects the fairness of the comparison.
  3. [Section II-D and Conclusions] The method assumes planar motion: Section II-B uses a 2D differential-drive kinematic model and Section II-D extracts a fixed-height horizontal slice (z in [0.0,0.2] m), ignoring roll and pitch. The Conclusions acknowledge that 'vibrations transmitted to the sensors... as well as sudden changes in terrain slope increase the chances for matching errors.' On a quadruped walking on uneven grass, this assumption is not guaranteed to hold, so the general claim of robustness in arboreal environments needs a quantitative bound; please report the platform's pitch and roll ranges during the experiments and either restrict the robustness claim to near-planar conditions or add a mitigation such as IMU-based tilt compensation.
  4. [Section II-E and Section II-F] The MHD scan-matching module depends on several free parameters that are never specified in the paper, including the number k of best-matching points in the modified Hausdorff distance, the angular bin resolution in the 2D polar projection, the distance transform resolution, and the EKF noise matrices Q and R. Without these values, the reported results are not reproducible, and without a sensitivity analysis, the claimed robustness of MHD matching to outliers and repetitive geometry is not established. Please add the parameter values used for all experiments and a simple sensitivity study.
  5. [Section III-B and Conclusions] The conclusion that the method 'maintained high map accuracy and alignment across all tests' is not supported by the orchard experiments: Table III reports only pose errors (position, angular, local consistency, loop closure), while the map accuracy metrics in Table II (mean map error, precision, recall, F1) are reported only for the controlled field environment. If the claim covers the orchard datasets, map-accuracy metrics computed against the available RTK/GNSS ground truth should be provided for each dataset.
minor comments (5)
  1. [Section II-F, Eq. (6)] Equation (6) appears to have a typo: the Kalman gain is written as K_{t+1} = P_{t+1|t} C_t^T + S^{-1}_{t+1|t}, which is dimensionally inconsistent; it should be a multiplication. Also, the matrices A, C, G, H, Q, and R are not defined in the text.
  2. [Section II-G, Eqs. (9)-(11)] The recursive Bayesian formulation in Eqs. (9)-(11) is presented without a clear connection to the occupancy-grid update described in Section II-C; please clarify how these equations are used in the implementation.
  3. [Section III-A] The text says the trajectory is a '16 m closed-loop trajectory' and also a '6 m x 2 m rectangular path, repeated five times'; please clarify whether the total path length is 16 m or five times that length.
  4. [References / Section I] Because the MHD machinery is drawn from the authors' earlier work in [6], the paper should explicitly state what is new in the present contribution beyond applying [6] to orchard environments, in order to sharpen the novelty claim.
  5. [Throughout] Use a consistent name for the baseline (A-LOAM, not ALOAM) and fix typographical errors such as 'quadrupped' in Section I and 'Scena 3' in Fig. 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: MHD scan matching and EKF pose estimation are defined in-paper and validated against external RTK ground truth.

full rationale

The paper's derivation chain is self-contained rather than circular. Section II-E defines the scan-matching objective directly: q^* = arg min h_k(M, M_t(q, Z^w_t)), with the MHD metric described in the same section, so the core matching criterion does not depend on a fitted parameter or on an imported uniqueness theorem. Section II-F fuses the scan-match result with a differential-drive motion model through an EKF, and Sections II-C and II-G specify the occupancy-grid and Bayesian map-update rules. Performance is then evaluated against external centimeter-accurate GNSS-RTK ground truth (Section III-A) and against public orchard datasets (Section III-B), so the reported pose errors are not constructed from the method's own outputs. The authors cite their prior work [6] as the origin of the MHD-based localization idea, but the present paper re-derives the MHD formulation and does not rely on [6] for any mathematical premise; thus the self-citation is not load-bearing. No fitted input is renamed as a prediction, no uniqueness theorem from the authors is invoked to force the chosen approach, and the MHD technique is not merely a renamed known result presented as an independent derivation. The paper does contain an internal inconsistency: the Abstract claims lower positional and angular errors than A-LOAM, while Table III shows A-LOAM with substantially lower positional error on CitrusFarm (0.31 m vs. 1.48 m mean, 0.36 m vs. 1.83 m RMS), and Section III-B.1 concedes this. That inconsistency undermines the strength of the comparative claim, but it is a correctness/interpretation issue rather than a circularity: the measurements still come from external benchmarks. Therefore, no circular step is exhibited, and the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities. It depends on a set of hand-chosen parameters (k, slice height, bin width, noise covariances, grid resolution) whose values are not reported, and on strong planar-motion and static-map assumptions. The MHD scan-matching machinery itself is imported from the authors' prior work [6], so the ledger is dominated by imported technique and unstated tuning rather than new derivation.

free parameters (5)
  • k (best-matching count in MHD) = Not specified
    Eq. (1) uses 'k best matching object coordinates' in the modified Hausdorff distance; k controls outlier rejection and affects all pose estimates, but no value or selection rule is given.
  • Angular bin resolution for 2D scan projection = Not specified
    Section II-D says the 3D slice is 'discretized into angular bins' but the bin width is not stated; it determines the effective scan resolution.
  • Lidar slice height range = 0.0 to 0.2 m
    Section II-D extracts this fixed-height slice from the 3D point cloud; it is a hand-chosen threshold that the method's performance depends on.
  • EKF process and measurement noise (Q, R) = Not specified
    Section II-F uses Q and R in the covariance update but gives no values or tuning procedure; EKF pose output depends directly on these.
  • Occupancy grid cell size and distance transform resolution = Not specified
    The map is a probabilistic occupancy grid (Section II-C) and the MHD cost uses a distance transform of it (Section II-E); neither resolution is given, yet they set the scale of the map error metric in pixels.
assumptions (6)
  • standard math The distance transform and modified Hausdorff distance define a cost whose local minimum is the correct pose.
    The scan matching solves argmin over pose of h_k (Eq. 1); the paper does not prove convexity or smoothness, so the solver is assumed to find the right minimum.
  • domain assumption A fixed horizontal slice of the 3D lidar scan corresponds to a consistent world plane.
    Section II-D extracts z in [0.0,0.2] m; on uneven terrain or when the robot pitches or rolls, this slice plane is not fixed in the world.
  • domain assumption The robot's motion is described by a 2D differential-drive kinematic model.
    Section II-B uses v_t cos(theta), v_t sin(theta), omega; the Unitree Go1 is a quadruped, and no justification is given for ignoring roll and pitch and leg dynamics.
  • domain assumption Foliage and dynamic objects can be treated as outliers in the MHD matching.
    The k-best matching is intended to reject outliers (Section II-E), but the paper does not quantify how much foliage motion is tolerated.
  • domain assumption The initial pose x0 is known.
    Section II-B sets x0 = x_init; the method does not address global localization or kidnap recovery.
  • ad hoc to paper The scan-matching optimization converges from the previous pose estimate.
    No optimization method, search bounds, or initialization strategy is given for Eq. (1); the paper assumes the EKF prediction is close enough.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust 2D lidar-based SLAM in arboreal environments without IMU/GNSS." pith.science (2026). https://pith.science/paper/5QCDOKO5

@misc{pith2026250510847,
  author       = {Pith},
  title        = {Pith review of: Robust 2D lidar-based SLAM in arboreal environments without IMU/GNSS},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5QCDOKO5}},
  note         = {Machine review of arXiv:2505.10847}
}
read the original abstract

Simultaneous localization and mapping (SLAM) approaches for mobile robots remains challenging in forest or arboreal fruit farming environments, where tree canopies obstruct Global Navigation Satellite Systems (GNSS) signals. Unlike indoor settings, these agricultural environments possess additional challenges due to outdoor variables such as foliage motion and illumination variability. This paper proposes a solution based on 2D lidar measurements, which requires less processing and storage, and is more cost-effective, than approaches that employ 3D lidars. Utilizing the modified Hausdorff distance (MHD) metric, the method can solve the scan matching robustly and with high accuracy without needing sophisticated feature extraction. The method's robustness was validated using public datasets and considering various metrics, facilitating meaningful comparisons for future research. Comparative evaluations against state-of-the-art algorithms, particularly A-LOAM, show that the proposed approach achieves lower positional and angular errors while maintaining higher accuracy and resilience in GNSS-denied settings. This work contributes to the advancement of precision agriculture by enabling reliable and autonomous navigation in challenging outdoor environments.

Figures

Figures reproduced from arXiv: 2505.10847 by the authors.

Figure 1
Figure 1. Fruit orchard with dense canopies: RGB camera view [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Data filtering and scan matching with MHD. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The mobile robot’s pose qt and lidar measurements zt,k in an agricultural environment. at time t, shown in [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Slicing of 3D lidar measurements and 2D projection onto “ground” plane. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Online SLAM navigation based on scan matching. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Pullally fruit orchard and the robot trajectory (red line) [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: CitrusFarm dataset results obtained with the proposed [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 7 canonical work pages

  1. [6]

    Torres-Torriti , P

    M. Torres-Torriti , P. Nazate-Burgos , F. Paredes-Lizama , J. Guevara, and F. Auat Cheein, ``Passive landmark geometry optimization and evaluation for reliable autonomous navigation in mining tunnels using 2D lidars,'' Sensors, vol. 22, no. 8, 2022. [Online]. Available: https://www.mdpi.com/1424-8220/22/8/3038

  2. [1]

    A. S. Aguiar, F. Neves dos Santos, H. Sobreira, J. Boaventura-Cunha, and A. J. Sousa, ``Localization and mapping on agriculture based on point-feature extraction and semiplanes segmentation from 3D LiDAR data,'' Frontiers in Robotics and AI, vol. 9, 2022. [Online]. Available: https://www.frontiersin.org/articles/10.3389/frobt.2022.832165

  3. [2]

    Debeunne and D

    C. Debeunne and D. Vivet, ``A review of visual-lidar fusion based simultaneous localization and mapping,'' Sensors, vol. 20, no. 7, 2020. [Online]. Available: https://www.mdpi.com/1424-8220/20/7/2068

  4. [3]

    X. Liu, G. V. Nardari, F. C. Ojeda, Y. Tao, A. Zhou, T. Donnelly, C. Qu, S. W. Chen, R. A. F. Romero, C. J. Taylor, and V. Kumar, ``Large-scale autonomous flight with real-time semantic SLAM under dense forest canopy,'' IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 5512--5519, April 2022

  5. [4]

    Cremona, R

    J. Cremona, R. Comelli, and T. Pire, ``Experimental evaluation of visual-inertial odometry systems for arable farming,'' Journal of Field Robotics, vol. 39, no. 7, pp. 1121--1135, 2022. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.22099

  6. [5]

    Auat Cheein , G

    F. Auat Cheein , G. Steiner, G. Perez Paina , and R. Carelli, ``Optimized EIF - SLAM algorithm for precision agriculture mapping based on stems detection,'' Computers and Electronics in Agriculture, vol. 78, no. 2, pp. 195--207, 2011. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0168169911001542

  7. [7]

    Torres-Torriti and P

    M. Torres-Torriti and P. Nazate-Burgos, SLAM in Agriculture. 1em plus 0.5em minus 0.4em Cham: Springer International Publishing, 2022, pp. 1--22

  8. [8]

    write newline

    " write newline "" initialize.prev.this.status FUNCTION begin.bib " write newline preamble empty 'skip preamble write newline if " thebibliography " longest.label * " " * write newline " [1] #1 " write newline " url@samestyle " write newline " " write newline " [2] #2 " write newline " =0pt " write newline " " ALTinterwordstretchfactor * " " * write newli...

Show all 9 references
  1. [9]

    86 ,bK ?a˅MϮc8jD3 \/a/86ǰo 5ŴFdm( & TX܄iF J#ARH GRfoa aq nUEj, ;7 M@t>r!mapR#5 <f

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.