{"id":"777ea026-32ca-4788-92aa-752ed079676c","arxiv_id":"2505.10847","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modified-Hausdorff-distance 2D lidar scan matcher fused with an EKF achieves reasonably accurate SLAM in orchards without IMU or GNSS, but results against A-LOAM are mixed.","lead":"The paper presents a 2D lidar SLAM method that matches scans with a modified Hausdorff distance, needing no IMU, GNSS, or feature extraction. It reports reliable pose estimates in orchards and on a legged robot, though a standard 3D SLAM baseline outperforms it on one dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's blanket 'lower positional and angular errors' claim is contradicted by the paper's own CitrusFarm results in Table III.","rationale":"The reader's weakest_assumption was the planarity/roll-pitch limitation, which is a genuine concern but is explicitly acknowledged in the Conclusions. My stress-test identifies a more direct and decisive problem: the paper's own Table III contradicts the abstract's comparative claim. The proposed method loses to A-LOAM on CitrusFarm by a large margin in positional error and loop closure, and the paper admits this in Section III-B.1. Yet the Abstract states without qualification that the proposed approach 'achieves lower positional and angular errors' than A-LOAM, and the Conclusions repeat that the method 'demonstrates that accurate and reliable SLAM is achievable' without the caveat that one of three real-orchard datasets shows the opposite. Since the headline contribution is framed comparatively, this internal inconsistency is load-bearing. The fix is not necessarily to reject the method; the evidence may still support a qualified feasibility claim for 2D MHD scan matching in specific orchard settings. But the abstract and conclusions must be revised to report per-dataset results honestly, and the A-LOAM baseline should be run in its intended configuration (with IMU) as well as the IMU-free configuration to disentangle the effect of the handicap. This does not alter the reader's conditional verdict: the paper still needs correction and more complete reporting before its central claim can be accepted as stated.","tokens_in":7523,"tokens_out":4151,"duration_ms":42629,"concrete_test":"Reproduce Table III on the CitrusFarm dataset: run the proposed method and A-LOAM on the same VLP-16 bag with the same RTK ground truth, with A-LOAM in both its standard IMU-fused configuration and the IMU-free configuration described in the paper. If the proposed method's mean/RMS positional errors remain about 1.5/1.8 m while A-LOAM's are about 0.3/0.4 m, the abstract's blanket 'lower positional errors' claim is refuted and must be qualified; if the numbers in Table III are transcription errors, the corrected comparison should be reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central comparative claim — that the proposed method 'achieves lower positional and angular errors' than A-LOAM (Abstract) — is internally contradicted by Table III. On CitrusFarm, the proposed method has mean positional error 1.48 m vs 0.31 m for A-LOAM, RMS positional error 1.83 m vs 0.36 m, and loop-closure error 2.56 m vs 0.10 m. Section III-B.1 itself concedes that 'A-LOAM outperformed due to its reliance on rich geometric features.' The Abstract and Conclusions present the comparative result as uniform, but the data show the claimed superiority holds only on Bacchus and Pullally. Moreover, the A-LOAM baseline is reportedly run without IMU/lidar stabilization, a configuration outside its intended design, so 'lower errors than A-LOAM' may reflect a handicapped baseline rather than an intrinsic advantage. Because the central claim is comparative, this inconsistency is load-bearing: the stated conclusion cannot be accepted as written, and the evidence supports at most the weaker claim that 2D MHD scan matching can provide useful localization in some orchard conditions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7771,"tokens_out":6967,"duration_ms":64896,"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":[{"comment":"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.","section":"Abstract and Section III-B.1 / Table III"},{"comment":"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.","section":"Section III-A and Section III-B"},{"comment":"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.","section":"Section II-D and Conclusions"},{"comment":"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.","section":"Section II-E and Section II-F"},{"comment":"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.","section":"Section III-B and Conclusions"}],"minor_comments":[{"comment":"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.","section":"Section II-F, Eq. (6)"},{"comment":"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.","section":"Section II-G, Eqs. (9)-(11)"},{"comment":"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.","section":"Section III-A"},{"comment":"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.","section":"References / Section I"},{"comment":"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.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main concern is overclaiming: the abstract's comparative statement needs to be corrected before publication, and the A-LOAM baseline configuration needs to be documented carefully. The CitrusFarm contradiction and the acknowledged sensitivity to terrain slope are the key load-bearing issues. I would not reject: the external RTK ground truth, public-dataset validation, and the controlled-field results are useful, and the manuscript can be made accurate with a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid workshop-level demonstration that a 2D lidar plus MHD scan matching can keep a robot localized in orchard rows without IMU or GNSS, backed by real RTK ground truth on multiple datasets. The central comparative claim, though, is oversold: the abstract says the method achieves 'lower positional and angular errors' than A-LOAM, but Table III shows A-LOAM beating it badly on CitrusFarm (0.31 m vs 1.48 m mean position error). The authors themselves admit A-LOAM wins there. So the honest conclusion is the weaker one: the method is competitive in some arboreal settings, not uniformly better.\n\nWhat's genuinely new is the integration: slicing a 3D lidar to a 2D scan, matching via MHD with a distance transform, and fusing through an EKF, all without IMU. The prior work [6] was in mining tunnels, so transferring to orchards with foliage and uneven terrain is a real extension. The experiments cover a controlled field test, three real orchards, and a Unitree Go1 quadruped, with RTK ground truth. The errors on Bacchus and Pullally are respectable (0.84 m and 0.35 m RMSE), and the method degrades gracefully where A-LOAM fails without IMU stabilization. That is useful evidence for precision-ag applications.\n\nThe soft spots are real but not fatal for a workshop paper. The method description is too thin to reproduce: the MHD parameter k, angular bin resolution, slice height, EKF noise matrices, and grid resolution are not specified. No code or data are released, which limits the claim to 'reported to work.' The A-LOAM baseline is run without IMU, outside its intended design; that is a fair test of sensor-minimal operation, but it means 'beats A-LOAM' is really 'beats a handicapped A-LOAM.' The planar assumption is a genuine limitation, especially for the quadruped on slopes; the authors acknowledge it in the conclusions. None of these sink the paper, but they cap the strength of its claims.\n\nWho this is for: people building low-cost autonomous navigation for orchards and forests, especially on legged platforms. A serious referee should engage with it, mostly to push for specification and data release. I would not desk reject it, but I would require revisions to tone down the comparative claim and disclose parameters.\n\nRecommendation: conditional accept, with required revisions to align the abstract with the actual results and to release enough detail for reproducibility.","headline":"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.","tokens_in":8340,"tokens_out":1860,"would_cite":false,"duration_ms":18162,"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":"A 2D lidar plus modified Hausdorff scan matching can localize and map a robot in orchards with no GNSS, IMU, or handcrafted features.","keywords":["SLAM","2D lidar","scan matching","modified Hausdorff distance","orchard navigation","GNSS-denied environments","agricultural robotics","Extended Kalman Filter"],"falsifier":"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.","tokens_in":7336,"feed_emoji":"🌳","tokens_out":6933,"duration_ms":66554,"temperature":0.7,"pith_summary":"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.","feed_headline":"Orchard SLAM works with 2D lidar alone","feed_subtitle":"Modified Hausdorff scan matching beats A-LOAM on most orchard runs, enabling GNSS-free navigation.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the earlier landmark-based SLAM in tree groves that this method's whole-scan matching replaces.","marker":"[5]"},{"why":"Provides the authors' earlier 2D lidar robust localization work that the proposed scan-matching approach builds on.","marker":"[6]"},{"why":"Represents the feature-based 3D-lidar alternative the method seeks to outperform in cost and simplicity.","marker":"[1]"},{"why":"Maps the visual-lidar sensor-fusion landscape that the GNSS-free, IMU-free method deliberately avoids.","marker":"[2]"},{"why":"Documents the IMU-dependent odometry baseline that the paper claims to replace in GNSS-denied fields.","marker":"[4]"}],"fun_headline_variants":["2D lidar SLAM for orchards, no IMU or GNSS needed","Robust orchard SLAM with just a 2D lidar","Forest SLAM without IMU/GNSS using 2D lidar","Cheap 2D lidar beats A-LOAM in GNSS-free orchards","Scan-matching SLAM that works where GPS fails"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["2D lidar SLAM for orchards, no IMU or GNSS needed","Robust orchard SLAM with just a 2D lidar","Forest SLAM without IMU/GNSS using 2D lidar","Cheap 2D lidar beats A-LOAM in GNSS-free orchards","Scan-matching SLAM that works where GPS fails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000786,"raw_usage":{"total_tokens":3474,"prompt_tokens":956,"completion_tokens":2518,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2419}},"tokens_in":572,"tokens_out":2518,"duration_ms":17167,"temperature":1.0,"reasoning_tokens":2419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:02:12.709576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Auat Cheein , G","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier landmark-based SLAM in tree groves that this method's whole-scan matching replaces."},{"cited_title":"Torres-Torriti , P","cited_arxiv_id":null,"evidence_quote":"Provides the authors' earlier 2D lidar robust localization work that the proposed scan-matching approach builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the feature-based 3D-lidar alternative the method seeks to outperform in cost and simplicity."},{"cited_title":"Debeunne and D","cited_arxiv_id":null,"evidence_quote":"Maps the visual-lidar sensor-fusion landscape that the GNSS-free, IMU-free method deliberately avoids."},{"cited_title":"Cremona, R","cited_arxiv_id":null,"evidence_quote":"Documents the IMU-dependent odometry baseline that the paper claims to replace in GNSS-denied fields."}],"review_version":1}