{"id":"2800e579-c47b-4b78-bc22-2cd2ff19ae86","arxiv_id":"2505.08388","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN-based object detection module is added to a 2D LiDAR/IMU SLAM system, with claimed reductions in indoor localization error, but the supporting evidence is inconsistent and incomplete.","lead":"This paper describes a handheld indoor localization system that combines a 2D LiDAR, an IMU, and a CNN object detector inside a SLAM pipeline, and reports lower trajectory error than two LiDAR SLAM baselines. The reported accuracy numbers are internally inconsistent, and the CNN component is described without a training dataset or code, so the central performance claims are not verified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported ATE values are mutually inconsistent: Table II's 0.25 m cannot coexist with Table V's mean 1.97 m and RMSE 2.00 m, so the claimed 21%–26% improvements rest on contradictory metrics.","rationale":"The Reader's verdict is REJECT, and my analysis supports keeping that verdict. I agree with the Reader's broader conclusion that the central accuracy claim is internally inconsistent; the same system is reported with ATE 0.25 m, mean errors near 2 m, and RMSE values up to 3.5 m. However, the Reader's stated weakest_assumption is the unreliability/unreproducibility of the CNN object-detection module, whereas the most load-bearing problem in my reading is the direct arithmetical contradiction between Table II and Table V. This is a stronger and more fundamental objection: it does not depend on external dataset availability or training details, and it invalidates the headline improvement percentages even if every CNN component worked exactly as described. The concrete test I propose is therefore a re-derivation of ATE from the paper's own reported error statistics, which would settle whether the contradiction is real or whether the tables refer to different error definitions that were not disclosed. Since the Reader's final verdict is already REJECT and my concern reinforces it, I recommend no change to the verdict.","tokens_in":14946,"tokens_out":5081,"duration_ms":44769,"concrete_test":"Recompute ATE from the per-pose error vectors summarized in Table V: apply the same trajectory alignment and RMSE formula used for Table II, and check whether the result equals 0.25 m or approximately 1.97 m for Round 1. If ATE ≈ 0.25 m, then Table V's mean/RMSE/max values must refer to a different error definition or a different run, and the paper must disclose which. If ATE ≈ 1.97 m, then Table II's headline value is wrong and the claimed 21.03% and 26.09% improvements do not follow from the reported data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that CNN-LiDAR-SLAM achieves ATE 0.25 m and outperforms SC-ALOAM by 21.03%/26.09%—is undercut by an internal metric contradiction. Table II reports ATE = 0.25 m for the proposed system. Table V, labeled 'Statistical Localization Performance,' reports for the same Round 1 an ATE distribution with mean 1.97 m, median 1.87 m, RMSE 2.00 m, and maximum 2.974 m; Round 2 gives mean 2.67 m and RMSE 3.50 m. Under the standard definition (Umeyama-aligned absolute trajectory error, as stated in Section IV-B), ATE is an RMSE over per-pose errors, so it cannot be 0.25 m when the RMSE for the same run is 2.00 m. Moreover, Table IV's RMSE values of 2.00 m and 2.67 m match Table V's ATE rows, suggesting that Table V is reporting trajectory RMSE under another name. The Abstract's mean x-position error of -0.884 m (±1.976 m) and Section IV-B's 'mean localization error of 1.02 m' introduce yet another incompatible scale. Because every improvement percentage derives from these numbers, the central claim is not merely unproven; as stated, it is arithmetically inconsistent. The CNN-reproducibility concern raised by the Reader is secondary: even if the CNN were perfectly reliable, the reported outcome metrics do not establish the claimed improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CNN-LiDAR-SLAM, a handheld indoor localization system combining a 2D LiDAR, a BNO055 IMU, and a CNN-based object detection module that predicts up to four object centroids and radii from a 26-dimensional feature vector. These predicted landmarks are fed into an IEKF-based pose optimization pipeline built on Cartographer and FAST-LIO2. The authors report that the system achieves ATE values of 0.25 m and 0.3 m in two indoor corridors without loop closure, runs in real time at 15 FPS on a standard laptop, reduces ATE by 21.03% with respect to SC-ALOAM, and improves localization accuracy by 26.09% over a baseline method. The experiments were conducted in two university corridors with ground truth obtained by manual tape-measure tracing of the handheld path.","tokens_in":15342,"tokens_out":5158,"duration_ms":42979,"significance":"The motivating problem is practically relevant: a low-cost 2D LiDAR and IMU setup augmented with learned landmarks could, if substantiated, offer a cheaper and faster alternative to 3D LiDAR SLAM in structured indoor environments, with clear applications in emergency response and industrial automation. The paper also makes a positive contribution by candidly listing limitations in Section V, including the absence of dynamic obstacles, loop closure, and environmental diversity. However, the manuscript does not ship reproducible code or datasets, provides no training protocol for the CNN, reports no ablation isolating the CNN's contribution, and contains internally contradictory accuracy metrics. Because the central quantitative claims rest on these unsupported and inconsistent numbers, the significance of the contribution as demonstrated is low, even though the underlying idea has potential.","major_comments":[{"comment":"The reported ATE values are mutually inconsistent. Table II lists ATE = 0.25 m for CNN-LiDAR-SLAM, and Section IV-B claims ATE values of 0.25 m and 0.3 m for the two corridors. Table V, however, reports for Round 1 an ATE with mean 1.97 m, RMSE 2.00 m, and maximum 2.974 m, and for Round 2 a mean of 2.67 m and RMSE 3.50 m. Since ATE is defined as an RMSE over per-pose errors (as the APE/Umeyama description in Section IV-B confirms), a single run cannot have ATE 0.25 m and RMSE 2.00 m simultaneously. Moreover, the numbers in Table V exactly match the RMSE rows in Table IV (2.00, 2.67, 1.97), suggesting that Table V's 'ATE' is actually the trajectory RMSE. Every improvement percentage in the abstract and conclusions (21.03%, 26.09%) is derived from these inconsistent figures, so the central quantitative claim is not supported.","section":"§IV-B, Tables II, IV, V"},{"comment":"The CNN-based object detection module, which is the paper's core novelty, is not reproducible and its contribution is not isolated. No training dataset, label source, loss function, hyperparameters, or train/test split are given; Table VII reports precision 92.5%, recall 91.3%, and F1 0.919 without stating the evaluation set, the ground-truth labeling procedure, or the matching criterion. Because the predicted landmarks from this CNN are fed into pose optimization via Eq. (23), the claimed 26.09% improvement in Table VI depends entirely on the reliability of these predictions. Without a documented training protocol, an external benchmark, or an ablation of the SLAM system with and without the CNN landmarks, the direction and magnitude of the effect of object detection on localization are unestablished.","section":"§III-C, Eq. (23), Table VII"},{"comment":"The baseline comparison is inadequately described and likely unfair. A-LOAM and SC-ALOAM are 3D LiDAR odometry methods; the authors say they were 'adapted to process the 2D LiDAR dataset by simulating pseudo-3D input', but no details of this simulation are given. In addition, Table IV lists sequence numbers 05, 06, 07, and 09, which are not explained in relation to the two corridor runs described in Table III (100 m and 90 m); the reported RMSE values (2.00, 2.67, 1.97, 2.50 m) are also not reconciled with the '0.25 m and 0.3 m' ATE claims. As presented, the comparison does not provide a valid basis for the claimed 21.03% ATE reduction.","section":"§IV-A, Table IV"},{"comment":"The manual tape-measure tracing of the handheld path cannot provide the per-pose ground truth required for ATE and APE computation. ATE and APE are defined relative to a time-synchronized pose trajectory; tracing a path on a floor plan with a tape measure yields at best a coarse path-level reference, with no heading or timestamp alignment. The sensitivity of the reported sub-meter ATE values to this ground-truth uncertainty is not analyzed, so the precision implied by '0.25 m' is not credible.","section":"§IV-A, ground truth"}],"minor_comments":[{"comment":"Sections III-B and III-D are both titled 'Data Preprocessing'; the second should be retitled (e.g., 'Object Detection Data Preprocessing') to avoid confusion.","section":"Section III"},{"comment":"The abbreviation 'IKEF' is presumably a typo for 'IEKF'; please correct it throughout.","section":"Section III-B"},{"comment":"Equation (24), 'T = T', is tautological and does not define an iterative update; please replace it with the actual update rule used in the optimization.","section":"Eq. (24)"},{"comment":"The abstract reports a 'mean x-position error of -0.884 meters (1.976 meters)' while Section IV-B reports a 'mean localization error of 1.02 m'; the relationship between these quantities is never defined.","section":"Abstract and Section IV-B"},{"comment":"The caption refers to 'baseline [14]', but the text compares against [28] (Mochurad et al.); the reference is inconsistent.","section":"Fig. 3 caption"},{"comment":"The sentence 'we have introduced a confusion matrix VII' is followed by no confusion matrix; either add the matrix or remove the sentence.","section":"Table VII"},{"comment":"Several references have formatting or citation inconsistencies (e.g., [15] is listed as LOAM but the reference entry appears under a different title in [5]); please standardize the bibliography.","section":"Table I and references"},{"comment":"The text describing Fig. 8 appears to swap the captions for parts (a) and (b) relative to the discussion of Round 1 and Round 2; please verify the mapping.","section":"Fig. 8"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a very early draft; the number of internal inconsistencies in the reported metrics suggests the results have not been carefully checked. I recommend rejection, but if the authors can supply corrected and consistent metrics, full training details for the CNN, and a fair baseline comparison, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know that the central accuracy claim of this paper is not merely unproven; it is arithmetically inconsistent. Table II reports ATE 0.25 m for the proposed system, while Table V reports for the same runs ATE mean 1.97 m and 2.67 m with RMSE 2.00 m and 3.50 m. Since ATE is an RMSE over per-pose errors, these numbers cannot coexist. The abstract says mean x-position error -0.884 ± 1.976 m, and Section IV-B claims mean localization error 1.02 m. The improvement percentages (21.03%, 26.09%) are derived from this incoherent set, so they do not establish anything.\n\nWhat is actually new: the specific integration of a CNN regressor that outputs up to four object centroids and radii from a 26-dimensional feature vector and feeds them into FAST-LIO2/Cartographer as landmarks via Eq. 23. Each component is known, so this is an incremental extension, but it is a legitimate one. The paper is also honest about limitations (no loop closure, limited environments, 2D LiDAR weaknesses) in the Discussion.\n\nThe soft spots are substantial. The CNN training is undocumented: no dataset, no label source, no loss, no ablation. The landmark term in Eq. 23 just aligns consecutive CNN predictions; it is not tied to an external or global map, so it could be fitting noise. The baseline comparison is vague: A-LOAM and SC-ALOAM are adapted to 2D LiDAR via 'pseudo-3D input', which is underspecified. Ground truth is a manual tape-measure trace, so sub-meter claims are optimistic. There are also mismatched table labels: Table IV's RMSE values look like Table V's ATE rows, suggesting the authors confused trajectory RMSE with ATE.\n\nDoes the paper deserve a serious referee? The contradiction is load-bearing, not a minor typo. A referee would need the authors to re-run experiments and report consistent metrics, provide the training details, and add an ablation. That is a major rewrite, not a revision. I would not send it to review as is.\n\nMy recommendation: desk reject, with an invitation to resubmit if they can provide a coherent evaluation. The system concept is worth exploring, but the current manuscript does not support its conclusions.","headline":"The paper's central ATE numbers are mutually contradictory, so its improvement claims collapse despite a plausible system integration.","tokens_in":15855,"tokens_out":3261,"would_cite":false,"duration_ms":28578,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CNN-LiDAR-SLAM fuses 2D LiDAR, IMU, and CNN-predicted object landmarks to cut Absolute Trajectory Error by 21.03% over SC-ALOAM and reach 0.25–0.3 m accuracy in corridors without loop closure.","keywords":["2D LiDAR SLAM","indoor localization","CNN object detection","IMU sensor fusion","Absolute Trajectory Error","object landmarks","real-time navigation","pose graph optimization"],"falsifier":"Rerun the same two corridor sequences with the CNN landmark term switched off and all other modules identical: if ATE does not rise by roughly the claimed 21–26%, or if disabling object detection improves accuracy, the central claim is wrong. Independently, re-train the CNN on a public point-cloud dataset and check its predicted centroids against a laser-scanned ground-truth map of a held-out corridor.","tokens_in":14767,"feed_emoji":"📍","tokens_out":17296,"duration_ms":146396,"temperature":0.7,"pith_summary":"CNN-LiDAR-SLAM is a handheld indoor localization system that fuses a low-cost 2D LiDAR and an inertial measurement unit with CNN-based object detection. The paper's central claim is that feeding CNN-predicted object centroids and radii into the SLAM backend as landmarks reduces Absolute Trajectory Error by 21.03% over SC-ALOAM and by 26.09% over an existing obstacle-detection baseline, reaching ATE values of 0.25 m and 0.3 m in two 90-meter and 100-meter corridors without loop closure. The system also runs in real time at about 15 frames per second with CPU utilization below 40%. If these numbers hold, they would show that inexpensive 2D LiDAR combined with learned semantic landmarks can provide accuracy close to or better than costlier 3D LiDAR SLAM in structured indoor environments.","feed_headline":"CNN-LiDAR-SLAM reaches 0.25-m indoor accuracy","feed_subtitle":"A handheld 2D LiDAR and IMU rig with learned object landmarks matches costlier 3D SLAM in corridors.","key_machinery":"The load-bearing mechanism is the CNN landmark predictor and its coupling equation. The CNN takes a 26-dimensional feature vector (LiDAR scan statistics, preliminary object geometry, and IMU acceleration, gyroscope, and quaternion data), processes it through three 1D convolutional layers with batch normalization, a bidirectional long short-term memory (BiLSTM), and an attention-weighted temporal pooling, and outputs twelve numbers: predicted center coordinates and radii for up to four objects. Those predicted centroids enter the pose correction step of Eq. 23, which minimizes the distance between landmark predictions in consecutive frames to estimate the transformation $\\hat{T}$; this correction is iterated together with the iterated extended Kalman filter state update in Eqs. 15–16 and a final pose-graph optimization in Eqs. 25–27.","core_discovery":"The discovery the authors are trying to establish is that object landmarks extracted by a CNN from fused LiDAR and IMU signals can serve as stable geometric anchors for pose optimization in 2D LiDAR SLAM, reducing drift enough that sub-0.3 meter trajectory accuracy is possible in long indoor corridors even when loop closure is unavailable. The system couples a tightly coupled iterated extended Kalman filter odometry with a base 2D SLAM backend, and the CNN's predicted landmark centers are aligned across frames by minimizing the residual in Eq. 23, which the authors treat as a correction that prevents landmark drift. Relative to SC-ALOAM they report a 21.03% ATE reduction, and relative to the obstacle-detection baseline [28] they report a 26.09% improvement in mean localization error, with smaller CPU and memory footprints than the compared methods.","pith_inferences":["If the CNN landmark correction is genuinely load-bearing, the same 26-dimensional input recipe should transfer to other 2D LiDAR platforms and buildings with retrained detection heads; that is a direct test of the claim.","The predicted object centroids could be reused as loop-closure candidates in longer deployments, potentially extending the no-loop-closure accuracy shown here to larger floor plans.","Because the system predicts only up to four landmarks per frame, an immediate extension is to make the predictor output a variable number of objects per frame, letting the method exploit densely furnished rooms.","A public benchmark with the same ground-truth measurement protocol would let other groups reproduce the reported 0.25–0.3 m ATE values and the 21.03% reduction."],"forward_implications":["A 2D LiDAR plus IMU SLAM system with learned landmarks can achieve roughly 0.25–0.3 m trajectory error over 90–100 m corridors without loop closure, a regime where 2D LiDAR SLAM is often considered drift-prone.","The reported 15 FPS at under 40% CPU utilization implies the approach is deployable on a standard laptop or embedded computer for emergency response and mobile robotics.","If the object-landmark correction is genuinely responsible for the gain, then semantic object predictions can serve as a substitute for geometric loop closure in repetitive indoor structures.","The authors note that accuracy would degrade in featureless or open spaces and over longer trajectories without loop closure, so the claimed advantage is specific to structured corridors."],"supporting_citations":[{"why":"Supplies the tightly coupled iterated extended Kalman filter odometry that forms the estimation backbone of the system.","marker":"[16]"},{"why":"Supplies the base 2D SLAM and pose-graph optimization that the CNN landmarks refine.","marker":"[42]"},{"why":"Supplies the orientation filter used to stabilize IMU data before state estimation.","marker":"[41]"},{"why":"Serves as the obstacle-detection baseline whose localization accuracy is reported to improve by 26.09%.","marker":"[28]"},{"why":"Serves as the detection baseline compared in the precision/recall/F1 evaluation.","marker":"[31]"},{"why":"Supplies the least-squares alignment procedure used to compute the reported absolute pose error.","marker":"[48]"},{"why":"Supplies the SC-ALOAM baseline that the main trajectory-error comparison is measured against.","marker":"[6]"}],"fun_headline_variants":["CNN object anchors slash LiDAR-SLAM drift by 21%","Handheld LiDAR plus CNN landmarks hits 0.25-m accuracy","Object landmarks from CNN cut 2D LiDAR SLAM error by 26%","CNN-derived landmarks boost indoor LiDAR SLAM accuracy","Fusing CNN objects with LiDAR-IMU trims SLAM error 21%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the CNN's predicted object centroids and radii are stable enough to act as landmarks, something the paper does not support with training data, label sources, or an ablation; if those predictions are unreliable, the claimed gains from object detection collapse.","fun_headline_variants_meta":{"raw":{"variants":["CNN object anchors slash LiDAR-SLAM drift by 21%","Handheld LiDAR plus CNN landmarks hits 0.25-m accuracy","Object landmarks from CNN cut 2D LiDAR SLAM error by 26%","CNN-derived landmarks boost indoor LiDAR SLAM accuracy","Fusing CNN objects with LiDAR-IMU trims SLAM error 21%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000653,"raw_usage":{"total_tokens":2986,"prompt_tokens":934,"completion_tokens":2052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1953}},"tokens_in":550,"tokens_out":2052,"duration_ms":13214,"temperature":1.0,"reasoning_tokens":1953,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:55:36.924150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same two corridor sequences with the CNN landmark term switched off and all other modules identical: if ATE does not rise by roughly the claimed 21–26%, or if disabling object detection improves accuracy, the central claim is wrong. Independently, re-train the CNN on a public point-cloud dataset and check its predicted centroids against a laser-scanned ground-truth map of a held-out corridor.","supporting_citations":[{"cited_title":"Real-time loop closure in 2d lidar slam,","cited_arxiv_id":null,"evidence_quote":"Supplies the base 2D SLAM and pose-graph optimization that the CNN landmarks refine."},{"cited_title":"An obstacle-finding approach for autonomous mobile robots using 2d lidar data,","cited_arxiv_id":null,"evidence_quote":"Serves as the obstacle-detection baseline whose localization accuracy is reported to improve by 26.09%."},{"cited_title":"Real-time 3d object detection and slam fusion in a low-cost lidar test vehicle setup,","cited_arxiv_id":null,"evidence_quote":"Serves as the detection baseline compared in the precision/recall/F1 evaluation."},{"cited_title":"Lidar-imu tightly- coupled slam method based on iekf and loop closure detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the SC-ALOAM baseline that the main trajectory-error comparison is measured against."}],"review_version":1}