{"id":"b8f76d14-3099-4743-850c-f3b45fb98b0d","arxiv_id":"2504.19624","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ARMOR couples scanblock-based spatio-temporal normal smoothing with a PPO agent that selects SDF sampling and meshing parameters, reporting improved underground mesh quality over PIN-SLAM.","lead":"This paper presents ARMOR, a real-time 3D meshing system for underground environments that uses reinforcement learning to adapt its reconstruction parameters on the fly. It reports accuracy gains over one baseline, but the headline 3.96% improvement is not reproduced in the paper's own tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3.96% headline gain is an F-score percentage-point difference, not a reduction in geometric error; the central claim misstates the metric it reports.","rationale":"The reader's weakest assumption (sim-to-real transfer, §5.2/Remark 2) is legitimate: the RL policy is trained in MARSIM with Livox Mid-360 and deployed on VLP-16 and Mid-360 real sequences without a quantitative domain-gap analysis. However, I found a more immediate defect in the paper's own data: the flagship 3.96% figure is not a geometric-error reduction. The arithmetic above shows it exactly matches the average F-score gain in percentage points. This is verifiable from the tables alone and therefore does not depend on any external baseline or domain-transfer argument. I therefore partially agree with the reader: both concerns point to overclaiming, but the metric misattribution is more load-bearing because it undermines the strongest claim as worded even under the most favorable reading of the experiments. The correct fix is not to reject the work outright—the tables do show consistent improvements on most rows—but to condition acceptance on correcting the quantitative claim, reporting F-scores per dataset at their stated thresholds, and either demonstrating sim-to-real transfer or explicitly limiting the claim to the evaluated sensor/scene configuration. Since the reader already issued CONDITIONAL, my read does not change the verdict.","tokens_in":18456,"tokens_out":5396,"duration_ms":49679,"concrete_test":"Recompute the pooled metrics from Tables 2–4: if the mean F-score gap is 3.96 percentage points while the pooled Chamfer-L1 reduction is ≈9.2% and no accuracy/completeness/Chamfer-L1 calculation yields 3.96%, then the abstract's \"reducing geometric error by 3.96%\" is a metric misattribution. As a further check, recompute F-scores separately per dataset with their stated thresholds; if the pooled 3.96 pp gain is an artifact of mixing 10 cm, 15 cm, and 30 cm thresholds, the headline claim should be reworded to report per-dataset F-score changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that ARMOR \"reduc[es] geometric error by 3.96%\" (Abstract; also Conclusion) is not the number the experiments produce. Averaging F-scores over the six benchmark rows in Tables 2–4 gives PIN-SLAM (88.60+88.51+86.66+89.94+87.71+86.06)/6 = 87.91% and ARMOR (92.53+92.36+91.08+93.50+90.36+91.40)/6 = 91.87%, a difference of exactly 3.96 percentage points. The same tables give Chamfer-L1 reductions of roughly 4.3% (synthetic, Table 2), 10.3% (SuperLoc, Table 3), and 12.9% (WHU-Helmet, Table 4), pooled about 9.2%, and no pooled error metric equals 3.96%. So the central quantitative claim substitutes an F-score improvement for an error reduction. F-scores are precision/recall harmonic means at fixed thresholds, not geometric errors, and the thresholds differ across datasets (15 cm, 30 cm, 10 cm), so averaging them into one \"geometric error\" percentage is not meaningful. This is load-bearing: the paper's central promise is a specific error reduction, but the supplied data only support a weaker, differently scoped claim; the sim-to-real transfer issue is secondary because even the simulation tables are misread.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ARMOR, an online 3D meshing framework for underground and other unexposed environments, built on a point-based neural implicit signed distance field (SDF) representation. It contributes three components: (1) spatio-temporal geometry smoothing using temporally aggregated scanblocks, L0-regularized normal refinement, and a multi-segment centroid normal orientation method; (2) normal-guided sampling to replace projective SDF labels; and (3) a PPO-trained reinforcement learning agent that selects six reconstruction parameters based on local neural point features, with a composite reward of accuracy, completeness, Chamfer-L1, and F-score. The system is evaluated on three synthetic MARSIM scenes and on real SuperLoc cave, WHU-Helmet tunnel, and an in-house lava-tube sequence, comparing against PIN-SLAM and an ablation without the RL module. The abstract claims a 3.96% reduction in geometric error relative to state-of-the-art baselines while maintaining real-time efficiency.","tokens_in":18795,"tokens_out":7572,"duration_ms":75171,"significance":"If the results hold, the paper addresses an operational need: automatic, real-time mesh generation in GNSS-denied underground environments without manual parameter tuning. The consistent quantitative improvements over PIN-SLAM across all six scenes, the ablation isolating the RL component, and the real-world evaluation against TLS reference data are genuine strengths. The volumetric tunnel-construction application is also a useful demonstration. However, the headline quantitative claim is misreported, the comparison set is narrow, no variance information is given, and the real-time efficiency claim is not directly measured.","major_comments":[{"comment":"The claim that ARMOR reduces geometric error by 3.96% is not supported by the reported data. Averaging the six F-score rows in Tables 2-4 gives 87.91% for PIN-SLAM and 91.87% for ARMOR, a difference of 3.96 percentage points; this is the only computation that reproduces the headline number. F-score is a harmonic mean of precision and recall at a fixed distance threshold, not a geometric error in centimeters, and the thresholds differ across the tables (15 cm in Table 2, 30 cm in Table 3, and 10 cm in Table 4), so averaging these into a single geometric-error percentage is not meaningful. The pooled Chamfer-L1 reduction from the same tables is roughly 9.2%, and per-scene reductions range from about 4.8% to 15.6%; none equals 3.96%. The abstract and conclusion must either state an F-score improvement of 3.96 percentage points or report a properly computed pooled geometric-error metric.","section":"Abstract and Conclusion; Tables 2-4"},{"comment":"The evaluation uses only PIN-SLAM as a baseline, despite the abstract referring to state-of-the-art baselines in the plural, and no variance or number of independent runs is reported for any of Tables 2-4. Without repeated runs or standard deviations, the observed differences cannot be distinguished from run-to-run noise, and the absence of a second recent implicit LiDAR mapping baseline (e.g., SHINE-Mapping, LONER, or N3-Mapping) leaves the state-of-the-art claim unsubstantiated. Please add at least one additional relevant baseline and report means and standard deviations over multiple runs.","section":"Sections 5.3-5.5"},{"comment":"The training reward in Eq. (9) is a weighted sum of accuracy, completeness, Chamfer distance, and F-score, and the reported evaluation metrics are exactly those four quantities, with the synthetic F-score threshold of 15 cm matching the F-score reward threshold described in Section 4.2.4. This is not full circularity because the real-world evaluation uses TLS reference data not present during training, but it does mean the simulation results partly reflect optimization of the evaluation metric itself. The paper should acknowledge this overlap explicitly and report at least one quantitative criterion that is not part of the reward; the volumetric error in Section 5.6 is a useful step and could be extended to the other real scenes.","section":"Section 4.2.4, Eq. (9); Tables 2-4"},{"comment":"The RL policy is trained only in MARSIM with a Livox Mid-360 sensor, yet it is deployed on SuperLoc sequences acquired with a Velodyne VLP-16 and on WHU-Helmet data from a Livox Mid-360. Remark 2 asserts that simulation-trained strategies generalize effectively to previously unseen real-world scenarios, but no direct evidence about the learned policy's behavior on real data is provided, such as distributions of selected actions, comparison with fixed tuned parameters, or sensitivity to sensor type. The aggregate F-score improvements are consistent with successful transfer, but since adaptive parameter selection is a central contribution, the paper should provide direct evidence of this adaptive behavior rather than only its downstream effect.","section":"Section 5.2 and Remark 2"},{"comment":"The abstract and title emphasize real-time 3D monitoring, but no runtime or latency measurements are reported anywhere in Section 5. There is no table comparing per-frame or per-scanblock processing time between ARMOR, the w/o RL ablation, and PIN-SLAM, and no discussion of the computational cost of the RL agent or the sparse-convolution encoder. Please include quantitative timing results, ideally on the same hardware, to support the real-time efficiency claim.","section":"Section 5"}],"minor_comments":[{"comment":"The text says the authors select the Cave01 and Cave02 sequences from SuperLoc, while Table 3 is headed with cave01 and cave04; the sequence labels should be reconciled.","section":"Section 5.4 and Table 3"},{"comment":"The notation for the L0 smoothing hyperparameters is inconsistent: Eq. (7) uses eta as the L0 weight, Eq. (10) introduces beta for the convergence weight, and Section 5.1 fixes both a smoothing weight beta and a preservation weight eta. Please define each symbol once and align the notation between the equations and the experimental setup.","section":"Eqs. (7), (10), and Section 5.1"},{"comment":"The scanblock construction in Eq. (5) would be clearer if the concatenation were written as an explicit sequence of transformed frames rather than a sum, and if the concatenation operator were defined in the text.","section":"Eq. (5)"},{"comment":"There are several typographical errors, including Trajctory in the Figure 3 caption, 'the the Cave01 and Cave02' in Section 5.4, and 'artifacts cased by SDF label error' in Section 5.4; these should be corrected.","section":"Figures and text"},{"comment":"The F-score thresholds differ across the tables (15 cm, 30 cm, and 10 cm) without explanation; a brief justification or a fixed-threshold sensitivity analysis would help the reader interpret the results.","section":"Tables 2-4"},{"comment":"The Xianren Lava Tube evaluation reports that 94.8% of points fall within a +/-5 cm threshold of a reference produced with Leica Cyclone 3DR, but the reference itself is not described; please specify what the mesh-to-mesh distance was computed against and how the reference model was generated.","section":"Section 5.5"}],"recommendation":"major_revision","confidential_remarks":"The 3.96% claim in the abstract and conclusion is a misreporting of an F-score percentage-point difference as a geometric-error reduction, and it needs to be corrected. The underlying method appears sound and the requested additional experiments are feasible within the scope of the paper, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThere is a real engineering contribution here: a scanblock-based spatio-temporal normal smoothing step plus a PPO agent that selects six meshing and sampling parameters online, aimed at real-time mesh reconstruction in underground scenes. The method is described clearly enough to reproduce, and the ablations show both components help. The real-world evaluations on SuperLoc, WHU-Helmet, and a lava tube are a plus.\n\nBut the headline is wrong. The abstract and conclusion claim ARMOR reduces geometric error by 3.96%. That number is the difference in average F-score between ARMOR (91.87%) and PIN-SLAM (87.91%) over the six rows of Tables 2–4, expressed in percentage points. F-score is not a geometric error; it is a harmonic mean of precision and recall at a fixed threshold, and the thresholds differ across the datasets (15, 30, and 10 cm). Calling that an error reduction is misleading. The actual per-scene Chamfer-L1 reductions are roughly 5–16%, which is a stronger result and the one they should report.\n\nOther soft spots. Only one baseline is tested (PIN-SLAM); no standard deviations, no repeated runs, so the gains are not statistically grounded. The reward in Eq. 9 is a weighted sum of the same accuracy, completeness, Chamfer, and F-score metrics used for evaluation, so the reported improvements are partly the training objective just re-measured. The ablation 'Ours w/o RL' shows the smoothing module matters, but there is no comparison against a per-scene parameter grid search on the full pipeline, so we cannot tell whether the learned policy adds value over fixed tuned parameters. Sim-to-real transfer is assumed: the policy is trained in MARSIM with a Livox Mid-360 and deployed on different sensors and scenes with no analysis of policy behavior changes. Also, Table 3 labels columns 'cave01' and 'cave04' while the text says Cave01 and Cave02, and no code or data are released.\n\nNone of this kills the paper. The core idea is new and the experiments are consistent in direction. But the central numerical claim misstates what was measured, and the evaluation is too thin to support the 'adaptive' contribution as it stands. The paper deserves a serious referee, and I would send it out, but I would expect major revision: rewrite the abstract around the actual F-score and Chamfer numbers, add at least one more baseline, report variance, and include a fixed-parameter upper bound. After that, it would be worth citing.\n\nCheers.","headline":"Useful adaptive meshing system, but the headline '3.96% error reduction' is actually an F-score percentage-point gain, and the single-baseline evaluation leaves the RL contribution under-supported.","tokens_in":19306,"tokens_out":5763,"would_cite":false,"duration_ms":50187,"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":"Adaptive, reinforcement-learned parameter selection enables real-time neural SDF meshing in unexposed scenes with lower geometric error than fixed-parameter baselines.","keywords":["LiDAR","SLAM","real-time meshing","implicit neural fields","signed distance fields","reinforcement learning","adaptive parameter optimization","underground mapping"],"falsifier":"Re-run the real-world evaluations with the learned policy replaced by a per-scene grid-searched fixed parameter set while keeping the smoothing module unchanged; if the best fixed parameters match or beat ARMOR's F-score and Chamfer distance on the same sequences, the claim that adaptive RL is what reduces error by 3.96% is falsified. A cheaper check would compare the policy's chosen action distribution on simulated versus real scanblocks and test whether any systematic mismatch correlates with metric drops.","tokens_in":18289,"feed_emoji":"🕳️","tokens_out":7165,"duration_ms":70759,"temperature":0.7,"pith_summary":"ARMOR claims that the quality bottleneck in real-time underground meshing is not the neural field itself but the fixed reconstruction parameters used to train it, and that reinforcement learning can replace manual tuning by choosing parameters from local scene structure. The paper builds an online pipeline that aggregates LiDAR frames into temporal blocks, smooths normals with an L0-regularized orientation procedure, and trains a PPO agent, entirely in simulation, to select six sampling and meshing parameters for a point-based neural signed distance field (SDF) at each step. On more than 3,000 meters of tunnels, caves, and lava tubes, the complete system reports lower Chamfer distance, higher F-scores, and a 3.96% reduction in geometric error compared with a strong fixed-parameter implicit-SLAM baseline, while keeping reconstruction real-time. The sympathetic reading is that adaptive parameter selection plus projection-error-free SDF labels together make implicit neural meshing usable in places where manual tuning is impractical.","feed_headline":"RL picks mesh parameters live, cutting cave/tunnel error ~4%","feed_subtitle":"Adaptive policy plus geometry smoothing gives real-time neural 3D meshes across 3,000+ meters of unexposed scenes.","key_machinery":"The load-bearing mechanism has three cooperating parts. First, the scanblock: fixed-duration aggregations of LiDAR frames that densify the cloud and enable multi-view normal orientation through a multi-segment centroid line, followed by an L0-minimization normal smoother that suppresses orientation noise while preserving detail. Second, normal-guided sampling: SDF training labels are computed as true distances along smoothed normals inside a truncation band, replacing projective distances that systematically overestimate surface proximity. Third, the reinforcement meshing network: a sparse-convolution encoder converts a downsampled local neural map into a state embedding, an actor-critic policy with a PPO clipped objective maps that embedding to six discrete reconstruction parameters, and a composite reward of accuracy, completeness, Chamfer distance, and F-score shapes the policy. The agent is what makes the pipeline adaptive; the smoothing and sampling modules are what make the reward landscape trustworthy in sparse, noisy, geometry-heavy scenes.","core_discovery":"The central discovery claimed by the paper is that scene-adaptive, learned parameter control and true-distance SDF supervision, not a more powerful network, are what let implicit neural meshing work online in unexposed environments. ARMOR represents the scene as neural points with an auto-decoder SDF, samples training points along smoothed surface normals instead of along LiDAR beams so that labels are true signed distances rather than projective approximations, and lets a six-dimensional RL action space, governing sampling spread, surface and free-space sample counts, ray-ratio bounds, and nearest-neighbor support threshold, respond to the current local map. The agent is trained with PPO in a LiDAR simulator against geometric rewards derived from accuracy, completeness, Chamfer distance, and F-score, then run on real handheld and helmet-mounted scans. Across three synthetic scenes, two open cave datasets, and a lava-tube field deployment, the complete method outperforms the point-based implicit baseline and its own ablation without RL, with the largest gains in geometrically complex regions, and with 94.8% of the lava-tube mesh points falling within five centimeters of a comparison reconstruction.","pith_inferences":["If the sim-to-real transfer generalizes, the same Markov-decision-process formulation could be applied to other parameter-sensitive reconstruction systems, such as outdoor LiDAR meshing or dense visual SLAM, wherever a simulator can supply ground truth for reward.","The fixed 15 cm F-score threshold in the reward means the policy is tuned for fidelity at that scale; a monitoring task that cares about centimeter-level structural detail would need the threshold moved and the policy retrained, so the reported 3.96% improvement is partly reward-design-dependent.","The helmet deployment on a low-power processor hints that the adaptive policy is cheap enough to run on embedded hardware, which would make real-time monitoring a wearable, continuous process rather than a scheduled survey.","The ablation leaves room for a decomposition experiment that the paper does not isolate: measuring how much of the gain comes from the learned policy versus the smoothing and sampling modules alone, which would clarify which component future systems should build on."],"forward_implications":["If the reported gains hold, real-time mesh reconstructions in cave surveys, tunnel inspections, and lava-tube monitoring can be produced on-site by a single operator with a helmet or handheld scanner, without a post-processing step that takes hours.","A policy trained in simulation appears to transfer across LiDAR types and site categories in the paper's experiments, suggesting that one learned parameter policy may serve many unexposed sites instead of per-site manual tuning.","Because sampling and meshing parameters are chosen online, the system can trade geometric fidelity against frame rate on the fly, reducing free-space sampling once the map is stable, which is useful for battery- or compute-limited field robots.","Removing projective-distance SDF labels corrects the systematic volume inflation seen in baseline meshes; in the tunnel construction test, the reported excavation volume error drops to 1.11% relative to ground truth, making the mesh usable for over- and under-excavation analysis.","The real-time property means safety monitoring and structural assessment can use current geometry rather than delayed survey products, which directly addresses the response-time gap that motivates the work."],"supporting_citations":[{"why":"The point-based implicit neural LiDAR SLAM method that serves as the baseline and whose representation and projection-error limitation ARMOR extends.","marker":"Pan et al., 2024"},{"why":"Supplies the clipped PPO objective that stabilizes the policy updates used to train the parameter-selection agent.","marker":"Schulman et al., 2017"},{"why":"Provides the LiDAR simulator in which the agent is trained and in which synthetic evaluation scenes are generated.","marker":"Kong et al., 2023"},{"why":"Defines the auto-decoder point-based SDF representation that ARMOR adopts for its neural map.","marker":"Park et al., 2019"},{"why":"Contributes the progressive spatial smoothing idea with L0 minimization that the normal-smoothing step adapts.","marker":"Li et al., 2024"},{"why":"Supplies the reinforcement learning implementation used to run the PPO training routine.","marker":"Raffin et al., 2021"},{"why":"Motivates normal-guided, non-projective SDF supervision that ARMOR combines with temporal smoothing.","marker":"Song et al., 2024"},{"why":"Provides the helmet-mounted dataset used as one of the real-world evaluations of the complete system.","marker":"Li et al., 2023"}],"fun_headline_variants":["RL adapts mesh parameters live to reduce error in caves and tunnels","RL tunes mesh sampling in real time, reducing cave/tunnel error ~4%","Adaptive RL meshing reduces 3D error in unexposed scenes by ~4%","RL-driven adaptive mesh reduces real-time 3D error underground ~4%","Reinforcement learning picks mesh settings, reducing underground mapping error ~4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned parameter policy is trained only in a simulator and is then trusted, with no closed-loop retraining, to transfer to real-world scans from other LiDAR sensors; if that transfer fails, the reinforcement-learning module is no better than a fixed parameter choice and the reported improvement would have to come entirely from the smoothing module.","fun_headline_variants_meta":{"raw":{"variants":["RL adapts mesh parameters live to reduce error in caves and tunnels","RL tunes mesh sampling in real time, reducing cave/tunnel error ~4%","Adaptive RL meshing reduces 3D error in unexposed scenes by ~4%","RL-driven adaptive mesh reduces real-time 3D error underground ~4%","Reinforcement learning picks mesh settings, reducing underground mapping error ~4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001336,"raw_usage":{"total_tokens":5475,"prompt_tokens":1030,"completion_tokens":4445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":4341}},"tokens_in":646,"tokens_out":4445,"duration_ms":30657,"temperature":1.0,"reasoning_tokens":4341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:47:30.805594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the real-world evaluations with the learned policy replaced by a per-scene grid-searched fixed parameter set while keeping the smoothing module unchanged; if the best fixed parameters match or beat ARMOR's F-score and Chamfer distance on the same sequences, the claim that adaptive RL is what reduces error by 3.96% is falsified. A cheaper check would compare the policy's chosen action distribution on simulated versus real scanblocks and test whether any systematic mismatch correlates with metric drops.","supporting_citations":[{"cited_title":"PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency","cited_arxiv_id":"2401.09101","evidence_quote":"The point-based implicit neural LiDAR SLAM method that serves as the baseline and whose representation and projection-error limitation ARMOR extends."},{"cited_title":", author Liu, X","cited_arxiv_id":null,"evidence_quote":"Provides the LiDAR simulator in which the agent is trained and in which synthetic evaluation scenes are generated."},{"cited_title":", author Hill, A","cited_arxiv_id":null,"evidence_quote":"Supplies the reinforcement learning implementation used to run the PPO training routine."},{"cited_title":", author Zhao, J","cited_arxiv_id":null,"evidence_quote":"Motivates normal-guided, non-projective SDF supervision that ARMOR combines with temporal smoothing."}],"review_version":1}