{"id":"c4a50c99-16e5-40a0-ac4c-9b93ca39746f","arxiv_id":"2412.13873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Uncertainty-aware model predictive control of motor speed reduces LiDAR odometry error by over 60% in simulation while keeping scanning completeness roughly unchanged.","lead":"A new control method for motorized LiDAR systems adjusts motor speed based on predicted scanning uncertainty, aiming to improve 3D odometry accuracy without losing coverage. If it works as reported, it offers a smarter, scene-adaptive alternative to constant-speed spinning for mapping and robot navigation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline '>60% improvement, <2% efficiency loss' is not reproducible from Table I: the 60% aggregate is dominated by the KTH outlier (45.03 m constant-speed ATE), and the NTU site shows a 10.3% CMPLT drop.","rationale":"Good-faith reading: UA-MPC's mechanism, predicting per-orientation LO uncertainty from a rendered local map and adjusting rotor speed to dwell in informative directions, is coherent, and the qualitative direction of the experiments is positive. The load-bearing problem is that the paper's central quantitative claim is internally contradicted by the paper's own evidence. The 63.7% aggregate ATE improvement is an average of an 8.3% gain, an 80% gain driven by a single 45 m drift failure, and a 28.9% gain; excluding the outlier the gain is 23.3%, and the real-world gain is 47%. The efficiency claim fails on the NTU site (-10.3% CMPLT) and on the total (-5.5%), so 'less than 2%' is not a faithful summary of Table I. I am not claiming the method does not work; I am claiming the headline is not supported by the reported data. This is more direct and decisive than questioning the U(theta) surrogate, because even if the predictor is imperfect, the paper must first present accurate empirical claims. The reader flagged the same efficiency contradiction and outlier dominance in the rationale, but chose the uncertainty predictor as the weakest assumption; hence partial agreement. A recomputation with revised reporting, repeated trials, and confidence intervals would settle whether the central claim survives in corrected form.","tokens_in":10716,"tokens_out":9838,"duration_ms":99147,"concrete_test":"Recompute every headline percentage strictly from Table I and Section IV.D: per-site ATE ratios (NTU 4.55/4.96 = -8.3%; KTH 9.01/45.03 = -80.0%; TUHH 9.39/13.21 = -28.9%), aggregate with and without KTH (63.7% vs 23.3%), real-world Spine (1.51 to 0.80 m = 47%), and per-site plus total CMPLT decreases (NTU 10.3%, KTH 1.2%, TUHH 0.25%, total 5.5%). If the corrected aggregates do not support 'over 60% with less than 2% efficiency loss,' revise the abstract and Section IV.C to report per-site values, repeated trials, and error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is internally contradicted by the paper's own Table I. Aggregating all three MCD sites gives an ATE reduction of 1 - (7.65/21.07) = 63.7%, but that average is dominated by the KTH site, where constant-speed control drifts to 45.03 m while UA-MPC achieves 9.01 m. Excluding KTH, the mean ATE improvement is only 23.3% (NTU: 4.96 to 4.55 m, 8.3%; TUHH: 13.21 to 9.39 m, 28.9%), and the real-world NTU Spine result is a 47% reduction (1.51 to 0.80 m), not over 60%. The efficiency claim also fails on the reported data: NTU CMPLT falls from 69,237 to 62,068 voxels, a 10.3% decrease, and the total simulated CMPLT falls 5.5% (132,670 vs 140,345). Thus the abstract and Section IV.C overstate both the size and the uniformity of the benefit. The underlying mechanism may still be sound and the qualitative direction is positive, but the headline claim as written is not supported by the paper's own experiments, and no repeated trials or confidence intervals are provided to show that the KTH drift is a stable property rather than a single failure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UA-MPC, an uncertainty-aware model predictive controller that adjusts the rotation speed of a motorized LiDAR to improve LiDAR odometry accuracy while preserving scanning efficiency. The uncertainty of future scans is predicted by ray tracing the current local map, computing an A-optimal trace of the inverse information matrix, and then a piecewise-linear surrogate model is used to solve the MPC problem. The authors also contribute a ROS-based motorized LiDAR simulation environment built on MARSIM and evaluate UA-MPC against constant-speed and zero-speed control on the MCD dataset and on two real-world handheld scanning routes. The central claim, stated in the abstract and Section IV.C, is that UA-MPC achieves over 60% reduction in positioning error with less than 2% decrease in efficiency compared to constant-speed control.","tokens_in":11025,"tokens_out":4564,"duration_ms":41292,"significance":"If the claims are supported, the paper would make a useful contribution: adaptive motor-speed control for motorized LiDAR is a practical and underexplored direction, the proposed uncertainty predictor based on rendered observability is principled, and the open-sourced simulation environment is a valuable resource for the community. The real-world deployment on an embedded platform (Orin-NX) and the use of the MCD dataset are also positive aspects. However, the paper's headline quantitative claims are not supported by the reported experiments, and the predictive mechanism is not directly validated. The core idea is promising, but the evidence as presented is not yet sufficient for the strength of the claims.","major_comments":[{"comment":"The headline claim of 'over a 60% reduction in positioning error with less than 2% decrease in efficiency' is not supported by Table I. Per-site ATE improvements are 8.3% for NTU (4.96 to 4.55 m), 80.0% for KTH (45.03 to 9.01 m), and 28.9% for TUHH (13.21 to 9.39 m); the aggregate improvement of about 63.7% is dominated by the KTH constant-speed drift. The efficiency claim is also contradicted by the table: NTU CMPLT drops by 10.35% (69,237 to 62,068 voxels) and the total simulated CMPLT drops by 5.5% (140,345 to 132,670). The real-world Spine result is a 47% reduction (1.51 to 0.80 m), not over 60%. The abstract and Section IV.C should be corrected to report per-site results and to avoid the aggregate claim, or the experiments need to be extended to support it.","section":"Abstract and Section IV.C"},{"comment":"All comparisons are based on a single simulation run per site and a single real-world pass per route, with no repeated trials, standard deviations, or confidence intervals. Because the aggregate improvement depends heavily on one large drift in the KTH constant-speed run (45.03 m), it is unclear whether this is a stable property of the environment or a single failure realization. I request repeated runs (or, for the real-world experiments, at least a few passes) with reported spread, and an analysis of how the KTH result varies with initialization and noise.","section":"Section IV.C and Section IV.D"},{"comment":"The proposed mechanism rests on U(θ_i) in Eq. (8) being a faithful predictor of future LO accuracy, but no evidence is provided that the predicted uncertainty correlates with actual ATE. Since U is computed by ray tracing the local map produced by the same LO that is being evaluated, an incomplete or drifting local map can bias the prediction, and Remark 2's static-body rendering may mis-rank orientations during fast motion. I suggest adding a validation experiment that compares predicted U against actual per-orientation odometry error in the simulator (where ground truth is available), and an ablation that replaces U with a constant or a map-independent heuristic to isolate the benefit of the uncertainty prediction.","section":"Section II.C"},{"comment":"Several parameters needed to reproduce the MPC are not specified: the control horizon M in Eq. (4), the sampling interval Δθ in Eq. (9), the solver used, and the optimization termination criteria. The weights α and β are set to balance magnitudes (α=1000, β=1) rather than derived; a sensitivity study over α/β and M would clarify how robust the reported improvement is. Please add these details and, if possible, a sensitivity table.","section":"Section II.D and Section IV.A"}],"minor_comments":[{"comment":"The text says 'the ATE is improved by 60% using UA-PMC'; this should read 'UA-MPC'.","section":"Section IV.C"},{"comment":"The caption lists subfigures (a), (b), (d), and (e), but the text refers to 'Fig. 8(c)' for the NTU Spine point cloud; either add subfigure (c) or renumber the subfigures.","section":"Fig. 8"},{"comment":"The sentence 'The evaluations of the UA-MAP are shown in Fig. 9(e) and (f)' should read 'UA-MPC'.","section":"Section IV.D"},{"comment":"The interpolation formula in Eq. (10) does not specify how the indices behave when θ approaches 2π; please state whether U is periodic and how the wrap-around is handled.","section":"Eq. (9)-(10)"},{"comment":"The text states that 2π/Δθ samples are collected, which assumes Δθ divides 2π; please state the chosen Δθ value (apparently 5 degrees from Fig. 3, but not stated in the text).","section":"Section II.D"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the open-source simulation environment is a genuine contribution, but the headline claims need to be corrected and the experiments need to be strengthened before publication. The paper appears suitable for a robotics venue such as RA-L or IROS. I have no concerns about novelty relative to prior work; the distinction from [26] is adequately drawn."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read UA-MPC. Short version: the control idea is sensible, the experiments point in the right direction, but the paper oversells its own results. The 60% ATE improvement and the less-than-2% efficiency loss in the Abstract and Section IV.C are not what Table I shows.\n\nWhat's new: nobody else, to the authors' credit, has cast motor speed control for a motorized LiDAR as an MPC problem with a ray-traced A-optimal observability surrogate. That is a genuine contribution, and the MARSIM-based simulator, while an adaptation rather than a from-scratch platform, is useful. The real-world evaluation on the NTU Spine with a Leica MS60 reference is a good piece of evidence, and the system ran on an Orin-NX, so the compute story is credible.\n\nThe soft spots are the usual ones. First, the headline. Aggregating the three MCD sites, the ATE drop is 63.7%, but that number is carried entirely by KTH, where constant-speed drifts to 45 m. Excluding KTH, the average improvement is about 23%, and the real-world gain is 47%. The CMPLT claim is worse: NTU drops 10.3%, and the total simulated completeness falls 5.5%, not 'less than 2%'. A reader skimming the abstract will carry away a wrong impression of effect size and robustness.\n\nSecond, there are no repeated trials or error bars anywhere. The KTH constant-speed run looks like a single failure mode; it may be stable, but with one run per condition we can't tell. The lack of error bars is a minor issue for a systems paper, but it becomes a major one when the headline claim rests on a single outlier site.\n\nThird, the uncertainty predictor assumes a static body and a complete local map. The authors flag the static-body assumption in Remark 2, which is honest, but the deeper coupling is that U is computed from the same LO the controller is trying to improve. That is not circular in a fatal sense—the reported result is an empirical closed-loop outcome—but it does mean the predictor inherits the odometry's biases, and the paper does not discuss when that could actively mis-rank orientations.\n\nThe weights alpha, beta, and the reference speed are hand-set and scene-independent; a sensitivity study would strengthen the claims. None of this kills the paper. The mechanism is plausible, the implementation is real, and a careful revision that reports per-site numbers honestly would make it a solid contribution. I'd send it to review. The right referee will want the headline rewritten and an additional repeated-trial experiment, not a new research program.","headline":"A sound control idea and a useful simulator, undermined by headline numbers that the paper's own Table I contradicts.","tokens_in":11575,"tokens_out":2716,"would_cite":true,"duration_ms":24942,"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 motor speeds cut LiDAR odometry error by over 60 percent in the paper's tests.","keywords":["model predictive control","motorized LiDAR","LiDAR odometry","uncertainty prediction","ray tracing","observability","scanning efficiency","active sensing"],"falsifier":"Run the controller in a scene where a large unvisited or occluded region lies within the prediction horizon, and compare the controller's predicted uncertainty ranking at each orientation with the actual ATE contribution of scans taken at those orientations; a near-zero or negative correlation between predicted uncertainty and realized error would show the proxy is not faithful and that speed adaptation is not the cause of the reported gains.","tokens_in":10514,"feed_emoji":"📡","tokens_out":6660,"duration_ms":59277,"temperature":0.7,"pith_summary":"The paper claims that a motorized LiDAR spinning at a scene-dependent speed, rather than a constant speed, can significantly improve the accuracy of LiDAR odometry without sacrificing scanning coverage. The proposed controller predicts how observable the environment will be from each motor angle by ray tracing through the current local map, then adjusts the motor speed through a model-predictive loop to dwell on feature-rich directions while keeping an efficiency penalty close to zero. Across three simulated campus sites and two real handheld scans, the authors report over a 60% reduction in absolute positioning error compared with constant-speed rotation, with less than a 2% drop in the number of scanned voxels. If this holds, adaptive motor control becomes a low-cost, hardware-free way to make lightweight motorized scanners reliable in feature-sparse or complex scenes.","feed_headline":"Adaptive motor speeds cut LiDAR odometry error by 60%","feed_subtitle":"Scene-aware motor control dwells on feature-rich views while keeping scan coverage almost unchanged.","key_machinery":"The load-bearing object is the uncertainty function $U(\\theta_i)$, the A-optimal trace of the inverse information matrix for the LiDAR odometry's point-to-plane residuals at motor angle $\\theta_i$. The paper renders a single panoramic depth image from the current local map, samples the LiDAR beams at discrete motor angles by ray tracing, computes the residual Jacobians, and sums their outer products to form the information matrix. A piecewise linear surrogate function is fit to samples of $U$ and used inside a model-predictive controller whose objective trades odometry uncertainty against deviation from the preset speed, so the optimization remains solvable on an edge computer.","core_discovery":"The central discovery is that the quality of LiDAR odometry from a motorized scanner is not fixed by the scanner geometry; it can be actively steered by choosing how long the motor dwells at each orientation. The authors define an observability score for an orientation as the trace of the inverse information matrix built from simulated point-to-plane residuals at that angle, then solve a short-horizon model-predictive control problem that minimizes this score together with deviation from a preset speed. In the reported experiments, this reduces absolute trajectory error from 45.03 m to 9.01 m on a challenging simulated campus route and from 1.51 m to 0.80 m on a real 300-meter corridor, while the fraction of scanned space stays within 2% of the constant-speed baseline. The paper treats this as evidence that uncertainty-aware control, rather than faster spinning or additional sensors, is the effective lever for accurate motorized 3D sensing.","pith_inferences":["Because the uncertainty function is defined from geometric residuals only, the same surrogate could be attached to other estimation backends, not just the specific odometry used in the paper, provided their residual Jacobians are available.","The static single-render assumption is the most likely failure point at higher vehicle speed or with moving objects; a motion-compensated multi-render variant is a natural testable extension.","The largest simulation improvements occur on the most degenerate route, suggesting the benefit may concentrate in feature-sparse scenes while feature-rich scenes gain little; this pattern can guide where to deploy the method."],"forward_implications":["Motorized LiDAR odometry accuracy can be improved by software-only motor-speed adaptation, without adding sensors or changing the scanner geometry.","The controller runs in real time on an edge processor, so the approach is deployable on lightweight handheld and aerial platforms.","The simulation setup, built from existing ground-truth maps, allows any motor-control policy to be benchmarked before hardware deployment.","The reported trade-off implies that constant-speed rotation underuses the motorized scanner's potential accuracy in feature-rich environments."],"supporting_citations":[{"why":"Supplies the point-realistic LiDAR simulation backend that generates simulated scans from ground-truth maps for the motorized scanner evaluation.","marker":"[31]"},{"why":"Provides the LiDAR odometry algorithm whose state-estimation information matrix is predicted and optimized by the controller.","marker":"[32]"},{"why":"Supplies the A-optimal design criterion (trace of the inverse information matrix) used to define the per-orientation uncertainty score.","marker":"[33]"},{"why":"Provides the multi-campus dataset with ground-truth trajectories and point-cloud maps from which the simulation scenes are built.","marker":"[34]"},{"why":"Documents observability degeneracy in LiDAR mapping under high dynamic sensor motion, motivating the need for scene-adaptive motor control.","marker":"[15]"}],"fun_headline_variants":["Adaptive motor speeds slash LiDAR odometry error by 60%","Uncertainty-aware control cuts LiDAR error 60% with same coverage","Smart dwell times boost motorized LiDAR accuracy by 60%","Scene-aware motor control improves LiDAR odometry by 60%","Active scanning: 60% lower odometry error via adaptive dwell"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest premise is that one static render of the current local map correctly predicts how well the future LiDAR scan at each motor angle will constrain the odometry estimate; if the map is incomplete, stale, or the sensor moves during the horizon, the predicted uncertainty can rank orientations incorrectly and the speed choices can hurt accuracy instead of helping.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive motor speeds slash LiDAR odometry error by 60%","Uncertainty-aware control cuts LiDAR error 60% with same coverage","Smart dwell times boost motorized LiDAR accuracy by 60%","Scene-aware motor control improves LiDAR odometry by 60%","Active scanning: 60% lower odometry error via adaptive dwell"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000734,"raw_usage":{"total_tokens":3304,"prompt_tokens":988,"completion_tokens":2316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2222}},"tokens_in":604,"tokens_out":2316,"duration_ms":15671,"temperature":1.0,"reasoning_tokens":2222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:41:59.433797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller in a scene where a large unvisited or occluded region lies within the prediction horizon, and compare the controller's predicted uncertainty ranking at each orientation with the actual ATE contribution of scans taken at those orientations; a near-zero or negative correlation between predicted uncertainty and realized error would show the proxy is not faithful and that speed adaptation is not the cause of the reported gains.","supporting_citations":[{"cited_title":"Marsim: A light-weight point-realistic simulator for lidar- based uavs,","cited_arxiv_id":null,"evidence_quote":"Supplies the point-realistic LiDAR simulation backend that generates simulated scans from ground-truth maps for the motorized scanner evaluation."},{"cited_title":"I2EKF-LO: A Dual-Iteration Extended Kalman Filter Based LiDAR Odometry","cited_arxiv_id":"2407.02190","evidence_quote":"Provides the LiDAR odometry algorithm whose state-estimation information matrix is predicted and optimized by the controller."},{"cited_title":"Pukelsheim, Optimal design of experiments","cited_arxiv_id":null,"evidence_quote":"Supplies the A-optimal design criterion (trace of the inverse information matrix) used to define the per-orientation uncertainty score."},{"cited_title":"Mcd: Diverse large-scale multi-campus dataset for robot perception,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-campus dataset with ground-truth trajectories and point-cloud maps from which the simulation scenes are built."},{"cited_title":"Or-lim: Observability-aware robust lidar-inertial- mapping under high dynamic sensor motion,","cited_arxiv_id":null,"evidence_quote":"Documents observability degeneracy in LiDAR mapping under high dynamic sensor motion, motivating the need for scene-adaptive motor control."}],"review_version":1}