{"id":"ddd151d5-f124-48c4-b452-17f64898562c","arxiv_id":"2508.12729","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MCTR improves Delaunay-triangulation-based reactive racing by using opposing-edge midpoints and CCMA curvature smoothing, and validates the approach in a CARLA digital twin with 2D and 3D LiDAR.","lead":"This paper introduces a LiDAR-based reactive racing controller, MCTR, that extracts a smooth centerline by triangulating track boundaries and filtering it with a curvature-corrected moving average, and validates it in a CARLA digital twin with both 2D and 3D LiDAR. The work is aimed at making reactive controllers for autonomous racing more robust and easier to transfer from simulation to real cars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M-track result in Table III contradicts the paper's 'consistently outperforming' real-world claim; lack of error bars makes the reported O/F gains unreliable.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict. However, the single most load-bearing concern is not the structured-track dependence or the CCMA parameter sensitivity, though both are real secondary issues. It is the M-track row of Table III: DTR beats MCTR by 0.31 s, which directly contradicts the conclusion that MCTR consistently outperforms baselines across diverse track configurations. Combined with the absence of variance or significance reporting in the real-world table, the paper's strongest real-world claim is not established. The simulation evidence is more supportive: Table II includes standard deviations and shows a substantial lap-time improvement from DTR to MCTR, and the ablation without CCMA still beats DTR, so the core algorithmic idea has independent support. A focused re-run of the M-track comparison would settle whether the counterexample is a tuning artifact, a genuine failure mode, or a typo. If the M-track result is a genuine failure mode, the paper should report it as a limitation and qualify the claim; if it is noise, the paper should provide the per-lap data to show that. Either way, the reader's CONDITIONAL verdict remains appropriate, with the M-track issue added as a required clarification.","tokens_in":10418,"tokens_out":4515,"duration_ms":51221,"concrete_test":"Rerun the real-world M-shaped track comparison between MCTR and DTR with at least 10 recorded laps per algorithm, using the same parameters/code released by the authors, and report the full per-lap list plus mean±std and a paired difference test. If DTR's mean remains below MCTR's (or the difference is not statistically significant), the conclusion that MCTR 'consistently outperforms baselines across diverse track configurations' is not supported and must be restricted/qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Table III, MCTR's real-world lap time on the M-shaped track is 9.59 s, while DTR records 9.28 s—DTR is 3.3% faster. The paper never explains or acknowledges this counterexample, yet the conclusion states that MCTR 'consistently outperforms baselines across diverse track configurations.' This is an internal inconsistency in the central empirical claim, not merely a scope limitation. The problem is compounded because Table III reports only point values ('Each track was tested for five laps') with no per-lap times, standard deviations, or significance tests, so the O/F-track advantages (7.39 vs 7.82; 9.77 vs 9.88) could be within run-to-run variability. The simulation ablation in Table II does show a large lap-time improvement over DTR (23.39 vs 28.73 s) and includes mean±std, so the method may be sound; but the real-world 'consistent' superiority claim is exactly what the M-track row falsifies as reported. The missing CCMA λ/µ values and the structured-track limitation are secondary: they affect scope, whereas the M-track datum affects whether the headline real-world claim is true as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCTR, a reactive LiDAR-based racing controller. MCTR builds on DTR by replacing circumcircle-based centerline extraction with opposing-edge midpoint selection and applying a Curvature Corrected Moving Average (CCMA) smoother; control is then performed with pure pursuit and a curvature-dependent speed law. The authors also construct a CARLA digital twin that supports both 2D and 3D LiDAR, including a projection from 3D point clouds to pseudo-2D scans. In F1TENTH simulation on an F-shaped track, MCTR reports a lap time of 23.39 s versus 28.73 s for DTR, with 20% lower curvature error and 24.83% lower lateral jerk cost. Real-world experiments on 1:10-scale tracks report faster lap times for MCTR than DTR on O- and F-shaped tracks, a slower lap time on an M-shaped track, and passability of 95% versus 80% (DTR) and 45% (FTG). The paper concludes that MCTR \"consistently outperforms baselines across diverse track configurations.\"","tokens_in":10694,"tokens_out":6176,"duration_ms":76685,"significance":"If the empirical claims hold, the paper makes a useful contribution: a simple mapless reactive controller that produces smoother centerlines than DTR, combined with a digital-twin pipeline for 3D LiDAR validation. The open-source release is a concrete strength, as are the simulation and physical experiments with two LiDAR modalities. The core idea is plausible and the simulation ablation suggests that the CCMA component helps. However, the paper's central real-world claim is currently undermined by incomplete reporting: the M-track result contradicts the \"consistently outperforming\" conclusion, and the absence of variance or significance information makes the reported O/F-track advantages unreliable. The method's load-bearing details, especially the definitions of the smoothness metrics and the implementation of the CCMA objective, are also missing.","major_comments":[{"comment":"The M-shaped track row directly contradicts the conclusion that MCTR \"consistently outperforms baselines across diverse track configurations\": MCTR's lap time is 9.59 s while DTR's is 9.28 s, so DTR is 3.3% faster. The table also reports only point values with no per-lap times, standard deviations, or significance tests; the text says each track was tested for five laps, but the PASS metric is aggregated over 20 laps. The authors should either report the full distribution and show that the O/F differences are meaningful, or explicitly acknowledge and analyze the M-track exception and soften the conclusion.","section":"§IV-B, Table III and §V"},{"comment":"The smoothness metrics ε_k and J_lat are never defined. Table II reports them as \"curvature error\" and \"centerline lateral error jerk,\" but the reference path, the error definition, and the computation method are unspecified, making the 20.00% and 24.83% improvements unverifiable. In addition, Eq. (4) as typeset, v_target = sqrt(μ a_y^max) κ^{-1}, is dimensionally inconsistent; if the intended formula is v_target = sqrt(μ a_y^max / κ), it should be corrected. Since the velocity command is central to the lap-time argument, this needs clarification.","section":"§IV-A, Table II and Eq. (4)"},{"comment":"The CCMA objective is the paper's main algorithmic contribution, but it is incompletely specified. Eq. (1) is presented as an optimization problem, yet no solver, boundary conditions, or λ and μ values are given, and no sensitivity analysis is reported. The surrounding text also mislabels the terms: the second term is a second finite difference, not a first-derivative/velocity term, and the third term is a fourth finite difference, not a second-derivative/curvature term. This makes the method difficult to reproduce and the description internally inconsistent.","section":"§III-A.1, Eq. (1)"},{"comment":"The passability claim (95% vs 80% vs 45%) is not interpretable as reported. The table says each track was tested for five laps, but PASS is described as \"out of a total of 20\"; the criterion for a \"minor collision\" is not defined, and no breakdown by track or algorithm is given. The authors should clarify the number of laps per condition, the failure criterion, and whether PASS is aggregated across tracks.","section":"§IV-B, Table III"}],"minor_comments":[{"comment":"Reference [9] contains a malformed arXiv identifier, 'arXiv:2505.2505.24320'; this should be corrected to 2505.24320.","section":"References"},{"comment":"The text compares a CARLA 3D-LiDAR lap time of 7.90 s with a real-world time of 7.23 s, but 7.23 s does not appear in Table III. The track and conditions for this comparison should be identified.","section":"§IV-B"},{"comment":"The PID gain notation in Eq. (5) is ambiguous: K^i_p, K^i_i, K^i_d with i=t/b is not clearly defined, and the sampling interval Δt and error e(t) should be stated explicitly.","section":"§III-A.2, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the open-source release is a positive feature. The main concerns are empirical reporting and reproducibility rather than fundamental conceptual errors. If the authors provide the missing definitions, full real-world data with variance, and a revised conclusion that does not overclaim on the M-track result, I would support publication. The current version needs a substantive revision before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a reasonable incremental successor to DTR, with two genuinely useful additions: opposing-edge midpoint extraction that avoids DTR's geometric segmentation, and a CARLA digital twin that accepts 3D LiDAR and projects it to pseudo-2D. They release code. The simulation ablation is the best part — MCTR beats DTR decisively on the F-shaped track (23.39±0.60 vs 28.73±0.38 s), and the no-CCMA variant still beats DTR, so the core algorithmic claim is supported even if the CCMA contribution is modest.\n\nThe problems are in the real-world section. Table III gives single lap-time values with no error bars, and the caption states five laps per track while the PASS metric is described as percentage of 20 laps — that inconsistency is easy to fix but should have been caught. More seriously, on the M-shaped track DTR is faster (9.28 vs 9.59 s), which contradicts the conclusion's 'consistently outperforms baselines across diverse track configurations.' The authors never mention that row. That is a load-bearing overstatement, even if the method is still competitive overall.\n\nSeveral definitions are missing: curvature error ε_k and lateral jerk cost J_lat are never given formulas, so the simulation improvements on those metrics can't be checked. The CCMA weights λ and μ are never reported, nor the pure pursuit gains or PID gains, which makes the result hard to reproduce despite the code release. The structured-track limitation is acknowledged in the conclusion, so that is fine as a scope statement.\n\nThe digital twin comparison (7.90 s simulated vs 7.23 s real for 3D LiDAR on the O-track) is a useful single data point, but it is indeed only one point and the real-world 2D vs 3D parity claim is supported only by a figure.\n\nI'd send this to peer review. It is an honest engineering paper with an ablation and open code; a good referee can push them to report the parameters, add error bars, and fix the M-track contradiction. That is a revision, not a rejection. I would not cite it myself unless I were actively working on reactive racing centerlines, but I'd put it in a reading group.","headline":"A solid incremental extension of DTR with a useful CARLA digital twin, but the real-world 'consistently outperforms' claim is contradicted by its own M-track result and too many tuning parameters are unreported.","tokens_in":11238,"tokens_out":2434,"would_cite":false,"duration_ms":26998,"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":"MCTR extracts track centerlines from LiDAR midpoints between opposing boundaries, cutting lap time by 18.6% versus DTR while smoothing curvature and jerk.","keywords":["autonomous racing","reactive control","Delaunay triangulation","LiDAR","centerline extraction","curvature smoothing","digital twin","sim-to-real transfer"],"falsifier":"Record lap times on a track with a long straight open on one side (no opposite wall within LiDAR range) or with a partial occlusion; if MCTR cannot complete the lap or its centerline diverges, the opposing-edge assumption fails. Alternatively, remove the CCMA module and compute the lap-time change; the paper's own ablation shows only a small drop, so a careful experiment at higher speeds could test whether smoothness, not centerline accuracy, is the main source of the gain.","tokens_in":10279,"feed_emoji":"🏎️","tokens_out":4984,"duration_ms":48030,"temperature":0.7,"pith_summary":"This paper claims that the Delaunay-triangulation racing (DTR) controller can be made substantially faster and smoother by changing how the centerline is extracted from LiDAR scans. Instead of using circumcircles, MCTR looks for triangle edges whose endpoints lie on opposite sides of the driving direction and takes their midpoints as waypoints, then applies a curvature-aware moving-average filter (CCMA). In simulation on the F1TENTH F-shaped track, lap time drops from 28.73 s to 23.39 s (18.6%), with a 20% reduction in curvature error and a 24.8% reduction in lateral jerk. The paper also presents a CARLA-based digital twin that supports 3D LiDAR, and shows real-world lap times that beat DTR and FTG with a 95% passability rate on a 1:10 scale car. The point is that reactive, map-free racing can achieve near-smooth racing lines from raw 3D LiDAR without localization or planning.","feed_headline":"Midpoint centerlines make reactive racing 18.6% faster","feed_subtitle":"A map-free LiDAR controller gains smooth curvature through midpoint extraction and curvature-aware smoothing, tested in CARLA and on real ca","key_machinery":"The load-bearing object is the opposing-edge midpoint extraction: after Delaunay triangulation of sampled LiDAR boundary points, edges whose two vertices lie on opposite sides of the current driving direction are selected, and their midpoints form the initial centerline. This replaces DTR's circumcircle-based centerline and eliminates the need for geometric track segmentation. The second component, CCMA, is an optimization over waypoints $q_i$ minimizing fidelity to original points plus $\\lambda$-scaled second-difference (velocity-continuity) and $\\mu$-scaled fourth-difference (curvature-smoothness) penalties.","core_discovery":"The central discovery is that the quality of a triangulation-based racing line is determined less by the triangulation itself than by which edges are used to define the centerline. DTR relies on circumcircles and pre-segmented track regions; MCTR shows that selecting only opposing edges—edges with one endpoint on each side of the heading—and taking their midpoints yields a more stable and smoother centerline implicitly. Adding a CCMA smoothing step that penalizes both first and second discrete derivatives of the waypoints produces curvature-continuous paths that allow higher cornering speeds. The paper demonstrates this in simulation and on physical vehicles, with both 2D and 3D LiDAR.","pith_inferences":["The 18.6% lap-time gain likely comes mostly from higher cornering speed enabled by smoother curvature, not from a shorter path; the paper does not separate path-length from speed effects, so an experiment holding centerline length fixed would isolate the mechanism.","Because MCTR relies on detecting opposite boundaries, it should fail gracefully on open or occluded tracks; a natural extension is to fuse stereo vision or prior maps to fill boundary gaps, which the authors mention as future work.","The CCMA parameters $\\lambda$ and $\\mu$ are not swept in the paper; testing sensitivity to these weights would reveal how much tuning is needed for a new track.","The digital twin uses a simplified vehicle dynamics model; the sim-to-real gap for higher-speed maneuvers with tire slip is untested, so the 95% passability may degrade near the friction limit."],"forward_implications":["Reactive mapless controllers can achieve lap times close to planning-based methods on structured tracks, since smoother curvature directly raises attainable cornering speed.","The pipeline transfers from 2D to 3D LiDAR by projecting the 3D point cloud to a pseudo-2D scan, so algorithms developed in 2D simulation can be deployed on full-scale cars with 3D sensors without retraining.","The CARLA digital twin can serve as a validation environment for perception-driven racing controllers, reducing the need for expensive on-track testing.","The CCMA smoothing may generalize beyond racing to any local path-planning task where curvature continuity matters, e.g., warehouse robots or search-and-rescue."],"supporting_citations":[{"why":"Supplies the DTR baseline that MCTR improves upon; the comparison against it provides the main performance numbers.","marker":"[9]"},{"why":"Supplies the Follow-The-Gap baseline used as an earlier reactive approach and a benchmark in real-world tests.","marker":"[10]"},{"why":"Provides the R-CARLA method for reconstructing a digital-twin mesh from point clouds, used for 3D LiDAR simulation.","marker":"[12]"},{"why":"Provides the 'boxed method' for secondary sampling of LiDAR scans before triangulation.","marker":"[21]"},{"why":"Defines the Pure Pursuit tracking law used to convert the centerline into steering commands.","marker":"[22]"},{"why":"Supplies the curvature-to-velocity formula used to set speed limits from local curvature and friction.","marker":"[23]"},{"why":"Establishes the use of Delaunay triangulation for centerline extraction in racing, which MCTR adapts.","marker":"[3]"}],"fun_headline_variants":["Opposing-edge midpoints beat circumcircles for racing lines","Curvature-corrected smoothing yields smoother reactive racing","MCTR: midpoint lines and CCMA for faster autonomous racing","CARLA digital twin tests 3D LiDAR racing controller","Smoother centerlines from midpoint extraction, not circumcircles"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole method assumes the track is a structured corridor whose opposite boundaries are continuously visible in the LiDAR scan; if a boundary is occluded or too far to be sampled, the opposing-edge midpoint extraction has nothing to anchor the centerline.","fun_headline_variants_meta":{"raw":{"variants":["Opposing-edge midpoints beat circumcircles for racing lines","Curvature-corrected smoothing yields smoother reactive racing","MCTR: midpoint lines and CCMA for faster autonomous racing","CARLA digital twin tests 3D LiDAR racing controller","Smoother centerlines from midpoint extraction, not circumcircles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000379,"raw_usage":{"total_tokens":1842,"prompt_tokens":725,"completion_tokens":1117,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1044}},"tokens_in":469,"tokens_out":1117,"duration_ms":11261,"temperature":1.0,"reasoning_tokens":1044,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:16:12.464362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record lap times on a track with a long straight open on one side (no opposite wall within LiDAR range) or with a partial occlusion; if MCTR cannot complete the lap or its centerline diverges, the opposing-edge assumption fails. Alternatively, remove the CCMA module and compute the lap-time change; the paper's own ablation shows only a small drop, so a careful experiment at higher speeds could test whether smoothness, not centerline accuracy, is the main source of the gain.","supporting_citations":[{"cited_title":"Ros-based localization of a race vehicle at high-speed using lidar,","cited_arxiv_id":null,"evidence_quote":"Provides the 'boxed method' for secondary sampling of LiDAR scans before triangulation."},{"cited_title":"Dtr: Delaunay triangulation-based racing for scaled autonomous racing,","cited_arxiv_id":null,"evidence_quote":"Supplies the DTR baseline that MCTR improves upon; the comparison against it provides the main performance numbers."},{"cited_title":"A novel obstacle avoidance algorithm: “follow the gap method","cited_arxiv_id":null,"evidence_quote":"Supplies the Follow-The-Gap baseline used as an earlier reactive approach and a benchmark in real-world tests."},{"cited_title":"R-carla: High-fidelity sensor simulations with inter- changeable dynamics for autonomous racing,","cited_arxiv_id":null,"evidence_quote":"Provides the R-CARLA method for reconstructing a digital-twin mesh from point clouds, used for 3D LiDAR simulation."},{"cited_title":"Implementation of the Pure Pursuit Path Tracking Algorithm,","cited_arxiv_id":null,"evidence_quote":"Defines the Pure Pursuit tracking law used to convert the centerline into steering commands."},{"cited_title":"Minimum cur- vature trajectory planning and control for an autonomous race car,","cited_arxiv_id":null,"evidence_quote":"Supplies the curvature-to-velocity formula used to set speed limits from local curvature and friction."},{"cited_title":"Amz driverless: The full autonomous racing system,","cited_arxiv_id":null,"evidence_quote":"Establishes the use of Delaunay triangulation for centerline extraction in racing, which MCTR adapts."}],"review_version":1}