{"id":"2196fbd1-ad44-4b0e-b28f-c6a2f4547959","arxiv_id":"2505.24320","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"DTR, a mapless LiDAR reactive controller, extracts a track centerline with Delaunay triangulation plus boundary segmentation, avoiding dead-ends and cutting lap time from 9.45s to 5.79s versus Follow-The-Gap.","lead":"A new reactive controller for scaled autonomous race cars uses Delaunay triangulation of LiDAR points and track-boundary segmentation to extract a centerline and avoid dead-end traps, improving lap time from 9.45s to 5.79s over Follow-The-Gap. It runs on the car's onboard computer with 8.95ms average latency and closes part of the gap to map-based controllers without needing a map.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'systematic FTG-trap avoidance' claim in Sec. III-B rests on an unvalidated distance-threshold segmentation invariant; no proof or ablation shows that dead-end circumcenters are always rejected, so the central advantage over FTG is not established.","rationale":"The reader's weakest assumption concerns distance-threshold segmentation reliability, and my review independently converges on that same point: the paper's core novelty is the claim that dead-end regions always produce single-class triangles and are therefore rejected. In good faith, the method is plausible and the field results are suggestive, but the argument does not contain a proof or a sensitivity analysis for the segmentation invariant. The closest the paper comes to support is one qualitative dead-end scenario and five lap-time measurements, which cannot establish 'systematic' avoidance. I considered whether a different issue, such as the absence of code or the small number of laps, is more load-bearing; those limit reproducibility but do not attack the mechanism itself. The segmentation invariant is more fundamental because if it fails, the central claimed advantage over FTG disappears. My proposed test would settle the concern by measuring how wide a safe operating region exists for the threshold across varied dead-end geometries. This does not move the verdict away from CONDITIONAL; it reinforces the condition that the authors must provide code, thresholds, and a robustness evaluation before the claim can be treated as a benchmark result.","tokens_in":8302,"tokens_out":10606,"duration_ms":156047,"concrete_test":"On the recorded LiDAR rosbag or a synthetic replica of the test track, sweep the segmentation distance threshold T over a range that includes the (unreported) deployed value. For each T, run the Sec. III-B filtering and record whether any accepted circumcenter lies inside the dead-end or whether the greedy centerline enters the dead-end. Report the interval of T values that produce dead-end-free centerlines. Then repeat with synthetic dead-end layouts varying depth, width, and entrance angle. If no T interval is safely nonempty across layouts, or the deployed threshold has zero margin, the 'systematically avoiding FTG-traps' claim is not supported and should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DTR 'systematically' avoids FTG-traps via the Sec. III-B constraint that only circumcenters of triangles whose vertices span at least two boundary-segment classes are used for centerline extraction. That constraint is sound only if wall segmentation by distance thresholding has a specific invariant: dead-end interiors are always labeled as a single class, and every accepted triangle whose circumcenter lies in a dead-end must have all three vertices in that single class. The paper gives no proof of this invariant and reports no threshold values. The invariant is fragile in at least two concrete ways. First, point spacing in a LiDAR scan grows with range; a far wall in a deep dead-end can have inter-point spacing larger than the threshold, splitting one physical wall into several classes, so triangles inside the dead-end can become mixed-class and be accepted. Second, even with perfect single-class labeling of a dead-end, triangles near its mouth can have one or two vertices on the adjacent outer track wall (a different class) and a circumcenter that lies inside or leads the greedy nearest-neighbor centerline into the dead-end; the single-class filter alone does not rule this out. Because the claimed 70% lap-time improvement and the comparison to FTG are presented as consequences of trap avoidance, this unvalidated heuristic is load-bearing. The one qualitative scenario in Fig. 3 and five collision-free laps do not establish 'systematic' avoidance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DTR, a reactive controller for scaled autonomous racing that applies Delaunay triangulation to raw 2D LiDAR scans, segments the track boundaries by distance thresholding, and uses the circumcenters of triangles whose vertices span multiple boundary classes to extract a centerline without a map. Steering is computed via a kinematic bicycle model with a speed-dependent lookahead, and the velocity target is set from local curvature and an estimated friction coefficient. The method is deployed on a 1:10-scale car and compared with the classical FTG controller and the map-based MAP and MPC controllers. Reported results are a mean lap time of 5.79 s versus 9.45 s for FTG and 4.19 s for MAP, with a controller latency of 8.95 ms and CPU utilization of 38.85%.","tokens_in":8598,"tokens_out":3947,"duration_ms":49878,"significance":"If the reported performance is correct, DTR is a meaningful contribution to mapless reactive racing: it approximately halves the lap-time gap to full map-based approaches while running onboard in real time, and it does so with a geometric, interpretable algorithm. The paper's strengths are its real-world closed-loop deployment on standard hardware, the use of an open-source baseline (FTG from [6]) for direct comparison, and the inclusion of computational metrics (latency and CPU) alongside lap times. The main limitation is that the central claimed advantage—systematic avoidance of FTG-traps—rests on a heuristic segmentation invariant that is not formally analyzed, and the experimental evidence for it is a single qualitative trajectory and a small number of laps.","major_comments":[{"comment":"The abstract's claim that DTR achieves lap times that are '70% faster' than FTG is not supported by the reported numbers. From Table II, the reduction from 9.45 s to 5.79 s is a 38.7% decrease relative to FTG's lap time, or a 63% decrease relative to DTR's own lap time; neither equals 70%. This headline metric should be corrected and the definition of 'faster' (e.g., relative to which baseline) should be stated explicitly.","section":"Abstract and Table II"},{"comment":"The claim that the wall-segmentation constraint 'ensures that dead-ends are avoided' and enables 'systematic avoidance of FTG-traps' is not established. No proof or ablation is provided for the invariant that triangles inside dead-ends always have all three vertices in a single boundary-segment class, nor for the exclusion of triangles near a dead-end mouth whose vertices lie on different classes and whose circumcenters lead into the dead-end. Because LiDAR point spacing grows with range, a distant wall can be split into multiple classes by the distance threshold, allowing mixed-class triangles inside the dead-end to pass the filter. The evidence offered—one qualitative trajectory in Fig. 3 and five collision-free laps—is insufficient to support the word 'systematic.' Please provide either a formal argument under explicit assumptions about the segmentation or an ablation/sensitivity study over a range of dead-end geometries.","section":"Section III-B"},{"comment":"The experimental protocol is too limited to support the quantitative comparisons. Only 5 laps per controller are reported, lap times are computed from 'laps without collisions' without stating how many collision laps were discarded, and all controllers were manually tuned. With five samples, the reported standard deviations give little statistical confidence, and no significance test is provided for the lap-time differences. Please report the total number of runs, the number of excluded collision laps, the tuning procedure, and ideally results on more than one track or repeated trials across different days.","section":"Section IV-A and Table II"}],"minor_comments":[{"comment":"The velocity formula is typeset ambiguously as vadm = q µ amaxy κ−1; it should be written as vadm = sqrt(µ · amax_y / κ), and all variables (especially amax_y and κ) should be defined immediately before the equation.","section":"Section III-C"},{"comment":"The caption states that the inner track segment is light-blue and the outer segment is orange, but these colors are not visible in the grayscale printing and the figure does not clearly label the segmentation classes; please use distinct labels or a colorblind-safe palette.","section":"Figure 2"},{"comment":"The caption notes that 'the maximum amount for U is 800%,' but this is unexplained; please clarify that this reflects multi-core CPU accounting or a different normalization, and ensure the metric is defined in the text.","section":"Table II caption"},{"comment":"The text claims the approach requires 'minimal tuning,' yet Section IV-A states that all controllers were manually tuned to achieve the lowest lap time; please reconcile these statements, for instance by reporting the number and sensitivity of the hand-tuned thresholds.","section":"Section II (Our Contribution)"},{"comment":"In Table I, the MAP and MPC rows have a dash under 'Speed Deficit,' but the caption says this metric is limited to mapless methods; please clarify whether these rows are intended as reference entries only and remove the ambiguity.","section":"Table I"},{"comment":"The conclusion states that DTR 'effectively avoiding dead-ends' is demonstrated, but the paper shows only a single dead-end scenario; please temper this claim to match the empirical evidence.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The circularity concern raised by the stress-test note does not land: the controller is not derived from its own outputs, and the cited works [6,7] provide the comparison baseline and the platform, not the claimed result. The key issue is the strength of the 'systematic FTG-trap avoidance' claim relative to the thin experimental and theoretical support; this should be resolved before acceptance. The paper is within scope for a conference-level robotics venue, provided the load-bearing assertions are either proven or appropriately qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: DTR is a worthwhile incremental step for reactive autonomous racing. The core idea—Delaunay triangulation of LiDAR points plus a wall-segmentation filter that only keeps triangles spanning two boundary classes—is a legitimate extension of the centerline extraction in Kabzan et al. [5]. The paper is honest about the trade-off: it is slower than map-based MAP/MPC but much faster than FTG on the same open-source stack, and the real-car deployment with latency and CPU numbers is concrete evidence that the method runs onboard. The friction-based velocity limiting is also sensible.\n\nThe main soft spot is exactly what the stress-test note flags. The claim of \"systematically avoiding FTG-traps\" rests on an unproven invariant: that dead-end interiors always produce triangles with all vertices in a single boundary class, and that mixed-class triangles near a dead-end mouth never lead the greedy centerline into it. Given that LiDAR point spacing grows with range, the distance-threshold segmentation can easily split one physical wall into several classes, so the filter's guarantee is fragile. The paper gives no proof, no threshold values, and only one qualitative scenario (Fig. 3) to support the claim. That is a real weakness, and the \"systematic\" language should be softened unless the authors add an ablation or a formal argument.\n\nThe evaluation also has the usual limitations: five laps, collision-free laps only, all controllers manually tuned, no code or data released. The \"70% faster\" phrasing in the abstract is inaccurate: 9.45 s to 5.79 s is a 38.7% lap-time reduction (about 63% higher average speed), not 70%. Minor, but it should be corrected.\n\nThese issues do not sink the paper. The lap-time improvement over FTG is real on the published numbers, and the main algorithmic contribution is novel enough to be worth discussing. The paper is best for researchers working on mapless reactive navigation or resource-constrained racing platforms. It deserves a serious referee who can push for more data and a more careful claim, not a desk reject. I would engage with it, though I would not cite it in the next year without seeing an ablation or broader evaluation.","headline":"A plausible incremental win for mapless reactive racing, with a real field demo and a genuine new filter idea, but the central trap-avoidance claim is overreaching and the evaluation is thinner than the abstract suggests.","tokens_in":732,"tokens_out":950,"would_cite":false,"duration_ms":32970,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A purely reactive controller can triangulate raw LiDAR scans into a race-track centerline, avoid dead-end traps, and cut the lap-time gap to map-based racing from 126% to 38%.","keywords":["autonomous racing","reactive control","Delaunay triangulation","LiDAR","centerline extraction","follow-the-gap","mapless navigation","dead-end avoidance"],"falsifier":"Run the controller on a track with a dead-end whose opening is widened step by step (or whose walls meet at a shallow angle) until the distance-threshold segmentation merges the dead-end walls with the visible track wall into a single segment class; if the car then enters the dead-end, the class-spanning constraint has failed. A cheaper in-silico version: sweep the dead-end opening width in simulated LiDAR scans, count the accepted triangles whose circumcenters lie inside the dead-end, and find the smallest opening at which that count becomes nonzero.","tokens_in":8044,"feed_emoji":"🏎️","tokens_out":11928,"duration_ms":123998,"temperature":0.7,"pith_summary":"The paper sets out to show that a mapless, reactive controller can drive a 1:10-scale race car near the pace of map-based systems by extracting the track centerline directly from each LiDAR scan, using Delaunay triangulation together with a distance-threshold segmentation of the track walls. The central move is to keep only triangles whose vertices span at least two distinct wall-segment classes, so that dead-end regions, whose triangles have vertices in a single class, are systematically rejected; this fixes the well-known failure mode of Follow-The-Gap (FTG). Field results on the test course give 5.79 s laps for DTR against 9.45 s for FTG (the authors report 70% faster lap times) and 4.19 s for a map-based pursuit controller, meaning the lap-time deficit to map-dependent racing falls from 126% to 38%. Because the whole pipeline runs on the car's onboard computer at 8.95 ms average latency and 38.85% CPU usage, the claim is that reactive racing no longer has to sacrifice map-free autonomy for competitiveness.","feed_headline":"Mapless race car dodges dead-ends, nears map-based lap times","feed_subtitle":"Delaunay-triangulated LiDAR cuts the lap-time gap to mapped racers from 126 percent to 38.","key_machinery":"The load-bearing object is the class-filtered Delaunay triangulation. Delaunay triangulation connects the subsampled LiDAR points into triangles with empty circumcircles; the paper partitions the scan into consecutive wall segments by distance thresholding and admits a triangle only if its vertices belong to at least two distinct segment classes, on top of geometric heuristics (isosceles-like, pointedness, area) that keep only triangles spanning the track width. That single constraint is what makes dead-ends impossible: a dead-end corridor produces triangles whose vertices all sit on the same wall class, so their circumcenters are rejected and the greedy nearest-neighbor centerline never extends into the trap. The chain that carries the rest of the argument is: circumcenters of accepted triangles, greedy nearest-neighbor ordering, Savitzky-Golay smoothing, spline with curvature, lookahead steering by kinematic bicycle model, and a speed target $v_{\\mathrm{adm}} = \\sqrt{\\mu\\, a^{\\max}_y\\, \\kappa^{-1}}$ from curvature and estimated friction.","core_discovery":"On the paper's own terms, the discovery is that Delaunay triangulation of raw 2D LiDAR points, restricted by a wall-segmentation constraint, yields a centerline good enough to race on. The pipeline subsamples the scan, triangulates it, keeps triangles whose geometry suggests they span the track width (isosceles-like and pointedness conditions, or a large area), and then, as the key step, requires every kept triangle to have vertices in at least two distinct boundary classes obtained by distance-threshold segmentation of the walls. Circumcenters of the surviving triangles are chained by a greedy nearest-neighbor search into an ordered centerline, smoothed with a Savitzky-Golay filter and fitted with a spline; a lookahead point on that spline sets the steering via a kinematic bicycle model, and the admissible speed is derived from curvature and a hand-measured friction coefficient ($v_{\\mathrm{adm}} = \\sqrt{\\mu\\, a^{\\max}_y\\, \\kappa^{-1}}$). The consequence, demonstrated on one course, is that the controller never enters the dead-end that traps FTG, produces consistent laps at 5.79 s against 9.45 s for FTG and 4.19 s for the map-based MAP controller, and closes the map-based performance deficit from 126% to 38%.","pith_inferences":["An editor's read: the distance-threshold segmentation is validated on only one track, so the property worth stress-testing is a dead-end whose opening is wide enough that thresholding merges its walls with the main track wall into a single class; in that regime the class-spanning filter would admit trap triangles and the car would enter the dead-end.","The friction coefficient is measured by hand with a spring scale before the run, so the claimed speed profile silently depends on a per-surface calibration step; on a different surface the same controller would need a new pull-test to reproduce the lap time.","Part of the residual 38% gap to the map-based controller is structural: DTR follows the geometric centerline, whereas the map-based baseline tracks a globally optimized minimum-curvature racing line, so offsetting the extracted line toward corner apexes is a natural next lever for closing the gap.","The authors list obstacle avoidance as a limitation; a concrete extension their mechanism suggests is to feed detected obstacles into the triangulation as extra segment classes, so the same class-spanning rule that rejects dead-ends could also reject cells occupied by opponents."],"forward_implications":["On the tested course DTR laps in 5.79 s against 9.45 s for FTG, a 70% lap-time improvement as the authors report, so a purely reactive, mapless controller beats gap-following by a wide margin.","The lap-time deficit to map-based controllers falls from 126% to 38% relative to the MAP baseline (4.19 s), while requiring no localization, no prior map, and no state estimator.","The pipeline is real-time capable on the car's onboard computer: 8.95 ms average latency and 38.85% CPU utilization, versus 16.86 ms and 68.27% CPU for the MPC controller, which also needs roughly 180% additional CPU for state estimation.","DTR produces smooth, repeatable trajectories across laps, whereas FTG oscillates laterally and falls into the dead-end trap located in the middle of the track.","Because the method consumes range measurements rather than intensity images, it should transfer to other 2D range sensors; the paper names radar as a candidate."],"supporting_citations":[{"why":"Supplies the 1:10-scale race platform, the FTG baseline implementation, and the map-based MAP/MPC comparison numbers and setup.","marker":"[6]"},{"why":"Defines Follow-The-Gap, the reactive baseline whose dead-end failure mode (FTG-traps) DTR is designed to fix.","marker":"[9]"},{"why":"Provides the map-based pursuit controller that serves as the performance reference (4.19 s laps) for the 126%/38% deficit comparison.","marker":"[7]"},{"why":"Introduces Delaunay-triangulation-based centerline extraction from range measurements, the technique DTR adapts and extends with wall segmentation.","marker":"[5]"},{"why":"Supplies the boxed subsampling of LiDAR scans used before triangulation.","marker":"[21]"},{"why":"Provides the pure-pursuit kinematic bicycle lookahead steering law DTR uses for control.","marker":"[22]"},{"why":"Supplies the friction- and curvature-based admissible-velocity formula that DTR adopts for speed control.","marker":"[23]"}],"fun_headline_variants":["Delaunay-triangulated LiDAR steers racer around dead-end traps","Reactive racer avoids FTG traps, laps 70% faster with Delaunay","Mapless reactive controller closes lap-time gap to mapped racers","Delaunay triangulation of LiDAR yields racing centerline without maps","DTR: reactive steering avoids dead-ends, nears map-based lap times"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole dead-end-avoidance argument rests on the assumption that the distance-threshold step splits the LiDAR scan into wall segments cleanly enough that every dead-end produces only triangles whose vertices belong to a single segment class; if the threshold merges a dead-end's walls with the main track wall or splits one wall into two classes, the filter either lets a trap triangle through or throws away a good one.","fun_headline_variants_meta":{"raw":{"variants":["Delaunay-triangulated LiDAR steers racer around dead-end traps","Reactive racer avoids FTG traps, laps 70% faster with Delaunay","Mapless reactive controller closes lap-time gap to mapped racers","Delaunay triangulation of LiDAR yields racing centerline without maps","DTR: reactive steering avoids dead-ends, nears map-based lap times"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2567,"prompt_tokens":1002,"completion_tokens":1565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":618,"tokens_out":1565,"duration_ms":13028,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:25:53.847497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller on a track with a dead-end whose opening is widened step by step (or whose walls meet at a shallow angle) until the distance-threshold segmentation merges the dead-end walls with the visible track wall into a single segment class; if the car then enters the dead-end, the class-spanning constraint has failed. A cheaper in-silico version: sweep the dead-end opening width in simulated LiDAR scans, count the accepted triangles whose circumcenters lie inside the dead-end, and find the smallest opening at which that count becomes nonzero.","supporting_citations":[{"cited_title":"Forzaeth race stack—scaled autonomous head-to-head racing on fully commercial off-the-shelf hardware,","cited_arxiv_id":null,"evidence_quote":"Supplies the 1:10-scale race platform, the FTG baseline implementation, and the map-based MAP/MPC comparison numbers and setup."},{"cited_title":"A novel obstacle avoidance algorithm: “follow the gap method","cited_arxiv_id":null,"evidence_quote":"Defines Follow-The-Gap, the reactive baseline whose dead-end failure mode (FTG-traps) DTR is designed to fix."},{"cited_title":"Model- and acceleration- based pursuit controller for high-performance au- tonomous racing,","cited_arxiv_id":null,"evidence_quote":"Provides the map-based pursuit controller that serves as the performance reference (4.19 s laps) for the 126%/38% deficit comparison."},{"cited_title":"AMZ driverless: The full autonomous racing system,","cited_arxiv_id":null,"evidence_quote":"Introduces Delaunay-triangulation-based centerline extraction from range measurements, the technique DTR adapts and extends with wall segmentation."},{"cited_title":"Ros-based localization of a race vehicle at high-speed using lidar,","cited_arxiv_id":null,"evidence_quote":"Supplies the boxed subsampling of LiDAR scans used before triangulation."},{"cited_title":"Implementation of the Pure Pursuit Path Tracking Algorithm,","cited_arxiv_id":null,"evidence_quote":"Provides the pure-pursuit kinematic bicycle lookahead steering law DTR uses for control."},{"cited_title":"Minimum cur- vature trajectory planning and control for an autonomous race car,","cited_arxiv_id":null,"evidence_quote":"Supplies the friction- and curvature-based admissible-velocity formula that DTR adopts for speed control."}],"review_version":1}