Pith. sign in

REVIEW 3 major objections 6 minor 25 references

DTR: Delaunay Triangulation-based Racing for Scaled Autonomous Racing

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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%.

desk verdict 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. read the letter →

arxiv 2505.24320 v1 pith:LP4NYUF5 submitted 2025-05-30 cs.RO

Luca Tognoni , Neil Reichlin , Edoardo Ghignone , Nicolas Baumann , Steven Marty , Liam Boyle , Michele Magno This is my paper · ORCID
classification cs.RO
keywords autonomousracingreactivecontrolDelaunaytriangulationLiDARcenterlineextractionfollow-the-gapmaplessnavigationdead-endavoidance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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%.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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%.

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 (3)
  1. [Abstract and Table II] 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.
  2. [Section III-B] 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.
  3. [Section IV-A and Table II] 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.
minor comments (6)
  1. [Section III-C] 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.
  2. [Figure 2] 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.
  3. [Table II caption] 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.
  4. [Section II (Our Contribution)] 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.
  5. [Table I] 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.
  6. [Section V] 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.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; DTR's centerline extraction and trap avoidance are computed from raw LiDAR geometry rather than derived from the reported lap times.

full rationale

DTR's derivation chain is self-contained: the centerline is obtained from raw LiDAR points via Delaunay triangulation with geometric filters (Sec. III-A), distance-threshold wall segmentation (Sec. III-B), greedy nearest-neighbor connection, and spline smoothing, and the velocity command is the standard curvature-limited friction bound vadm = sqrt(mu amax_y / kappa). No stage fits or defines its output in terms of the target lap time, FTG performance, or the final trajectory, so no prediction reduces to a fitted input. The Sec. III-B claim that dead-ends are systematically avoided rests on the geometric assumption that dead-end regions produce triangles whose vertices belong to a single boundary-segment class. That is an empirical invariant about LiDAR segmentation, not a definitional reduction: the paper does not define a dead-end as 'a region whose triangles are single-class', and the experimental results in Table II are external measurements that could in principle falsify the invariant. A failure of this assumption would be a correctness or robustness risk, not circular reasoning. The self-citations [6] and [7] supply the hardware/testbed, the FTG baseline implementation, and the MAP comparison reference, but they do not supply the central DTR algorithm or the trap-avoidance result; the central claim retains independent content and is benchmarked against externally known FTG [9]. The stated limitation on obstacle avoidance is acknowledged in Sec. V and does not bear on circularity.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

The central claim depends on several hand-chosen geometric thresholds and domain assumptions about track geometry. No values are reported for segmentation and triangle filtering thresholds, so the exact procedure is underdetermined. There are no invented entities; the method reuses standard geometry and vehicle models.

free parameters (9)
  • boundary_segmentation_distance_threshold
    Distance threshold in Section III-B used to split consecutive LiDAR points into boundary classes; value not reported, and the trap-avoidance behavior depends on it.
  • isosceles_like_ratio_threshold
    Threshold for two longer sides being 'of similar length' in Section III-A; no value given.
  • pointedness_ratio_threshold
    Ratio between longest and shortest side above which a triangle is considered pointed in Section III-A; no value given.
  • area_threshold
    Minimum triangle area condition in Section III-A; no value given.
  • circumcenter_connection_max_distance
    Maximum allowed distance and behind-angle constraint in the greedy nearest-neighbor centerline construction in Section III-A; no values given.
  • lookahead_distance_scale
    Linear scaling coefficient of lookahead with speed in Section III-C; no value given.
  • friction_coefficient_mu
    Estimated by pulling the car laterally with a spring scale in Section III-C; used in the velocity formula, value not reported.
  • max_lateral_acceleration_underestimate_amaxy
    The amax_y term in the velocity formula in Section III-C is an underestimation chosen by hand; value not reported.
  • longitudinal_accel_decel_limits
    Predefined limits constraining speed transitions in Section III-C; values not reported.
assumptions (6)
  • domain assumption Delaunay circumcenters of track-spanning triangles approximate the track centerline.
    Central to Section III-A; follows [5], but assumed for this setup.
  • domain assumption Track boundaries are separable by a distance threshold on consecutive LiDAR range readings.
    Wall segmentation in Section III-B assumes boundary classes arise from thresholded distance jumps.
  • domain assumption Dead-end regions produce triangles whose vertices lie in a single boundary class, so rejecting such triangles avoids dead-ends.
    Core mechanism of Section III-B; not proven, only demonstrated on one track.
  • domain assumption The kinematic bicycle model adequately describes vehicle steering.
    Used in Section III-C to convert angular deviation to steering command.
  • domain assumption The formula vadm = sqrt(mu * amax_y / kappa) bounds achievable speed by lateral grip.
    Velocity profile uses this first-order grip model from [23]; assumes constant mu and no dynamic obstacle constraints.
  • domain assumption Laps with collisions can be excluded without biasing the lap-time comparison.
    Table II caption states only collision-free laps are counted; assumes the excluded laps are not systematically favorable to one controller.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DTR: Delaunay Triangulation-based Racing for Scaled Autonomous Racing." pith.science (2026). https://pith.science/paper/LP4NYUF5

@misc{pith2026250524320,
  author       = {Pith},
  title        = {Pith review of: DTR: Delaunay Triangulation-based Racing for Scaled Autonomous Racing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LP4NYUF5}},
  note         = {Machine review of arXiv:2505.24320}
}
read the original abstract

Reactive controllers for autonomous racing avoid the computational overhead of full ee-Think-Act autonomy stacks by directly mapping sensor input to control actions, eliminating the need for localization and planning. A widely used reactive strategy is FTG, which identifies gaps in LiDAR range measurements and steers toward a chosen one. While effective on fully bounded circuits, FTG fails in scenarios with incomplete boundaries and is prone to driving into dead-ends, known as FTG-traps. This work presents DTR, a reactive controller that combines Delaunay triangulation, from raw LiDAR readings, with track boundary segmentation to extract a centerline while systematically avoiding FTG-traps. Compared to FTG, the proposed method achieves lap times that are 70\% faster and approaches the performance of map-dependent methods. With a latency of 8.95 ms and CPU usage of only 38.85\% on the robot's OBC, DTR is real-time capable and has been successfully deployed and evaluated in field experiments.

Figures

Figures reproduced from arXiv: 2505.24320 by the authors.

Figure 1
Figure 1. Graphical illustration of the proposed Delaunay Triangulation-based Racing (DTR) reactive controller, which uses [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the Delaunay triangulation of the same setup as in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of the trajectories from left to right, driven by DTR (red), FTG (blue), and the non-reactive [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 18 canonical work pages

  1. [5]

    AMZ driverless: The full autonomous racing system,

    J. Kabzan, M. I. Valls, V . J. F. Reijgwart, H. F. C. Hendrikx, C. Ehmke, M. Prajapat, A. Bühler, N. Gos- ala, M. Gupta, R. Sivanesan, A. Dhall, E. Chisari, N. Karnchanachari, S. Brits, M. Dangel, I. Sa, R. Dubé, A. Gawel, M. Pfeiffer, A. Liniger, J. Lygeros, and R. Siegwart, “AMZ driverless: The full autonomous racing system,” J. field robot., no. rob.21...

    work page 2020

  2. [6]

    Forzaeth race stack—scaled autonomous head-to-head racing on fully commercial off-the-shelf hardware,

    N. Baumann, E. Ghignone, J. Kühne, N. Bastuck, J. Becker, N. Imholz, T. Kränzlin, T. Y . Lim, M. Lötscher, L. Schwarzenbach et al. , “Forzaeth race stack—scaled autonomous head-to-head racing on fully commercial off-the-shelf hardware,” Journal of Field Robotics, 2024

    work page 2024

  3. [1]

    Emergent ecosystem for radical innovation: Entrepreneurial probing at formula e,

    S. Jarvenpaa and W. Standaert, “Emergent ecosystem for radical innovation: Entrepreneurial probing at formula e,” in Proceedings of the 50th HICSS , 2017

    work page 2017

  4. [2]

    A survey of autonomous driving: Common practices and emerging technologies,

    E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda, “A survey of autonomous driving: Common practices and emerging technologies,” IEEE Access, vol. 8, pp. 58 443– 58 469, 2020

    work page 2020

  5. [3]

    F1tenth: An open-source evaluation environment for continuous control and reinforcement learning,

    M. O’Kelly, H. Zheng, D. Karthik, and R. Mang- haram, “F1tenth: An open-source evaluation environment for continuous control and reinforcement learning,” in NeurIPS 2019 Competition and Demonstration Track . PMLR, 2020, pp. 77–89

    work page 2019

  6. [4]

    Indy autonomous challenge – autonomous race cars at the handling limits,

    A. Wischnewski, M. Geisslinger, J. Betz, T. Betz, F. Fent, A. Heilmeier, L. Hermansdorfer, T. Herrmann, S. Huch, P. Karle, F. Nobis, L. Ögretmen, M. Rowold, F. Sauerbeck, T. Stahl, R. Trauth, M. Lienkamp, and B. Lohmann, “Indy autonomous challenge – autonomous race cars at the handling limits,” 2022. [Online]. Available: https://arxiv.org/abs/2202.03807

    arXiv 2022

  7. [7]

    Model- and acceleration- based pursuit controller for high-performance au- tonomous racing,

    J. Becker, N. Imholz, L. Schwarzenbach, E. Ghignone, N. Baumann, and M. Magno, “Model- and acceleration- based pursuit controller for high-performance au- tonomous racing,” in 2023 IEEE International Confer- ence on Robotics and Automation (ICRA) , 2023, pp. 5276–5283

    work page 2023

  8. [8]

    Siegwart, I

    R. Siegwart, I. R. Nourbakhsh, and D. Scaramuzza, Introduction to Autonomous Mobile Robots, 2nd ed. The MIT Press, 2011

    2011

Show all 25 references
  1. [9]

    A novel obstacle avoidance algorithm: “follow the gap method

    V . Sezer and M. Gokasan, “A novel obstacle avoidance algorithm: “follow the gap method”,” Robotics and Autonomous Systems , vol. 60, no. 9, pp. 1123–1134,

  2. [10]

    Information theoretic mpc for model-based reinforcement learning,

    G. Williams, N. Wagener, B. Goldfain, P. Drews, J. M. Rehg, B. Boots, and E. A. Theodorou, “Information theoretic mpc for model-based reinforcement learning,” pp. 1714–1721, 2017

    2017

  3. [11]

    Optimization- based autonomous racing of 1:43 scale rc cars,

    A. Liniger, A. Domahidi, and M. Morari, “Optimization- based autonomous racing of 1:43 scale rc cars,” Optimal Control Applications and Methods , vol. 36, no. 5, pp. 628–647, 2015. [Online]. Available: https: //onlinelibrary.wiley.com/doi/abs/10.1002/oca.2123

    2015 doi

  4. [12]

    Cautious model predictive control using gaussian process regres- sion,

    L. Hewing, J. Kabzan, and M. N. Zeilinger, “Cautious model predictive control using gaussian process regres- sion,” IEEE Transactions on Control Systems Technology, vol. 28, no. 6, pp. 2736–2743, 2020

    2020

  5. [13]

    Learning to drive in a day,

    A. Kendall, J. Hawke, D. Janz, P. Mazur, D. Reda, J.-M. Allen, V .-D. Lam, A. Bewley, and A. Shah, “Learning to drive in a day,” in 2019 International Conference on Robotics and Automation (ICRA) , 2019, pp. 8248–8254

    2019

  6. [14]

    Autonomous vehicles on the edge: A survey on autonomous vehicle racing,

    J. Betz, H. Zheng, A. Liniger, U. Rosolia, P. Karle, M. Behl, V . Krovi, and R. Mangharam, “Autonomous vehicles on the edge: A survey on autonomous vehicle racing,” IEEE Open Journal of Intelligent Transportation Systems, vol. 3, pp. 458–488, 2022

    2022

  7. [15]

    Residual policy learning facilitates efficient model-free autonomous racing,

    R. Zhang, J. Hou, G. Chen, Z. Li, J. Chen, and A. Knoll, “Residual policy learning facilitates efficient model-free autonomous racing,” IEEE Robotics and Automation Let- ters, vol. 7, no. 4, pp. 11 625–11 632, 2022

    2022

  8. [16]

    Deep Drone Acrobat- ics,

    E. Kaufmann, A. Loquercio, R. Ranftl, M. Müller, V . Koltun, and D. Scaramuzza, “Deep Drone Acrobat- ics,” in Proceedings of Robotics: Science and Systems , Corvalis, Oregon, USA, July 2020

    2020

  9. [17]

    Learning high- speed flight in the wild,

    A. Loquercio, E. Kaufmann, R. Ranftl, M. Müller, V . Koltun, and D. Scaramuzza, “Learning high- speed flight in the wild,” Science Robotics , vol. 6, no. 59, p. eabg5810, 2021. [Online]. Available: https: //www.science.org/doi/abs/10.1126/scirobotics.abg5810

    2021 doi

  10. [18]

    The dynamic win- dow approach to collision avoidance,

    D. Fox, W. Burgard, and S. Thrun, “The dynamic win- dow approach to collision avoidance,” IEEE Robotics & Automation Magazine, vol. 4, no. 1, pp. 23–33, 1997

    1997

  11. [19]

    A local trajectory planning and control method for autonomous vehicles based on the rrt al- gorithm,

    S. Feraco, S. Luciani, A. Bonfitto, N. Amati, and A. Tonoli, “A local trajectory planning and control method for autonomous vehicles based on the rrt al- gorithm,” in 2020 AEIT International Conference of Electrical and Electronic Technologies for Automotive (AEIT AUTOMOTIVE),...

    2020

  12. [20]

    Incremental sampling- based algorithms for optimal motion planning,

    S. Karaman and E. Frazzoli, “Incremental sampling- based algorithms for optimal motion planning,” in Pro- ceedings of Robotics: Science and Systems , Zaragoza, Spain, June 2010

    2010

  13. [21]

    Ros-based localization of a race vehicle at high-speed using lidar,

    T. Stahl, A. Wischnewski, J. Betz, and M. Lienkamp, “Ros-based localization of a race vehicle at high-speed using lidar,” in E3S Web of Conferences , vol. 95. EDP Sciences, 2019, p. 04002

    2019

  14. [22]

    Implementation of the Pure Pursuit Path Tracking Algorithm,

    R. Coulter, “Implementation of the Pure Pursuit Path Tracking Algorithm,” DTIC Document, 1992

    1992

  15. [23]

    Minimum cur- vature trajectory planning and control for an autonomous race car,

    A. Heilmeier, A. Wischnewski, L. Hermansdorfer, J. Betz, M. Lienkamp, and B. Lohmann, “Minimum cur- vature trajectory planning and control for an autonomous race car,” Vehicle System Dynamics , vol. 58, no. 10, pp. 1497–1527, Oct. 2020, publisher Copyright: © 2019 Informa UK L...

    2020

  16. [24]

    Learning-based on-track system identification for scaled autonomous rac- ing in under a minute,

    O. Dikici, E. Ghignone, C. Hu, N. Baumann, L. Xie, A. Carron, M. Magno, and M. Corno, “Learning-based on-track system identification for scaled autonomous rac- ing in under a minute,” IEEE Robotics and Automation Letters, vol. 10, no. 2, pp. 1984–1991, 2025

    1984

  17. [2012]

    Available: https://www.sciencedirect

    [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0921889012000838

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.