pith. sign in

arxiv: 2603.07126 · v2 · pith:AY3T5FF5new · submitted 2026-03-07 · 💻 cs.RO

Efficient Trajectory Optimization for Autonomous Racing via Formula-1 Data-Driven Initialization

Samir Shehadeh , Lukas Kutsch , Nils Dengler , Sicong Pan , Maren Bennewitz This is my paper

Pith reviewed 2026-05-21 11:54 UTC · model grok-4.3

classification 💻 cs.RO
keywords autonomous racingtrajectory optimizationneural network initializationFormula 1 telemetryraceline predictionoptimal control solverdata-driven prior
0
0 comments X

The pith

A neural network trained on reconstructed Formula 1 telemetry initializes trajectory optimizers for autonomous racing, speeding convergence while keeping the same minimum lap times.

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

The paper aims to show that seeding a minimum-time optimal control solver with a predicted raceline from a neural network trained on expert F1 driving data leads to faster solver performance than using simple geometric paths like the centerline. This addresses the sensitivity of trajectory optimization to poor initial guesses, which can cause slow convergence or suboptimal results in autonomous racing. By building a dataset from GPS telemetry across 17 tracks and training the network to output raceline offsets based only on local geometry, the method transfers real-world expert behavior without needing vehicle-specific models. A sympathetic reader would care because faster optimization enables more responsive planning in high-speed racing scenarios.

Core claim

The central claim is that predicting an expert-like raceline offset directly from local track geometry using a neural network trained on Formula 1 telemetry provides an effective initialization for minimum-time optimal control solvers, resulting in accelerated convergence and reduced runtime on all 17 tested tracks without altering the quality of the final optimized trajectory.

What carries the argument

The neural network that predicts raceline offsets from local track geometry, serving as a data-driven prior for the optimization solver.

Load-bearing premise

The initializations from the neural network trained only on F1 telemetry remain effective even when the solver applies the specific dynamics and constraints of the autonomous vehicle being controlled.

What would settle it

Running the optimization on a new track or with a substantially different vehicle and finding that the learned initialization leads to slower convergence or longer runtimes than the minimum-curvature baseline would disprove the effectiveness claim.

Figures

Figures reproduced from arXiv: 2603.07126 by Lukas Kutsch, Maren Bennewitz, Nils Dengler, Samir Shehadeh, Sicong Pan.

Figure 1
Figure 1. Figure 1: Overview of our proposed learning-informed initialization [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The racing track is described by a centerline parameterized [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Network architecture for raceline prediction. History and future windows are encoded using dilated temporal convolutional networks [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: a shows smooth and sta [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Neural network training behavior and qualitative raceline prediction results. (a) Training and validation loss curves show stable [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative example of the optimized raceline on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Trajectory optimization is a central component of fast and efficient autonomous racing. However practical optimization pipelines remain highly sensitive to initialization and may converge slowly or to suboptimal local solutions when seeded with heuristic trajectories such as the centerline or minimum-curvature paths. To address this limitation, we leverage expert driving behavior as a initialization prior and propose a learning-informed initialization strategy based on real-world Formula~1 telemetry. To this end, we first construct a multi-track Formula~1 trajectory dataset by reconstructing and aligning noisy GPS telemetry to a standardized reference-line representation across 17 tracks. Building on this, we present a neural network that predicts an expert-like raceline offset directly from local track geometry, without explicitly modeling vehicle dynamics or forces. The predicted raceline is then used as an informed seed for a minimum-time optimal control solver. Experiments on all 17 tracks demonstrate that the learned initialization accelerates solver convergence and significantly reduces runtime compared to traditional geometric baselines, while preserving the final optimized lap time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that a neural network trained on reconstructed Formula 1 GPS telemetry across 17 tracks can predict expert-like raceline offsets from local track geometry alone; when used to initialize a minimum-time optimal control solver, this yields faster convergence and lower runtime than geometric baselines (centerline, minimum-curvature) while preserving the final optimized lap time.

Significance. If the transfer from F1-derived initializations to the target autonomous-vehicle dynamics holds, the approach would meaningfully reduce the practical barrier of poor initialization in real-time racing optimization. The multi-track dataset construction and end-to-end evaluation on 17 tracks constitute a concrete empirical contribution; reproducible code or parameter-free derivations are not reported.

major comments (3)
  1. [Abstract and §4] Abstract and §4 (Experiments): the central claim that the learned initialization 'accelerates solver convergence and significantly reduces runtime' on all 17 tracks is load-bearing, yet the manuscript supplies no quantitative metrics (e.g., iteration counts, wall-clock deltas, feasibility margins, or distance to the solver optimum) nor statistical significance tests; without these the reported speedup cannot be evaluated.
  2. [§3] §3 (Method): the network is trained solely on F1 telemetry offsets and contains no vehicle parameters, tire models, or mass distribution. The claim that the resulting initialization remains effective for a downstream OCP whose dynamics and constraints belong to a different autonomous platform therefore rests on an untested transfer assumption; no ablation across vehicle models or quantification of basin-of-attraction distance is provided, which directly threatens the runtime benefit if mismatch occurs.
  3. [§3.2 and §4] §3.2 and §4: the paper states that the network predicts 'directly from local track geometry, without explicitly modeling vehicle dynamics or forces,' yet the solver uses the target vehicle's full dynamics. This mismatch is not addressed by any cross-validation experiment, leaving the independence of the initialization benefit from circular fitting unverified.
minor comments (2)
  1. [§2] §2 (Related Work): a brief discussion of prior learning-based warm-starting techniques in model-predictive control would help situate the contribution.
  2. [Figure captions and §4] Figure captions and §4: clarify whether reported runtimes are averaged over multiple solver initializations or single runs, and include standard deviations.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and valuable suggestions. We address the major comments point by point below, providing clarifications and indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the central claim that the learned initialization 'accelerates solver convergence and significantly reduces runtime' on all 17 tracks is load-bearing, yet the manuscript supplies no quantitative metrics (e.g., iteration counts, wall-clock deltas, feasibility margins, or distance to the solver optimum) nor statistical significance tests; without these the reported speedup cannot be evaluated.

    Authors: We acknowledge that while convergence behavior is illustrated through figures in §4, explicit quantitative summaries such as average iteration counts, wall-clock time reductions, and statistical significance were not tabulated. In the revised manuscript, we will include a new table in §4 reporting these metrics across all 17 tracks, including mean and standard deviation of runtime improvements and results from statistical tests to allow proper evaluation of the speedup. revision: yes

  2. Referee: [§3] §3 (Method): the network is trained solely on F1 telemetry offsets and contains no vehicle parameters, tire models, or mass distribution. The claim that the resulting initialization remains effective for a downstream OCP whose dynamics and constraints belong to a different autonomous platform therefore rests on an untested transfer assumption; no ablation across vehicle models or quantification of basin-of-attraction distance is provided, which directly threatens the runtime benefit if mismatch occurs.

    Authors: The proposed method intentionally decouples the initialization from vehicle-specific parameters by learning a general expert-like offset from track geometry alone. This design choice enables transfer to different platforms, as demonstrated by the consistent performance gains observed when applying the F1-derived initialization to our target autonomous racing dynamics across 17 diverse tracks. We agree that additional ablations would be beneficial; however, given the scope of the current work focusing on one platform, we will expand the discussion in §3 and §5 to explicitly address the transfer assumption and its empirical support from the multi-track results, while noting the absence of cross-model ablations as a limitation. revision: partial

  3. Referee: [§3.2 and §4] §3.2 and §4: the paper states that the network predicts 'directly from local track geometry, without explicitly modeling vehicle dynamics or forces,' yet the solver uses the target vehicle's full dynamics. This mismatch is not addressed by any cross-validation experiment, leaving the independence of the initialization benefit from circular fitting unverified.

    Authors: The network provides a dynamics-agnostic warm-start based on expert geometry, which the full-dynamics OCP then optimizes. The independence from circular fitting is supported by the superior performance relative to purely geometric baselines (centerline and minimum-curvature) that also do not model dynamics. To further verify this, we will add a cross-validation experiment in the revised §4, where we compare the learned initialization against baselines under the same dynamics, and include analysis showing that the benefit persists. revision: yes

Circularity Check

0 steps flagged

No significant circularity; data-driven initialization evaluated on independent solver metrics

full rationale

The paper constructs an F1 telemetry dataset and trains a neural network to predict raceline offsets from local geometry. This output seeds a separate minimum-time OCP solver whose convergence runtime and final cost are measured directly. No equations, fitted parameters, or self-citations reduce the reported speedup to a quantity defined by the same inputs; the central empirical claim is tested against external solver behavior and remains independent of circular fitting. Minor self-citation risk is possible in related work but is not load-bearing here.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Limited information available from abstract only; the approach implicitly assumes accurate reconstruction and alignment of noisy GPS data across tracks and that expert F1 behavior transfers to the target vehicle without dynamics modeling.

axioms (2)
  • domain assumption Reconstructed and aligned GPS telemetry accurately represents expert driving behavior transferable to autonomous vehicle optimization.
    Invoked when building the dataset and training the network to predict racelines without vehicle dynamics.
  • domain assumption Local track geometry alone is sufficient input for predicting useful initialization offsets.
    Stated in the description of the neural network that predicts directly from geometry.

pith-pipeline@v0.9.0 · 5708 in / 1466 out tokens · 40019 ms · 2026-05-21T11:54:17.403546+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

  1. [1]

    A Survey of Motion Planning and Control Techniques for Self-Driving Urban Vehicles,

    B. Paden, M. ˇC´ap, S. Z. Yong, D. Yershov, and E. Frazzoli, “A Survey of Motion Planning and Control Techniques for Self-Driving Urban Vehicles,”IEEE Transactions on intelligent vehicles, vol. 1, 2016

    work page 2016

  2. [2]

    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, 2022

    work page 2022

  3. [3]

    Pacejka,Tire and vehicle dynamics

    H. Pacejka,Tire and vehicle dynamics. Elsevier, 2005

    work page 2005

  4. [4]

    Rajamani,Vehicle dynamics and control

    R. Rajamani,Vehicle dynamics and control. Springer, 2006

    work page 2006

  5. [5]

    Optimal control for a formula one car with variable parameters,

    G. Perantoni and D. J. Limebeer, “Optimal control for a formula one car with variable parameters,”Vehicle System Dynamics, vol. 52, 2014

    work page 2014

  6. [6]

    J. T. Betts,Practical methods for optimal control and estimation using nonlinear programming. SIAM, 2010

    work page 2010

  7. [7]

    Minimum curvature trajectory planning and control for an autonomous race car,

    A. Heilmeier, A. Wischnewski, L. Hermansdorfer, J. Betz, M. Lienkamp, and B. Lohmann, “Minimum curvature trajectory planning and control for an autonomous race car,”Vehicle System Dynamics, 2020

    work page 2020

  8. [8]

    A sequential two-step algorithm for fast generation of vehicle racing trajectories,

    N. R. Kapania, J. Subosits, and J. Christian Gerdes, “A sequential two-step algorithm for fast generation of vehicle racing trajectories,” Journal of Dynamic Systems, Measurement, and Control, vol. 138, no. 9, 2016

    work page 2016

  9. [9]

    Learning model predictive control for iterative tasks: A data-driven control framework,

    U. Rosolia and F. Borrelli, “Learning model predictive control for iterative tasks: A data-driven control framework,”IEEE Transactions on Automatic Control, vol. 63, no. 7, 2017

    work page 2017

  10. [10]

    Advancing autonomous racing: A comprehensive survey of the roboracer (f1tenth) platform,

    I. Charles, H. Maghsoumi, and Y . Fallah, “Advancing autonomous racing: A comprehensive survey of the roboracer (f1tenth) platform,” inProc. of the IEEE Intl. Conf. on Artificial Intelligence, Robotics and Control (AIRC), 2025

    work page 2025

  11. [11]

    Time- optimal trajectory planning for a race car considering variable tyre- road friction coefficients,

    F. Christ, A. Wischnewski, A. Heilmeier, and B. Lohmann, “Time- optimal trajectory planning for a race car considering variable tyre- road friction coefficients,”Vehicle System Dynamics, vol. 59, 2021

    work page 2021

  12. [12]

    A Sequential Convex Programming Approach to Free-trajectory Minimum-lap-time Optimization of Racing Cars

    E. v. d. Eshof, J. van Kampen, and M. Salazar, “A computationally efficient framework for free-trajectory minimum-lap-time optimization of racing cars,”arXiv preprint arXiv:2511.13522, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Optimal control of a formula one car on a three-dimensional track—part 1: Track modeling and iden- tification,

    G. Perantoni and D. J. Limebeer, “Optimal control of a formula one car on a three-dimensional track—part 1: Track modeling and iden- tification,”Journal of Dynamic Systems, Measurement, and Control, vol. 137, 2015

    work page 2015

  14. [14]

    A simple mono-dimensional approach for lap time optimisation,

    B. Lenzo and V . Rossi, “A simple mono-dimensional approach for lap time optimisation,”Applied Sciences, vol. 10, no. 4, 2020

    work page 2020

  15. [15]

    Optimal control of a formula one car on a three-dimensional track—part 2: Optimal control,

    D. J. Limebeer and G. Perantoni, “Optimal control of a formula one car on a three-dimensional track—part 2: Optimal control,”Journal of Dynamic Systems, Measurement, and Control, vol. 137, no. 5, 2015

    work page 2015

  16. [16]

    An introduction to trajectory optimization: How to do your own direct collocation,

    M. Kelly, “An introduction to trajectory optimization: How to do your own direct collocation,”Society for Industrial and Applied Mathematics Review, vol. 59, no. 4, 2017

    work page 2017

  17. [17]

    Spline-based minimum-curvature trajectory optimization for autonomous racing,

    H. Xue, T. Yue, and J. M. Dolan, “Spline-based minimum-curvature trajectory optimization for autonomous racing,”arXiv preprint arXiv:2309.09186, 2023

    work page arXiv 2023

  18. [18]

    Deep learning warm starts for trajectory optimization on the international space station,

    S. Banerjee, A. Cauligi, and M. Pavone, “Deep learning warm starts for trajectory optimization on the international space station,” inSpace Robotics Workshop at Robotics: Science and Systems (RSS), 2025

    work page 2025

  19. [19]

    Real-time optimal trajectory planning for autonomous vehicles and lap time simulation using machine learning,

    S. Garlick and A. Bradley, “Real-time optimal trajectory planning for autonomous vehicles and lap time simulation using machine learning,” Vehicle System Dynamics, vol. 60, no. 12, 2022

    work page 2022

  20. [20]

    A simulation benchmark for autonomous racing with large-scale human data,

    A. Remonda, N. Hansen, A. Raji, N. Musiu, M. Bertogna, E. E. Veas, and X. Wang, “A simulation benchmark for autonomous racing with large-scale human data,”Advances in Neural Information Processing Systems, vol. 37, 2024

    work page 2024

  21. [21]

    Data-driven pit stop decision support for formula 1 using deep learning models,

    A. Sasikumar, A. A. Leema, and P. Balakrishnan, “Data-driven pit stop decision support for formula 1 using deep learning models,”Frontiers in Artificial Intelligence, vol. 8, 2025

    work page 2025

  22. [22]

    Autof1: An rnn-based approach to simulating strategic decision-making in formula 1,

    V . Mistry and W. A. P. Smith, “Autof1: An rnn-based approach to simulating strategic decision-making in formula 1,” inProc. of the IEEE Intl. Conf. on Intelligent Transportation Systems (ITSC), 2025

    work page 2025

  23. [23]

    On the comparison of interpolation methods,

    E. Maeland, “On the comparison of interpolation methods,”IEEE Transactions on Medical Imaging, vol. 7, no. 3, 1988

    work page 1988

  24. [24]

    Tem- poral convolutional networks for action segmentation and detection,

    C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager, “Tem- poral convolutional networks for action segmentation and detection,” inProc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2017

    work page 2017

  25. [25]

    Forza- ETH Race Stack—Scaled Autonomous Head-to-Head Racing on Fully Commercial Off-the-Shelf Hardware,

    N. Baumann, E. Ghignone, J. K ¨uhne, N. Bastuck, J. Becker, N. Imholz, T. Kr¨anzlin, T. Y . Lim, M. L¨otscher, L. Schwarzenbach,et al., “Forza- ETH Race Stack—Scaled Autonomous Head-to-Head Racing on Fully Commercial Off-the-Shelf Hardware,”Journal of Field Robotics, 2024

    work page 2024