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

This repository contains the official implementation for the paper:

Efficient Trajectory Optimization for Autonomous Racing via Formula-1 Data-Driven Initialization
Samir Shehadeh, Lukas Kutsch, Nils Dengler, Sicong Pan, Maren Bennewitz
arXiv:2603.07126

Formula 1 photograph from Pexels.

Overview

Trajectory optimization is a key component of autonomous racing, but practical minimum-time optimization pipelines can be highly sensitive to the initial trajectory. Poor initialization, such as using the track centerline or purely geometric baselines, may lead to slow convergence or suboptimal local solutions.

This work proposes a learning-informed initialization strategy that uses expert Formula 1 telemetry as a prior for autonomous racing trajectory optimization. The method reconstructs and aligns real-world Formula 1 racelines across multiple tracks, trains a neural network to predict expert-like raceline offsets from local track geometry, and uses the predicted raceline as an initialization seed for a minimum-time optimal control solver.

This repository contains the dataset of 17 tracks, the network with a trained model, and the scripts to download telemetry, build a raceline from it, train, and predict. The predicted raceline (predict.py output) is the initialization handed to the optimizer.

Key Ideas

• Reconstruction of expert Formula 1 racelines from telemetry data.
• Standardized track-centered representation using centerline arc length, curvature, track boundaries, and lateral raceline offsets.
• Neural raceline prediction from local track geometry.
• Learned initialization for minimum-time trajectory optimization.
• Evaluation across 17 Formula 1 tracks.
• Hardware validation on a 1:10 RoboRacer platform.

Method Summary

The full pipeline consists of three main stages:

1. Formula 1 telemetry processing
   Real-world Formula 1 telemetry is reconstructed, aligned to track geometry, and converted into a Frenet-frame representation.

2. Learning-based raceline prediction
   A neural network predicts expert-like raceline offsets from local track features such as curvature and track boundary offsets.

3. Minimum-time trajectory optimization
   The predicted raceline is used as an informed initialization for a physics-based minimum-time optimal control solver.

Install

Tested with Python 3.12. A GPU is optional.

pip install -r requirements.txt

Quick start

Predict the raceline of a track the released model has never seen, and plot it:

python scripts/predict.py --track data/split_01/test/barcelona.csv --mode recursive --plot
python scripts/predict.py --track data/split_01/test/barcelona.csv --mode non-recursive --out barcelona_nonrec_nn.csv --plot

Train a model on the 14 training tracks:

python scripts/train.py --config configs/train.yaml

Build a raceline from F1 telemetry yourself (download a driver's laps, align them to the track, average them):

python scripts/download_telemetry.py --year 2025 --event "Italian Grand Prix" --driver VER --laps 1 4-12 --out telemetry/monza
python scripts/align_track.py --track data/split_01/train/monza.csv --telemetry telemetry/monza --out monza_aligned.csv --plot

Repository

folder	contents
data/	17 tracks with the expert F1 raceline, split into 14 training and 3 test tracks; sources and disclaimer
checkpoints/	the trained model for this split
scripts/	predict.py, train.py, download_telemetry.py, align_track.py
f1init/	geometry, alignment, dataset and model code used by the scripts
configs/	training configuration
assets/	images used in the READMEs

Not included yet

• Minimum-time optimization initialization interface. The optimizer is the minimum-time planner of TUM's global_racetrajectory_optimization; we added a small patch that lets it start from any given line instead of only the centerline.
• RoboRacer-related validation utilities

Citation

If you use this work, please cite the paper:

@inproceedings{shehadeh2026efficient,
  title     = {Efficient Trajectory Optimization for Autonomous Racing via Formula-1 Data-Driven Initialization},
  author    = {Shehadeh, Samir and Kutsch, Lukas and Dengler, Nils and Pan, Sicong and Bennewitz, Maren},
  booktitle = {IEEE International Conference on Intelligent Transportation Systems (ITSC)},
  year      = {2026}
}

License

The code is released under the MIT License. For the data sources and their terms, see data/README.md.

Contact

For questions, please open an issue or contact the authors directly.