Machine Learning Assisted Formula 1 Racing

Abstract

Automated digitization of the racing track from satellite images is challenging due to noisy, obstructive natural environments and widespread confusion with other road segments. We propose a semi-automated method for track extraction and Machine Learning (ML) based detection by pixel-level classification guided by a small sample of track ground truth patches. Our methodology builds on and strengthens the approach proposed in [1] by exploring a wider variety of new engineered features and refined ML models that utilize low-effort human guidance to identify the area of interest. That way, the digitized YAS Marina Formula 1 track is then used as a simulation platform, on which we first develop a simplified physics-based model of optimal driving control of an F1 car and then build supervised ML models trained on the physics-model optimally labeled simulated data to try to learn the optimal driving control actions without the knowledge of physics-based purely on the simulated driving experience data. A family of explainable decision tree-based ML models has been developed and tested in various simulated weather scenarios accommodating wet/snowy track surfaces and different visibility depths. Although the cross-validation-based evaluation indicates encouraging prediction results across various track/visibility conditions, we still observe the small misclassification rates, most critically including a few instances of failing to break at high speed well before the sharp turn. We believe, however, that these errors stem from very limited testing conditions and can be eliminated by further training on multiple tracks capturing diverse road trajectories with wider driving control patterns. In conclusion, the research contributes to the field by offering a semi-automated approach for digitizing F1 racing tracks from satellite images, a physics-based model that represents optimal driving scenarios considering different weather conditions and friction constants on the digitized track, and a supervised learning approach to automate vehicle actions under different visibility conditions. Our overall inference is that ML methodology offers an untapped level of support to extracting and modeling road networks and assisting or eventually autonomously controlling driving actions, whether in the challenging F1 race or casual driving mode, especially under limited visibility or slippery road conditions.

Date of Award	Aug 2023
Original language	American English
Supervisor	Chan Yeun (Supervisor)