SplineTM: B-Spline Tire Modeling for Autonomous Racing

Piotr Kicki, Jan Węgrzynowski, Grzegorz Czechmanowski, Krzysztof Walas

Keywords: autonomous racing, learning dynamics models, reinfrocement learning, model predictive control

TL;DR: SplineTM - a novel tire modeling framework that represents tire forces using B-splines

Abstract:

High-performance autonomous racing requires accurate modeling of vehicle dynamics, particularly in limit-handling regimes where tire-road interactions dominate vehicle behavior.
While traditional semi-empirical tire models provide interpretability, they often lack the flexibility to capture complex tire characteristics, especially when combined with simplified vehicle dynamics models. Conversely, purely data-driven methods offer high expressiveness and flexibility, but may suffer from physical inconsistencies and poor extrapolation in extreme regimes, leading to unsafe exploitation by control policies.
To address these issues, we introduce SplineTM, a novel tire modeling framework that represents tire forces using B-splines. SplineTM bridges the gap between rigid empirical structures and flexible data-driven approaches by maintaining physical grounding while providing scalable representational capacity. We evaluate the model against strong baselines across three tasks: trajectory prediction, Model Predictive Control, and Reinforcement Learning based sim-to-real transfer. Results on small-scale and full-scale racing platforms demonstrate that SplineTM achieves superior prediction accuracy and significantly faster lap times, providing a robust, differentiable, and interpretable alternative for safe, high-speed autonomous vehicle control.

Submission Number: 5