Physics-Informed Neural Networks for Vehicle Lateral Dynamics Modeling

Abstract

Accurate estimation of vehicle lateral dynamics is important for advanced driver assistance systems (ADAS) and autonomous driving applications. This thesis proposes a Physics-informed Neural Network (PINN) framework that combines fundamental vehicle dynamics, based on simplified single-track models and Pacejka tire formulations, with data-driven recurrent neural networks that incorporate attention mechanisms. By embedding physical constraints within the learning process, the hybrid model aims to improve interpretability, robustness, and real-time performance without relying on additional sensor inputs. Evaluations using high-fidelity simulations and real-world datasets suggest that the model can estimate latent states, such as lateral velocity, with promising accuracy. While the approach generalizes well across a variety of simulated driving scenarios, it faces challenges in maintaining comparable performance on real-world data mainly due to modeling uncertainties and measurement noise. These limitations highlight the need for further investigation into robustness and domain adaptation techniques. Future work will explore on-board deployment, adaptation to individual vehicle characteristics, and the integration of more detailed physical models. These developments could enhance the model’s reliability and utility in practical applications, contributing to improved vehicle safety and operational efficiency.