Trajectory Tracking Control for Autonomous Vehicles with Physicsinformed Neural Network Vehicle Model

Vehicle dynamics modeling is a crucial task for the trajectory tracking of autonomous vehicles. Due to the variable road conditions and the high cost of data acquisition, both physics-based and data-driven models face significant challenges with sparse data. To address this issue, we propose a vehicle model that combines the single-track model with physics-informed neural network methods. Firstly, we encode the single-track model in a physics-based computational graph to identify the key parameters in the physical model. Then, we use these parameters to calibrate the single-track model, which is again encoded in a physics-based computational graph to design a physics-informed neural network for vehicle dynamics modeling (PINN-VM). The PINN-VM leverages the advantages of model-driven and data-driven approaches to achieve accurate vehicle dynamics modeling. Compared to models without physics, PINN-VM performs better, especially in situations with sparse data. Finally, the PINN-VM is incorporated into a novel model predictive control framework to design a trajectory tracking controller. Simulation experiments using CarSim and Simulink demonstrate that the proposed controller enhances trajectory tracking performance and lateral stability under the condition of large curvature at medium-high velocity with different friction coefficients.