Integration of Partially Prior Physics and Data for Longitudinal Vehicle Dynamics Restoration

Vehicle dynamics model is critical for autonomous vehicles since it provides certain rules of motion in response to the surrounding rapidly changing environment. However, current pure physics-based models and data-driven approaches are set against each other in constructing the dynamics model. In this work, we propose a vehicle dynamics restoration approach by integrating physics principles and data based on physics-informed machine learning. The vehicle dynamics model parameters are estimated by combining data and limited prior physical knowledge. The results of a case study demonstrate that (1) the proposed methods can accurately identify the dynamics model parameters and restore the vehicle dynamics model and (2) the neural network with physical knowledge can accurately predict the state of vehicle. The proposed method has the potential to be applied in quantifying the dynamics model uncertainty due to parameter uncertainty, and stochastic motion planning and control due to uncertainty, which will be further investigated in the future.