Modeling of vehicle dynamics from real vehicle measurements using a neural network with two-stage hybrid learning for accurate long-term prediction

摘要

This paper describes the neural network model of an actual vehicle and the associated hybrid learning scheme. The neural vehicle models the actual vehicle dynamics with the structure of a real-time recurrent network. The neural network was trained to predict the next state of the vehicle given the current state, the current input steering angle of the front wheel, and the velocity of the vehicle. A hybrid learning scheme is proposed which consists of open-loop training for stabilization and closed-loop training for prediction. The open-loop training is necessary to avoid instability at an initial stage. The closed-loop training follows in such a way that the neural network predicts the vehicle's sequence of state change given the initial state, the velocity, and the steering sequence. Furthermore, after this training procedure, it not only learns the vehicle's lateral dynamics along the trained trajectories, but can also generalize to similar trajectories.