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

Abstract

This paper describes a neural network (NN) model of a real vehicle and the associated hybrid learning scheme. The NN vehicle models the actual vehicle dynamic behavior with the architecture of a real-time recurrent network. The NN was trained to predict the next state of the vehicle, given the current vehicle state, the current input steering angle of the front wheel, and the vehicle's speed. A hybrid training scheme for the network has been proposed, which consists of two phases: open-loop training for stabilization of the NN weight learning and closed-loop training for long-term prediction of the vehicle behavior. The open-loop training is necessary to avoid learning instability at initial stages. The closed-loop training then follows in such a way that the NN correctly predicts the vehicle's next state in a recursive mode. The outcome is that the model can correctly generate the vehicle trajectory, given the initial state and the steering and speed sequence of the vehicle. 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. This modeling technique has been successfully applied to model the actual dynamics of a Daewoo Leganza vehicle. It is an intelligent vehicle that is fully autonomous in that steering, braking, and accelerating were all done via computer control. The training data were taken from a four-vehicle platoon demonstration in which four vehicles were automatically controlled in a convoy mode.