A Longitudinal/Lateral Coupled Neural Network Model Predictive Controller for Path Tracking of Self-Driving Vehicle
Sibing Yang, Cong Geng
Abstract
In recent years, the model predictive control(MPC) algorithm has been increasingly applied to the path tracking of self-driving vehicles due to its capacity to deal with dynamic constraints explicitly. The control performance of MPC is highly dependent on the dynamic model. However, as vehicles are strongly coupled nonlinear systems, the prediction accuracy of the classical mechanism model decreases greatly at high-speed conditions, and the control error also increases. This paper proposes replacing the classical mechanism model with a recurrent neural network(RNN) for vehicle dynamical state prediction under the framework of MPC to realize higher prediction accuracy. RNN uses historical control and state variables to predict future state variables. The input and output of the RNN model include longitudinal and lateral variables, and based on this, longitudinal/lateral coupled control is realized. The differential evolution algorithm is proposed to solve the optimization problem. Finally, the prediction accuracy of the RNN model is verified on the real vehicle dataset and compared with linear/nonlinear mechanism models. The control algorithm proposed in this paper is compared with classical MPC against low and high speeds on the ADAMS/ Python/ Simulink joint simulation platform. The results show that the control accuracy and stability of the longitudinal/lateral coupled neural network MPC are higher than that of the MPC based on the mechanism model, especially at high speed.