Neural Network-Aided Fast Iterative Model Predictive Control for Autonomous Vehicle Motion Control

This paper proposes an innovative data-physics hybrid framework, termed neural network-aided fast iterative model predictive control (NN-FIMPC), aimed at enhancing the motion control accuracy and robustness of autonomous vehicles (AVs) under challenging conditions while addressing the real-time constraints of onboard control units. This framework integrates a neural network dynamics model to effectively capture the vehicle’s complex dynamic behavior across diverse environments, compensating for the modeling errors of the linear controller model. The core concept of NN-FIMPC lies in reintroducing this modeling error as a known additional disturbance into the linear MPC framework. By sequentially solving two linear MPC problems, the framework balances control accuracy with computation time, overcoming limitations of conventional MPC-based motion control algorithms and enabling high-precision, real-time motion control for AVs. To validate the accuracy of the neural network dynamics model and the effectiveness of the NN-FIMPC framework, extensive co-simulations and hardware-in-the-loop tests were conducted. The results demonstrated that, on one hand, the neural network dynamics model exhibits superior modeling accuracy compared to conventional nonlinear physical models. On the other hand, compared with traditional linear and nonlinear MPC algorithms, the NN-FIMPC framework achieves higher control accuracy while maintaining millisecond-level computation time.