Robust path-tracking controller for autonomous vehicles incorporating neural networks and model predictive control

This paper proposes a neural-network-assisted model predictive controller for improving the path-tracking performance and lateral stability of autonomous vehicles on low-friction roads. First, phase-plane analysis is performed under various driving conditions and vehicle states to identify the stable region in the phase plane of lateral velocity and yaw rate. The phase-plane analysis results are then used to train a neural network that estimates the admissible yaw-rate range without requiring accurate estimation of lateral velocity or tire slip angle. In addition, a preview-based reference state is generated by accounting for vehicle-speed variations over the model predictive control horizon. A linear time-varying model predictive controller is subsequently formulated using a two-degree-of-freedom linear bicycle model to track the reference state. The yaw-rate bounds predicted by the neural-network-based stable-region estimator are incorporated into the controller as stability constraints. The proposed controller is evaluated through CarSim-MATLAB/Simulink co-simulations under low-friction road conditions. The simulation results demonstrate that the proposed controller improves path-tracking accuracy while maintaining the vehicle states within the estimated stable region.