Real-time model predictive control of path-following for autonomous vehicles towards model mismatch and uncertainty

Abstract

The path following function is a critical component of functional safety for autonomous vehicles, and following precision has garnered increased attention in practical applications. However, control performance can be compromised due to uncertainties in vehicle parameters and discrepancies between the control model and the actual vehicle to be controlled. To address this, a real-time model predictive control for path following of autonomous vehicles is proposed, incorporating an estimation of model mismatch. An adaptive extended Kalman filter is developed to estimate the potential model mismatch terms, and state deviations are compensated accordingly. Subsequently, a parameter-varying model predictive controller is formulated to achieve unbiased path-following control while maintaining robustness to parameter variations. Simulation results demonstrate a significant improvement in lateral following accuracy, with enhancements of 53.85%, 47.83%, and 42.86% compared to the nonlinear model predictive control, robust model predictive control, and learning-based control, respectively. The hardware-in-the-loop and real-road experiments further validate the excellent real-time executability, with a maximum time cost of 12.4 ms, accounting for 62% of the sampling period.