Data-Driven Hierarchical Model Predictive Control for Automated Overtaking Maneuver via Gaussian Process Regression

The effectiveness of model-driven trajectory generation and control methods for an automated vehicle is closely related to the accuracy of vehicle kinematics and dynamics. However, the dynamic behavior of a vehicle is time-varying and uncertain, especially during high mobility driving. For the data-driven end-to-end trajectory generation and control approach, the transferability and robustness of the trained neural network cannot easily be guaranteed. In this paper, a combined model- and learning-based trajectory generation and control framework for automated overtaking maneuver is presented. Unlike the model-driven approach and end-to-end learning approach, the proposed framework blends the single-track vehicle model of the automated vehicle with time-series driving data to generate safe and high-quality overtaking motions. Specifically, an efficient vectorized implementation of multivariate Gaussian process regression (MGPR) is applied to learn the unmodeled dynamics in hopes of shortening the deviation between the single-track dynamic model and the true vehicle dynamics. A hierarchical nonlinear model predictive control with MGPR is designed for realizing local trajectory generation and control in the overtaking maneuvers. Multi-phases automated overtaking maneuvers at different speed ranges are conducted on roads with different friction coefficients to illustrate the effectiveness of the proposed trajectory generation and control framework.