Learning-Based Hierarchical Model Predictive Control for Drift Vehicles

Drift-driving technique offers valuable insights to support safe autonomous driving in extreme conditions. Model predictive control (MPC) has become the dominant choice for drift vehicle control with a superior ability to handle the changing system dynamics and constraints. Existing control strategies necessitate a precise system model to calculate the reference drift equilibriums, which can be more intractable due to the highly nonlinear dynamics and sensitive vehicle parameters. Furthermore, MPC performance strictly contingents on appropriate controller parameters, presenting an additional hurdle for drift vehicles. To solve these problems, Bayesian optimization (BO) is first applied to compensate for modeling errors and optimize controller design for drift vehicles. A learning-based hierarchical model predictive control (LHMPC) strategy is proposed in this paper, where an upper-level BO supervisor provides learned drift equilibriums and controller parameters for a lower-level MPC. This hierarchical system architecture also effectively resolves the inherent conflict between path tracking and drifting. The LHMPC strategy is verified on the Matlab-Carsim platform, and simulation results demonstrate its effectiveness in guiding the vehicle following the reference trajectory while maintaining the drift states.