Sparse Gaussian process-based strategies for two-layer model predictive control in autonomous vehicle drifting

Abstract

Vehicle safety is paramount in autonomous driving, particularly when managing vehicles at extreme side-slip angles—a challenge often overlooked by conventional controllers. Recent studies have focused on vehicle drift control under such extreme conditions. However, tracking complex trajectories while drifting is challenging, especially in the presence of a model mismatch. This paper proposes a two-layer model predictive controller based on sparse variational Gaussian processes. The first layer is responsible for computing the optimal drift equilibrium points, while the second layer is tasked with tracking these points. A variational free energy-based Gaussian process is utilized to compensate for errors in the upper-layer drift equilibrium point calculations and mismatches in the lower-layer controller model. Moreover, the vehicle’s state is determined to be either in transit drift or deep drift based on whether the slip angle and steering angle have reached critical values. Gaussian models are established for each state to enhance prediction accuracy. The effectiveness of the controller is demonstrated through joint simulations on MATLAB and CarSim platforms. First, the proposed two-layer model predictive controller was compared with three state-of-the-art drift controllers, demonstrating at least a 48.64% reduction in average lateral error when tracking trajectories with varying curvature. Second, when combined with sparse Gaussian processes, the controller’s learning ability was validated in scenarios with a 5% to 20% friction coefficient mismatch. Specifically, in the scenario with a 20% friction coefficient mismatch, its average lateral error was reduced by 95.09% after model error learning. Additionally, the controller was compared with both Fully Independent Training Conditional (FITC) GP-based MPC and Full GP-based MPC controllers, demonstrating better trajectory tracking capability and model error learning ability.