Learning-based Motion Control of Autonomous Vehicles Considering Varying Adhesion Road Surfaces

Advanced motion control of intelligent vehicles, e.g., trajectory tracking still faces challenges in enhancing stability on varying adhesion roads, particularly for driving at high speed with strong uncertainty in tire-road adhesion. The learning-based Gaussian process (GP) model can be used to mitigate the adverse effect arising from the state noise and model mismatch. However, it is challenging to cope with strong variation under diverse road conditions. This research proposes a multi-model control framework to enhance motion stability by capturing and reducing tracking errors on varying adhesion roads. We build three error models for high, middle and low adhesion road surfaces separately, and propose a switching strategy that can choose an appropriate model for the current road adhesion condition. Simulation results show that the proposed control strategy enhances the stability during trajectory tracking on complex roads with large curves and strong uncertainty, e.g., in case of a sudden change of road adhesion.