A Learning-Based Controller for Trajectory Tracking of Autonomous Vehicles in Complex and Uncertain Scenarios

This paper proposes a learning-based controller for autonomous driving in dynamic and uncertain environments. The controller network parameters from demonstrations of human expert drivers and then updates the policy with online data samples using incremental learning methods. The controller aims to fit the vehicle's inverse dynamics and cope with external disturbances such as varying adhesion coefficients. To avoid catastrophic forgetting and fit the optimal policy, a knowledge evaluation method and a gradient constraint scheme are introduced. The effectiveness and robustness of the controller are demonstrated by a vehicle dynamics simulation model in MATLAB/Simulink. The experimental results show that the proposed method can adapt to complex curve environments with varying adhesion coefficients under high-speed driving conditions and is able to continuously improve control performance through incremental learning.