Deviation Sequence Neural Network Control for Path Tracking of Autonomous Vehicles

Abstract

Despite its excellent performance in path tracking control, the model predictive control (MPC) is limited by computational complexity in practical applications. The neural network control (NNC) is another attractive solution by learning the historical driving data to approximate optimal control law, but a concern is that the NNC lacks security guarantees when encountering new scenarios that it has never been trained on. Inspired by the prediction process of MPC, the deviation sequence neural network control (DS-NNC) separates the vehicle dynamic model from the approximation process and rebuilds the input of the neural network (NN). Taking full use of the deviation sequence architecture and the real-time vehicle dynamic model, the DS-NNC is expected to enhance the adaptability and the training efficiency of NN. Finally, the effectiveness of the proposed controller is verified through simulations in Matlab/Simulink. The simulation results indicate that the proposed path tracking NN controller possesses adaptability and learning capabilities, enabling it to generate optimal control variables within a shorter computation time and handle variations in vehicle models and driving scenarios.