Ensuring safe driving under real-time uncertainties remains a critical challenge in autonomous vehicle control. To address this issue for a collision avoidance task, this study proposes a robust model predictive control (RMPC) framework that handles parametric uncertainties using optimization-based linear matrix inequality (LMI). By incorporating system parametric uncertainties, the RMPC enhances driving stability and safety through the use of input-state constraints. However, due to its computational complexity, we employ a data-driven approach by collecting measurements under different road adhesion conditions to train deep neural networks with a long short-term memory layer (DNN-LSTM). The proposed DNN-LSTM effectively captures temporal dependencies, outperforming existing DNNs when using the same hyperparameters in accuracy and generalization. All comparative simulations are conducted and verified using the high-fidelity CarSim/Simulink co-simulation platform. Therefore, the proposed DNN-LSTM approach approximates the RMPC policy with high training performance and significantly reduces computational complexity, which is more beneficial for real-time implementation. Using the DNN-LSTM is further emphasized to maintain the ability to drive stability of autonomous vehicles compared with online and offline RMPCs, which show a stable region violation at some fixed operation points.
