This study proposes an integrated control architecture for end-to-end autonomous driving by combining Generative Adversarial Imitation Learning (GAIL), Proximal Policy Optimization (PPO), and Nonlinear Model Predictive Control (NMPC). To address the l...
This study proposes an integrated control architecture for end-to-end autonomous driving by combining Generative Adversarial Imitation Learning (GAIL), Proximal Policy Optimization (PPO), and Nonlinear Model Predictive Control (NMPC). To address the limitations of reinforcement learning in real-world applications—such as instability during early exploration and the sim-to-real gap—GAIL is first used to generate a stable initial policy from expert driving data. This policy is further refined through PPO to improve adaptability and generalization across dynamic urban environments. However, since learning-based policies alone may overlook physical constraints of real vehicles, NMPC is employed to translate high-level policies into optimal control commands while respecting vehicle dynamics and safety limits. The proposed framework effectively combines the flexibility of learning-based methods with the robustness of model-based control. Simulation results using CARLA and digital twin environments demonstrate superior path tracking accuracy and driving stability compared to conventional BC+NMPC methods, highlighting the feasibility of hybrid learning-control integration for real-world autonomous driving systems.