This presentation analyzes the architecture of end-to-end autonomous driving, which integrates perception and control into a single neural network, and the technical value of virtual simulation for large-scale data learning. It delves into high-precision digital twin-based data augmentation techniques and edge-case scenario generation technologies, and examines global trends in simulation solutions for implementing a closed data loop. Through this, Morai proposes essential requirements and technical challenges for a simulation infrastructure that will advance domestic autonomous driving technology.