Automated driving requires robust and reliable perception systems, but rare and dangerous scenarios are often missing from real-world data. Kun Gao proposes an approach to scene data augmentation that combines real and virtual data to improve the performance of perception systems in complex environments. The goal is to reduce the limitations caused by insufficient training data for AI models. The method first analyzes important risk factors that influence perception performance. A scene data augmentation framework is then developed, integrating the realism of real data with the flexibility of virtual data. Using computer graphics and reinforcement learning, the approach generates a large number of challenging scenes and efficiently explores high-risk parameter combinations. The experimental results show that the proposed method improves robustness in rare and hazardous situations and increases the performance of AI-based object detection. The study also demonstrates that combining real and virtual data helps reduce the domain gap between them.