Efficient Domain Augmentation for Autonomous Driving Testing Using Diffusion Models

This paper, published at the 47th International Conference on Software Engineering (ICSE 2025), presents an original contribution at the intersection of software testing and generative AI for autonomous driving systems. We propose an innovative framework that integrates physics-based simulators with diffusion models to enhance system-level testing of Autonomous Driving Systems (ADS). The approach introduces three generative strategies—instruction editing, inpainting, and inpainting with refinement—to expand the Operational Design Domain (ODD) of driving simulators. An automated semantic validator ensures realism and validity of generated images, while a knowledge-distilled CycleGAN model achieves efficient online rendering during simulation. Experiments across 52 ODDs and multiple ADS models demonstrate that the approach improves fault detection and domain coverage by up to 20×, with minimal runtime overhead. This work advances software engineering for safety-critical AI by enabling scalable, realistic, and continuous testing of autonomous vehicles.