Can autonomous cars learn to drive in new cities without retraining?

Autonomous vehicles trained in one city often fail in another due to different road layouts, appearances, and traffic patterns. CityGen uses diffusion models guided by HD maps and city-level visual cues to synthesize realistic training data for new cities without any manual annotation. Tested on a new cross-city benchmark covering perception, segmentation, and planning, the method consistently improves robustness across all tasks—tackling a real scalability bottleneck for deploying autonomous driving systems globally.