Zero-Shot Cross-City Generalization in End-to-End Autonomous Driving: Self-Supervised versus Supervised Representations
The study tested zero-shot cross-city transfer in autonomous driving to see how visual pretraining affects robustness when models are trained only in one city and tested in another. Using self-supervised backbones like I-JEPA, DINOv2, and MAE, the research found that domain-specific self-supervised pretraining significantly reduces performance degradation compared to standard supervised ImageNet models when transferring between cities with different road layouts.