Zero-Shot Generalization of Vision-Based RL Without Data Augmentation