In this paper, we propose four easy-to-implement methods: consistency learning, edge prediction, color quantization, and gray-world algorithm which are compatible with the existing UDA-SST works. The proposed methods combined with the existing works outperform state-of-the-art methods on the GTA5 $\to$ Cityscapes benchmark significantly. Furthermore, we integrate our UDA-SST method with RL agent and deploy it to the edge device Husky A200. Ultimately, we develop a sim-to-real autonomous driving car. With a robust vision model and intelligent agent, our end-to-end self-driving system finally shows great adaptability to the NTHU campus and performs exceptionally in the autonomous driving task.
