Autonomous navigation in open, dynamic urban environments poses unique challenges due to unstructured instructions, complex layouts, and moving obstacles. We propose Real-Nav,a unified vision-and-language navigation framework that operates seamlessly indoors and outdoors by tightly integrating semantic mapping with multimodal alignment. A simple simulation-to-reality adaptation strategy based on social-aware decision modules is employed for real-world deployment. Furthermore, in order to utilize the 3D semantic information of the space to be explored efficiently, we propose an additional pre-exploration stage in our model. We constructed a virtual environment simulator based on real photograph data, Tsinghua-roads, from Tsinghua University and completed the training on this simulator, then we evaluate Real-Nav on challenging vision-and-language navigation benchmarks and in a real-world campus setting. Our work demonstrate that building and exploiting semantic maps and employing curiosity-driven target candidate screening can significantly boost embodied navigation performance in both simulated and real-world environments.