Mobile robots have become a powerful tool for executing transportation tasks in various industries, such as automatic warehouse and intelligent dock. However, autonomous navigation in open environment with crowd dynamic objects is still a huge challenge. To address this challenge, we propose an end-to-end navigation framework. In which, we design a relationships learning approach based on graph neural networks (GNNs) for inferring interaction relationships. We propose a scalable transform approach to switch semantic information with gird maps. A local RL navigation outputs navigation actions through local information. We demonstrate the efficiency and scalability of our method in navigation across various scenarios.