We propose a zero-shot object-goal navigation framework by constructing an online 3D scene graph to prompt LLMs. Our method can be directly applied to different kinds of scenes and categories without ...