Topological Map Composed of Text Information

Abstract

In recent years, research on vision-and-language navigation has made significant progress, although it typically requires costly user instructions for each navigation step. To address this problem, we explored a method that creates a map using the user’s language path instructions. This study introduces two approaches using a large language model: one where mapping within a large language model and another where it’s done externally. We tested these methods on graph maps and language navigation instructions, revealing the capacity limits of the large language model and the success of the external mapping approach.