멀티모달 LLM 기반 Topological Map 작성 방법

Abstract

This paper presents a novel approach to building topological maps based on a multi-modal large language model (MLLM). Traditional methods for building topological maps compare the features of the previous and current scenes and separate places by comparing their similarity. However, their reliance on features has made them difficult to use in environments with repetitive patterns. In addition, traditional methods only provide a simple link between places without relative directional information. To overcome these limitations, we propose a method for building a topological map using directional signboards. By using text and directional information from directional signboards, a robot can overcome ambiguity in repetitive patterns and can identify places. Furthermore, the robot can build a topological map without having to visit all the places. To obtain directional signboard data and transform it into a suitable format for further processing, we use a MLLM, which can simultaneously perform directional signboard recognition, text recognition, and orientation recognition within an image. We evaluate the efficiency of MLLM-based topological map building methods in real environments.