메뉴 건너뛰기
소속 기관 / 학교 인증
인증하면 논문, 학술자료 등을  무료로 열람할 수 있어요.
한국대학교, 누리자동차, 시립도서관 등 나의 기관을 확인해보세요
(국내 대학 90% 이상 구독 중)
고객센터 ENG
주제분류

논문 기본 정보

저자정보
(Kwangwoon University) (Kwangwoon University) (Kwangwoon University) (Kwangwoon University)
저널정보
제어로봇시스템학회 제어로봇시스템학회 논문지 제어로봇시스템학회 논문지 제31권 제8호
오류 신고하기

피인용 1

검색

    초록·키워드

    Object-based goal navigation tasks require embodied agents to locate specified objects in unfamiliar environments, a key challenge in vision-language navigation. Several works use end-to-end models to project observations into actions, but they often suffer from poor interpretability. Conversely, modular approaches provide interpretability by separating perception, planning, and action. There is a notable modular framework that predicts object locations in unexplored areas via supervised learning. However, it is prone to false positives, which can degrade navigation performance. In this work, we propose an enhanced method that integrates large language model assistance into the object prediction pipeline. Specifically, we introduce a validation step that uses GPT-4o to confirm the existence of objects predicted using mask region-based convolutional neural network (Mask R-CNN) before retrieving them in semantic maps. This validation effectively filters false positives, leading to more accurate goal identification. Experiments on the HM3D dataset demonstrate improved success rates and trajectory efficiency, validating the effectiveness of our approach in complex indoor navigation scenarios.

    최근 본 자료 전체보기

      UCI(KEPA) : I410-151-25-02-093823949