In this study, a novel approach is presented to automatically synthesize wayfinding instructions for embodied robot agents. The method uses in-context learning with LLMs to generate instructions across multiple simulation platforms. A user study showed that 83.3% of users found the synthesized instructions accurately captured environment details. Zero-shot navigation experiments demonstrated comparable performance with human-annotated data. The generated instructions were found to be usable and scalable for creating wayfinding instructions.
Sang ngôn ngữ khác
từ nội dung nguồn
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Vishnu Sasha... lúc arxiv.org 03-19-2024
https://arxiv.org/pdf/2403.11487.pdfYêu cầu sâu hơn