Scholarlink

This paper focuses on converting high level language instructions into navigation commands for robotic systems functioning within an office environment. A method is offered based on the capabilities of a large language model, GPT-4o-mini, which is first evaluated in a zero-shot scenario and later fine-tuned to domain-specific data. We build a synthetic dataset from a simulated office structure that entails natural language instructions and navigation steps that were generated using breadth-first search algorithms. The zero-shot performance is bound to be low with a 10% accuracy rate, but fine-tuning using smaller datasets of 1,000 observations noticeably improved the token accuracy to 72% while only accomplishing an 18% correct full-path match. In contrast, floundering with 10,000 observations visibly demonstrated much higher success where token level accuracy reached 94% while complete path accuracy achieved 85.5%. These results demonstrate the effectiveness of controlled data modification and prove that a sufficiently tuned large language model can accurately comprehend user commands and provide optimal short-path navigation instructions to achieve set goals. It offers the possibility of lessening the efforts required in manual control design, thus facilitating robotic systems to operate more efficiently and responsively to audio instructions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging Large Language Models (LLMs) for Robot Navigation: A Case Study of GPT-4o-mini on Constrained Office Environment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sari Masri,
  • Abdelrahem Atawnih,
  • Huthaifa I. Ashqar

摘要

This paper focuses on converting high level language instructions into navigation commands for robotic systems functioning within an office environment. A method is offered based on the capabilities of a large language model, GPT-4o-mini, which is first evaluated in a zero-shot scenario and later fine-tuned to domain-specific data. We build a synthetic dataset from a simulated office structure that entails natural language instructions and navigation steps that were generated using breadth-first search algorithms. The zero-shot performance is bound to be low with a 10% accuracy rate, but fine-tuning using smaller datasets of 1,000 observations noticeably improved the token accuracy to 72% while only accomplishing an 18% correct full-path match. In contrast, floundering with 10,000 observations visibly demonstrated much higher success where token level accuracy reached 94% while complete path accuracy achieved 85.5%. These results demonstrate the effectiveness of controlled data modification and prove that a sufficiently tuned large language model can accurately comprehend user commands and provide optimal short-path navigation instructions to achieve set goals. It offers the possibility of lessening the efforts required in manual control design, thus facilitating robotic systems to operate more efficiently and responsively to audio instructions.