Enhancement of long-horizon task planning via active and passive modification in large language models.

Abstract

This study proposes a method for generating complex and long-horizon off-line task plans using large language models (LLMs). Although several studies have been conducted in recent years on robot task planning using LLMs, the planning results tend to be simple, consisting of ten or fewer action commands, depending on the task. In the proposed method, the LLM actively collects missing information by asking questions, and the task plan is upgraded with one dialog example. One of the contributions of this study is a Q&A process in which ambiguity judgment is left to the LLM. By sequentially eliminating ambiguities contained in long-horizon tasks through dialogue, our method increases the amount of information included in movement plans. This study aims to further refine action plans obtained from active modification through dialogue by passive modification, and few studies have addressed these issues for long-horizon robot tasks. In our experiments, we define the number of items in the task planning as information for robot task execution, and we demonstrate the effectiveness of the proposed method through dialogue experiments using a cooking task as the subject. And as a result of the experiment, the amount of information could be increased by the proposed method.