Enhancement of Long-Horizon Task Planning via Active and Passive Modification in Large Language Model

Abstract This study proposes a method for generating complex, 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 are often simple. In the proposed method, the LLM actively collects missing information by asking questions, and the task plan is upgraded with a one-shot dialog example. In addition, the robustness of the task planning is enhanced by using the method that modifies the plan results based on human instructions. The effectiveness of the proposed method is demonstrated through dialogue experiments using a cooking task as the subject. Furthermore, the proposed method can be applied to instructions that include images.