Task Planning for Robot Manipulator Using Natural Language Task Input with Large Language Models

Abstract

The utilization of robots is expanding beyond the industrial sector and reaching into society. This paper considers the implementation of robots in retail stores for stocking products on shelves. To optimize the task input for users unfamiliar with robots minimizing the robot’s operational time is essential. This paper presents a novel task planning method using natural language input. Our proposed method converts natural language tasks into symbolic sequence representations using Large Language Models (LLM) and then the optimal task procedures are derived by executing task planning based on Monte Carlo Tree Search (MCTS). To improve the accuracy of the conversion, we propose an interactive method that allows users to confirm the conversion results to improve the correctness of the response of LLM. We incorporate a mechanism for handling error messages caused by unexpected outputs from LLM. A few-shot prompting method is adopted to guide the LLM to higher performance. Computational experiments demonstrate that the proposed method can successfully identify the intended operational procedures for approximately 90% of 18 natural language tasks.