ChatGPT for Robotics: A New Approach to Human-Robot Interaction and Task Planning

Bing Xie, Xiangming Xi, Xinan Zhao, Yuhan Wang, Wei Song, Jianjun Gu, Shiqiang Zhu

Abstract: ChatGPT, released by OpenAI, has garnered academic interest due to its powerful natural language processing capabilities. It can accurately understand conversations and generate high-quality responses. This paper explores the potential of applying ChatGPT to two significant research topics in robotics: Human-Robot Interaction and Task Planning. Human-Robot Interaction involves studying the interactions between humans and robots. ChatGPT is well-suited for this purpose as it enables robots to communicate with humans. However, ChatGPT has shortcomings such as outdated knowledge and fabricating answers, making it unsuitable for direct use in Q&A. To address these issues, we propose an architecture called FRC that combines FAQ, retrieval module, and ChatGPT. In this architecture, ChatGPT is used for rephrasing questions and reading comprehension. Experiments show that the architecture can combine multiple rounds of dialogue to answer incomplete questions or questions that need coreference resolution. Task planning involves using an internal model to reason about the world and create a plan of actions to achieve a specific goal. Traditional approaches to robotic task planning rely on search while using ChatGPT for task planning is a novel approach based on generation. Existing methods using ChatGPT do not consider the state of the robot. We propose a method that enables ChatGPT to perform task planning based on the state of the robot and have verified its feasibility through experiments. We also discuss the limitations of ChatGPT in multi-party talk and motion control.