KGGPT: Empowering Robots with OpenAI's ChatGPT and Knowledge Graph.

This paper presents a study on using knowledge graph with ChatGP for robotics applications, called KGGPT. Traditional planning methods for robot tasks based on structured data and sequential actions, such as rosplan, have limitations such as limited data range and lack of flexibility to modify behaviors based on user feedback. Recent research has focused on combining AI planning with large language models (LLMs) to overcome these limitations, but generated text may not always be consistent with real-world physics and the robot skills to perform physical actions. To address these challenges, we propose KGGPT, a system that incorporates prior knowledge to enable ChatGPT for a variety of robotic tasks. KGGPT extracts relevant knowledge from the knowledge graph, generates a semantic description of the knowledge, and connects it to ChatGPT. The gap between the knowledge of ChatGPT and actual service environments is addressed by using the knowledge graph to model robot skills, task rules, and environmental constraints. The output is a behavior tree based on robot skills. We evaluate our method in an office setting and show that it outperforms traditional PDDL planning and a separate ChatGPT planning scheme. Additionally, our system reduces programming effort for applications when new task requirements arise. This research has the potential to significantly advance the field of robotics.