FireRobBrain: Planning for a Firefighting Robot using Knowledge Graph and Large Language Model

Jialei Zhang, Shubin Cai, Zhiwei Jiang, Jian Xiao, Zhong Ming

Abstract: Firefighting robots play a crucial role in improving fire safety. However, these robots have limitations in understanding their surroundings and adapting to changing situations. In this paper, we propose the “FireRobBrain”, a combination of a knowledge graph and a large language model, to act as the “brain” of firefighting robots. This approach aims to tackle challenges related to planning robots in dynamic environments. The FireRobBrain consists of a KG with a dynamic information base and a relatively static knowledge base. It also includes a prompting module that helps the language model generate suggestions for the robot's reactions. We evaluated the framework using a dataset of 864 samples and discovered that the combination of LLM and KG, facilitated by a well-designed prompt module, significantly improves the quality of answers, particularly for tasks involving specific contexts and structured information. Furthermore, we noted that providing a task scope in the input prefix contributes to a better understanding of the robot's task, resulting in enhanced performance.