This paper presents a new conversational model for human-robot interaction that uses a graph-based representation of the dialogue state. The knowledge graph is continuously updated with new observations from the robot’s sensors and is converted into a natural language form. This conversion is performed using a set of parameterized functions optimized based on a set of interactions. The text representation of the dialogue state graph is then used as a prompt for a large language model to decode the agent response. The approach was evaluated through a user study with a humanoid robot and showed a significant improvement in the perceived factuality of the robot responses compared to a baseline.
Publication date: 29 Nov 2023
Project Page: http://creativecommons.org/licenses/by/3.0/
Paper: https://arxiv.org/pdf/2311.16137
