Modeling Conversation as a Prompt Planning Problem for Adaptive Human–Robot Interaction

Abstract

Adaptive behavior is a fundamental requirement for social robots operating in real-world environments, where interactions must dynamically respond to both observable actions and inferred user states. In this work, we propose a novel framework that models human–robot interaction as a planning problem, where adaptive conversational behavior is driven by a symbolic planner based on the Planning Domain Definition Language (PDDL). Unlike other approaches that map actions onto sentences, we model actions as corresponding to high-level verbal interaction strategies, represented by a specific prompt fed to a Large Language Model (LLM) to generate dialogue in real-time. Additionally, the system continuously updates its representation of the interaction state by grounding user mental and nonmental states from dialogue using LLMs. This enables dynamic replanning, allowing the robot to adapt its behavior in real-time according to the evolving interaction context while maintaining a natural flow. We provide a descriptive analysis of the framework through a case study demonstrating behavior adaptivity, involving a robot acting as a tutor and a child asked to perform a task. The results illustrate the potential of combining symbolic planning and LLM-based dialogue generation to achieve flexible, context-aware, and adaptive human–robot interaction.