Behavior tree generation and adaptation for a social robot control with LLMs

Large Language Models have recently emerged as a powerful tool for generating flexible and context-aware robotic behavior. However, adapting to unforeseen events and ensuring robust task completion remain significant challenges. This paper presents a novel system that leverages LLMs and Behavior Trees to enable robots to generate, execute, and adapt task plans based on natural language commands. The system employs ChatGPT to process user instructions, generating initial Behavior Trees that encapsulate the required task steps. A modular architecture, combining the BT planner and a Failure Interpreter module, allows the system to dynamically adjust Behavior Trees when execution challenges or environmental changes arise.

Unlike conventional methods that rely on static Behavior Trees or predefined state machines, our approach ensures adaptability by integrating a Failure Interpreter capable of identifying execution issues and proposing alternative plans or user clarifications in real time. This adaptability makes the system robust to disturbances and allows for seamless human–robot interaction. We validate the proposed methodology using experiments on a social robot across various scenarios in our workplace, demonstrating its effectiveness in generating executable Behavior Trees and responding to execution failures. The approach achieves an 89.6% success rate in a realistic home environment, highlighting the effectiveness of LLM-powered Behavior Trees in enabling robust and flexible robot behavior from natural language input.