Boosting Behavior Tree Generation for Robots with Large Language Models and Genetic Programming

Mobile robots are becoming increasingly ubiquitous in modern society, requiring more human-like interaction capabilities, such as following operator instructions and collaborating with humans. Conventional robot programming methods often fall short in achieving these complex behaviors. Behavior Trees (BTs) offer a promising alternative due to their modularity, scalability and reactivity. We propose utilizing general purpose, system-prompted Large Language Model (LLM) assistants to decompose task descriptions into executable BTs, which are subsequently refined using Genetic Programming (GP) and a state machine like low-resource BT execution simulator, where they will be tested for task completion. This approach eliminates the need for fine-tuning LLMs, thereby reducing computational costs and saving time and energy. Our method successfully solves all proposed scenarios, enhances applicability across diverse environments, and democratizes behavior generation for non-experts, outperforming baseline methods in efficiency.