STRAP-LLM: structured task allocation and planning for heterogeneous robots using large language models

Abstract

When a system using large language models (LLMs) with reasoning abilities performs task allocation and skill planning with multiple robots, key challenges arise: generating only executable output and adapting to new robots. We address these challenges with STRAP-LLM: structured task allocation and planning for heterogeneous robots using large language models. It is a structured prompting framework that (1) enforces strict rules to restrict the LLMs’ generative capabilities, (2) modularizes robot information to support seamless integration of new robots without additional examples, and (3) separates reasoning plans for task allocation and execution planning in natural language and formalizing the plans into skill sequences in robot’s execution language. STRAP-LLM achieves consistently high execution accuracy and adaptability even when new robots are added. We evaluate the system by providing various commands and incrementally adding robots to verify its accuracy and executability. We also compare its accuracy, token usage, and execution time with the current state-of-the-art framework. The system is applied to a real robot platform to confirm that the generated plans can be directly executed in a real-world control environment without post-processing. Code is available at https://github.com/jaebum98/STRAP-LLM.