Abstract
In this paper, we present a novel framework that combines large languagemodels (LLMs), digital twins and industrial automation system to enableintelligent planning and control of production processes. We retrofit theautomation system for a modular production facility and create executablecontrol interfaces of fine-granular functionalities and coarse-granular skills.Low-level functionalities are executed by automation components, and high-levelskills are performed by automation modules. Subsequently, a digital twin systemis developed, registering these interfaces and containing additionaldescriptive information about the production system. Based on the retrofittedautomation system and the created digital twins, LLM-agents are designed tointerpret descriptive information in the digital twins and control the physicalsystem through service interfaces. These LLM-agents serve as intelligent agentson different levels within an automation system, enabling autonomous planningand control of flexible production. Given a task instruction as input, theLLM-agents orchestrate a sequence of atomic functionalities and skills toaccomplish the task. We demonstrate how our implemented prototype can handleun-predefined tasks, plan a production process, and execute the operations.This research highlights the potential of integrating LLMs into industrialautomation systems for more agile, flexible, and adaptive production processes,while it also underscores the critical insights and limitations for futurework.