An LLM-Assisted Multi-Agent Control Framework for Roll-to-Roll Manufacturing Systems

Jiachen Li, Shihao Li, Christopher Martin, Zijun Chen, Dongmei Chen, Li Wei

Abstract

Roll-to-roll manufacturing requires precise tension and velocity control to ensure product quality, yet controller commissioning and adaptation remain time-intensive processes dependent on expert knowledge. This paper presents an LLM-assisted multi-agent framework that automates control system design and adaptation for R2R systems while maintaining safety. The framework operates through five phases: system identification from operational data, automated controller selection and tuning, sim-to-real adaptation with safety verification, continuous monitoring with diagnostic capabilities, and periodic model refinement. Experimental validation on a R2R system demonstrates successful tension regulation and velocity tracking under significant model uncertainty, with the framework achieving performance convergence through iterative adaptation. The approach reduces manual tuning effort while providing transparent diagnostic information for maintenance planning, offering a practical pathway for integrating AI-assisted automation in manufacturing control systems.

Introduction

Roll-to-roll (R2R) manufacturing is the backbone of producing flexible electronics, thin films, and printed sensors. The process involves moving a continuous "web" of material through various rollers at high speeds. The primary challenge is maintaining precise web tension and linear velocity. These factors are strongly coupled; changing the speed of one roller instantly affects the tension in the material, which can lead to wrinkles, misalignment, or material breakage.

Currently, setting up these control systems and adjusting them for different materials or changing roll diameters requires highly experienced control engineers. This manual tuning is time-consuming, expensive, and difficult to scale. This paper introduces a multi-agent framework that leverages Large Language Models (LLMs) to automate these complex tasks while ensuring the resulting control actions are physically safe for industrial equipment.

Methodology

The proposed framework employs specialized LLM "agents" that collaborate across five distinct phases to manage the control lifecycle.

Overview of the LLM-assisted multi-agent framework architecture and its five operational phases.

The framework's operation is divided into:

  1. Phase 0 (SysID Agent): Identifies system parameters from historical data using physics-informed structures.
  2. Phase 1 (Initial Control Agent): Selects the optimal controller architecture (e.g., PID, LQR, or MPC) and performs initial tuning in simulation.
  3. Phase 2 (Adaptation Agent): Bridges the "sim-to-real" gap by analyzing hardware performance and iteratively suggesting parameter adjustments.
  4. Phase 3 (Monitoring Agent): Provides continuous vigilance, detecting performance degradation and diagnosing root causes such as mechanical wear.
  5. Phase 4 (Model Refinement): Periodically updates the simulation model using new operational data.

Technical Innovations

A key technical innovation is the integration of Retrieval-Augmented Generation (RAG). Instead of relying solely on an LLM's internal training data, the agents query a curated database of control theory fundamentals and R2R best practices. This grounding significantly improves the reliability of the suggested parameters.

The Initial Control Agent integrates RAG to ground LLM reasoning in verified control theory and design history.

Additionally, the framework incorporates a Safety Filter. Every parameter change suggested by an LLM must pass a simulation-based validation check for stability and operational constraints before deployment to real hardware. A specialized Code Agent translates LLM reasoning into executable Python or MATLAB code for numerical solvers.

Experimental Results

The framework was validated using a high-fidelity simulation of a lab-scale multi-span R2R system. The system included six spans with specific physical properties, and testing involved step changes in tension setpoints and aggressive web acceleration.

The LLM-assisted framework significantly outperformed a standard Model Predictive Control (MPC) baseline:

  • Tension Regulation: Achieved a 55.7% reduction in Root Mean Square Error (RMSE).
  • Velocity Tracking: Achieved an 82.4% reduction in RMSE during aggressive acceleration transients.
  • Efficiency: The system demonstrated rapid adaptation, converging to optimal performance within a few iterations.
  • Diagnostics: The Monitoring Agent successfully distinguished between control-adjustable issues and physical maintenance needs, providing transparent reasoning for its decisions.

Conclusion

This research demonstrates that LLMs can function as effective reasoning engines for complex manufacturing control when combined with domain-specific knowledge and safety guardrails. While the current results are based on high-fidelity simulations, they provide a practical pathway for AI adoption in risk-averse manufacturing environments. Future work will focus on validation on physical R2R hardware and addressing the computational overhead of real-time LLM inference.

Summary updated

Abstract

Roll-to-roll manufacturing requires precise tension and velocity control to ensure product quality, yet controller commissioning and adaptation remain time-intensive processes dependent on expert knowledge. This paper presents an LLM-assisted multi-agent framework that automates control system design and adaptation for R2R systems while maintaining safety. The framework operates through five phases: system identification from operational data, automated controller selection and tuning, sim-to-real adaptation with safety verification, continuous monitoring with diagnostic capabilities, and periodic model refinement. Experimental validation on a R2R system demonstrates successful tension regulation and velocity tracking under significant model uncertainty, with the framework achieving performance convergence through iterative adaptation. The approach reduces manual tuning effort while providing transparent diagnostic information for maintenance planning, offering a practical pathway for integrating AI-assisted automation in manufacturing control systems.
BibTeX
@misc{li2025llm,
  author = {Jiachen Li and Shihao Li and Christopher Martin and Zijun Chen and Dongmei Chen and Li Wei},
  title = {{An LLM-Assisted Multi-Agent Control Framework for Roll-to-Roll Manufacturing Systems}},
  year = {2025},
  eprint = {2511.22975},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.22975},
}
Figure