This repository implements the system described in the paper
"Agent based Autonomous Manufacturing from Planning to Production: Experimental Study on Automatic Tool Generation and Control Using Large-Scale AAS Process Information".
It presents a fully functional agent-based autonomous manufacturing pipeline powered by LLMs, AAS, and MCP.
A2M/
├── AAS/ # Asset Administration Shell files (partial public subset - AFPM only)
├── Evaluation/ # LLM code evaluation framework & prompts
├── FastMCP_Tools/ # Automatically generated MCP tools
├── AI_Agent/ # Agent interface for orchestration & command execution
├── AAS_Server/ # BaSyx-based API server for AAS registry, repo, and discovery
├── configs/ # IP, network and connection configurations
├── examples/ # Demonstration flows and template interactions
├── README.md
└── LICENSE # Apache 2.0
A2M is a practical framework that bridges standardized industrial metadata (AAS) with AI planning and control (LLM Agents).
The system:
- Parses and classifies large-scale AAS files (4,588 total, 1,005 with common operations)
- Prompts LLMs (Claude, GPT, Gemini, Gemma) to generate MCP tools for device control
- Evaluates tool quality using structured, multi-dimensional criteria
- Operates end-to-end AFPM motor production based on user natural language commands
🔒 Only AAS files related to AFPM Motor Production are included publicly. The full dataset belongs to Kyungnam University and is available upon request.
The following diagram shows the high-level architecture of A2M,
connecting the AI Agent, AAS Server, MCP Server, and real manufacturing facility via a local network.
- Based on BaSyxMinimal example
- Provides AAS Registry, Submodel Repo, and Discovery endpoints
- Acts as an API metadata server for the AI Agent and FastMCP
- Tools automatically generated by LLMs from AAS Operation submodels
- Implements dummy PLC control logic using
pymodbus - Controlled via the AI Agent at runtime
- Orchestrates:
- Process discovery via AAS registry
- Tool availability check via FastMCP
- Tool invocation from natural language
- Implements real-time monitoring and parameter mapping
- Prompts and scores generated tools across 10 criteria
- Visualization support for heatmaps and comparison tables
- Uses Claude and GPT-based evaluators
- Conducted at ANYTOY Inc. with actual AFPM motor line
- Used Mitsubishi PLC and real AAS-mapped equipment
- Full workflow from
calculate_required_turns()tostart_manufacturing()validated
The A2M framework was deployed on a real AFPM motor manufacturing line at ANYTOY Inc.,
integrating Mitsubishi PLCs, AAS-based tool modeling, and LLM-based orchestration.
Watch the full demo of A2M in operation on the actual AFPM motor line:
This project only includes AFPM-related digital twins.
To access the full 4,588-file AAS dataset:
📩 Email: jms663100@kyungnam.ac.kr
| Name | Role |
|---|---|
| Namhyun Yoon | Project Director |
| Jinhong Yang | Overall System Architecture Design |
| HoanSuk Choi | Generation of Process-specific AAS Files & AAS API Design and Development |
| Minsu Jo | Agent Module Development & MCP Server Implementation |
| Name | Role |
|---|---|
| Gisu Sin | AFPM Process—Overall System Architecture Design |
| Changyong Kwak | AFPM Process—PLC Code Development |
Apache License 2.0
This work was supported by the Institute of Information & Communications Technology Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2024-00436773).
Supported by IITP, MSIT (IITP-2025-RS-2024-00436773)