【Control】🧩 01. AITL-controller

— An Integrated Three-Layer Control Architecture: PID × FSM × LLM

---

This article introduces AITL-controller,
a lightweight framework that integrates a three-layer control architecture:

• PID (real-time control)
• FSM (state supervision)
• LLM (intelligent control and redesign)

The framework is designed for education and research use,
making advanced control architectures accessible and reproducible.

• Official site:
  https://samizo-aitl.github.io/aitl-controller-a-type/

• GitHub repository:
  https://github.com/Samizo-AITL/aitl-controller-a-type

---

🧠 What Is AITL-controller?

AITL-controller is a framework that unifies classical control, formal control, and intelligent control
into a single, coherent architecture suitable for learning and experimentation.

The three layers are clearly separated by responsibility:

• PID: numerical stability and real-time performance
• FSM: explicit system states and transitions
• LLM: diagnosis, adaptation, and redesign

This separation is intentional and fundamental.

---

🧱 Three-Layer Architecture (PID × FSM × LLM)

1️⃣ Inner Loop: PID (Real-Time Control)

• The innermost layer directly controls the plant or robot
• Examples: angle stabilization, velocity control, posture regulation
• Responsible for stability and dynamic response

PID runs at high frequency and handles
millisecond-level real-time behavior,
exactly as in classical control systems.

---

2️⃣ Middle Loop: FSM (Finite State Machine)

• The supervisory layer managing system modes and state transitions
• Examples:
  • Idle → Running → Fault
  • Lift → Hold → Place

FSM determines:

• Which PID controller is active
• Which parameters are enabled
• When transitions should occur

It acts as an explicit and inspectable control supervisor.

---

3️⃣ Outer Loop: LLM (Intelligent Control)

• Detection of faults, anomalies, and performance degradation
• Responsible for reasoning, diagnosis, and redesign

Typical roles include:

• Automatic re-identification of PID parameters
• Proposing modifications to state-transition rules
• Inferring abnormal trends from operation logs

This layer leverages LLM strengths such as:

• Adaptation to out-of-model behavior
• Natural-language explanation and design rationale generation

---

❓ Why a Three-Layer Architecture?

Traditional control systems can be built using PID and FSM alone,
but they suffer from inherent limitations:

• Poor adaptability to environmental changes
• Manual fault analysis by human experts
• High skill requirements for parameter tuning

AITL-controller assigns these responsibilities to the LLM layer,
with the goal of enabling:

Self-improving control systems

at the education and research level.

---

📦 Repository Structure (Overview)

aitl-controller/
├── core/
│   ├── pid/             # PID modules (stability & performance)
│   ├── fsm/             # FSM modules (state transitions)
│   └── llm/             # LLM modules (redesign & reasoning)
│
├── demo/
│   ├── inverted_pendulum/   # Inverted pendulum demo
│   ├── quadrotor/           # Quadrotor control example
│   └── simple_robot/        # Small-scale robot systems
│
├── docs/
│   ├── architecture/        # Three-layer architecture explanation
│   ├── math/                # PID & FSM mathematics
│   ├── llm/                 # Design guidelines for LLM control
│   └── examples/            # Tutorials
│
└── assets/                  # Figures and workflow materials

The framework core resides in core/,
while demo/ provides immediately runnable examples.

---

🔄 Typical Control Loop Structure

AITL-controller assumes the following control loop:

Sensor → PID → Actuator → Plant
        ↑        ↓
        FSM ← LLM (Outer loop)

Role Summary

• PID: real-time error correction
• FSM: orchestration of PID controllers and modes
• LLM: analysis and redesign of PID and FSM structures

---

🧪 Example: Inverted Pendulum Demo

In the inverted pendulum example, the three layers cooperate as follows:

1. PID: stabilizes the pendulum angle
2. FSM: manages states such as
   • Stabilize
   • Recover
   • Fault
3. LLM:
   • Analyzes the cause of failures
   • Suggests improved PID gains
   • Proposes refinements to state-transition logic

The LLM never directly drives actuators.
Its role is strictly to improve the control structure itself.

---

🧩 The Role of LLM Control (Design Policy)

In AITL-controller, the LLM layer is assigned three explicit tasks:

1. Monitoring

• Detect anomalies and performance degradation from logs

2. Diagnosis

• Identify root causes in an explainable manner

3. Redesign

• Improve PID gains
• Modify state-transition tables
• Refine supervisory rules

All improvements are expressed through
natural language, code generation, or structured design proposals.

---

⚙ Setup

git clone https://github.com/Samizo-AITL/aitl-controller.git
cd aitl-controller
pip install -r requirements.txt

Full documentation is available on the
official GitHub Pages site.

---

🎓 Educational and Research Applications

AITL-controller is suitable for:

• Learning classical control (PID, FSM)
• Research on AI/LLM integration into control systems
• Advanced robotics control education
• Validation of autonomous redesign algorithms
• Explainable AI (XAI) for fault handling

In education, it naturally supports the progression:

Understanding PID/FSM fundamentals
→ Improving them using LLMs

---

📝 Summary

AITL-controller integrates three distinct layers:

• PID: real-time control
• FSM: explicit state supervision
• LLM: intelligent redesign and reasoning

Built on classical control principles, it aims to become
an educational and research platform for next-generation control systems
in the age of AI.

---

• Official site:
  https://samizo-aitl.github.io/aitl-controller-a-type/

• GitHub repository:
  https://github.com/Samizo-AITL/aitl-controller-a-type