37.【Physical AI Design】🏗️ How to Build Systems That Don’t Break

The Three-Layer Architecture: PID × FSM × LLM

tags:

• Physical AI
• AITL
• PID Control
• FSM
• Architecture Design

🏗️ How to Build Physical AI That Doesn’t Break

The Three-Layer Architecture: PID × FSM × LLM

In the previous article, we explained why
directly connecting an LLM to a control loop inevitably causes failure.

The next natural questions are:

“So where should we place the LLM?”
“How do we arrange the system so it doesn’t break?”

The answer is simple—yet unless it is explicitly designed, it will be violated.

That answer is the three-layer architecture:

PID × FSM × LLM

🧠 The Conclusion Up Front

To keep Physical AI systems from breaking, you must clearly separate:

• Real-time control
• State management and safety
• Reasoning and redesign

The moment you let a single component handle all of them,
the system collapses.

🧱 The Three-Layer Architecture at a Glance

[ LLM ]        ← Reasoning / Re-design (non-real-time)
   ↑
[ FSM ]        ← Mode management / Safety transitions
   ↑
[ PID ]        ← Real-time control / Stability

The key rule is simple:

The deeper the layer, the stricter the requirements.
Upper layers may be slow. Lower layers may not fail.

🟢 Inner Layer: PID (Real-Time and Stability)

The Role of PID

• Real-time control
• Stability guarantees
• Immediate response

PID control may look old-fashioned, but it is
the only layer directly connected to the physical world.

• Response times on the order of milliseconds
• Oscillation or divergence leads directly to damage
• Behavior can be analyzed mathematically

What this layer demands is not intelligence, but certainty.

👉 AI must never be placed here.

🟡 Middle Layer: FSM (State and Safety)

The Role of FSM

• Mode management
• Explicit state transitions
• Exception and error handling
• Safety guarantees

Most Physical AI systems already have an FSM—
even if it is never written down.

Typical states include:

• Initializing
• Idle
• Active
• Error
• Safe stop

If these states are not explicitly coded,
LLMs or ad-hoc human logic will eventually step over them.

The FSM decides:

• “Is this action allowed now?”
• “This action is never allowed.”

It is the system’s last line of defense.

🔵 Outer Layer: LLM (Reasoning and Re-design Only)

The Correct Role of an LLM

• Situation assessment
• Policy decisions
• Parameter re-design
• Interpretation of logs and history

The critical mindset here is:

Use the LLM to think—not to act.

In this layer:

• Real-time performance is unnecessary
• Non-determinism is acceptable
• Delayed responses are fine

That is exactly why the LLM belongs on the outside.

🚫 Forbidden Placements

❌ Putting an LLM inside the PID loop

• Latency causes immediate instability
• Zero reproducibility

❌ Letting an LLM decide safety

• Decisions fluctuate with context
• “It seemed okay this time” becomes an accident

❌ Omitting the FSM

• State management collapses
• Everything becomes ad hoc and brittle

✅ The One Design Rule to Remember

Let the LLM think.
Do not let it move.

That single sentence prevents
90% of Physical AI accidents.

🧭 What This Architecture Really Means

This three-layer structure makes one thing explicit:

• Physical AI is not an AI algorithm problem
• It is a system architecture problem

LLMs are powerful—
but they are not universal.

Only when placed in the correct layer
do they become truly usable intelligence.

🔜 Next Article

In the next piece, we will show—visually and concretely—
that this architecture actually works by comparing:

• PID only
• PID + FSM (AITL)

side by side in a running demo.

Not theory.
Just behavior.