Physical AI and AIoT: Designing the Data Chain from Sensors to Smart Actions
Artificial intelligence becomes far more interesting when it leaves the digital realm and starts interacting with the physical one.
A factory machine emits telemetry. A vehicle updates its location. A sensor vibrates. A robot reports its coordinates. A production line changes its mode.
As individual events, these are only data.
The engineering task ahead is to convert them into reliable context, decisions, and ultimately physical actions.
This is essentially the domain in which physical AI and AIoT intersect.
AIoT can be thought of as connected physical systems augmented with artificial intelligence and analytics, while physical AI goes a step further to think about these systems from an understanding-and-interaction perspective.
A useful architecture would tie the following key stages together:
Identify → Sense → Transport → Understand → Decide → Act → Verify
The interesting engineering tasks are between them.
Before an intelligent system can reason about a physical event, it needs to classify it correctly.
Consider a manufacturing environment.
A sensor might be reporting:
temperature = 82.4C
This is a useful observation by itself, but not particularly actionable.
The system would also need to understand:
asset_id = motor-204
location = assembly-line-3
timestamp = 2026-09-30T14:21:03Z
operating_mode = high_load
Now it can infer that this is a machine in high load conditions which is exhibiting a temperature anomaly.
Identification can take many different forms, including RFID, barcodes, BLE, UWB, device identity, vehicle identifiers or other means depending on the physical environment.
The essential insight remains the same, however: that to have value, physical data needs to have identity and context.
The next stage is sensing.
An industrial environment has an extraordinary amount of telemetry that can be collected from various sources:
Temperature
Pressure
Vibration
Position
Speed
Current
Voltage
Machine state
Environmental conditions
Production events
Only a small fraction of which should be transported to the cloud for processing.
Which is why edge computing is a critical enabler of AIoT systems.
An edge application might process sensor information from vibrations and only report high-value events.
A simplified architecture would look like:
Sensor
↓
Edge Gateway
Filtering / Feature Extraction
Event
Message Broker
AI / Analytics
Application
AIoT systems often involve a large number of independent devices and applications.
A tightly coupled architecture becomes unmanageable as the number of individual elements grows.
A different approach is to use event-driven communication.
A machine would publish an event like:
{
"asset": "motor-204",
"event": "temperature_anomaly",
"value": 82.4,
"timestamp": "2026-09-30T14:21:03Z"
}
And a message broker would relay it to interested parties.
An analytics application can store said event for later processing.
A maintenance application can notify operators.
An AI system can update a model.
A support system can create a work order.
This decoupling is valuable, particularly in industrial architecture.
A variety of message brokers, MQTT and Kafka among them, can provide the underlying infrastructure for such systems.
Feeding raw telemetry data into an AI model is not likely to yield valuable results.
The model requires context in order to make accurate predictions and inferences.
If an AI model receives: vibration = 8.7
It would not know which machine, sensor, environment, operating condition or load caused this value.
It would also not understand this value in isolation, without knowing the expected range or previous conditions leading up to this reading.
Context is needed for operational decisions, and it can originate in different systems:
Sensor data
Asset identity
Location
Production state
Historical data
Operational context
AI inference
This is one reason why the AIoT architecture is not just simply tying an AI model into an IoT database.
One of the most critical aspects of AIoT design is the separation of AI predictions and physical operations.
An AI model might conclude:
Probability of equipment anomaly: 87%
But should not directly control physical equipment.
A safer design is to apply decision policies and human oversight:
Policy / Rules
Risk evaluation
Human approval or automation
Action
For low-risk scenarios, operational automatons might be acceptable.
For safety-critical or mission-critical equipment, additional oversight might be needed.
The architecture should therefore treat AI intelligence and operational control as separate responsibilities.
A physical AI system should ideally have a way of knowing that an action had the desired effect.
Imagine you had an automated system that suggested a parameter change in a production machine.
The system should not declare success after submitting the request.
The system needs to observe an outcome after the state change:
Observe
Infer
Decide
Act
Observe again
The verification step can be as simple as observing a new device state, or more involved with multi-sensor and production data analysis.
This is one reason why verification is vital, given the non-deterministic nature of many physical environments.
One of the frequent mistakes when designing AIoT applications is to assume that existing industrial infrastructure can be replaced.
A real-world setting rarely resembles something one builds in isolation.
It may include:
PLCs
SCADA systems
MES platforms
ERP systems
Industrial databases
Legacy equipment
Multiple network protocols
Existing safety systems
And so on.
A practical AIoT architecture needs to integrate with this infrastructure, rather than replace it.
Industrial integration technologies such as OPC UA, MQTT, APIs and edge gateways can be valuable enablers.
The objective is rarely to build something from scratch, but rather to define a robust intermediary layer that can facilitate useful information flow between legacy equipment and new applications.
The most interesting aspect of Physical AI may very well be the system around the model, not the model itself.
A production-ready architecture has to address the following issues:
Data quality
Device identity
Connectivity
Edge processing
Message delivery
Storage
Cybersecurity
Reliability
Observability
Human factors
Physical safety
Interoperability
This is why Physical AI intersects several different fields of engineering:
Software engineers: Build applications and APIs
Data engineers: Build processing pipelines
IoT engineers: Connect physical devices
Automation engineers: Understand industrial control systems
AI engineers: Build inference and learning systems
Security engineers: Ensure the integrity of infrastructure
The entire system has to come together for the whole to be greater than the sum of its parts.
A practical architecture
A simplified design for a Physical AI system might look like this:
┌───────────────────────────────┐
│ Physical Environment │
│ Machines / Vehicles / People │
└───────────────┬───────────────┘
│ Identification + Sensors │
│ RFID / BLE / UWB / IoT │
│ Edge Layer │
│ Filtering / Processing │
│ Event & Data Infrastructure │
│ MQTT / Kafka / APIs / Storage │
│ AI + Analytics │
│ Detection / Prediction │
│ Decision Layer │
│ Policies / Human Oversight │
│ Physical / Operational Action │
Verify
│
└──────→ Feedback
The architecture is deliberately generic, and different industries would implement specific layers in different ways.
Where venture building fits in
Building this type of infrastructure can often uncover valuable opportunities for software companies.
A specific industry might have a recurring problem tied to asset visibility, production intelligence, sensing or decision automation.
A venture-building approach would begin by identifying the operational problem, building the necessary technology system, verifying the concepts in a physical environment and determining if the capabilities could evolve into an economically viable repeatable product.
Aperture Venture Studio's published architecture speaks to building companies around the concept of tying identification, sensing, AI decision-making and physical-world actions into systems. This approach highlights an essential realization about Physical AI:
While the value is in the application itself, it is not enough to merely build a digital AI model. The product is the system that enables intelligence in the physical environment to achieve tangible outcomes.
Final Thoughts
Physical AI and AIoT tend to talk a lot about futuristic robots and machines.
The more immediate engineering opportunity may be the infrastructure behind connected events.
When a system can identify an asset, sense its properties, understand context, decide on an appropriate action, perform it in a controlled way, and verify the outcome, artificial intelligence becomes a part of an operational feedback loop.
That is a fundamentally different problem than building an AI application on purely digital data.
For engineers building the next systems, the important question may not be: "Where can we apply AI?"
but rather:
"What is the process we need to understand, and what information does the system need to have in order to support and enhance that process?"