When developers think about an AI system, the architecture often begins with data flowing into a model.
Physical world systems need a broader awareness.
A useful industrial AI architecture must take into account what is, what happens, what it means, and what should happen.
One possible view of that flow is:
Identification
↓
Sensing
↓
AI Decision
↓
Physical Action
↓
Feedback
Aperture Venture Studio is currently using a similar four-engine architecture for their Physical AI and AIoT work.
1. Identification
Prior to analysis of an event, the system needs to know the context of what was sensed.
Which asset?
Where?
Who or what caused the event?
What processes are involved?
Identification technologies define the identity, location, motion, and traceability of any physical entity; its digital "DNA".
Sensor data without that context can be like unstructured data.
2. Sensing
The question then is:
What is happening?
Some form of sensors can provide the information about condition, environmental factors, movement, performance, or operations.
But the key engineering challenge isn't to provide more telemetry; rather, it's to deliver the telemetry that matters to a particular decision.
3. AI Decision
When you can get relevant information about the physical world, AI can begin to add value in transforming that information into something more useful.
Possible functions include:
Anomaly detection
Forecasting
Diagnosis
Root-cause analysis
Optimization
Risk prioritization
Recommendations
The AI Decision Engine described by Aperture takes physical-world data as well as enterprise information and operational constraints to make a recommendation.
That last factor is critical: an AI engine making suggestions for action in context of enterprise constraints.
Otherwise, there's little value in a prediction.
4. Physical Action
Decisions can take many forms, depending on the application. But the question the system must answer is:
What should happen?
The result might require that humans, processes, a machine-to-machine exchange, a control system, or robots perform specific bounded tasks, such as altering production lines, opening a switch, or directing a forklift.
But an important engineering principle is that AI decisions should not be converted automatically to free-reign physical actions.
The Aperture architecture includes advisory, human-approved and fully automated decision-making, guided by limits and operating rules which define how decisions get converted into action.
5. Feedback
Actions shouldn't be taken in isolation, but evaluated for effectiveness and their impact on other activities. Hence the next question must be:
Did what happen match the intended outcome?
Feedback creates a loop in an AI system, comparing what was expected against what actually occurred and how the real-world changed as a result of specific actions.
The end result is a pattern like this:
Observe → Understand → Decide → Act → Verify
↑ |
└─────────────┘
Such a closed-loop system in manufacturing, logistics, construction, mining, and other fields involving physical forces, can be extremely valuable in influencing operations and outcomes.
The engineering takeaway
Engineering Physical AI is not about finding a better model, but about defining systems that take into account:
Data quality,
identity, context, delays,
constraints,
security,
human authority and judgment,
integration,
feedback
That systems approach is what makes industrial AI challenging and rewarding.
A good place to begin further research into the architecture is at Aperture Venture Studio.
Top comments (0)