Skip to main content

Your data, your model: Make physical AI understand your world

Physical AI delivers real value when a model learns how your world works — and what to do when something doesn’t

17 November 2026

Physical AI introduces a new set of terms to get our heads around. Open and closed models are already familiar from the world of large language models (LLMs), but now we also have world models to consider. And, as happened with LLMs, we’ll spend plenty of time debating which approach works best.

Whether you’re a manufacturer, airport operator or city authority, I’d begin somewhere else: What are you actually trying to solve?

Choosing the right model is only the starting point. What makes it valuable is teaching it to understand your world and applying that understanding to the business outcomes you want to achieve.

What exactly is a world model?

An LLM learns patterns in language. A world model learns how the physical world behaves by identifying relationships between objects, actions and outcomes.

The physical world is governed by laws. When you supply an input, a world model predicts the most probable outcome based on the equations we were taught in high school. It uses inputs such as video, images and sensor data — temperature, vibration, radar, lidar, sound and movement, among others — to understand what’s happening in a physical environment.

This becomes particularly interesting when the model encounters something that shouldn’t happen. When someone misses a step in a manufacturing process or enters a restricted area in a building, that anomaly becomes a trigger to act.

World models and physical AI go beyond isolated computer-vision applications to understand and reason based on inputs from heterogeneous sensors. A sufficiently capable model can be adapted to multiple physical use cases rather than requiring a separate AI system for every situation.

Open or closed? It’s not a new debate

Then there’s the question of open and closed models. Broadly speaking, a closed model is controlled by its provider: Organizations can use its capabilities without having full access to the underlying model. Open-source and open-weight approaches make more of the model available for organizations to inspect or adapt themselves, although exactly what is “open” varies considerably between models.

We’ve already seen this debate play out with LLMs. Closed models initially drove innovation and capabilities, while open models increasingly gave organizations another option, particularly when they wanted more control over where models ran and how they were adapted.

I expect the same open-or-closed debate to play out in physical AI. Some of the most sophisticated world models are now being trained on vast quantities of sensor and physical-world data. Over time, I expect more open alternatives that organizations can adapt for their particular environments.

Teach the model to understand your world

So how do you take a pretrained world model and adapt it to your world?

That’s essentially what we did with a multinational materials-handling solutions provider. We took a world model from Archetype and adapted it to understand a specific manufacturing environment, including whether individual steps in the production process had been completed correctly.

Consider an assembly step that requires an O-ring to be installed before two components are joined. Once the assembly is complete, it may look exactly the same, whether the O-ring is present or not. A conventional vision system inspecting the finished product may therefore have no visible defect to detect.

A physical AI model can instead understand the sequence of actions that led to the finished assembly. By recognizing whether the O-ring installation step actually occurred, it can identify a potential defect even when the final product looks correct. Catching that omission during production could prevent a leak or failure later and avoid costly rework, returns or repairs.

Much of the value lies in that kind of process-specific understanding. We expect manufacturers to use world models built on similar foundations but adapted to their own environments. Over time, each model can develop a deeper understanding of the processes, equipment and operating conditions around it — becoming, in effect, a local world model.

The more a model learns, the more useful it becomes

You achieve success with a world model through a combination of training and adaptation. World-model developers do the heavy lifting of pretraining models on enormous amounts of real-time sensor data. You then make this foundation model relevant to your operations by exposing it to the right industrial data and teaching it what is expected and what abnormal looks like.

Over time, the model becomes more specific to your operations or industry, and increasingly capable of spotting anomalies:

  • For a manufacturer, that could mean detecting missed assembly steps, unusual vibrations or rising temperatures before they become quality or maintenance problems.
  • For a city, the same underlying idea could support spotting illegal dumping, recognizing license plates or detecting fires early.
  • At an airport, it could mean recognizing that someone has crossed a security boundary.
  • In a hospital, specialized models could assist with medical imaging and monitoring physical environments for unusual activities.
  • In insurance, satellite and other real-world data could help models identify patterns and anomalies that inform risk predictions.

Sometimes the model needs to live where the world is

There’s another practical consideration: Physical AI can consume an enormous amount of real-time data.

Video is data-heavy, and sensors may generate readings every second, millisecond or even microsecond. Sending all of that data continuously to the cloud can be expensive, while the round trip introduces latency that delays results.

This makes the idea of adapting a world model to run inside your operating environment particularly interesting: putting it inside the factory network, closer to the cameras, machines and sensors producing the data. Combined with technologies such as private 5G and edge infrastructure, this approach can reduce the amount of data that needs to leave the OT environment, supporting lower-latency responses and stronger control over sensitive operational data. It also supports a sovereign AI strategy by giving greater control over where data is processed and where the model runs.

If a production-line model sees that a worker forgot to install an O-ring, you don’t necessarily want to send the video to a distant cloud before an alert is issued. The intelligence can sit much closer to the action.

Start with what moves the needle

None of this means putting a world model everywhere simply because you can. GPUs and compute are expensive. Physical environments generate huge amounts of data, and security becomes critical as you add more models, sensors and devices to the network. And the old rule still applies: garbage in, garbage out.

So, the route from pilot to scale starts with a use case that justifies the investment.

For the materials-handling solutions provider we’ve worked with, maintaining high quality standards and delivering the most reliable products to their customers are key. If another “watchtower” along the production line can catch manufacturing errors before a forklift leaves the factory, the value can be measured against costs the business already understands, such as rework, warranty claims and returns.

At an airport, the equation might be different. One person entering a restricted area could have enormous consequences, making security a natural place to deploy physical AI use cases.

That brings us back to where I started. Before solutioning and diving deep on technical requirements, understand your business. What are you trying to improve? Where do physical-world anomalies cost you money or create risk? What would change if you could recognize those anomalies sooner?

Then find a world model you can adapt to that problem — not the one that knows the most about the world, but the one that can learn the most about yours.

What to do next

Read more about NTT DATA’s Edge AI and Physical AI services to see how we automate data discovery for real-time decision-making.

Author

 Paul Bloudoff

Paul Bloudoff

Paul Bloudoff is a key player in driving innovation and strategic decision-making across NTT DATA's Edge as a Service and Private Wireless Network ecosystems. He has over two decades of expertise spanning various industries, driving innovation and fostering strategic growth within the rapidly evolving landscapes of business operations, corporate development, corporate strategy and legal affairs management. Most recently he was Director of Operations at Momentum Dashboard, a platform designed to enhance focus and productivity. Before Momentum Dashboard, Paul ran his own consulting company, providing strategic advice to a diverse portfolio of clients. He also has extensive experience in drafting and prosecuting patent and trademark applications, contributing to the development of the world’s leading and largest quantum-computing patent portfolio. Outside of work you’ll find Paul travelling to one exotic destination or another and spending time with friends and family. He is based in Vancouver, Canada.