In a Driving Tomorrow podcast interview with AiNews.com, Archetype AI co-founder and chief scientist Jaime Lien explained how physical AI can help industrial teams turn sensor data into operational intelligence across machines, infrastructure, and environments.

Many businesses already collect data from sensors, machines, cameras, buildings, and industrial systems. The harder challenge is understanding what those measurements mean in real time, especially when companies still rely heavily on human expertise to interpret changing physical systems. For companies operating factories, energy systems, transportation networks, infrastructure, or connected buildings, the decision is whether existing physical data can become a usable intelligence layer for safety, maintenance, uptime, and efficiency.

Lien said physical AI is broader than robotics because it is grounded in real-world measurements, including vibration, pressure, temperature, electricity, radar, acoustics, and other sensor signals. Archetype AI is applying that idea through Newton, a foundation model for the physical world that combines real-time sensor data with natural language so people can ask questions about what is happening across systems, environments, and assets.

In short, Archetype AI is making the case that the next stage of AI adoption in physical industries depends on turning real-world behavior into something businesses can understand and act on. Archetype AI’s approach uses locally adapted world models to help machines and infrastructure become easier for people to interpret, monitor, and act on.

The operational intelligence gap is the distance between having real-world sensor data and being able to understand what a physical system is doing, why it is behaving that way, and what action may be needed next.

The operational intelligence gap is the challenge of turning sensor data from machines, infrastructure, and environments into actionable understanding of what physical systems are doing and what may happen next.

In the interview, Lien described physical AI as a way to help machines, infrastructure, and environments become more understandable through data they already produce. That begins with a problem many industrial companies already face: physical systems generate measurements constantly, but those measurements do not automatically explain what is happening.

Many factories, energy systems, transportation networks, buildings, and infrastructure assets already produce large amounts of data through sensors, machines, cameras, meters, and other measurement tools. The challenge is turning those measurements into a clear understanding of system behavior.

Lien said industrial manufacturing is one area where Archetype AI has seen strong interest because companies often operate tens or hundreds of machines on a factory floor. Those machines may already track internal conditions and environmental conditions, but that data does not automatically give operators a clear picture of how each system is behaving.

She described the core problem as “turning all of these diverse streams of different physical measurements into some sort of actionable insight about what the system is actually doing, how it got into its current state and what is going to happen next”.

That is the operational intelligence gap. A business may know a machine’s temperature, vibration, electrical behavior, pressure readings, or location, while still lacking a clear answer about the machine’s condition. The same measurements could point to normal operation, early degradation, an unsafe condition, or a pattern that requires human attention.

Archetype AI’s supporting blog post, “Adaptive World Models: Closing the Operational Intelligence Gap,” describes the same challenge from a technical perspective. The company says physical industries, including energy, manufacturing, transportation, and infrastructure, account for roughly 85% of global economic activity, yet many of those systems remain difficult for AI systems trained mainly on text, images, and other digital content to understand.

The difficulty comes from how physical systems express their behavior. Unlike text or images, machines and environments often produce many streams of sensor data at once, including vibration, pressure, temperature, acoustics, electrical current, telemetry, radar, and other signals. Each signal gives only a partial view of the system, so understanding what is happening requires interpreting many measurements together over time.

That is why the business problem is larger than collecting more operational data. The physical economy depends on systems that already generate evidence about their own behavior, but much of that evidence still requires experienced people to interpret it before companies can act.

For many people, the phrase physical AI still brings robotics to mind first. Lien said robotics belongs in the category, but it does not define the whole field.

“So to me, robotics is a subset of physical AI,” Lien said. “Physical AI in the bigger picture just means AI that is grounded in actual physical measurements and observations of the real world.”

That distinction is important because many widely used AI systems are trained mainly on text, images, code, documents, and online content. Physical AI has a different starting point. It works from measurements generated by real-world systems.

Camera data can be one modality, but Lien emphasized that physical AI is not limited to vision. It can involve gas, pressure, electricity, radar, weather data, vibration, acoustics, and other measurements that capture how machines, environments, and infrastructure behave.

That is one reason physical AI matters for industries outside robotics. A factory machine, HVAC system, wind turbine, construction site, traffic intersection, or electrical system may produce data that reveals something important, but the pattern may be too complex, too subtle, too large, or too small for people to detect manually.

Lien said some sensor modalities can capture phenomena people cannot directly sense. Humans cannot feel electrical current running through a wire or see radio frequency, or RF, waves moving through an environment. Sensors can measure those phenomena, and AI can help interpret them into insight that teams can use for decisions about safety, maintenance, operations, or repairs.

Physical AI also helps when the scale of the data is beyond human perception. Lien pointed to Archetype AI’s work with the city of Bellevue to analyze pedestrian safety at traffic intersections. In that use case, AI combined information from cameras and traffic systems to detect patterns around when people step into intersections or stay back.

The model used those patterns to recommend when unsafe situations might be about to occur, giving city teams a way to adjust traffic behavior before risk turns into an accident.

Archetype AI built Newton as a foundation model for the physical world. In the interview, Lien explained that Newton combines sensor data with natural language so people can ask questions about real-world systems, environments, and assets.

Lien said Newton is built around locally adapted world models, a term that describes two parts of Archetype AI’s approach. The model needs to understand real-world behavior, and it also needs to adjust to the specific machine, environment, and operating history it is observing.

“World models because they understand what is happening and the behavior of the real world and locally adapted because they can actually adjust based on local sensor observations to the particular physics and operating history and contextual information of each individual machine,” Lien said.

That local adaptation matters because physical systems do not behave the same way everywhere. Lien said even machines of the same asset type can differ because of hardware configuration, deployment environment, operating history, wear and tear, and degradation.

She described the physical world as “completely dynamic” and said the technical term is non-stationary, meaning the system’s behavior changes over time rather than staying fixed.

A drilling rig, for example, may be the same nominal type as another drilling rig, but its hardware configuration, degradation path, location, and environmental conditions may cause it to behave differently. A model that treats both machines as identical could miss important operational differences.

The key point is that Newton is designed to combine a general foundation model with local adaptation, so the system can learn from real-time observations and adjust to the specific physics, behavior, and operating history of each deployment.

Lien said Newton also provides a language interface on top of that machine-native understanding.

“So now this AI native, this machine native understanding of physical behavior can be translated and interacted with by a human through natural language,” she said.

That means a team could query or prompt the model about a machine or system, then receive an answer in language people can understand. In practice, the interface is meant to move operational intelligence from specialized interpretation toward a more accessible question-and-answer workflow.

Archetype AI’s blog adds technical context to that idea. The company describes Adaptive World Models as systems that learn patterns from sensor data instead of relying only on predefined rules or labels. In practical terms, the model needs to recognize the different states a machine moves through, such as normal operation, unusual behavior, or a possible transition toward failure.

The blog uses the term operational ontology for that structure. In simpler terms, it means the model builds a usable map of how a specific machine or system behaves over time.

That explanation reinforces Lien’s interview point. Physical AI needs to discover how a specific system behaves in its own environment rather than relying only on generic rules, fixed labels, or one-size-fits-all machine assumptions. That makes physical AI different from systems built around static rules because the model has to keep learning from real-world conditions as machines, environments, and operating patterns change.

That need to keep learning from real-world conditions leads to how Archetype AI trains its physical AI models. Lien said a core part of the company’s approach is that models should learn from sensor data itself instead of depending on humans to label every condition or define the system through equations in advance.

Lien said models should learn from sensor data “so that humans don’t have to go in and provide a lot of labeled data. They shouldn’t have to inject equations of the physical laws as we understand them.”

That means the model is not trained only to recognize categories humans have already named. Archetype AI wants the system to learn common patterns of physical behavior from the measurements themselves.

“Basically from sensor observations alone, the machine should be able to understand what these common underlying physical behaviors and dynamics actually are,” Lien said.

Archetype AI trains models by combining observations across a large, sparse, but diverse set of sensor modalities, physical assets, and physical domains. That training exposes the model to many kinds of physical measurements, while local adaptation helps the model adjust to the particular system it is observing.

This is where the company’s approach differs from traditional industrial analytics or classical machine learning. Lien said people have tried to solve operational intelligence problems for a long time, but previous techniques have not scaled well across the long tail of sensors, machines, and physical environments.

A foundation model approach is meant to reduce the need for custom engineering on every machine or system. That matters because operational intelligence becomes difficult to scale if each factory line, wind turbine, HVAC system, drilling rig, or industrial asset needs its own separate model or rule set.

Lien described that as the promise of building one solution that can scale across companies, physical environments, and assets without requiring heavy per-system or per-machine engineering.

In the blog, Archetype AI argues that structure should be learned rather than specified because important operational patterns are often unknown in advance. Degradation pathways, anomalous transitions, and machine-specific behaviors may not fit neatly into predefined labels.

That is one reason the operational intelligence gap has remained difficult. A system built only around known fault conditions could miss unusual or emerging behavior, leaving operators without early warning when a machine begins moving toward failure, unsafe operation, or costly downtime.

When asked where physical AI can create practical value, Lien pointed to three categories: real-time monitoring, anomaly detection, and safety.

Real-time monitoring can help teams understand whether systems are operating within normal or safe ranges before problems create downstream failures. Anomaly detection can help identify when a machine or system is behaving abnormally in a way that may predict a future fault.

Safety adds another layer because machines often operate near people. Lien said physical AI can combine sensor measurements with human behavior to help keep operators and nearby workers out of risky situations.

“How do we ensure that human operators and humans in the vicinity of these machines are kept in a safe environment and not exposing themselves to any sort of risk or hazard,” she said.

Those use cases are not limited to heavy manufacturing. When asked about commercial HVAC systems, where large air conditioning and cooling units often already contain sensors that track whether equipment is running properly, Lien said that area is a major field of interest because the sensing layer already exists in many parts of the physical world.

About 15 years ago, Lien said, the IoT wave embedded sensors into machinery and environments to measure physical quantities. What is changing now is the ability to connect those existing measurements with large models and self-supervised learning.

“I think what we’re seeing now is that the combination of large models and self-supervised learning and the commoditization of these sensors are all converging so that AI can actually be the common intelligence layer across all of these different things,” Lien said.

For many organizations, the sensing layer is already in place. Machines, buildings, equipment, and environments are already producing measurements. The next step is adding intelligence that can interpret those measurements across systems and over time.

Lien’s experience leading Soli Radar at Google’s Advanced Technology and Projects group also shaped how she thinks about deploying physical AI. Soli used radar sensing for gesture and motion detection in consumer products, which required moving advanced sensing technology from the lab into real-world environments.

She said that experience reinforced how much real-world diversity matters. A system may work in a lab, but it still has to perform across different deployment areas, user behaviors, materials, and environmental conditions.

Radar, for example, can be sensitive to the enclosure or materials around it. Those details matter because physical AI systems must work in environments that include weather, hardware differences, location-specific conditions, and changing user behavior.

That is also why Archetype AI emphasizes the connection between technology and design. Lien said the company considers user experience from the beginning because people need to understand and act on the technology’s output.

“It can’t be technology first and then design second,” she said. “It’s really a co-development.”

For operational AI, clear communication is part of the product’s value. If a model identifies machine behavior but cannot explain that behavior in a usable way, operators and business teams may not know what to do next. The system needs to turn machine-native understanding into language, alerts, explanations, or recommendations that operators, engineers, safety teams, and business leaders can act on.

World models have become a major AI topic, but Lien said Archetype AI is using the concept in a specific way. Many world models focus on predicting what comes next or generating realistic spatial simulations for gaming, robotics training, or autonomous systems.

Archetype AI’s world models focus on operational understanding. The goal is to help teams understand the current state of a machine or system, how it reached that state, and what that state may indicate.

“Can we explain it? Can we understand how it got there? Can we understand what that means in terms of what is going to happen next?” Lien said.

Archetype AI’s world models focus less on simulating future environments and more on helping organizations understand current operational states, how a system reached that state, and what that may mean for operations, risk, or maintenance.

The blog’s discussion of operational ontology supports the same idea. If a model can represent machine behavior as understandable states and transitions, people can reason about system behavior more clearly, compare today’s operating state against previous states, identify uncertainty, and communicate with the AI system in more practical terms.

Industrial decisions often depend on context. A vibration pattern, temperature reading, or pressure change may mean different things depending on the machine, the environment, the operating history, and what happened before. Operational world models are valuable because they are designed to preserve that context instead of treating each measurement as an isolated signal.

One of the most interesting parts of the conversation came near the end, when Lien described what may surprise people about physical AI over the next few years.

“To me, what is most exciting and what will be most impactful is when physical AI can actually surface phenomena or insights that we didn’t know already,” she said.

She described the physical world as something people can see and hear, while much more remains beneath direct perception. AI could expand human understanding by making those hidden patterns visible.

Archetype AI has already seen examples of this in energy. Lien described work with an energy company that operates a fleet of wind turbines. Each turbine had roughly 40 to 60 sensors, creating a large amount of nonbiological measurement data.

The value came from discovering when turbines were behaving abnormally and identifying commonalities in normal operation that were not obvious from looking at the raw data traces.

The wind turbine example expands the operational intelligence gap beyond faster monitoring. A system that learns from physical observations may also support discovery by identifying behavior that experts did not know to label in advance.

Physical AI still depends on human expertise, especially in industrial settings where operators understand equipment, context, and risk. Its value comes from helping experts see more clearly, ask better questions, and detect conditions that would otherwise remain buried in operational data.

Archetype AI’s approach points to a major opportunity for physical industries, but adoption will depend on how well companies connect physical AI to their existing systems, teams, and operational decisions.

Newton is designed to address one of the biggest barriers to scale. Lien said the promise of a foundation model approach is that one solution can work across companies, physical environments, and assets without requiring heavy per-system or per-machine engineering.

That does not remove the work of deployment. Companies will still need to identify where they already have useful sensor data, which machines or systems produce the highest-value signals, and which operational decisions could improve if teams had a clearer understanding of system behavior.

Trust will also depend on how the technology fits into real workflows. Physical AI may support decisions about safety, downtime, maintenance, energy efficiency, or infrastructure behavior. In those settings, teams need outputs they can understand, compare with existing expertise, and use when deciding what action to take.

That makes human expertise part of the adoption path. Experienced operators already understand how machines behave through years of observation, and physical AI will be most useful when it helps those experts see patterns earlier, test assumptions, and respond with more confidence.

The adoption question is not whether physical AI can collect more data. The more important question is where companies can turn existing physical data into operational decisions that are safer, faster, more efficient, or easier to explain.

Q: What is Archetype AI trying to solve with physical AI?
A: Archetype AI is using physical AI to help close the operational intelligence gap by turning sensor data from machines, infrastructure, and environments into usable insight about real-world behavior.

Q: How does Newton help people understand what machines are doing?
A: Newton combines real-time sensor data with natural language so people can ask questions about machines, environments, systems, and assets. The model uses locally adapted world models to adjust to each system’s physics, operating history, and local conditions.

Q: Is physical AI just another term for robotics?
A: No. Physical AI includes robotics, but it also covers AI grounded in physical measurements such as vibration, pressure, temperature, electricity, radar, acoustics, traffic signals, and other sensor data. That makes it relevant to factories, infrastructure, buildings, energy systems, and industrial equipment.

Q: Why is the operational intelligence gap becoming more important?
A: The operational intelligence gap is becoming more important because many organizations already collect sensor data, but they still need better ways to understand what those measurements mean in real time. Large models, self-supervised learning, and widely available sensors make it more practical to turn physical data into operational intelligence.

Q: Where would companies actually use physical AI?
A: Companies could use physical AI for real-time monitoring, anomaly detection, safety, maintenance, and operational optimization. In the interview, Lien discussed industrial manufacturing, traffic intersections, HVAC systems, and wind turbines as examples of physical systems where AI can help interpret measurement data.

Q: What should companies be careful about before adopting physical AI?
A: Companies should not assume physical AI is already proven across every industrial environment. Archetype AI’s approach is designed for local adaptation, but companies still need to validate performance against their own assets, sensors, data quality, safety requirements, and operational workflows.

Jaime Lien’s Driving Tomorrow interview with AiNews.com highlights a core challenge for physical industries. Many organizations have data-rich systems, but the operational meaning of that data is still difficult to extract in real time.

The key point is that sensor data alone does not create operational intelligence. Companies need AI systems that can interpret physical measurements in context, adapt to local machine behavior, and explain system behavior in language people can use.

Manufacturers, energy companies, transportation teams, building operators, city planners, infrastructure leaders, and safety teams should pay attention because advances in physical AI could affect how they monitor assets, identify risk, plan maintenance, and improve resource use.

Physical AI is becoming more practical because many machines and environments already have sensors in place. As large models and self-supervised learning improve, companies have a clearer path to learn from unlabeled physical data and use AI as an intelligence layer across physical operations.

For business leaders, the question is where physical AI can create measurable value first. Organizations will need to identify where they already have useful sensor data, where human interpretation is stretched thin, and where better operational understanding could improve safety, uptime, maintenance, sustainability, or cost.

In short, Archetype AI is making the case that physical AI should be understood as an operational intelligence layer for the real world. Newton and locally adapted world models are the company’s proposed path for helping machines, infrastructure, and environments become easier to understand through AI.

The real test for physical AI will be whether it can turn the world’s machine data into decisions people can trust.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.