How Physical AI Agents Are Transforming Manufacturing

A model trained on one machine usually fails on the identical machine one aisle over. Same make, same sensor package, different humidity, different structural vibration, different rate of wear. The model that worked yesterday starts drifting today. That single fact is why most Physical AI projects in manufacturing stall before they leave the pilot stage.

In our recent webinar, "How Physical AI Agents Are Transforming Manufacturing," Sisinio Baldis, Head of Solutions Engineering at Archetype AI, and Raleigh Murch, Managing Director of the Physical AI Practice at NTT Data, walked through what changes when the model adapts to the plant instead of the plant adapting to the model. The discussion, moderated by Archetype's Aristo Chang, centered on a live deployment with forklift manufacturer Hyster-Yale. This post recaps the core ideas. The full recording is linked at the end.

Why manufacturing breaks traditional industrial AI

Every plant is a collection of assets from different manufacturers, with different sensor packages, retrofitted at different times. Environmental conditions vary from one site to the next, and components degrade at different rates while production schedules shift underneath them. Conditions end up being inherently local.

Traditional industrial AI answers that variability the only way it can: one model per application, per machine, per fault type. You collect data, train a model, deploy it, then retrain every few months as it drifts. Raleigh, who led computer vision implementations across the Fortune 200 at AWS before NTT Data, described the pattern from experience. A model gets trained to recognize one specific defect or one pallet type, the environment changes, and you are left with a brittle model and a retraining schedule. It works, but it does not scale across a plant, let alone across sites.

The result is the 100-models-by-50-plants problem: an approach where coverage grows only as fast as you can hand-build and maintain each model.

What Physical AI Agents do differently

Newton is a foundation model pre-trained on billions of sensor measurements across many domains and modalities. Because it learns the structure of physical signals rather than one machine's labeled history, it can read a sensor stream it has never seen and often needs only a few examples of what you are looking for, with no model training required, to start working on your specific case.

Those capabilities reach the floor as Physical AI Agents deployed on the Archetype platform. Two details mattered most to the manufacturers in the room. First, the platform runs entirely inside your own infrastructure, so data never leaves your security perimeter and there is no dependence on an external cloud. Second, the Physical Agents run on the edge, whether on a rack server inside the plant or directly on-device, next to the operation they monitor.

Sisinio framed the shift plainly. For decades, vision systems and signal analysis worked well in narrow applications but stayed brittle and expensive to maintain. A generalized model that you can pull out of the box, apply to your assets, and improve over time removes most of that maintenance burden. The unlock is not a better version of the old task. It is a new capability.

Inside the Hyster-Yale deployment

Hyster-Yale came to NTT Data with a quality problem showing up at the end of the line. Rather than inspect finished units, the team looked upstream. Sisinio and Raleigh ran a workshop and site assessment at the Berea, Kentucky plant, walked the line, and picked a single assembly work cell for a proof of concept.

They installed cameras, configured edge compute and connectivity, and captured roughly 40 hours of video to establish what was actually happening in the cell. Newton then compared that footage against the standard operating procedure: what was happening versus what should be happening.

The findings split into two kinds. Some were genuine quality issues. Others were deviations from the SOP that turned out to be intentional, cases where senior assemblers had developed more efficient movements that had outpaced the written procedure. That second category pointed to an opportunity nobody had put in the original scope.

From compliance checks to capturing expert knowledge

Once you can observe expert work at the cell level, you can do more than flag errors. You can reverse-engineer what your best people already know. Raleigh described a request now coming from customers that was never part of the pitch: with an aging workforce, can the technology observe skilled operators and help document an optimized set of work instructions? That maps directly to one of the five pre-built agents Archetype announced recently, which generates documentation and training material by watching demonstrations on video.

This is also why change management matters more in Physical AI than in back-office AI. Putting a camera in front of a work cell only succeeds when the people it affects understand the intent. Framed as assistive rather than punitive, the deployment becomes collaborative. Framed poorly, the cameras quietly stop working.

What to know before you start

Two questions came up repeatedly, and the answers are worth carrying out of the webinar.

On generalized versus specialized models, both speakers agreed on sequence: start with the generalized model, benchmark what it delivers out of the box, and layer in specialized capability or light fine-tuning only where a specific domain demands it. A generalized model that covers more assets with less upkeep usually buys more than a specialized model that needs constant care and feeding.

On getting started, a proof-of-value engagement can move quickly because the analysis Archetype runs internally is the same work you run on your own data. The platform deploys inside your infrastructure, you point it at raw signal data or video for a chosen use case, tune the prompt and model parameters, and evaluate against your own datasets. A model that once took months to build can be stood up against your data in hours.

About this webinar

This post is based on "How Physical AI Agents Are Transforming Manufacturing," featuring Sisinio Baldis of Archetype AI and Raleigh Murch of NTT Data, moderated by Aristo Chang.

To explore the Archetype platform and Physical AI Agents, visit Archetype AI or connect on LinkedIn and X (@PhysicalAI).

Frequently asked questions

What data was Newton trained on?

Newton was pre-trained on billions of sensor measurements drawn largely from open datasets, including atmospheric and scientific data published by agencies like NOAA and NASA, spanning modalities such as temperature and vibration. That breadth is what gives it a foundation to read sensor types it has not seen before. Archetype's founding team has published a paper detailing the initial datasets, and the model has been scaled substantially since.

How does Newton handle a machine or part it has never seen before?

Its foundation model can read unfamiliar signals and separate distinct operating regimes without labeled history. Where a domain includes parts or processes the model has not encountered, a small amount of fine-tuning closes the gap. In one construction-equipment engagement, Archetype fine-tuned Newton to recognize specialized machinery it had not seen before.

Can Newton tell the difference between a real fault and a planned human intervention?

Distinguishing a genuine equipment failure from a manual override, schedule change, or planned maintenance is an active area of work. Newton ingests multimodal sensor data and separates it into distinct clusters, and correlating that with human feedback and maintenance-system records is how the intent behind a change in telemetry gets resolved. Merging the model's signal-level view with operator domain knowledge is what enables deeper root-cause analysis.

What is the biggest blocker to adopting Physical AI on the floor?

Change management, more than the technology. Because deployment can mean placing a camera in front of someone's work cell, the people affected need to understand the intent up front. When the effort is communicated as assistive and quality-focused, adoption is collaborative; when it is not, deployments get resisted or quietly disabled.

Should manufacturers use a generalized model or a specialized one?

Start generalized. A single model applied across assets scales with far less maintenance, and it improves over time without a retraining cycle. Introduce specialized capability or fine-tuning only where a specific domain requires accuracy the generalized model does not yet reach. Every added specialization carries ongoing upkeep, so treat it as a deliberate trade-off.

What does a proof-of-value engagement actually involve?

The platform is deployed inside your own infrastructure rather than run as a hosted service. From there you process raw signal data or video for a specific use case, tune the prompt and model parameters, and evaluate results against your own datasets. The core analysis often takes about half a day, which is how a use case that once needed months of model building can be tested in hours.

‍