NTT DATA and Hyster-Yale Materials Handling announced on July 7 that they have deployed a physical AI system directly into the assembly workflow at Hyster-Yale’s manufacturing facility in Berea, Kentucky. The co-developed approach represents a first-of-its-kind use case of how physical AI can be applied in an industrial assembly environment by embedding intelligence into production workflows, according to the companies’ joint announcement. The system integrates vision sensors, edge AI processing, and advanced analytics into a critical assembly stage of Hyster and Yale lift truck production, validating product quality in real time without sending data to external cloud servers.
Key Takeaways
- NTT DATA and Hyster-Yale Materials Handling deployed a physical AI system at HYMH’s Berea, Kentucky, manufacturing plant that validates assembly quality in real time using vision sensors and edge computing.
- The system was co-developed with Archetype AI, a Palo Alto-based startup whose Newton foundation model processes multimodal sensor data to detect deviations during production.
- Early results showed that physical AI compressed deployment timelines from months to weeks compared with traditional machine learning integration methods.
- All data processing runs locally on-site, eliminating the need for cloud connectivity and enabling faster decision-making on the factory floor.
- NTT DATA and Hyster-Yale plan to evaluate how the architecture can support repeatable quality assurance across additional manufacturing operations globally.
What Does Physical AI Actually Do on the Assembly Line?
The term “physical AI” describes artificial intelligence systems that process real-world sensor data — from cameras, accelerometers, vibration monitors, temperature gauges, and other instruments — to perceive, interpret, and respond to conditions in physical environments. Unlike generative AI models that operate on text and images in digital spaces, physical AI systems are designed to interact with tangible production processes where deviations have immediate material consequences.
At the Berea facility, NTT DATA, in collaboration with HYMH, adapted a physical AI model that analyzes assembly activity against expected production steps, validating that all parts are installed and assembly stages are completed, flagging deviations before the product moves to the next stage. In practical terms, the system watches what is happening on the line, compares it against what should be happening at each step, and alerts production teams when something does not match. That feedback loop occurs in real time, meaning a missing component or an out-of-sequence installation gets caught during the build rather than during a post-production inspection.
The technology partner behind the AI model is Archetype AI, a Palo Alto-based company that raised $35 million in Series A funding in November 2025 from investors including Bezos Expeditions, Hitachi Ventures, and Amazon Industrial Innovation Fund. Archetype AI introduced Newton, a first-of-its-kind foundation model that is capable of perceiving, understanding and reasoning about the world by fusing multimodal sensor data with natural language processing. Newton can run on a single off-the-shelf GPU on local machines, which is why the Berea deployment does not require cloud infrastructure — all inference and data analysis happen on-site at the edge.
Why Does Edge Processing Matter for Manufacturing?
The decision to run the entire AI system locally is not incidental to the announcement — it is central to its significance. Traditional approaches to embedding machine learning into factory workflows have typically required months of custom model development, extensive cloud connectivity, and dedicated teams of ML engineers. Traditional approaches require building bespoke machine learning models for every use case and sensor type, a process that can take up over 12 months per model and 5 or more ML engineers for each application, according to Archetype AI’s own documentation.
The Berea deployment compressed that timeline substantially. The companies said early results reduced deployment timelines from months to weeks compared with legacy techniques. That acceleration matters because manufacturing facilities cannot afford extended downtime to install and calibrate AI systems, and the faster a quality assurance tool goes live, the sooner it begins catching defects that would otherwise reach the end of the line.
Edge processing also addresses a practical security and latency concern. Manufacturing data — including proprietary assembly sequences, production rates, and defect patterns — is competitively sensitive. Running inference locally means that data never leaves the facility, reducing exposure to cloud-based security risks while eliminating the latency that comes with transmitting sensor feeds to remote data centers for processing.
What Is the Berea Facility and Why Was It Selected?
Hyster-Yale opened its Berea facility in 1973 with annual production of approximately 5,000 trucks. The plant has since undergone significant expansion, including a $25.7 million, 160,000-square-foot addition in 2019 that substantially increased capacity. Its largest manufacturing plant is in Berea, Kentucky, making it the company’s primary Americas production hub and a logical proving ground for new manufacturing technology.
The Berea plant spans over 500,000 square feet and houses applications engineering, special products engineering, product maintenance and design, service engineering, and supplier quality engineering teams alongside the production line. Hyster-Yale Materials Handling, a wholly owned subsidiary of NYSE-listed Hyster-Yale, Inc., designs and manufactures lift trucks, parts, and technology solutions under the Hyster, Yale, Nuvera, and Maximal brand names, with production facilities across five continents and roughly 7,900 employees worldwide.
The relationship between NTT DATA and Hyster-Yale predates this AI deployment. In November 2023, NTT DATA announced an agreement to develop robotic lift truck technology with Hyster-Yale Group, covering autonomous materials-handling equipment for warehousing and manufacturing facilities. The physical AI deployment represents a deepening of that technology partnership into the production process itself.
What Comes Next for Physical AI in Manufacturing?
NTT DATA and HYMH plan to evaluate how the architecture can support repeatable quality assurance across additional manufacturing operations. The modular nature of the edge AI platform means that once the framework is validated at Berea, it can be adapted to other assembly workflows and facilities without rebuilding the system from scratch.
The broader manufacturing sector has been investing heavily in industrial AI, but most deployments to date have focused on predictive maintenance — using sensor data to anticipate equipment failures before they occur. The Hyster-Yale deployment targets a different use case: real-time quality validation during active assembly, which sits further upstream in the production process and has the potential to reduce rework, warranty claims, and post-production inspection costs.
Despite massive investments in monitoring systems, 70-90% of industrial sensor data goes to waste, according to Archetype AI. The gap between the volume of data that factory sensors generate and the volume that actually informs decision-making is where physical AI platforms are positioning themselves, and the Berea deployment is the first publicly documented case of that technology running in a live heavy-equipment assembly environment.
The NTT DATA and Hyster-Yale deployment at Berea marks the first documented application of physical AI in a live industrial assembly environment, converting sensor data that was previously collected but unused into a real-time quality assurance tool embedded directly in the production line.
FAQs
What is physical AI? Physical AI refers to artificial intelligence systems that process real-world sensor data — including visual, vibration, temperature, and motion signals — to perceive, interpret, and act within physical environments in real time, as distinct from AI systems that operate only on digital text and images.
What did NTT DATA and Hyster-Yale deploy? The companies deployed an edge AI system at Hyster-Yale’s Berea, Kentucky, manufacturing plant that uses vision sensors and Archetype AI’s Newton foundation model to validate assembly quality in real time, flagging deviations before products advance to the next production stage.
What is Archetype AI’s role in the deployment? Archetype AI, a Palo Alto-based company backed by Bezos Expeditions and Amazon Industrial Innovation Fund, provided the Newton foundation model that processes multimodal sensor data locally at the edge. The model was adapted by NTT DATA for the specific assembly workflows at the Berea facility.
How fast was the deployment compared to traditional approaches? Early results showed that the physical AI system reduced deployment timelines from months to weeks compared with traditional machine learning integration methods, which can take over 12 months per model.
Where is the Hyster-Yale manufacturing facility? The Berea, Kentucky, facility is Hyster-Yale’s largest manufacturing plant globally. The facility opened in 1973 and now spans over 500,000 square feet, producing Hyster and Yale lift trucks.
Will this technology expand to other facilities? NTT DATA and Hyster-Yale have stated they plan to evaluate how the architecture can support repeatable quality assurance across additional manufacturing operations.
