Hitachi Bets on Physical AI, Arguing That Real-World Expertise Trumps Raw Computing Power

In the escalating race to dominate physical AI—the branch of artificial intelligence that controls robots and industrial machinery in the real world—a quiet but powerful argument is emerging from an unlikely corner. While OpenAI and Google scale multimodal foundation models and Nvidia builds the platforms to run them, industrial manufacturers like Hitachi and Siemens are making the case that you can’t train machines to navigate the physical world without first understanding it.

Hitachi, the Japanese conglomerate with deep roots in railways, power infrastructure, and industrial control systems, is moving that argument from boardroom strategy to factory floor deployment. In a recent interview with Nikkei Asia, the company revealed how its decades of engineering expertise are becoming the foundation for a new wave of physical AI systems.

Why Physical AI Needs More Than Just Models

Kosuke Yanai, deputy director of Hitachi’s Centre for Technology Innovation-Artificial Intelligence, cuts straight to the heart of the matter: “Physical AI cannot be implemented in society without a systematic understanding that begins with foundational knowledge of physics and industrial equipment.”

This isn’t just philosophical positioning. Hitachi argues it already possesses much of this foundational knowledge—accumulated through decades of building everything from thermal fluid simulation technology that models gas and liquid behavior to signal-processing tools for monitoring equipment condition. As Yanai puts it, this is the engineering foundation underpinning Hitachi’s “extensive knowledge of product design and control logic construction.”

Real-World Deployments That Matter

While Hitachi’s overarching physical AI architecture—the Integrated World Infrastructure Model (IWIM), described as a mixture-of-experts system integrating multiple specialized models and datasets—remains in the concept verification stage, two real-world deployments signal that the underlying approach is already producing results.

In collaboration with Daikin Industries, Hitachi has deployed an AI system that diagnoses malfunctions in commercial air-conditioner manufacturing equipment. The system, trained on equipment maintenance records, procedure manuals, and design drawings, can now identify which component is likely failing when an anomaly is detected—the kind of operational intuition that previously existed only in the heads of experienced engineers.

With East Japan Railway (JR East), Hitachi has built an AI that identifies the root cause of malfunctions in the control devices running the Tokyo metropolitan area’s railway traffic management system, and then assists operators in formulating a response plan. In a network where delays ripple through millions of daily journeys, the ability to accelerate fault diagnosis carries real operational weight.

Cutting Development Time by Nearly Half

Hitachi’s physical AI push is also showing up in its research output. In December 2025, the company published findings from two projects presented at ASE 2025, a top-tier software engineering conference, that address a persistent bottleneck in industrial AI: the time and effort required to write and adapt control software.

In the automotive sector, Hitachi and its subsidiary Astemo developed a system that uses retrieval-augmented generation to automatically produce integration test scripts for vehicle electronic control units (ECUs)—pulling from hardware-specific API information and frontline engineering knowledge. In a pilot involving multi-core ECU testing, the technology reduced integration testing man-hours by 43% compared to manual execution.

In logistics, the company developed variability management technology that modularizes robot control software into reusable components structured around a robot operating system (ROS). By mapping out the environmental variables and operational requirements of different warehouse settings in advance, the system lets operators adapt robotic picking-and-placing workflows to new products or layouts without rewriting software from scratch.

Safety as a Structural Requirement, Not a Checkbox

One thread that runs through all of Hitachi’s physical AI work is its emphasis on safety guardrails—not as a compliance checkbox, but as an engineering constraint baked into system design. Yanai told Nikkei that the company is integrating its control and reliability technology from social infrastructure development to prevent AI outputs from deviating from human-approved operating parameters.

This includes input validation to screen out data that models should not be trained on, output verification to ensure machine actions do not endanger people or property, and real-time monitoring of the AI model itself for operational anomalies.

It’s a crucial distinction. Physical AI systems fail in the real world, not in a sandbox. The stakes for an AI controlling railway signaling or factory robotics are categorically different from those governing a chatbot.

Building the Infrastructure to Match Ambition

On the infrastructure side, Hitachi Vantara—the group’s data and digital infrastructure arm—is positioning itself as an early adopter of NVIDIA’s RTX PRO Servers, built on the RTX PRO 6000 Blackwell Server Edition GPU, designed to accelerate agentic and physical AI workloads. The hardware is being paired with Hitachi’s iQ platform and used to build digital twins—virtual replicas of physical systems—that can simulate everything from grid fluctuations to robotic motion at scale.

The IWIM concept, meanwhile, is designed to connect Nvidia’s open-source Cosmos physical AI development platform with specialized Japanese-language LLMs and visual language models via the model context protocol (MCP)—essentially a framework to stitch together the models, simulation tools, and industrial datasets that physical AI systems require.

The Broader Race and Hitachi’s Position

The broader race in physical AI is far from settled. But Hitachi’s position—that domain expertise and operational data are as important as model architecture—is increasingly hard to dismiss, particularly as deployments with partners like Daikin and JR East begin to demonstrate what that expertise is actually worth in practice.

As Yanai’s comments suggest, the company sees itself not as competing with OpenAI or Google on raw model scale, but as providing the essential bridge between those models and the physical world they’re meant to control. In an era where AI companies are racing to build ever-larger models, Hitachi is betting that understanding how the world actually works might be the real competitive advantage.

Sources: Nikkei Asia (Feb 21, 2026); Hitachi R&D (Dec 24, 2025); Hitachi Vantara Blog (Aug 27, 2025)

See also: Alibaba enters physical AI race with open-source robot model RynnBrain

tags

physical AI, industrial AI, Hitachi, Siemens, OpenAI, Google, Nvidia, robotics, manufacturing, railway systems, AI safety, digital twins, IWIM, model context protocol, ROS, retrieval-augmented generation, ECU testing, logistics automation, Daikin, JR East, infrastructure AI, domain expertise, operational data, engineering knowledge, thermal fluid simulation, signal processing, AI guardrails, real-world AI deployment

viral sentences

Hitachi argues you can’t train machines to navigate the physical world without first understanding it.

Physical AI systems fail in the real world, not in a sandbox.

The stakes for an AI controlling railway signaling are categorically different from those governing a chatbot.

Hitachi’s position—that domain expertise and operational data are as important as model architecture—is increasingly hard to dismiss.

In an era where AI companies are racing to build ever-larger models, Hitachi is betting that understanding how the world actually works might be the real competitive advantage.

,