The physical AI race is usually framed around foundation models and accelerated computing. Hitachi is making a more industrial argument: machines cannot act reliably in the real world without deep knowledge of physics, equipment and control systems.
That position draws on the company’s experience in railways, power infrastructure and manufacturing. Hitachi has thermal-fluid simulation for modelling gases and liquids, signal processing for equipment monitoring and decades of product-control knowledge. It is using those assets to build AI around the realities of operating machinery rather than treating deployment as a final step after model development.
The strategy is already visible in factory and railway applications, even as its broader Integrated World Infrastructure Model remains at the concept-verification stage.
Operational knowledge becomes a model advantage
With Daikin Industries, Hitachi has deployed an AI system that diagnoses faults in commercial air-conditioner manufacturing equipment. It learns from maintenance records, procedure manuals and design drawings to identify the component most likely to be failing when an anomaly appears.
The capability captures a form of intuition often held by a small number of experienced engineers. Making that knowledge accessible can shorten downtime and help newer employees respond more consistently, although human expertise remains vital for unusual or hazardous situations.
At East Japan Railway, an AI identifies root causes of faults in the control devices behind Tokyo’s railway traffic-management system and assists operators with response planning. In a network serving millions of journeys, faster diagnosis can prevent a local technical problem from producing wider disruption.
Both examples show the value of combining operational records with engineering context. A general model may interpret a maintenance note, but it needs equipment relationships and approved procedures to recommend a useful action.
Software tools cut industrial development effort
Hitachi is also applying AI to the work required to build and adapt control systems. With automotive subsidiary Astemo, it created a retrieval-augmented system that generates integration-test scripts for electronic control units using hardware-specific APIs and engineering knowledge. A multi-core ECU pilot reduced testing labour by 43% compared with manual execution.
In logistics, Hitachi developed variability-management technology that turns robot-control software into reusable components structured around the Robot Operating System. By describing environmental differences and operating requirements in advance, teams can adapt picking workflows to new products or warehouse layouts without rewriting everything.
This kind of reuse may be more immediately valuable than a general-purpose humanoid. Industrial automation often stalls because every site and machine requires bespoke integration. Modular software and generated tests attack that cost directly.
Hitachi Vantara is pairing Nvidia RTX PRO Servers with its iQ platform to develop digital twins for systems including grids and robots. The proposed IWIM architecture would connect Nvidia’s Cosmos platform, specialised Japanese language models, visual-language models and industrial data through the Model Context Protocol.
Safety remains an engineering constraint
Physical AI errors affect machines, property and people. Hitachi consequently treats guardrails as part of system design: validating training inputs, verifying outputs against safe operating limits and monitoring models for anomalies during operation.
That approach differs from adding a policy after deployment. A railway or factory system needs deterministic boundaries around where learned behaviour is allowed. It must also fail safely when sensors, models or communications become unreliable.
Domain expertise does not make these systems automatically safe. It does, however, give developers a richer understanding of failure modes and the control logic needed to contain them. The combination of generative intelligence and conventional engineering may be more useful than asking either to carry the full burden.
The physical AI market is far from decided. Model providers offer general reasoning, chip companies supply computing platforms and manufacturers contribute operational environments. Hitachi’s wager is that the last category owns knowledge that cannot be recreated quickly from internet-scale data.
Its deployments with Daikin and JR East lend weight to that claim. They solve bounded, consequential problems with information grounded in actual equipment. If Hitachi can turn that expertise into reusable architecture rather than one-off projects, it could occupy a distinctive layer of the market.
Physical AI will ultimately be judged on uptime, safety and economic results. In those measures, understanding how a machine should behave may matter as much as building a model intelligent enough to control it.
