How Physical AI Is Redefining the Future of Intelligent Manufacturing

 

As manufacturing moves beyond traditional automation, Physical AI is emerging as a powerful force connecting artificial intelligence with operational technology, connected assets, and robotics. In this exclusive Robotics Business News conversation, Anupam Bhatnagar, Head of Americas Manufacturing & Consumer Industries at Hitachi Digital Services, discusses how Physical AI is helping manufacturers automate complex tasks, optimize operations in real time, improve sustainability, and build more adaptive smart factories. He also shares insights into OT-IT integration, workforce readiness, digital twins, and the path toward increasingly autonomous manufacturing environments.

Physical AI is increasingly being discussed as the next stage of artificial intelligence. How do you define Physical AI in the context of manufacturing, and what differentiates it from traditional AI and industrial automation?

At Hitachi, we define Physical AI as artificial intelligence that doesn't just analyze and act on digital data and digital workflows but AI that operates at the intersection of information technology, operational technology, and physical assets.  Traditional AI, including most generative and predictive models manufacturers have deployed over the past decade, works with digital data: it forecasts demand, flags anomalies in sensor readings, or summarizes a maintenance log.  Physical AI closes the loop between that intelligence and the physical world.  It perceives its environment through vision, force, and tactile sensing; reasons about what it's seeing; and then acts - precisely and adaptively - through robots, actuators, or automated equipment on the plant floor.

That's also what separates Physical AI from traditional industrial automation. Classical automation is deterministic: a robot arm repeats the same programmed motion regardless of what's actually in front of it, which is exactly why it struggles with tasks like wire-harness assembly or handling flexible, variable materials.  Physical AI systems, by contrast, continuously learn from real-world behavioral data and task-specific know-how generated on-site.  They refine their own motion behavior over time, adjusting the force, angle, and technique they apply based on what the sensors tell them in the moment.  In practice, that means we can now automate categories of complex, delicate manual work that were simply out of reach for rigid automation - while giving manufacturers a system that improves the longer it runs, rather than one that has to be reprogrammed every time conditions change.

How is the convergence of AI, industrial data, connected assets and automation changing the way manufacturers manage and optimize their operations in real time?

For most of the last twenty years, manufacturing has run on separate technology stacks: operational technology (OT) that runs machines and lines, and information technology (IT) that runs enterprise systems, largely disconnected from each other.  The real shift we're seeing now is OT-IT convergence at scale - connected assets, industrial data, automation, and AI operating as a single, continuous digital thread from design through manufacturing and into quality management.

Practically, this means a plant is no longer optimized in retrospect, through a weekly report or a monthly review. Field data from connected equipment flows continuously into AI models that are analyzing quality, cost, and delivery performance in real time, and that intelligence feeds back into automated workflows immediately - adjusting a process parameter, rerouting a production schedule, or flagging a deviation before it becomes a defect.  We are leveraging this first inside our own Hitachi Group factories - over 100 mission-critical manufacturing sites - before taking it to clients.  That "Customer Zero" approach matters: it means the real-time optimization capability we're describing isn't theoretical, it's running our own production lines today.

What are some of the most significant opportunities for Physical AI to improve productivity, operational resilience and decision-making across manufacturing environments?

The productivity opportunity is most visible in the tasks that have resisted automation until now - delicate assembly, materials handling, inspection work that requires judgment rather than just repetition.  Physical AI systems that can sense force and adjust their own behavior are opening entire categories of manual work to automation for the first time, which is a genuine step-change in throughput and consistency, not just an incremental gain.

The resilience opportunity is arguably just as significant, and it's underappreciated. A production environment built on agile, AI-informed workflows can respond instantly to fluctuations in demand, supply disruptions, or a line going down, rather than requiring a manual re-plan.  And because these systems generate a continuous stream of real-world operational data, they improve the quality of every decision layered on top of them - from shift-level scheduling up to plant-wide capital planning.  We see the biggest wins where manufacturers connect all three: using Physical AI at the task level, real-time data at the operations level, and AI-driven decision support at the management level, rather than treating them as separate initiatives.

Manufacturers are under pressure to improve performance while reducing energy consumption, waste and emissions. How can AI-driven automation help companies achieve both business and sustainability goals simultaneously?

We don't see performance and sustainability as competing priorities - in a well-instrumented plant, they're usually the same problem.  Waste, excess energy use, and scrap are almost always symptoms of variability: a process running slightly outside optimal parameters, a machine drifting out of calibration, a batch that has to be reworked. AI-driven automation attacks that variability directly.  When a system can sense conditions in real time and continuously optimize its own behavior, it's naturally reducing the energy, material, and rework waste that variability produces - not as a separate sustainability initiative, but as a byproduct of running better.

Hitachi has built and proven this on its own factory floor before taking it to clients.  The Hagerstown, Maryland rail manufacturing facility - a $100 million "lighthouse" digital factory - runs on 100% renewable electricity, including on-site solar generation, and operates with zero landfill waste from day one, using H-Vision, our AI analytics system, to continuously monitor and optimize electricity and water consumption alongside production performance.  On the reporting and decision-making side, we've paired that operational data with RitaONE, Hitachi Digital Services' SaaS platform for ESG. RitaONE uses agentic AI and a built-in carbon accounting engine to automate the collection and validation of Scope 1, 2, and 3 emissions data directly from the systems already running the business - energy management systems, health and safety platforms, supply chain tools - rather than requiring a separate manual reporting exercise. We implemented it first at Hitachi Rail, unifying ESG data collection and reporting across multiple international compliance frameworks, before bringing it to external clients.

That combination - real-time operational AI reducing waste and energy use at the source, and platforms like RitaONE turning the resulting data into audit-ready sustainability reporting automatically - is what lets manufacturers stop treating performance and sustainability as two separate programs.  The efficiency gain and the emissions reduction come from the same underlying data and the same operational discipline.

What does the next generation of smart factories look like, and how close are we to manufacturing environments where machines and AI systems can make more autonomous decisions?

The next generation of smart factories looks less like a collection of automated stations and more like a coordinated system - one where connected assets, real-time data, and AI-driven decision-making form a continuous loop, and where people, AI, and robots are working alongside each other rather than AI simply replacing fixed automation.  That's the vision behind what we call Industry 5.0: not just efficiency for its own sake, but resilient, adaptive production environments that can absorb disruption and still perform.

On autonomy specifically, we're closer than many people assume for well-defined tasks, and further away than the hype suggests for fully unsupervised plants.  Today, we're seeing Physical AI systems reliably make autonomous decisions within a defined scope - adjusting motion behavior on a specific assembly task, rerouting a workflow in response to a schedule change, catching a quality deviation in real time.  What's still emerging is autonomy at the scale of an entire plant making interconnected decisions across safety, quality, and throughput simultaneously.  We expect that to arrive incrementally, site by site and process by process, rather than as a single leap - which is exactly why we've prioritized proving these systems in our own mission-critical manufacturing operations first.

What are the biggest challenges manufacturers face when moving from AI pilots and isolated automation projects to large-scale deployment of Physical AI across their operations?

There are many challenges - data readiness, lack of sponsorship, frontline resistance but the most common failure point isn't the technology itself - it's integration.  A Physical AI pilot can look impressive on a single line and still fail to scale because it wasn't designed to work with the plant's existing OT systems, product lifecycle management tools, and IT infrastructure.  Interoperate with the diverse systems a manufacturer already has, rather than asking them to rip and replace.

Beyond integration, we see three recurring obstacles: data readiness, since Physical AI depends on high-quality, continuous field data that many plants aren't yet capturing consistently; organizational readiness, because scaling from a pilot to plant-wide deployment requires operations and IT teams to work from a shared plan rather than separate roadmaps; and workforce readiness, which is a change-management challenge as much as a technical one.  The manufacturers who scale successfully tend to treat Physical AI as an operating model change from day one - piloting with the end-state architecture in mind - rather than proving a narrow use case and then discovering it doesn't generalize. That is critical. 

How important will human workers remain as Physical AI and autonomous systems become more capable, and what new skills will the manufacturing workforce need to develop?

Human workers remain essential, and any framing that treats Physical AI adoption as a displacement story is an overreach.  The environments we're building toward are explicitly ones where people, AI, and robots evolve together - Physical AI is best suited to precise, repetitive, or physically demanding tasks, while judgment, exception-handling, and the kind of contextual problem-solving that comes from experience on the floor stay firmly in human hands. In practice, we're seeing productivity gains come from human-machine collaboration, not human replacement.

That said, the skills the workforce needs are shifting. Operators increasingly need to understand how to work alongside AI-driven systems - interpreting the data these systems generate, knowing when to trust an automated recommendation and when to override it, and troubleshooting at the intersection of OT and IT rather than one or the other.  We're also seeing rising demand for skills that didn't traditionally sit on a factory floor: data literacy, basic AI fluency, and comfort with digital twin and simulation tools. Manufacturers who invest in this upskilling alongside the technology deployment are the ones getting the full value out of Physical AI; the technology and the workforce capability have to scale together.

Looking ahead over the next five to ten years, which manufacturing applications do you believe will see the greatest impact from Physical AI, and what should industry leaders be doing today to prepare?

We expect the greatest near-term impact in the categories of work that have been hardest to automate conventionally: precision assembly involving flexible or delicate materials, equipment maintenance that currently depends on tacit expertise, and quality inspection that requires contextual judgment rather than simple pass/fail thresholds.  Digital twins will also play an increasingly central role - not just for visualization, but as the environment where manufacturers simulate and validate Physical AI behavior before it ever touches a physical line, which materially de-risks deployment.

For industry leaders, the most important thing to do today isn't to wait for the technology to mature - it's to get the foundation in place so they can adopt it quickly once it does. That means investing now in OT-IT integration and clean, continuous data capture, since Physical AI is only as good as the field data it learns from.  It means starting to pilot with an eye toward how a use case will integrate into the broader operation, not just whether it works in isolation.  And it means beginning the workforce transition now, so operators and engineers are building AI fluency in parallel with the technology rollout rather than playing catch-up later.  The manufacturers who treat this as infrastructure work today will be the ones positioned to move fastest when Physical AI moves from proven pilots to standard practice.