Physical AI is taking artificial intelligence out of the screen—and putting it into robots, factories, vehicles and the physical world.

For the last few years, artificial intelligence has lived mostly on screens.

It writes our emails.

Answers our questions.

Generates images.

Writes code.

Analyses documents.

But something fundamental is changing.

The next generation of AI may not simply talk to us.

It may move things, drive vehicles, operate machines, inspect factories, assist workers and interact with the physical world.

This emerging field is increasingly being called Physical AI.

And it could become one of the biggest shifts in technology after generative AI.

---

AI Has Learned to Think. Now It Has to Act.

Traditional AI works primarily with digital information.

Give it text, images, audio or numbers and it can analyse them.

But the physical world is different.

A robot cannot simply “understand” a chair. It has to locate the chair, estimate its shape, understand where it can safely place its hand, calculate movement, deal with friction and gravity—and then actually pick it up.

That requires an entirely different level of intelligence.

NVIDIA defines Physical AI as systems that can perceive, understand, reason and perform or orchestrate actions in the physical world. The concept applies not only to robots but also to autonomous vehicles, cameras and other intelligent machines.

IBM similarly describes Physical AI as the combination of AI models with sensors, actuators and control systems that allow machines to perceive their environment, make decisions and act on the real world.

In simple words:

Generative AI creates.
Physical AI acts.

And that difference could be enormous.

---

Why Now?

Robotics isn’t new.

Factories have used robots for decades.

The difference is that many traditional industrial robots are designed around highly controlled environments and predefined instructions.

They are extremely good at doing specific jobs repeatedly.

But ask them to deal with something unexpected—and the problem becomes much harder.

Modern AI changes that equation.

New vision-language-action models are being developed to allow robots to interpret instructions, understand their surroundings and translate perception into physical actions.

NVIDIA’s 2026 robotics work, for example, includes its GR00T models designed to help robots understand natural-language instructions and perform multi-step tasks using vision, while its Cosmos platform is aimed at creating physical-world simulation and synthetic training data.

The goal is not simply to make a better robot.

It is to make robots more adaptable.

---

From Chatbots to Machines

Consider the difference.

You tell a chatbot:

“Help me organise this information.”

It processes digital information and produces an answer.

Now imagine telling a robot:

“Clean this table and put everything where it belongs.”

The robot has to:

See → Understand → Plan → Move → Manipulate → Check → Correct

That is Physical AI.

The machine has to deal with a world that does not always behave predictably.

A glass can slip.

A box can be heavier than expected.

A person can suddenly walk into its path.

A door can be partially open.

Lighting can change.

Objects can be in places the robot has never seen before.

The real challenge is therefore not making a robot move.

It is making a robot understand what is happening while it moves.

---

The Humanoid Robot Moment

This is where humanoid robots enter the story.

The argument for a human-shaped robot is straightforward: much of the world has already been designed for humans.

Doors have handles.

Stairs have steps.

Factories have workstations.

Warehouses have shelves.

Tools are designed around human hands.

If a machine can operate in the same environment without requiring the entire environment to be redesigned, its usefulness could increase dramatically.

But there is an important reality check.

A robot performing a spectacular dance or running a 100-metre race is not the same thing as a robot reliably working an eight-hour shift.

Recent events in China have demonstrated both sides of the story.

At the 2026 World Humanoid Robot Games in Beijing, more than 2,000 robots competed across dozens of events, with the Tiangong Ultra reportedly running 100 metres in 8.64 seconds—faster than Usain Bolt’s 9.58-second human world record.

That sounds extraordinary.

It is.

But speed is only one part of intelligence.

Recent reporting has also highlighted the industry’s much harder problem: humanoids still struggle with general-purpose autonomy, dexterity and reliable performance in real industrial environments.

The robot can run.
The question is whether it can work.

That is the real test.

---

The Factory Could Be the First Battlefield

The first large-scale impact of Physical AI may not happen in our homes.

It may happen in factories and warehouses.

Why?

Because factories offer something the open world does not:

structure.

A factory can be mapped.

Tasks can be defined.

Safety zones can be established.

Objects can be standardised.

And the economic value of automation can be measured.

Industrial companies are already exploring AI-powered robots for manufacturing, logistics and warehouse operations.

NVIDIA says its Physical AI ecosystem is being developed with robotics companies and manufacturers working on industrial robots, humanoids and autonomous systems. Its 2026 announcements include applications involving factories, warehouses and logistics.

The likely progression may therefore look something like:

Controlled factory → Warehouse → Commercial environment → Home

Not the other way around.

---

India Wants a Seat at the Table

This story matters to India too.

India has traditionally been strong in software and IT services.

Physical AI creates a different opportunity.

It combines:

AI + robotics + manufacturing + electronics + sensors + semiconductors + engineering

And Indian companies are beginning to position themselves around this emerging market.

Recent reporting on India’s robotics ecosystem highlights companies such as Addverb, Ati Robotics and Novus Hi-Tech working on AI-enabled robotic systems for industrial and other applications. Ati Robotics is also reported to be planning manufacturing expansion in the United States.

India also has another potential advantage:

scale.

A country with enormous manufacturing, logistics, healthcare, agriculture and infrastructure requirements does not need to invent a robot merely because it looks futuristic.

It needs robots that solve expensive, difficult problems.

That could create a much more interesting Indian Physical AI story.

---

The Hidden Problem: Robots Need Experience

There is a reason building useful robots is harder than building chatbots.

The internet contains an enormous amount of text.

AI models can train on books, websites, conversations, documents and code.

But robots need something different:

experience in the physical world.

They need to learn what happens when a hand touches an object.

How much force is required to lift something.

What happens when an object slips.

How different surfaces behave.

How humans move around machines.

How the same instruction changes depending on the environment.

This is why synthetic data and simulation have become such important parts of Physical AI.

Instead of teaching a robot every physical task only in the real world, researchers can create virtual environments where machines can practise thousands—or millions—of scenarios.

NVIDIA’s Physical AI ecosystem includes simulation, digital twins, synthetic data generation and robot-policy training specifically to bridge the gap between virtual learning and real-world deployment.

In other words:

The robot may learn in a virtual world before it works in ours.

---

And Then There Is the Data Problem

There is another emerging industry hiding behind Physical AI:

robot training.

A robot needs physical demonstrations and real-world interaction data to become better at tasks.

That is creating businesses focused not simply on manufacturing robots, but on collecting and generating the data required to train them.

One recent example is Reimagine Robotics, founded by former DeepMind engineers, which is exploring ways for customers to teach robots tasks through demonstration and physical correction.

This suggests something important.

The Physical AI economy may not consist only of:

Robot manufacturers.

It could eventually include:

• Robot foundation-model companies
• Sensor manufacturers
• Simulation companies
• Robot-training platforms
• Physical-data companies
• Safety systems
• AI chips
• Industrial integrators
• Robot-as-a-service companies

The robot itself may become only one part of a much larger ecosystem.

---

Safety Is Not an Optional Feature

There is a major difference between an AI generating a wrong sentence and an AI-powered machine making a wrong physical movement.

A bad answer can be corrected.

A machine moving incorrectly around a human can cause injury.

That is why safety is becoming a major part of Physical AI development.

NVIDIA introduced Halos for Robotics in 2026 as a safety system aimed at industrial robots, humanoids and autonomous mobile robots, with Agility using the technology for humanoids deployed in industrial environments.

The bigger lesson is clear:

The smarter robots become, the more important their safety architecture becomes.

A future where machines can independently make physical decisions will require more than better AI models.

It will require:

AI + engineering + safety + regulation + accountability.

---

What About Jobs?

This is probably the question most people will ask.

Will Physical AI replace human workers?

The honest answer is:

Some jobs and tasks will almost certainly change.

But “robots will take all the jobs” is far too simplistic.

Automation usually starts with specific tasks.

A robot might handle repetitive lifting.

Another might perform inspection.

Another might transport material around a warehouse.

Another might assist with assembly.

Human workers may then move toward supervision, maintenance, quality control, exception handling, design and other responsibilities.

The more interesting question is therefore not:

“Will robots replace humans?”

It is:

Which human tasks will no longer need to be done by humans?

That is a much bigger economic question.

---

The Consumer Robot Is Still a Different Challenge

A factory is predictable.

Your home is not.

A household robot would have to deal with thousands of objects, different rooms, children, pets, clutter, unpredictable instructions and constantly changing environments.

It would need to understand not just:

“Pick up the cup.”

but potentially:

“Please clean up the kitchen before my parents arrive.”

That single sentence could involve dozens of decisions.

Which objects?

Where should they go?

What is fragile?

What is rubbish?

What belongs in another room?

What should not be touched?

This is why the home may be one of the hardest environments for Physical AI.

The dream is simple.

The engineering is not.

---

The Bigger Shift

Physical AI may ultimately change our understanding of what an AI system is.

For decades, computers lived behind screens.

Then smartphones put computing in our pockets.

Generative AI put intelligence into everyday software.

Physical AI could put intelligence into the environment itself.

Cameras could interpret what they see.

Vehicles could make decisions.

Machines could coordinate with other machines.

Robots could learn new tasks.

Factories could respond dynamically to changing conditions.

And eventually, intelligent machines could become part of everyday infrastructure.

That is a much bigger idea than humanoid robots.

---

The Reality Check

There is enormous excitement around Physical AI—and some of it is justified.

Investment is increasing. Business Insider, citing PitchBook data, reported that global robotics and Physical AI investment had risen from roughly $4 billion in 2019 to $26 billion in 2025, with more than $23 billion raised by companies in the sector in 2026 at the time of its reporting.

But money does not automatically create technological breakthroughs.

The hardest problems remain:

Reliability.
Training data.
Dexterity.
Battery life.
Cost.
Safety.
Manufacturing.
Real-world adaptability.

And there is another uncomfortable question:

Can a robot be intelligent enough to work in the real world without becoming too expensive to deploy?

That may ultimately determine whether Physical AI becomes a trillion-dollar industry—or remains an impressive collection of demonstrations.

---

What Happens Next?

The next phase of AI may therefore be less about making models that know more.

It may be about making models that do more.

The winning systems could be those that combine:

Perception + reasoning + planning + action + learning

all in one loop.

And the companies that solve that loop could become some of the most important technology companies of the next decade.

The race has already started.

The interesting part is that the finish line is not a better chatbot.

It is a machine that can enter the physical world—and actually understand what to do there.

---

THE UNPLANNED VERDICT

WATCH THIS SPACE

Physical AI is still an emerging technology, and the industry has a long way to go before general-purpose humanoid robots become reliable everyday workers.

But something important has changed.

AI is no longer being designed only to understand the world.

It is increasingly being designed to act in it.

That makes Physical AI more than another robotics trend.

It could become the bridge between the digital economy and the physical economy.

And if that bridge works—

the next industrial revolution may not happen on our screens.

It may walk through the door.

---

THE NEXT BIG THING

Artificial intelligence learned to speak.
Now it is learning to move.

THE UNPLANNED explores the ideas, people and technologies that could shape what comes next.