Robots do not just need better AI. They need people to show them the world

For the past two years, one sentence has kept returning in the debate about artificial intelligence: AI will take work from copywriters, designers, programmers, and analysts, but physical jobs are safe.

A robot will not empty your dishwasher. A robot will not clean your apartment. A robot will not fold a shirt as efficiently as a person.

It sounded reasonable. For a while, it was even true. But it was only half the truth.

The problem with humanoid robots was never simply a lack of computing power or a good AI model. It was more ordinary: robots did not know our world. They did not know the weight of a wet towel, the resistance of a drawer, the unpredictability of a dog running underfoot, or the difference between a plate that can be gripped firmly and a glass that cannot.

That is why another, less visible competition is growing alongside the race for language models: the race for data from the physical world.

Physical work was supposed to be safe. That was only a stage

When generative AI entered the mainstream, professions based on language, images, code, and analysis felt the first wave of anxiety. Suddenly, a machine could draft a text, prepare a graphic variant, generate code, or summarize a report.

Physical work appeared much harder to automate, and for good reason. The digital world is ordered. Text, images, sound, and code can be stored as data, copied, processed, and used to train models. The internet became an enormous training set for AI.

The physical world is different. It is irregular, messy, changeable, and full of exceptions. The same cup can stand on different surfaces. The same shirt can be folded in different ways. The same dishwasher can have a different layout, door resistance, and lighting.

For a person, these are trivial details. For a robot, they are an enormous problem. This is why a new market has emerged: people recording everyday activities so that robots can learn from them.

A new class of work: people as an interface for robots

MIT Technology Review described gig workers in Nigeria, India, and elsewhere who attach iPhones to their heads and record themselves doing household work. One of them is Zeus, a medical student in Nigeria. After a hospital shift, he returns home, mounts a phone on his forehead, and records simple tasks such as making the bed, ironing, and organizing his space.

He is not doing it for social media. He works as a data recorder for Micro1, a company that collects data about human movement and sells it to robotics companies.

It is a symbolic image of our time. For years we called data “the new oil,” usually meaning clicks, likes, purchase histories, location, and behavior inside apps and search engines. Now something even more fundamental is entering that category: the movement of the human body through everyday space.

How to pick up a shirt. How to open a cupboard. How to move a plate. How not to drop a cup. How to straighten a bedsheet. All of it becomes training material.

Moravec’s paradox returns in its most practical form

Moravec’s paradox helps explain this process. In simplified terms, tasks people consider intellectually difficult can be relatively easy for machines, while things humans find effortless can be extremely hard for them.

Solving an equation, analyzing text, generating an image, or writing code are tasks AI can perform increasingly well because vast digital datasets exist for training.

Walking across a room, avoiding a dog, lifting a plate, distinguishing a cloth from a shirt, and not knocking a glass from a counter are different challenges entirely.

Robotics therefore needs more than “more AI.” It needs data about a world that cannot be downloaded in full from the internet.

Here lies the irony. Many people believed physical work would be the last stronghold against automation. Today, part of that work is becoming training material for machines that may eventually perform it.

Before a robot can replace a cleaner, warehouse worker, or production employee, someone must show it what cleaning, carrying, folding, gripping, correcting, and responding to disorder actually look like.

Europe is also building Physical AI laboratories

Europe is trying to develop its own response to the Physical AI race. NEURA Robotics and the Technical University of Munich announced TUM RoboGym, a training center for robots and Physical AI systems in Munich. The project is expected to cover 2,300 square meters, involve €17 million in investment, and host a large fleet of humanoid robots working under controlled conditions.

This is an important signal. Humanoid robotics will not be only a technological question. It will also be a question of data, infrastructure, standards, privacy, labor, and technological sovereignty.

Who will have the best data from the physical world? Who can collect it legally, ethically, and at scale? Who will build the models of movement, grip, and spatial interaction? Who will own the “brains” of robots?

These are not abstract laboratory questions. They concern the future balance of power in industry, logistics, care, household services, and manufacturing.

The United States has companies, capital, and global data platforms. China has scale, industry, and strong state–industry coordination. Europe has regulation, excellent universities, and a strong industrial base—but also the risk of debating longer than others spend building.

Europe is not defenseless. The question is whether it can combine its strengths before the Physical AI market is dominated by a few non-European ecosystems.

The counterargument: this may be good work and a necessary stage

It is easy to tell this story only as digital colonialism: people in lower-income countries record their homes, movements, and daily lives to feed models built by wealthier technology companies.

That description is partly accurate, but incomplete. For some people, this work may be a genuine opportunity. A rate that looks ordinary in the United States may be attractive in Nigeria or India. Not every worker must feel exploited. Some may see participation in technological change, additional income, or entry into the global digital economy.

A paternalistic story in which every worker in the Global South is only a victim would also be an oversimplification.

The real questions begin elsewhere. Do workers know who receives their data? Do they understand how long it will be stored? Do they know whether recordings will train household, warehouse, industrial, or military robots? Can they influence how their work is later used? Do they receive only an hourly rate, or share in the value they help create?

This is not a simple conflict between progress and exploitation. It is a question of whether the new physical-data economy will repeat the mistakes of the old platform economy.

First we collect the data. Then we build an advantage. Only at the end do we ask whether everyone understood what they were participating in.

The most important question is not whether robots will replace us

The AI debate often asks which profession will disappear first. It is a convenient question, but no longer a sufficient one.

A better question is: which parts of our work, movement, knowledge, and daily lives will become training data for systems that later change the market?

Copywriters, designers, and programmers saw this first because their work was already largely digital. Text, images, and code were easy to copy, process, and use in models.

Physical work was more difficult, but it was not magically resistant. It simply required a different kind of data. That stage is beginning now.

This does not mean every home will have a humanoid robot in five years. Robots are more likely to appear first in partly controlled environments: warehouses, factories, laboratories, logistics centers, institutional care, and selected services.

But the direction is clear. AI first learned our language. Then our images. Now it is beginning to learn our movements.

The sentence “no job is safe” is too simple, but not entirely false. No kind of work understood as a set of repeatable activities is completely safe. What differs is the speed of automation, the cost of data, the level of risk, and the difficulty of entering the physical world.

The most important question is therefore not whether robots will take people’s jobs.

It is this: who will own the map of the human world on which those robots learn?