Physical Intelligence: When AI Breaks the Fourth Wall

December 2025

We have spent the last three years marveling at, or worrying about, brains in a jar.

Since the arrival of ChatGPT, we have been interacting with brilliant yet disembodied minds, able to write sonnets or code websites, but powerless to make a coffee or fold a shirt. This dichotomy, known as Moravec's Paradox, pointed out that it is infinitely easier for a machine to ace a hard IQ test than to possess the dexterity of a two-year-old child.

As 2025 draws to a close, this paradox is shattering.

Forget chatbots. The next great frontier is no longer textual, it is kinetic. 2026 will not be the year of a new version of GPT, it will be the year of the rise of Physical Intelligence: the ability of an AI to understand, learn and act on the material world.

The "Universal Brain" of robots is born

Until recently, programming a robot meant writing thousands of lines of code for each specific movement. A Herculean and rigid task. That is the lock physical intelligence, or "embodied AI", aims to blow open. The idea: apply to robots the same recipes that made large language models a success. Rather than programming each movement, you train a model on thousands of hours of interaction data with the real world. The robot learns to generalize, to adapt, to improvise.

One of the most striking examples comes to us from San Francisco, with the aptly named startup Physical Intelligence__, __founded in 2024 by researchers from Google DeepMind and Berkeley. In November 2025, the company raised 600 million dollars to develop its π0 ("pi-zero") model.

Picture an LLM (like GPT), but trained not on texts, but on physical data: trajectories, forces, contacts. The result? A model able to control any robot to perform any task. π0 does not merely run a script; it understands the physics of a cardboard box it folds or the delicacy required to insert a filter into an espresso machine.

The Battle of the Giants: Figure, Tesla and Chinese agility

If startups supply the brain, the industrial giants build the bodies. And the competition is fierce.

On one side, we have Figure, which made a big splash by deploying its humanoid robots, the Figure 02, directly on BMW's production lines. This is no longer R&D, it is the real thing. What impresses about Figure is the fine dexterity: the robot is able to handle complex sheet metal, to "feel" whether it has misplaced a part and to correct its error in real time thanks to autonomous reasoning. It reaches near-human speed and precision.

Facing it, Elon Musk keeps pushing his Optimus, with the promise of producing one million units a year from 2026. While Tesla's timeline is often optimistic, the vision stays clear: make Optimus the standard, the iPhone of robotics, leaning on the massive production capacity of its Gigafactories.

But the real spoiler could come from China, with Unitree Robotics. If you thought robotics was expensive, Unitree changes the game. Their G1 model, launched at an unbeatable price (around $16,000), is not only affordable: it is acrobatic. Able to get up from a fall in an instant, to perform backflips and to fold in two to be stored away, Unitree brings a dynamic agility that Western robots, often more rigid, still struggle to match.

The Android in your living room: 1X and the data loop

While the industry transforms, domestic robotics too is getting ready to cross the threshold of our homes. Norway's 1X, with its Neo robot, embodies this consumer approach. Neo is not rigid and dangerous the way the image of the industrial robot would have it. It is soft, silent, designed not to hurt. And it looks kind (or particularly creepy, depending on taste!)

Offered on preorder ($20,000 or $499 a month), Neo uses a "teleoperation" strategy. When it does not know how to do a task (folding your laundry, for instance), a human operator takes remote control to carry out the action. It is a masterstroke: every human intervention becomes training data, creating a "data flywheel" that accelerates its autonomy.

Sergey Levine, co-founder of Physical Intelligence and professor at Berkeley, reckons that robots able to fully run a household could appear as early as 2030. A bold prediction, but grounded in a virtuous circle: the more robots you deploy, the more data you collect, the better the models get, the more robots you can deploy.

In parallel, XPeng and its Iron robot push mimicry to the extreme with a walking fluidity so natural that observers thought it was a human in disguise. The company's boss had to cut, on stage, the fabric covering the robot to prove it was not a human in disguise… ! Iron proves that the boundary between biological and mechanical movement is fading away.

The factory that thinks

But the ambition goes far beyond the domestic robot. It is the whole of industry that is in the crosshairs. Jeff Bezos has just launched Project Prometheus with 6 billion dollars on the table to create AI-driven industrial systems in heavy industry, notably automotive and aerospace. Tesla is deploying its Optimus robots in its own factories. In China, "dark factories", factories with no lights because they have no humans, are multiplying.

The idea is no longer only to robotize tasks, but to make entire industrial systems able to learn and adapt. A factory that detects a bottleneck and spontaneously reorganizes its flows. A production line that integrates a new product without reprogramming. Physical intelligence is autonomy injected into matter.

Breaking the glass: the end of theoretical AI

Physical Intelligence does not only mark an extension of AI's capabilities, it signs the end of its isolation. Until now, AI was a creature of the screen, a mind locked inside a glass bottle, fed on texts and static images that were only representations of the world.

By accessing a body, AI breaks the fourth wall. It leaves Plato's cave. It is no longer content to describe the real or to simulate it, it experiences it directly. Friction, gravity, the resistance of materials and the surprise of physical chaos become its new teachers. This confrontation with reality offers a source of "ground truth" infinitely richer and more complex than the whole of the internet combined. It is the shift from theory to practice, from word to act.

If a robot like Neo can learn to do the dishes by watching you three times, and if Project Prometheus can run an entire factory without supervision, the notion of "workforce" is about to live through its greatest mutation since the invention of the steam engine.

Code is no longer only law, it is now movement. After the age of information, the age of action begins.