The Unsupervized Thesis

01

Robotics broke the pattern

Humanoid robots today learn mostly through supervized methods. That means teleoperated demonstrations, curated datasets, and carefully annotated trajectories. Each hour of useful training data costs human time, human attention, and human judgment.

This approach can teach a robot a task. It cannot teach a robot the world. A general-purpose humanoid brain will need more experience than humans could ever demonstrate or label.

For robotics to follow the path of language and vision, robots need a source of training data that is abundant, auto-verifiable, and usable with little to no pre-processing. They need to learn directly from their own experience.

The question is what that experience should contain.

02

The data robots are throwing away

The current Physical AI stack is built around vision. Robots record what they see, and models are trained to map pixels to actions.

But vision shows only the surface of an event. When a robot grips a cup, pushes a door, or steps onto uneven ground, it is taking part in physics: force, friction, pressure, torque, temperature, vibration, and mass in motion. All of that information exists in every action, and almost all of it is discarded.

Physics is the most complete explanation of the ordinary world. Everything a robot does and everything that happens to it can be described in physical terms. And physical measurements have a property that labels made by people never will. They are ground truth. A force reading does not need a person to interpret it. It is already correct.

This is the unlock. A robot that captures the physics of its own experience generates self-verifying, multimodal training data every second it operates.

Every action becomes a lesson, and no one has to label it.

03

Fewer sensors, more physics

Nature solved this problem first. The human body contains trillions of cells, and a large share of them are dedicated to sensing. Together they build a rich, continuous model of the body and its surroundings.

No robot will ever carry trillions of sensors. We don't need it to. Physics lets a small number of well-chosen measurements explain a great deal. From the right signals, a robot can infer what is happening inside its own body and in the world around it.

Our goal is not to copy biology's sensor density. It is to reach biology's depth of understanding through physics.

04

Why we build the full stack

Existing robot hardware was not designed to capture physics, and existing robot software was not designed to learn from it. Fixing either one alone is not enough.

So we build both.

Our hardware is designed to capture far more physical information from every experience. With our sensor suite, a robot collects 20x more information per second than conventional setups.

Our software is reimagined around that data. It does not try to understand the entire world. It focuses on understanding the robot's own sensor streams at runtime, processing them on board, and giving the robot high-fidelity physical perception at millimetre-level precision.

The hardware makes the data possible. The software makes the data useful.

Neither works without the other, and that is why we own the full stack.

05

Superhuman, not humanlike

It is tempting to imagine the humanoid brain as a copy of the human one. We think that sets the ceiling too low.

Humans cannot sense force to a fraction of a newton, judge distance to the millimetre, or notice subtle thermal shifts in real time. A robot built on physics can do all three. The goal is not a machine that perceives the world the way we do. It is a machine that perceives it more precisely than we ever could, and learns from that perception without limit.

06

Beyond Earth and within it

Physics is not constant. It changes beside a blast furnace, inside a uranium mine, and within a burning coal seam. It changes as you rise toward the sky or descend toward the Earth's core. On the deep-sea floor, it transforms almost completely.

A robot trained only on images from ordinary environments will struggle when the physics around it changes. A robot that understands physics at its core can adapt.

These extreme environments are also the places where humans face the greatest danger. Deep mines, industrial furnaces, disaster zones, the ocean floor, and eventually worlds beyond our own all need capable hands that people should not have to risk their lives to provide.

We intend to pioneer that work.

07

Affordable, sovereign hardware

Abundant data depends on abundant robots. A capable humanoid that costs too much to buy and too much to repair, or that depends on a fragile, geopolitically exposed supply chain, will never be deployed widely enough to matter.

If robot hardware stays expensive, closed, and bottlenecked by geopolitics, robots will stay rare. Rare robots generate scarce data, and scarce data keeps Physical AI where it is today.

The need is already here. Across many regions, industries that rely on hard physical work in harsh conditions struggle to find local workers. These jobs are demanding, dangerous, or simply unwanted, and the gap keeps growing. This is the moment for humanoids to start working alongside people, beginning with the work no one wants to do.

That can only happen if the hardware is built for the real world. It has to be affordable to own, easy to repair, and simple to maintain wherever it operates. It has to be trustworthy and sovereign, so that the countries and industries relying on it are not exposed to supply shocks or decisions made far beyond their control.

This is why we design our hardware for affordability, repairability, and supply-chain independence from the start.

Affordable hardware also powers the data flywheel at the heart of our thesis. More robots at work means more physical experience captured. More experience means better models. Better models make robots more useful, and more useful robots get deployed more widely. Each turn of the cycle strengthens the next.

Abundant hardware creates abundant data. Abundant data is how unsupervized learning finally comes to the physical world.

08

The next five years

We believe the next five years will decide how Physical AI scales. One path keeps robotics dependent on human demonstration and annotation, improving slowly, one task at a time. The other gives robots the ability to learn from the physics of their own experience, the way language models learned from text.

We are building the second path.

Our aim is a general-purpose AI brain for humanoids, grounded in physics, trained on self-verifying data, and capable of working anywhere from the factory floor to the places no human can safely reach.