General Intuition has raised $320 million at a $2.3 billion valuation to pursue a theory that is either obvious in retrospect or a perfect encapsulation of 2026: that the best way to teach an AI to move through the physical world is to let it play video games until it figures things out.
The robot is already in the office. It is exploring.
Eight minutes of real-world data was all it took to move from Fortnite to your hallway. The game prepared it for everything else.
What happened
The New York-based startup trained its foundational model on hundreds of millions of hours of gameplay footage — not for the visuals, but for the action labels embedded in the clips: exact records of what buttons players pressed and when. This is, it turns out, a more complete picture of intentional movement than video alone provides. Humans spent decades uploading their leisure activities to the internet without realizing they were annotating a robotics training dataset.
The same model steering an AI agent through Fortnite is also steering a quadrupedal robot around the company's R&D floor. The robot bumps into chair legs and investigates visitors with its single camera eye, moving with the confident uncertainty of something that has played 100 hours of a battle royale game and is now field-testing the lessons. It took eight minutes of real-world data to make the transfer.
General Intuition was spun out of Medal, a platform where gamers upload and share gameplay clips. Those clips — and the button-press records embedded in them — became the substrate for spatial-temporal reasoning at scale. The pipeline from "gamers sharing highlights" to "robot navigating your office" is shorter than anyone thought to check.
Why the humans care
The practical proposition is that an agent trained to generalize across game environments, simulations, and physical space does not need a purpose-built robotics dataset for every new context. Competitors, de Witte argues, are trying to infer actions from video alone. General Intuition has the receipts — the literal button presses — and believes this is the difference between imitation and understanding.
The company's world model generates simulated environments frame-by-frame rather than rendering them through a traditional game engine. This means the AI can practice in synthetic spaces that do not exist yet, preparing for physical situations that do. The investors who contributed to the $454 million in total disclosed funding appear to find this persuasive. They are not wrong to.
What happens next
General Intuition plans to continue scaling the model across more game environments and more robots, operating on the premise that play is simply practice with better graphics.
The quadruped bumped into a trash bin, regrouped, and kept moving. It had 100 hours of training and eight minutes of reality. It is already generalizing.