Most of the AI we use today, like ChatGPT, is essentially a very powerful librarian. It has read millions of books and websites, making it brilliant at language, but it has never walked across a room. If you asked it to navigate an office, it would be hopelessly lost because it lacks a fundamental understanding of how things move through physical space.
WHAT'S HAPPENING
A New York-based startup called General Intuition is trying to change this by teaching AI to understand the physical world using an unexpected source: video game data. The company is building a foundation model—a massive, versatile digital brain—trained on millions of hours of footage showing people playing video games. This data includes the video feed itself, along with the precise button presses the player made to control their character. By analyzing this, the AI learns the relationship between seeing a goal on screen and performing the necessary actions to reach it. Recently, the company showed that a robot equipped with this model could navigate an office environment safely after being shown only eight minutes of real-world movement data.
Why games are the perfect practice ground
HOW IT WORKS
Think of typical AI training like learning math from a textbook. It’s logical and precise, but it doesn't prepare you for the messy, unpredictable nature of real life. To teach a robot to interact with the world, it needs to understand physics—how things fall, how objects get in the way, and how time relates to movement.
Video games are ideal instructors because they are simulations of these exact concepts. When someone plays a game, they are constantly making decisions based on moving images in a virtual space. By processing millions of hours of this, the AI develops a form of physical intuition. It learns that if it wants to move from point A to point B without hitting an obstacle, it needs to perform a specific sequence of actions. This base knowledge is so flexible that when the model is applied to a physical robot, it doesn't need to be taught from scratch. It just needs a tiny final lesson—like those eight minutes of office footage—to understand the specific physical limitations of its own body. This is a leap forward from past methods where robots had to be individually programmed for every single environment they might encounter.
WHY IT MATTERS
If this approach works, we may stop needing to collect massive amounts of data for every new robot we build. Currently, teaching a machine is an expensive, slow process. If developers can instead use a pre-trained model that already knows how the world works, it could drastically lower the cost and complexity of building everything from automated delivery vehicles to household assistants. The goal isn't to build the robots themselves, but to provide the common brain that allows anyone to build them more easily. It shifts the challenge of robotics from years of custom engineering to a quick, intuitive adjustment.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy