Researchers trained AI models using 1 million hours of gaming data

Large language models are great with words but struggle to move in the real world. A British company called Worldmodeldata is working to bridge this gap by using data from video games to train new AI systems.

Games contain a mix of visual information and player actions. This combination creates a record of cause and consequence that is often missing from static internet data. The company has collected 1 million hours of gameplay to teach machines how to interact with their surroundings.

Worldmodeldata hopes this resource will act as a shortcut for building better world models. These models aim to help robots or autonomous systems handle tasks like picking up objects or moving through busy spaces.

Not everyone is convinced that gaming data is the best path forward. Experts at Nvidia suggest that video games often use shortcuts to mimic physics. These shortcuts might not provide the precision required for delicate motor tasks, such as how firmly a finger should grip a specific item.

Imagine you are teaching a robot to hold an apple without crushing it or dropping it. Would you trust a system that learned its physics from a digital environment designed for fun rather than reality?

Researchers are still debating whether gaming data can provide enough detail for high-stakes environments like factories or autonomous vehicles. For now, this is just one of many ways scientists are trying to give machines a sense of physical reality.