Britain’s WORLDMODELDATA raises €8 million: will video games train the robots of tomorrow?

Video games are no longer merely an entertainment market. They could become one of the most strategic resources for the next generation of artificial intelligence. While large language models learned from web data, robots, autonomous vehicles and future physical agents require a different form of training: they need to understand how their actions transform their environment. Virtual worlds provide precisely this kind of experimental ground.

The €8 million funding round raised by British startup Worldmodeldata illustrates this trend. The company develops neither language models, robots nor chips. Instead, it is building a library of training data derived from video games to power “world models”, systems designed not merely to understand the world, but to act within it. This innovation could prove transformative for an industry currently under severe pressure. Video games could become a strategic infrastructure for physical AI, while discovering a second business model based on the monetisation of their virtual worlds.

After the internet, where will AI find its next raw material?

Generative AI was built on an abundant resource harvested from the internet. Large language models learned to predict the next word by absorbing billions of web pages, books, lines of code and images. This strategy enabled spectacular progress, but it begins to reach its limits once the objective is no longer to generate text, but to interact with the physical world.

Recognising a hammer in a photograph is one thing. Knowing how to pick it up, adjust one’s grip to its weight, drive in a nail or anticipate the trajectory of a moving object is quite another. Future robots, autonomous vehicles and industrial agents will need to reason in terms of actions, consequences and physical dynamics. They will need to learn causality.

That is precisely the ambition behind world models. Unlike generative models, which predict the most likely content, they seek to anticipate how an environment will evolve following a particular action. What happens if a robot pushes a door? How will the surroundings change if a vehicle brakes on a wet road? This ability to mentally simulate the consequences of a decision has become one of the main areas of research across AI laboratories.

Why video games now interest researchers

The data available online is insufficient to train these models. A video may show a car turning left, but it does not reveal why it turned, the steering angle involved, the tyres’ level of grip, the vehicle’s speed or the forces acting upon it.

A video game engine, by contrast, knows every variable within the simulation. It calculates gravity, collisions, speeds, masses, trajectories, mechanical constraints and every interaction between objects. Each player action can be recorded together with its context and consequences.

This distinction is fundamental. Engines developed by Epic Games through Unreal Engine, or by Unity Technologies, do not merely produce images. They generate fully simulated environments in which every state is known. For laboratories working on robotics or autonomous vehicles, these worlds represent an unprecedented source of rich data.

This is precisely Worldmodeldata’s business model. The startup transforms gameplay from titles developed with Unreal or Unity into structured datasets for AI laboratories. Unlike the mass data-harvesting practices that accompanied the rise of LLMs, the company says it operates exclusively through licensing agreements with studios and creators, opening the way for content owners to be compensated.

Game engines are becoming critical infrastructure

This shift extends far beyond the video game industry. For several years, technologies developed for gaming have gradually been migrating into the industrial world.

CARLA, a project built on Unreal Engine, has become one of the leading simulation platforms used by researchers working on autonomous vehicles. Microsoft developed AirSim, also based on Unreal, to train drones and self-driving cars in virtual environments before deploying them in the real world.

NVIDIA, meanwhile, has considerably expanded the ambitions of its Omniverse platform. Initially designed for industrial digital twins, it is now used to train robots through Isaac Sim, a simulation environment based on the same principles as game engines: reproducing the laws of physics accurately enough for machines to learn without risk.

Google DeepMind is following a similar path with Genie and Genie 2, which can generate interactive worlds from images, while Meta is developing Habitat to train embodied agents in virtual environments.

The objective is to replace part of the learning that currently takes place in the physical world with billions of iterations performed in simulation.

The economic stakes are considerable. Training a robot in a factory requires expensive equipment, ties up machinery and carries operational risks. Training thousands of robots simultaneously in a virtual environment requires only computing resources.

Publishers discover a second business model

This shift could profoundly transform the video game industry itself. Until now, the economic value of a game has depended on sales, subscriptions, additional content and microtransactions. Tomorrow, virtual worlds could also become strategic assets for training artificial intelligence systems.

Studios already possess what AI laboratories are seeking: cities, roads, buildings, characters, human behaviours, physical interactions and millions of hours of gameplay.

Their assets would no longer be limited to intellectual property created for entertainment. They could become resources used to train the robots of the future.

Worldmodeldata is attempting to organise this new value chain. Its ambition is to build a library containing more than one million hours of data by the end of 2026, compared with around 40,000 hours in the largest datasets available today. If it delivers on that promise, the company will control a strategic asset that would be difficult to replicate.

This prospect could create an entirely new licensing market between game publishers and AI laboratories. Just as music platforms learned to monetise their catalogues through streaming services, studios could soon license their virtual worlds to developers of autonomous systems.

A new AI value chain is emerging

Worldmodeldata is not an isolated case. Several companies are already building the different layers of this emerging economy.

Encord develops platforms for annotating and managing data intended for physical AI. Scale AI continues to expand its activities around datasets used in defence, robotics and autonomous vehicles. Labelbox is also positioning itself as an infrastructure provider for organising industrial datasets.

Competition is accelerating in world models themselves. World Labs, founded by Fei-Fei Li, aims to build models capable of understanding and generating coherent three-dimensional environments. Stanhope AI is working on adaptive models for robotics and defence. Physical Intelligence is developing a foundation model designed to control robots made by different manufacturers. Skild AI is pursuing a similar ambition with a universal platform for embodied AI. BeyondMath is applying these approaches to physical simulation.

These companies are not directly building the robots of the future. They are developing the software infrastructure, data and training environments that will allow others to do so.

The limits of simulation

The idea that robots could learn exclusively inside video games nevertheless goes too far. Robotics has long faced what researchers call the sim-to-real gap. A system that performs well in simulation may fail as soon as it is deployed in the real world. Materials age, sensors drift, human behaviour is unpredictable and industrial environments are rarely as clean as their digital equivalents.

Video games also introduce biases of their own. Players take risks that no real driver would accept. Physics engines sometimes prioritise a smooth gaming experience over absolute scientific accuracy. Rare events, despite being essential to the training of safety-critical systems, also remain difficult to reproduce.

Simulation will therefore not replace real-world data. It could, however, substantially reduce the cost and duration of the training phases that precede validation under real-world conditions.

The next AI battle will be fought over data

For the past three years, global competition has focused on language models, semiconductors and computing capacity. A new layer of value is now emerging: the data that allows artificial intelligence systems to interact with their environment.

This development presents an unexpected opportunity for the video game industry. Engines created to build immersive worlds could become critical infrastructure for physical AI. Publishers’ catalogues could acquire a value extending far beyond entertainment. And the data generated by millions of players could become one of the most sought-after resources among laboratories working on robots, autonomous vehicles and industrial agents.

If large language models are the children of the internet, the artificial intelligence systems capable of acting in the physical world may well become the children of virtual worlds.