Worldmodeldata
Turning video games into training data for physical AI
What it does
Worldmodeldata extracts millions of multi-modal, action-conditioned trajectories from video games — safe and controlled environments where every interaction is captured — and delivers them as training-ready datasets for world models and embodied AI. The data includes temporally aligned video, telemetry inputs, and ground-truth 3D state, all synchronized and training-ready. The company builds toward one million hours of gameplay data, with 200+ structured properties extracted per frame, such as player position, velocity, view matrix, weapon state, and more.
Who it is for
Worldmodeldata serves AI research labs and companies building world models, vision-language-action models (VLAs), and physical AI systems. The datasets are designed for training models that need to understand the dynamic mechanics of the world — not just static scenes. The company also collaborates with academic researchers working on world models, physical AI, and self-supervised learning.
Why it matters
Generic video corpora teach a model what scenes look like, but worldmodeldata's data teaches a model what a world is: a coupled system of state, motion, intent, and consequence, frame by frame, provably aligned. By providing ground-truth 3D state straight from the game engine (not inferred from pixels), frame-aligned action conditioning (recorded controller inputs), and dense structured game state, the data enables AI to learn the true geometry of the world without perception noise or labeling error. This is critical for building AI that can operate safely and predictably in the real world.
Launch signal
Worldmodeldata published a position paper in 2025 titled "Game-Generated Data, An Untapped Resource for Advanced AI Training," which received interest from the research community. The company also offers an interactive browser demo where users can play a 60-second mini-game and see their gameplay become training data in real time, demonstrating the capture pipeline.
Brand and naming
The name "Worldmodeldata" directly communicates the company's focus: providing data for world models. The domain worldmodeldata.com reinforces this clarity. The brand positions itself as the "data layer for world models," emphasizing volume, diversity, and quality. The playful cat mascot and interactive demo make the technical offering approachable, while the emphasis on formal licensing agreements and zero-loss engine output signals trustworthiness for enterprise customers.
Founder
Rhea Loucas
Related
Get more like this in our weekly newsletter.