Worldmodeldata publishes position paper on game-generated AI training data
Worldmodeldata published a position paper examining game-generated data as a training resource for advanced AI and JEPA-style world models.
Evidence notes
- The paper examines game-generated trajectories as a source of structured AI training data.
- It discusses Joint-Embedding Predictive Architectures and world-model learning.
- The work connects gameplay data with physical-AI, planning and self-supervised-learning research.
Company context
Worldmodeldata produces licensed, action-conditioned gameplay datasets for training world models and embodied AI. Its synchronized stack combines multi-view video, action inputs, telemetry and ground-truth 3D state. The Cambridge company emerged from stealth in July 2026 with a GBP 7 million seed round led by Iona Star Capital.