Kyutai Open-Sources MIRA: A Stable World Model That Learns Rocket League From Raw Video
Key Takeaways
- ▸Kyutai open-sourced MIRA, a 5.6B-parameter world model that learns to simulate Rocket League entirely from video, with no physics or rendering engine
- ▸The model achieves unprecedented stability in long-horizon video rollouts, a persistent challenge in generative video prediction
- ▸This is framed as research toward physical AI (robotics, autonomous vehicles) using video games as a safer, cleaner data domain than real-world footage
Summary
Kyutai has open-sourced MIRA, an ambitious world model that learned to simulate Rocket League from raw gameplay footage. The 5.6B-parameter diffusion transformer can generate real-time, consistent gameplay for four players at 576p resolution with no explicit physics engine or 3D representation — learning all mechanics purely from data. The release includes the trained model, dataset, training code, and inference infrastructure.
MIRA demonstrates remarkable stability across indefinite rollouts without divergence, a longstanding challenge in video prediction models. The system handles complex multi-agent interactions including car physics, boost mechanics, demolitions, and event tracking, all without hand-coded game logic. Kyutai positioned this work as a stepping stone toward physical AI applications like robotics and autonomous vehicles, where data collection and testing are far more expensive and dangerous.
- The full model, dataset, training code, and inference code are being released as open source
Editorial Opinion
MIRA represents a fascinating proof-of-concept for learning world dynamics from pure data, but Kyutai is refreshingly clear about what it is — not a game engine replacement, but a research milestone toward physical AI. The stability claim is compelling if verified, though the real test will be whether lessons learned in a controlled game environment transfer to messier real-world robotics and autonomous systems. This kind of foundational, open-sourced research could meaningfully accelerate progress in sim-to-real transfer.
