ChainThink reports that on July 7, AI research firm General Intuition, in collaboration with French AI lab Kyutai and Epic Games, launched MIRA, an open-source 5-billion-parameter multi-agent interactive world model capable of real-time simulation of Rocket League 2v2 matches using only historical gameplay footage and player inputs—without requiring a physics engine, rendering engine, or explicit 3D representations during inference.
MIRA adopts a generative simulation approach based on video latent space, constructing its latent prediction space on top of the frozen general-purpose visual encoder DINOv3-L to mitigate frame drift and divergence in long-term prediction.
Its multi-perspective scheme stitches together the views of four players into a unified grid to enhance spatial consistency across perspectives. The action dropout mechanism introduced during training completes the behavior of vehicles not under command when part of the action stream is missing.
MIRA can currently run in real-time at 20 FPS on a single NVIDIA B200 GPU.
The team has open-sourced both the training and inference code, and released the Rocket Science dataset, which includes 1,000 hours of matches and approximately 4,000 hours of video, motion flow, and physics state data across four perspectives;
The full training used approximately 10,000 hours of clean game data.