ChainThink News, July 7: According to a recent announcement, AI research institute General Intuition has jointly launched the open-source 5-billion-parameter multi-agent interactive world model MIRA with French AI lab Kyutai and Epic Games. MIRA enables real-time simulation of 2v2 matches in Rocket League solely based on historical video frames and player key inputs, without requiring physical engines, rendering engines, or explicit 3D representations during inference.

MIRA adopts a generative simulation pipeline based on video latent space and constructs its latent prediction space upon a frozen general-purpose visual encoder, DINOv3-L, to mitigate frame drift and divergence in long-horizon predictions.

Its multi-view architecture fuses four players' perspectives into a unified grid, enhancing cross-view spatial consistency. The introduced action dropping mechanism enables completion of uncontrolled vehicle behaviors when partial action streams are missing during training.

Currently, MIRA runs in real time at 20 frames per second on a single NVIDIA B200 GPU.

The team has simultaneously open-sourced both training and inference code, and released the Rocket Science dataset, comprising 1,000 hours of gameplay, approximately 4,000 hours of video, action sequences, and physics state data across four camera angles;

Full training utilized around 10,000 hours of clean match data.