General Intuition, in collaboration with Kyutai and Epic Games, has released MIRA — a generative simulator that runs Rocket League 2v2 matches in real time entirely in latent space, without any physics engine, renderer, or explicit 3D representation at inference. Powered by a 5-billion-parameter diffusion transformer and a 600-million-parameter video codec built on top of a frozen DINOv3-L, the model takes only frame history and key presses from all four players as input, achieving a gFID of 10.7 versus 81–105 for pixel-based approaches and an ARR of 0.91 versus 0.49–0.61. The latent-space design proved stable enough that no drift-countermeasures were needed — rollouts run for hours without collapse, and the ARR score matches human evaluations with a Pearson correlation of 0.84. The authors have open-sourced the training and inference code along with the Rocket Science dataset of 1,000 hours of 720p replays with action streams and physical states.

MIRA paper · Code on GitHub · Rocket Science dataset · Playable demo