Rocket League is Becoming a Testbed for World Models

MIRA, a world model by Kyutai and General Intuition, simulates interactive 2v2 Rocket League matches in real-time on a single Nvidia B200 GPU.

With MIRA, Kyutai, and General Intuition, in collaboration with Epic Games, present what the technical report describes as the first interactive multiplayer world model for such a dynamic environment. The model learns to simulate a 2v2 match from the views and actions of four simultaneous players, then produces a playable game in real-time, at 20 frames per second on a single Nvidia B200 GPU.

Training relies on approximately 10,000 hours of self-play generated matches between instances of the public Nexto bot, without any human player data. The architecture combines a 5-billion-parameter diffusion transformer and a 600-million-parameter representation video codec, built on the DINOv3 visual encoder: this choice, according to the authors, ensures the stability of the simulations, capable of running for hours without diverging. All this without a physics engine, without 3D rendering, without an explicit scene representation. Yet, the model maintains consistency across the four views, collisions, boost gauge, goals, and demolitions, solely from pixels and actions. An action dropout strategy during training also enables an autopilot mode: the model itself drives the cars whose controls are removed, from one to four players.

The project comes with an online demo, a detailed technical report, the training and inference code, as well as Rocket Science, a subset of 1,000 hours of matches in 720p. The limitations are candidly documented: invented goal replays, a drifting timer, a ball that rolls on its own when no one touches it. Rocket League serves here as a closed arena: clean enough to test, dynamic enough to break fragile systems, with robotics, autonomous driving, and Physical AI in sight.