Open Source 5B People World Model MIRA: Using DINOv3 Representation to Alleviate Long-Tail Drift, Enabling Real-Time Simulation of "Rocket League" 2v2 Showdown

According to Vision One monitoring, the AI research institution General Intuition, in collaboration with the French AI lab Kyutai and Epic Games, has launched the multiplayer interactive world model MIRA. As a generative game simulator that supports multiplayer real-time interaction, MIRA can simulate a 2v2 showdown in "Rocket League" in real time based solely on historical frames and player inputs, without the need for a physics engine, rendering engine, or explicit 3D representation.

In contrast to the "logic computation and graphics rendering decoupling" approach adopted by companies like Odyssey, MIRA follows a generative simulation path based on video latent space. With 5 billion parameters, MIRA's core design is to establish the latent prediction space on top of the frozen universal visual encoder DINOv3-L. Leveraging pre-trained visual features, the generated latent state can more stably land in a valid representation space, significantly alleviating visual drift and divergence in long-term predictions.

For multi-screen alignment, MIRA stitches the latent frames of four players' perspectives into a unified grid, allowing the spatial attention mechanism to naturally operate across views, enhancing spatial consistency of vehicles, soccer balls, and key events across multiple perspectives. The introduction of Action Dropout during training also helps the system to complete game behaviors for vehicles not directly controlled by instructions when some action streams are missing.

Currently, MIRA can run in real-time at 20 frames per second on a single NVIDIA B200 graphics card. The team has open-sourced the training and inference code, and released the Rocket Science dataset, which includes 1000 hours of matches, approximately 4000 hours of multi-angle videos, action streams, and physical state data; the model was trained on around 10,000 hours of clean match data.