MIRA Enables Multiplayer Interactive World Models Trained on Rocket League

The MIRA project introduces AI agents that learn a shared interactive world model. Training is performed on data from the video game Rocket League. The agents can predict game dynamics and coordinate actions in a multiplayer setting. By using a common model, the system reduces the need for separate policies per player. Experiments show that the agents achieve competitive performance against human players. The approach demonstrates how game environments can serve as testbeds for collaborative AI. Researchers highlight the potential for extending the method to other multi‑agent domains. Future work will explore scaling the model and improving real‑time responsiveness.