MIRA has released a new family of world models that learn from thousands of Rocket League matches. The models capture the game’s physics, ball dynamics, and player coordination, enabling agents to predict future states in a multiplayer setting.

Training the models on the full Rocket League dataset allows them to internalize complex interactions—corner kicks, aerial passes, and collision responses—without hand‑crafted rules. The resulting simulations reproduce realistic ball trajectories and player behavior, making them useful for testing strategy bots in a high‑fidelity virtual environment.

Practical applications extend beyond game AI. Developers can use the world models to prototype它 AI agents before integrating them into live servers, reducing load on production systems. The approach also suggests a path for robotics and autonomous driving research, where physics‑accurate simulations are essential.

Overall, MIRA’s work highlights how large‑scale, data‑driven world models can bridge the gap between simulated training and real‑world multiplayer interactions.