A research team has built the first world model capable of simulating a multiplayer game in real time, using Rocket League as their test bed.
The model is a 5-billion-parameter latent diffusion system trained on 10,000 hours of gameplay from publicly available bots. It generates four-player matches at 20 frames per second on a single Nvidia B200 GPU. Most world models treat other agents as background scenery; this one conditions on the action streams of every player simultaneously, learning to correctly attribute scene changes to the right actor. Trained only on short clips, it stays stable well past the training horizon — rollout quality held up to five minutes under formal evaluation, with researchers observing hour-long runs without collapse.
This matters because multiplayer environments are where single-player world models quietly break down. Rocket League's fast, tightly coupled physics — mid-air collisions, ball trajectories that depend on simultaneous inputs from four players — is a stress test that exposes whether a model truly understands causality or is just pattern-matching visuals. A model that survives that without falling apart is a meaningful step toward simulated environments that could replace or augment real game engines for training, testing, or creative tooling.
The team is releasing the dataset, training and inference code, and a live demo — which puts this work in a stronger position than most academic world-model papers, where the gap between published results and reproducible artifacts is often wide enough to park a truck.