Introducing Multiplayer Interactive World Models with Representation Autoencoders

Researchers present a groundbreaking multiplayer world model designed for dynamic environments characterized by complex physical interactions. Unlike traditional single-player models, this innovative approach conditions on the actions of multiple agents, enabling the model to attribute environmental changes accurately. The study focuses on the game Rocket League, utilizing a 5-billion-parameter latent diffusion model trained on 10,000 hours of gameplay, achieving real-time match generation at 20 frames per second on a single Nvidia B200 GPU. Notably, the model maintains stable rollouts well beyond its training horizon, showing resilience for up to five minutes and even continuing for hours without collapse. The research also explores critical design choices and provides insights into model behavior related to scale, while releasing a dataset and codebase for further exploration in multiplayer world models.