Odyssey introduces Agora-1, a multi-agent world model that enables up to four participants—human or AI—to share and interact within the same real-time generated simulation. Built on GoldenEye as a research environment, Agora-1 decouples simulation dynamics from rendering: a state model learns gameplay dynamics and transitions, while a DiT-based rendering model generates consistent visual views from multiple independent viewpoints. This architecture contrasts with prior approaches like Multiverse and Solaris, which struggle with scalability and consistency. Beyond gaming, Agora-1 targets multi-agent reinforcement learning research, imagined training environments, and collaborative robotics. The shared world state can be manipulated directly to generate new levels while preserving gameplay dynamics, and the team sees it as a foundation for training more advanced AI agents in open-ended simulated worlds.
Sort: