Agora-1: A Multi-Agent World Model

Agora-1 is the first model in the multi-agent world model series released by Odyssey. In a shared deathmatch simulation based on GoldenEye, up to 4 players interact in real time, with all participants sharing the same world state.

The core idea is the separation of simulation and rendering. First, it directly learns the internal state of the game to create a shared world state, and based on that state, a DiT-based renderer generates the screen for each viewpoint. Both components are learned systems rather than hardcoded rules, and they maintain a consistent world even across different viewpoints.

Key points

• Unlike existing Multiverse, Solaris, and MultiGen approaches, it separates shared state from rendering.
• The state model learns structured game data such as position and health.
• By directly manipulating the state, it can also generate new levels that conform to the original game's rules.
• As the number of participants increases, interactions such as collisions, cooperation, and competition become richer, making it advantageous for generating multi-agent RL data.
• Odyssey views this as a research environment that can be extended to robotics and foundation models.

The team presents Agora-1 as a system close to a learned game engine, expanding the scope of generative simulation into multi-agent interaction.