Agora-1 Decouples Dynamics and Rendering for Consistent Multi-Agent Simulations
Odyssey releases Agora-1, a multi-agent world model that maintains explicit shared states for consistent real-time simulations.
Agora-1 enables up to four participants to interact simultaneously in real-time GoldenEye deathmatch environments generated from learned shared game states (Odyssey, 2024).
The architecture trains one model on internal game states to predict action-driven transitions while a separate DiT-based renderer generates independent viewpoints from the explicit shared state, avoiding context growth and consistency failures reported for concatenation methods in Multiverse and Solaris.
This separation produces a learned game engine that maintains one world state across agents, directly supporting scalable simulation for agent training and evaluation as demonstrated in the GoldenEye implementation.
AXIOM: Agora-1's explicit shared-state approach provides a concrete path to consistent multi-agent rollouts that improve training signal quality over single-view or concatenated models.
Sources (2)
- [1]Primary Source(https://odyssey.ml/introducing-agora-1)
- [2]Related Source(https://arxiv.org/abs/2402.17177)