Event

OverSeer: Multi-agent world model introduces fixed auxiliary viewpoint

1 reports · 1 sources

The study proposes a multi-agent world model, OverSeer, which adds an auxiliary camera viewpoint fixed in space in addition to each agent's first-person viewpoint, serving as a proxy for a world-centric representation to improve temporal and cross-agent consistency. The auxiliary viewpoint is redundant and approximately static, and can be compressed using a delta-VAE that encodes only changes relative to a static reference image, reducing the number of tokens required. Experiments in Minecraft and Doom show improved world and agent consistency; qualitative results on real-world multi-camera captures indicate that the method is not limited to game engines.