Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction
ME-World is built to generate synchronized first-person video streams for several interacting agents in the same environment.
The paper frames multi-agent egocentric world modeling as a shared-world generation problem, not just separate ego-view prediction. Its model jointly denoises multiple ego streams, conditions each view on all agents’ target poses, and uses shared environment memory to keep interactions consistent. The authors also introduce metrics for environment, update, and identity consistency. They report gains over existing methods in shared-world consistency, action control, identity preservation, and video quality. HF Daily Papers' note
The paper frames multi-agent egocentric world modeling as a shared-world generation problem, not just separate ego-view prediction. Its model jointly denoises multiple ego streams, conditions each view on all agents’ target poses, and uses shared environment memory to keep interactions consistent. The authors also introduce metrics for environment, update, and identity consistency. They report gains over existing methods in shared-world consistency, action control, identity preservation, and video quality. HF Daily Papers' note
score 4