MetaWorld: Scaling Multi-Agent Video World Model from Single-view Video Data

2026-06-01

AI Summary

Existing video world models struggle to scale to multi-agent scenarios due to the scarcity of multi-view data and inconsistencies in cross-view states. This work proposes MetaWorld, the first framework to construct an open-domain multi-agent video world model using only monocular videos. It introduces Monocular World-State Unrolling (MWSU) to disentangle camera and agent motion, enabling synchronized multi-agent trajectories within a shared 3D space. A Subject-Aware World Generator combined with identity-conditioned controls facilitates appearance-controllable generation. Furthermore, cross-branch cross-attention is integrated into each layer of the diffusion transformer, enforcing physical consistency of geometry and dynamics across views through a World-State Alignment (WSA) mechanism. Experiments demonstrate that the proposed method significantly outperforms existing approaches in cross-view consistency and identity fidelity, establishing a scalable, physics-driven paradigm for multi-agent video world modeling.

Abstract

Video world models are a foundational generative technology for embodied AI and the Metaverse, yet existing approaches are inherently limited to a single agent observing from a single perspective. Extending these models to multi-agent settings introduces two critical challenges: data scarcity (coordinated multi-view recordings are prohibitively expensive to collect for general open-domain scenarios) and world state alignment (independently generated video streams cannot ensure that shared physical environments and events evolve consistently across views). To address these challenges, we propose MetaWorld, a novel framework that scales multi-agent video world models to open-domain environments directly from single-view videos. First, we introduce Monocular World-State Unrolling (MWSU) to explicitly decompose monocular footage into the camera operator's ego-motion and the visible subject's spatial trajectory. This camera-trajectory decomposition naturally extracts synchronized multi-agent motion data within a shared 3D space, completely bypassing the need for multi-camera setups. Second, for precise visual control, we develop the Subject-Aware World Generator to enable appearance-driven simulation conditioned on per-agent identity images. Finally, to ensure both views are grounded in the identical physical reality, we propose World-State Alignment, a per-frame inter-branch cross-attention mechanism inserted at every transformer layer of the video DiT. By jointly synchronizing the denoising process, WSA enforces both static geometric consistency and dynamic motion consistency, encouraging that the shared 3D environment and physical events remain well-aligned across both egocentric views. Extensive experiments demonstrate that MetaWorld achieves superior cross-view consistency and identity fidelity, establishing a highly scalable, physics-driven paradigm for multi-agent video world modeling.

Problem

• multi-agent video world model
• data scarcity
• world state alignment
• single-view video
• cross-view consistency

Innovation

• multi-agent video world model
• monocular world-state unrolling
• world-state alignment
• subject-aware generation
• cross-view consistency