Alibaba· 16 JunQwen-RobotWorld: Boundless Worlds for Embodied Agents
Research paper
Alibaba introduces Qwen-RobotWorld, a dual-stream diffusion world model that uses natural language as a universal action interface to unify 20+ robot embodiments and 500+ action categories, enabling cross-scenario physical generalization across manipulation, driving, and navigation.
The model employs Qwen2.5-VL as an action encoder within an MMDiT architecture, and is trained on the Embodied World Knowledge (EWK) dataset of 8.6M video-text pairs.
It also supports multi-view geometrically consistent generation and human-to-robot transfer via Scene2Robot.
