李飞飞World Labs收购SceniX，物理AI训练正从“采数据”走向“造世界”

TL;DR - World Labs acquired robotics simulation company SceniX to advance real-to-sim-to-real training, signaling a shift from visually generating 3D worlds to simulating the physical consequences of robot actions.

• SceniX adds task reconstruction, physical-property recovery, interaction simulation, and robot policy evaluation to World Labs’ world-model capabilities.
• The R2S2R workflow reconstructs real tasks in simulation, expands them into controlled scenarios, trains and evaluates policies, then deploys them back to physical robots.
• Effective physical-AI infrastructure requires a closed loop spanning real-world data collection, generative scenario expansion, physics simulation, training, evaluation, and real-world feedback.
• The emerging competitive metric is not raw data or asset volume, but how many useful training worlds can be created and whether they measurably improve robot capabilities.