Fei-Fei Li's World Labs Acquires SceniX: Physical AI Training Shifts from 'Collecting Data' to 'Creating Worlds'

On July 21, World Labs announced the acquisition of SceniX, followed a week later by revealing the latest results of its R2S2R (Real-to-Sim-to-Real) system. This system moves a real robot task into simulation, scales the task scenario for training and evaluation, and then deploys the policy back to the real robot. The key is whether the policy's relative performance in simulation predicts real-world performance and whether failure regions in simulation can expose potential deployment issues. World Labs aims to create a training and evaluation engine that expands real tasks into numerous controllable, reusable training worlds. The acquisition underscores a strategic bet on robotics: world models will not just generate visually realistic environments but also allow robots to act, experiment, and learn within virtual worlds. In parallel, Chinese company Wuwen AI has articulated a similar vision with its 'Collect World-Generate World-Simulate World' architecture, aiming to build data infrastructure for physical AI. Wuwen AI emphasizes closed-loop data cycles: real-world data feeds world reconstruction (Scan2Sim), generative expansion (Gen2Sim), and physics simulation (via its 'Wuqiong' world simulator), with failures in reality driving new data generation. The company argues that the next competitive frontier is not just amassing data or assets but creating many 'useful worlds' that continuously improve model capabilities.