Fei-Fei Li’s World Labs finally has something concrete to show for its billion-dollar war chest. The startup just unveiled Real-to-Sim-to-Real, or R2S2R, an engine that takes a single real-world robot task and explodes it into thousands of simulated variations. Models trained entirely in these virtual worlds then transferred to real hardware and ran for an hour straight across four different robot platforms without any human intervention.
The pitch is blunt: the bottleneck in robotics isn’t model architecture, it’s experience. Real-world data is expensive, and even YouTube videos don’t systematically cover the failure states that actually matter. So instead of scraping more data, World Labs rebuilds the task itself — capturing the robot, sensors, environment, and demo, then regenerating it with changed lighting, object positions, friction properties, and camera angles. The simulation doesn’t just look like reality; it’s supposed to behave like it physically.
What makes this credible is the evaluation work. The team ran side-by-side comparisons of simulated and real cube handoffs on the open-source ALOHA platform, which costs a fraction of commercial systems. Simulation reproduced borderline grasps and failures that real hardware showed. Across GR00T N1.6 and π₀.₅ checkpoints, model rankings stayed consistent between simulation and reality, using 2,000 simulated and 100 real runs per checkpoint. That’s the exact kind of evidence that matters if you’re asking engineering teams to trust a virtual test harness.
This is the first time World Labs’ world-model taxonomy has translated into a robotics application you can point to. The long-term pitch — scale robot intelligence by scaling the worlds they learn in — mirrors what’s already worked in autonomous driving, where Level 3 and Level 4 systems mix real and simulated training data. Whether that transfer holds for more chaotic environments and less standardized robots is still open. But for now, World Labs has done something unusual in the sim-to-real conversation: they’ve shown receipts.
