World Labs: A single physical demo, thousands of simulated robot training scenarios

World Labs, founded by AI visionary Fei-Fei Li and funded with $1 billion, introduces R2S2R (Real-to-Sim-to-Real), which takes in robots, sensors, environments, and a single real demo of a task, and outputs an interactive virtual environment with physically plausible behavior, along with thousands of controlled variants for training.

1Results: Sim Rankings Match Real-World Performance

R2S2R has been tested through 2,000 simulations and 100 real-world deployments. Zero-shot transfer to four new platforms, including dual-arm ALOHA robots (Stanford) and Nvidia GR00T N1.6 and π₀.₅. Result: Best performing policies in simulation are also best performing on real hardware (same ranking)!

2Why This Matters for Robotics

For years, training robots has been held back by the high cost of collecting real-world data; it’s an order of magnitude more expensive to train on physical hardware than on simulated hardware. R2S2R is the first credible evidence that what works in simulation will work on real hardware, which could revolutionize the creation of industrial and household manipulation robots.