World Labs Shows How Robots Learn Complex Tasks Without Real Training Data

World Labs published first results of its R2S2R engine, which converts real robot tasks into simulations to train and evaluate control models without expensive real-hardware trials. The technology comes from SceniX, acquired by World Labs in July. The engine works in two steps: first, it records robots, sensors, environment, and task demonstrations to reconstruct an interactive virtual world that behaves physically like the original. From a single real task, it generates thousands of variations in lighting, object positions, physics properties, and camera perspectives. Then control models are trained in simulation and transferred to real robots. Tests were conducted on the ALOHA platform and four other robot platforms, where models ran for an hour without human intervention. The examples include cable routing, inserting an elastic cable end into a hole, and two-handed box packing. The system is model-agnostic and can be reused for new models and robots. World Labs argues that robot development lags behind language models because evaluation is tied to real-world tests; a robust simulation must answer the same questions about model failures and improvements. They tested with a two-handed cube transfer task, where the simulation reproduced edge cases and failures. The ranking of models (including GR00T N1.6 and π₀.₅) remained similar between simulation and real world across 2000 simulated and 100 real runs per checkpoint. This allows weak versions to be filtered out in simulation, saving real hardware tests for promising candidates. World Labs positions the simulator as central to its taxonomy of world models, citing parallels to autonomous driving. The long-term goal is to scale the worlds in which robots learn. World Labs, founded in 2024 by Fei-Fei Li, has raised a billion dollars in venture capital. The R2S2R engine is the first concrete application in robotics. The research contributes to the debate on world models in robotics, distinguishing them from video generators and contrasting with other approaches like World Action Models and China's Orca.