World Labs

Robotics · Spatial Intelligence AI · Simulation

Launch

World Labs unveils Real-to-Sim-to-Real engine to train robots in simulation

Robots have no internet-scale record of physical experience, so companies are now building virtual classrooms to manufacture it.

• World Labs unveiled its Real-to-Sim-to-Real (R2S2R) engine, which trains robot control policies entirely in virtual environments and lets the resulting models run reliably on physical hardware for hours.
• World Labs, founded by AI researcher Fei-Fei Li, builds large world models; its Marble model already turns a single image, multiple images, or text into a geometrically consistent 3D world.
• The engine converts one captured real-world robot task into thousands of simulated variations by randomizing lighting, friction, geometry, and object properties to broaden training coverage.
• The engine builds on World Labs' quiet acquisition of SceniX, a robot-simulation startup founded by Yunzhu Li, an MIT PhD who did a postdoc under Fei-Fei Li at Stanford, whose pipeline maps physical environments into high-fidelity digital twins.
• Language models learned from decades of internet text, but robots have no comparable dataset of physical interaction, making physically-consistent simulation the emerging substitute for scarce, costly, and dangerous real-world robot data collection.
• The launch arrives days after Google's Gemini Robotics 2 and alongside Boston Dynamics' large-scale simulation training for its Atlas humanoid, underscoring simulation as a new competitive front in robotics.