Fei-Fei Li's World Labs launches R2S2R engine to train robots in simulation

World Labs released the Real-to-sim-to-real (R2S2R) engine for training and evaluating robot policies. R2S2R consists of two parts: Real-to-Sim, which reconstructs real-world robots, sensors, objects, and interactions into a virtual environment, and Sim-to-Real, which trains and evaluates policies in simulation before deploying them to real robots. Policies trained solely in simulation achieved up to 1 hour of continuous autonomous operation on real hardware without human intervention. The engine also enables scalable evaluation by aligning simulation and real-world performance. R2S2R extends World Labs' spatial intelligence from generating interactive virtual worlds to training and deploying real robots. World Labs acquired SceniX, a startup specializing in robot simulation and evaluation, co-founded by Yunzhu Li, a former postdoc under Fei-Fei Li. SceniX was initially a customer of World Labs' generative world model Marble before the acquisition. The engine supports multiple robot platforms including ALOHA, RB-Y1, YAM, Flexiv, and xArm, and tasks such as bin packing, cable routing, test tube transfer, and single object grasping.