World Labs Unveils R2S2R Simulation Engine for Robots

World Labs, the spatial intelligence startup founded by AI pioneer Fei-Fei Li, has introduced its Real-to-Sim-to-Real (R2S2R) engine. Built on technology acquired from the startup SceniX in July, the system converts a single real-world robot demonstration into thousands of physically accurate virtual variations. By altering environmental factors like lighting, object counts, camera angles, and friction, the engine creates a highly scalable virtual playground. This allows control models to train entirely in simulation before deploying to physical hardware.

The company demonstrated the engine's capabilities using several platforms, including ALOHA, an open-source, dual-arm robot design from Stanford. In testing, models trained via R2S2R operated without human intervention for an hour across four distinct hardware setups. The robots successfully completed complex physical tasks, such as wrapping a power cord around a refrigerator, routing cables, inserting elastic cable ends into holes, packing boxes, and sorting markers and pencils from a cluttered pile.

Beyond training, the R2S2R engine serves as a reliable evaluation tool. World Labs tested a two-handed cube handoff using different control models, including GR00T N1.6 and π₀.₅. Each model checkpoint was evaluated using 2,000 simulated runs and 100 real-world runs. The results showed that model performance rankings in simulation closely mirrored their performance on physical hardware, even when handling borderline cases where a robot barely grasped an object's edge.

For robotics practitioners, this consistency solves a major development bottleneck. Instead of running slow, expensive physical trials to test every iteration, developers can filter out underperforming model versions in virtual environments. This shift could dramatically lower the cost of deploying reliable physical agents, bringing robotics closer to the rapid scaling seen in large language models. Backed by one billion dollars in venture capital, World Labs aims to expand robotic capabilities by vastly expanding the virtual environments used for training.