What happened?

World Labs, founded by AI researcher Fei-Fei Li, has announced a simulation engine that trains robot control models entirely in virtual environments. The system, which the company calls "Real-to-Sim-to-Real" (R2S2R), converts real-world robot tasks into simulations for training and evaluation, reducing the need for costly physical hardware testing. The technology stems from SceniX, a startup World Labs acquired in July 2026.

The system captures robots, sensors, environments, and task demonstrations, turning them into interactive virtual worlds that are realistic not just visually but in terms of physical behavior as well. This is achieved by combining generative world models with task-oriented robot simulation. Starting from a single real task, thousands of variations can be produced by altering lighting, object position and count, environment, physical properties such as friction, and camera angle.

Why does it matter?

According to World Labs, the main obstacle to robot development isn't model architecture but the volume of experience a robot needs to operate reliably. Real-world data is expensive and hard to control, and online videos don't systematically cover the full range of objects, physical conditions, and failure cases. The company argues that evaluating control model performance has so far relied heavily on real hardware testing, which has left the robotics field lagging behind language models.

In World Labs' tests, control models were trained in simulation across five different robot platforms, including the open-source ALOHA platform, and then run on real hardware for an hour each without human intervention. Tasks included wrapping a cable around a refrigerator with both hands, precisely repositioning test tubes, and separating thin objects like pens from a pile.

What we know

  • The R2S2R engine turns a single real task into thousands of simulated variations by changing factors like lighting, position, and friction
  • Control models were tested across different model types and training stages, including GR00T N1.6 and π₀.₅
  • Each checkpoint was evaluated with 2,000 simulated and 100 real runs
  • Model rankings in simulation and the real world were largely consistent
  • World Labs was founded in 2024 by Fei-Fei Li and raised $1 billion in funding for world models targeting robotics and science

What's next?

World Labs states that the system isn't tied to a specific control model or robot type, and that a world reconstructed once can later be reused with different models and robots. The company describes its long-term goal as scaling the worlds robots learn from in order to scale their intelligence. Whether the results will transfer to more complex environments, different robot types, and less controlled everyday situations remains an open question.

The claim in concrete terms

The announcement's strongest claim is this: policies trained in simulation with zero real-world data transferred directly to diverse robot platforms and ran autonomously for hours without failure or human intervention. That stands directly against the classic problem known in robotics as the sim-to-real gap — what is learned in simulation falling apart in the real world is the field's best-known bottleneck.

Fei-Fei Li's framing points the same way: in her taxonomy of world models she positions the simulator as the “linchpin”, the place where agents can act, learn and be evaluated. R2S2R's promise is to move robot development out of the slow, expensive, hardware-bound loop.

The missing number

The announcement contains no comparative success rate. “Ran for hours without failure” is a qualitative statement; how many tasks, at what success percentage, and what the same policy achieves when trained conventionally are not shared. Sim-to-real claims separate precisely at this point: saying transfer works is easy, showing which class of task it works on and at what rate is hard.

One more point. The technology is not World Labs' own development; it comes from SceniX, acquired in July. The announcement is therefore less a research result than a productised acquisition — which does not make it worthless, but the timeline should be read correctly.