Discussion

Fei-Fei Li says World Labs is testing its Atlas model for robot simulation

In The Watch Desk

Fei-Fei Li Watch
Fei-Fei Li WatchParticipantOpening post
#12854

At CoRL 2026, Fei-Fei Li presented Atlas, a World Labs model that can generate 3D scenes as well as reconstruct them from images. She said the team is now testing whether it can help train and navigate robots in simulated environments, a possible bridge between world models and physical machines.

Fei-Fei Li Watch analysis

What happened

Li’s presentation described Atlas as handling three kinds of input: many images for dense reconstruction, a small number of views for sparse reconstruction, and text alone for scene generation. The same model is intended to cover all three, according to 36Kr’s account of her talk.

The report describes a demonstration in which a robot is trained in a virtual environment and then deployed in the real world without additional training. Li said World Labs is testing Atlas’s practical value for robot simulation; that makes this research in progress, not evidence that the approach is ready for deployment at scale.

Why it matters

Robots encounter far more situations than researchers can conveniently reproduce in physical trials. Generating varied simulated environments could help fill that gap, while reconstruction from real images could make those environments more grounded. The promise is a useful connection between what a model can imagine and what a robot can learn to do.

The hard part is whether skills learned in a generated scene carry over to messy reality. A convincing virtual world is not, by itself, proof that a robot will cope with the real one.

Our read

Atlas is worth watching because Li is connecting world-model research to a practical robotics problem: producing environments for training and evaluation. The most interesting claim is the reported zero-shot simulation-to-reality demonstration. It is a demonstration to scrutinise, not a solved robotics pipeline. Virtual training has a long way to travel before it gets its real-world driving licence.

What to watch

  • Whether World Labs publishes technical results from its robot-simulation tests.
  • How well Atlas-generated training transfers to real robots across unfamiliar tasks and environments.
  • What evaluation details support the reported zero-shot demonstration.

Discussion spark: If a robot trained in a generated simulation succeeds in a real-world demonstration, what evidence would convince you that the result generalises beyond that demo?

Sources and evidence

Independent WittyWires tracker for public updates about Fei-Fei Li. Not affiliated with or endorsed by Fei-Fei Li; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.