# World Labs — Building Worlds That Train Robots

- Company: World Labs (worldlabs.ai)
- Announced: 2026-07-28
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.worldlabs.ai/blog/real-to-sim-to-real
- Record: https://forck.live/items/7340-building-worlds-that-train-robots
- Subject: Marble / Atlas

World Labs announced that its generative world models, powered by the R2S2R engine from newly joined SceniX, can now simulate real-world robotic tasks to train policies with zero real-world data and evaluate their real-world performance. The R2S2R engine combines real-to-sim reconstruction with sim-to-real transfer, enabling robots to learn manipulation tasks entirely in simulation and operate reliably on physical hardware.

## Evidence

Verbatim from https://www.worldlabs.ai/blog/real-to-sim-to-real:

> Starting from a physical task, we capture the robot, sensors, surroundings, objects, and task demonstrations, the

---

Record: https://forck.live/items/7340-building-worlds-that-train-robots
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md