When SceniX joined World Labs, we said spatial intelligence was never only about perceiving and generating virtual and physical worlds, but also interacting with them. Today, we’re sharing early results from that vision: building worlds that train robots. 🌎🤖↓

With the help of generative world models, a real-to-sim-to-real (R2S2R) simulation engine turns one physical task into many controllable, reusable worlds, helping robotics teams train policy models and test changes faster, uncover failures earlier, and reduce costly experimentation on hardware.

The Real-to-Sim part of R2S2R transforms physical robots, sensors, environments, objects, and interactions into simulations that preserve task-relevant observations and dynamics - not only how the world looks, but how it acts when the robot interacts with it. These results set a new bar for sim-real alignment in robot manipulation.

The Sim-to-Real part of R2S2R uses these aligned worlds for training. The policies were trained entirely in simulation with zero real-world data, transferred directly to diverse robot platforms, and operated autonomously for hours without failure or human intervention. Our engine is policy- and embodiment-agnostic, allowing us to serve customers with different robots, sensors, policy stacks, and deployment needs.

Sim-to-real aligned simulation also makes evaluation scalable. Here, the same failure behavior and outcome appear in both virtual and physical worlds. The simulation captures more than the task setup or final success label; it reproduces the conditions that push a policy toward success or failure. The result is faster iteration, broader coverage, and substantially lower cost.

In our taxonomy of world models, we called the simulator the linchpin: the place where agents can act, learn, and be evaluated. Our R2S2R engine can move robot development beyond slow, expensive, hardware-bound iteration toward more scalable and cheaper training and evaluation.

RL is all about envs. Real2sim2real is one of the best ways to scale envs for physical RL. Congrats!

@theworldlabs always keeping us well fed.

building synthetic worlds to train robots is only as good as how well those worlds lie about the real one.

When AI is no longer just "creating content", Instead, start to "create intelligence", The value of the world model has just begun to be released.

Simulation can predict whether a policy will succeed. Governed execution determines whether and when it may act in the real world. The deeper opportunity is a loop of simulation, execution outcomes and learning that builds evidence for policy, model trust and bounded autonomy.

the failure-finding line is the part that sells. training budgets get argued over, test budgets don't, because a failure you catch in sim has a number attached and the same one caught on a real arm has a bigger number.

this is genuinely very interesting! congrats on the release!

建数字世界训练机器人，可现实里连买个充电器都要实名，这平衡点在哪？

Been running staging gates for years, same logic. What changes is what a failure costs you: not a rollback, but someone having to manually reset the robot.

How i join world lab @drfeifei

飞飞教授，如果编程智能在自我迭代上够快，会不会取代空间智能的数据呢？

Literally, that picture looks so perfect. Absolute perfection. Speechless.

@andrewdsouza check this out, World Labs is building worlds to train robots - a sharp step from spatial intelligence into physical action

"Thank you, Fei, this is so inspiring! It really puts into perspective how we're just at the starting line when it comes to understanding the massive amount of data needed to create real-world AI and prepare for humanoid robots that will work alongside us. TY 4 the inspiration!"

Hey everyone — I just shared a post on LinkedIn about a simple but effective idea for memory management in real-time AI systems. If you work with AI models, streaming, or anything that involves long-running sessions, this topic is definitely worth checking out. It covers a common issue: AI models slowing down, freezing, or losing consistency after extended conversations.The post explains a lightweight “memory rotation” approach that keeps the system stable without needing bigger servers or additional resources. It’s written in both Arabic and American English so anyone can jump in.Feel free to check it out and join the discussion.

not just perceiving worlds or generating them, learning to interact with them, that's the breakthrough 💀

The future of AI is not just about creating smarter models. It's about teaching machines to understand and interact with the real world. Spatial intelligence could be the bridge between digital AI and physical reality. 🤖🌎

Worldbuilding, literally.

i've seen sim gains vanish at contact physics, which metric catches that early

The real-to-sim fidelity here is next-level — not just visual, but dynamics and failure modes too. Policy- and embodiment-agnostic simulation that actually closes the gap this well could become the new default for scaling robot learning. Excited to see how far this goes.

really curious how far the interaction layer can go once robots start learning inside these generated worlds

wow

This is the future we believe in, where spatial intelligence acts in the real world. Excited to see SceniX and World Labs building the foundation for robots!

👍👍💪🏻💪🏻

A generated world becomes useful when a robot can fail inside it before failing in a room. I’d want the test set to include the boring edge cases humans walk past.

This is the future Building worlds that train robots is next level spatial intelligence SceniX + World Labs cooking something huge

我更在意的不是感知、生成与互动。我关注的是它对物理世界规则的理解与运用。

When should we expect them to develop an accurate, internal spatial representation of a chess board?

Exciting progress bringing AI world models closer to robotics