Email us

We are a robot lab. We build World Action Models. Test your robot policies in the worlds they render.

scroll to explore

Solve Robotics.
Then Abundance.

Frontier models can already plan, reason, and recover. What they can't do is move things. We distill the physics locked inside internet-scale video into World Action Models: executors that carry out an LLM's plans in the real world. Models that move atoms.

Our Approach
Where World Models Evaluate Robots and Learn to Act

The lab · 02 · Atom 1

First, predict the world. Then act in it.

Atom 1 watches what your robot sees and the controls it sends, then rolls the world forward, step by step. Your checkpoints meet rain, glare, clutter and night shifts before the field does. A model that knows what happens next can learn what to do next. That model becomes our executor.

Conditioned on robot actions
The future it renders follows the controls your policy emits, step by step.
Physics, measured
#1 single-sample entry on DeepMind’s Physics-IQ Verified at 56.6, second overall. Every case is public; submission pending merge.
Trained on internet-scale video
The largest and cheapest source of physical interaction data on earth. No teleop fleet required.
Conditioned on your robot’s camera and actions
Rolled-out future — generated
Input · camerayaw −8.4° · pitch 2.1°
Input · gripper posex .41 · y −.12 · grip .63
And held to what reality does — case 0002, straight from the benchmark
The 3 s it saw
Atom 1’s next 5 s
What reality did
DeepMind Physics-IQ Verified · our entry: single sample, no best-of-N · pending merge (PR #71)All 66 cases →
1Magi-1 + GeoPhysbest-of-N harness58.2
2Atom 1OURS56.6
3Magi-1 24BSand AI48.4
4Cosmos3-SuperNVIDIA39.5
···ranks 5–10 hidden
11Sora 2OpenAI26.5

The lab · 03

Every rollout returns a typed verdict

Agents watch every rollout and grade it step by step: success, failure, and why. This feedback loop lets a frontier LLM orchestrate motion, and it hands you every failure mode with the clips that prove it.

Graded step by step
Reach, grasp, transfer, place — each stage of every episode scored on its own.
Failures, clustered
Grouped by what actually breaks them — the scene, the object, the light.
What to fix first
Modes ranked by what they cost in success rate, not by what happened most.
What’s costing this checkpointwatched across 1,142 rollouts
✗ the moment it drops
Drops the plate at handover
✓ reach✓ grasp✗ transfer
231 rollouts · costs 9.1 pts of success
✗ the moment it slips
Grasp slips on glossy objects
✓ reach✗ grasp· transfer
148 rollouts · costs 5.8 pts of success
✗ the moment it stalls
Stalls in dense clutter
✗ reach· grasp· transfer
87 rollouts · costs 3.2 pts of success

Every mode arrives as footage, graded stage by stage — watch the failure, don’t hunt for it.

Step into a new world
and let your
imagination run wild
Want to Solve Robotics With Us?
CONTINUE TO SCROLL
CONTINUE TO SCROLL
Keep Scrolling
to Learn More
About Us
Next Page
0
0
0
0
0
0