Virtual Reality
for Robots
Robotics is stuck: there isn’t enough data. Atom 1 is a world model — a virtual reality trained on trillions of video tokens, made physics-aware and conditioned on robot actions — where you test and train your robots. Every checkpoint, thousands of rollouts, judged automatically, in minutes.
The platform · 01
One checkpoint, a thousand worlds
Massively parallelized GPU evaluation: your checkpoint fans out across every world at once, each rollout comes back already judged, and the verdict lands in minutes — not after a week of bench time.
Real Atom 1 generations, each carrying its measured internal benchmark score — worlds that break physics are caught before they touch your numbers.
The platform · 02 · Atom 1
Test in worlds you could never film
Atom 1 watches what your robot sees and the controls it sends, then rolls the world forward — so checkpoints meet rain, glare, clutter and night shifts before the field does.
The platform · 03
Agentic performance analysis
Agents watch every rollout, grade it step by step and hand you the failure modes — each one backed by the clips that prove it.
Every mode arrives as footage, graded stage by stage — watch the failure, don’t hunt for it.
Coverage
Benchmarks & worlds
One harness for academic benchmarks, classical simulators and generated worlds. Bring your own scenarios — we codify them into evaluation tasks with success rubrics.
Check every checkpoint.
Become a design partnerOnboarding a small cohort now · [email protected] · a founder replies directly
About the team
We’re two brothers who studied maths at Oxford. We’ve spent our careers on the two halves of this problem — video models at the largest scale there is, and the mathematics of making them honest.
FAQ
What is the platform?
A checkpoint gate for robot policies. One harness runs your policy in Atom 1 and in conventional simulators — NVIDIA Isaac Sim, MuJoCo, Cosmos 3 — batched across GPUs and graded automatically, so every checkpoint gets a verdict in minutes instead of waiting days for human evaluations.
Is Atom 1 a simulator?
It’s a world model: a learned simulator trained on trillions of video tokens, made physics-aware and conditioned on robot actions. Where classical simulators need scenes built by hand, Atom 1 generates the world from footage and prompts — and both run side by side in the same harness, scored the same way.
How do I know the physics is real?
We measure it, publicly. Atom 1 is the strongest bare model on our internal physics benchmark — 56.7 against the runner-up’s 48.4, ahead of NVIDIA’s Cosmos 3 — with every case, prompt and score published. The same scorer grades every rollout we return.
What do we need to get started?
A policy checkpoint (or an inference endpoint) and a few example episodes. We wire it into the harness, run the first thousand rollouts and send back the graded report — you don’t stand up any infrastructure.