We are a frontier robotics lab. We build the world models and robot models to solve general robotics.
Solve Robotics. Then Solve Abundance.
Two theories drive the lab. One: world models built on video models drastically cut the robot data you need — the physics is already in internet video. Two: LLMs will play a major role in solving robotics — they can plan, sequence and recover, but they can't move things. So we build robot models designed to be micromanaged by an LLM: it plans and retries, our model owns contact, slip and timing at 50 Hz. Models that move atoms.
Our ResultsThe lab · 01 · Atom 1
First, learn physics from video.
Theory one, tested. Start from the strongest physics backbone in open video — MiniMax H3 — and train it into Atom 1: a world model that watches a scene and rolls it forward. The physics arrives free from internet video instead of a teleop fleet, and on DeepMind’s Physics-IQ, Atom 1 now sits above every video model measured, including its own backbone.
The lab · 02 · Atom 2 · with our first design partner
Then, condition the world on your robot’s actions.
Atom 2 watches what a robot sees and the controls it sends, then rolls the world forward under those actions. With our first design partner — an independent robot lab — we run the same policy on the filmed robot and inside the dream. It earns the same score, milestone for milestone.
Same partial credit, same lost milestone — the dream grades like reality does. Every rollout comes back as a typed verdict: reach, grasp, transfer, place — scored stage by stage.
Running now with our first design-partner cohort — become a design partner →
The lab · 03 · The World Action Model
Next: the robot model your LLM micromanages.
Theory two says the planner already exists: LLMs plan, sequence and recover — they just can’t move things, and today’s end-to-end robot models take no direction mid-task. Our first major architecture closes that gap: micro-control-labelled robot data turns Atom 2 into the World Action Model, a robot model an LLM can actually drive. Built to top RoboArena.
reach(mug marked in frame) · grasp(marked mug at handle) — an interface precise enough to micromanage.reach(mug marked in frame)grasp(marked mug at handle)The lab · 04 · People
Brothers, Oxford maths, TikTok ML.
Two brothers who kept scoring top of the class, then trained neural networks at planetary scale. We’re hiring the smartest people in the UK — if solving robotics is what you want your decade to be about, write to us.
together!