The gym where robot intelligence gets its reps.
We think the slow part of robotics is the physical world: the room, the robot, the reset between attempts. So we rebuild the room in software, put a model at the controls and record everything that happens. Then we publish the numbers, including the ones we wish were better.
Three habits.
Measured, not narrated
A model saying it is done is a stop request. The simulator reads the state and writes the verdict. We report single recorded trials as single recorded trials and never dress them up as rates.
Limits, declared
Every run states its budget before it starts and shows it beside the result. Physics pauses while a model thinks, and we say so wherever wall time appears.
Evidence you can hold
Every attempt produces the same archive: frames, commands, receipts, provider traffic, trajectory and grade. Failed runs stay in the history. Nothing is pruned to flatter a model.
Two brothers, Oxford maths.
We kept scoring top of the class, then trained neural networks at planetary scale. Now we build RobotGym. We hire in the United Kingdom.
Oxford Maths and Computer Science. Trained TikTok’s recommendation models, serving billions of people. Writes most of the words on this site and reads every message sent to the founders’ address.
Oxford Maths, first in college. British Mathematical Olympiad gold. Built neural networks for the UK Government. Owns the simulator, the harness and the evidence format.
Facts, briefly.
For anyone writing about us, the press kit has the long version, the logo files and product images.
Say hello.One address reaches both of us.
Whether you want to run a model on your bench, write about what we build or work with us.
Or write directly to [email protected]. There is also a Discord.