CustomersRobotGym is in its first release. We work closely with a small number of design partners and publish only what we can measure. Here is who the platform is for, what a partnership includes and the runs we can show in public.
We would rather show one honest recorded run than a wall of logos.
Embodied reasoning you can measure.
Run a task you designed, on a scene you control, with budgets you declared. Frozen scenes, seeds and limits. Independent grading. Keep every frame and command for the paper, and hand reviewers the archive.
The customer’s bench, before the hardware arrives.
Rebuild a customer workstation from the photographs they already sent you. Place the arm where it would stand, write the task in a sentence and run it in the twin. What you learn about reach, camera placement and the order of operations arrives weeks before the robot does.
One harness for every model.
Frontier models are evaluated on text and code every day and on physical tasks almost never. RobotGym gives them a robot, a room and a fixed set of primitives, then records the whole exchange. The harness is the same for every provider, so differences belong to the model.
A cell you can describe in a sentence.
The factory starter has a moving conveyor, loose parts, an inspection tray and an arm. You write the task. Run it, read the decision notes and see where the twin’s idea of the cell differs from yours. Then build the real cell from your own photographs and drawings.
A robotics lab that fits in a browser.
Students can build a scene from a photograph of their own desk, run a task in it and read exactly why it failed. The evidence archive is a complete lab notebook. Welcome credits cover a first project.
We take on a handful of design partners at a time. You get the two people who wrote the platform. We get real tasks and honest feedback. Nothing about your work is published without your written consent.
You describe the task and what you have to build it from. We say plainly whether RobotGym can do it today.
We rebuild the workspace with you from your material and agree the success check before anyone runs a model.
You run models against the scene. We review the evidence together and change what needs changing.
Keep going on a plan, take the scenes and archives with you, or both. Nothing is locked in.
Public environment · LogisticsA moving line. Three returns. An agent that has to figure it out.
Watch the run
Public environment · Wet labThin labware. A separate retaining frame. A 90° change of orientation. Just 0.5 mm clearance on each side of the imaging nest.
Watch the run
Public example · Precision electronicsSmall holes. Real collision geometry. Pick the populated board, align its mounting holes, engage the locating pins, release and inspect.
Watch the runSay what the robot needs to do, what you have to build the scene from and how you would judge success.
Or write directly to [email protected]. There is also a Discord.