The Proving
Ground
for Physical AI
Put your model to the test on real robots deployed in-the-wild.
In-the-wild
evaluations at scale
Apply for an eval
We operate one of the largest real-world robot benchmarks, with hundreds of robots deployed across 140+ scenes.
Evaluation tracks
Open track
Two unnamed models try the same new instruction in the same scene. The evaluator picks the better run, or calls a tie.
Fine-tune track
Every model is fine-tuned on a fixed task from the same 100 demonstrations, then makes 25 attempts.
You bring the model.
We handle the rest.
From robots to scoring, scenes, and operators, we handle everything to evaluate your model.
Apply now- 001
Apply
Tell us about your policy and the robots you want results on.
- 002
Connect
We help you link your inference server through our policy API, or serve your checkpoint on a machine next to the robot.
- 003
Evaluate
Operators set up each scene and run your model in Open track comparisons and Fine-tune track trials, recording every camera.
- 004
Track
See where your model ranks against top labs on that robot's leaderboard. Every run, failures included, is published on Hugging Face.
Active
embodiments
5 active
PoliciesFrom top labs
All instructions written by an evaluator on the spot, with more added daily.
Evaluation viewer
Launch viewerPreview the evaluation viewer. Watch a real blind comparison, both runs in sync, with the evaluator's pick and notes.
Data samples
Fine-tune datasets
Explore every recorded eval trialWe publish everything: training demonstrations and every evaluation run, failed attempts included, on Hugging Face in LeRobot format. Track scores or train on the data.
Get 10,000 hours
of free evaluation
You bring the policy. We put it to the test
on real robots, in the wild.
FAQ
Frequently asked
Who can apply?
Teams with a policy for one of our robots. The 10,000 free hours are allocated by the RoboColosseum team, who review every application and follow up with next steps.
What happens after I apply?
We read your application and get in touch to plan the evaluation with you: which robots and tasks, and how your model will connect. Then the runs start and results appear as they are scored.
Which robots are available?
Unitree G1 humanoids, bimanual YAM stations, Franka arms in the DROID setup and SO-ARM101 arms. Earth Rover Mini+ wheeled robots are being added. Your policy has to take the robot's camera inputs and produce its action format.
How do I connect my model?
Two ways. Run your own inference server and connect it with our policy API, and the robot streams observations to it during each run. Or hand over a checkpoint and we serve it on a machine next to the robot. We help you set up either one.
How is a model scored?
Open track: two unnamed models try the same fresh instruction and the evaluator picks the better run or calls a tie. Ratings are fitted from all comparisons on that robot and shown with their uncertainty. Fine-tune track: every model is fine-tuned on the same 100 demonstrations, then makes 25 attempts, and we report full successes and milestone progress.
How is it kept fair?
Evaluators see trial A and trial B, never the model names. Both models in a comparison face the same robot, scene and instruction, one after the other. Every fine-tuned model trains on identical demonstrations.
What gets published, and where?
Every run from both kinds of evaluation, failures included, with its videos, in LeRobot format on Hugging Face. Scores go on the public leaderboard.