OopsieData hits 15k robot episodes from 54 labs
The shared pool of real-world manipulation rollouts keeps the messy runs most datasets throw away.
OopsieData, a multi-lab effort to pool real-world robot manipulation rollouts (recordings of a robot attempting a task), now holds more than 15,000 episodes, about 140 hours. Paul Zhou posted the update, saying that as of last week the project had 54 signed up labs and 10,000+ contributed episodes, and this week it is past 15,000. Signups are still open.
As of last week we had 54 signed up labs with 10k+ contributed episodes, and as of this week we have 15k+ episodes (140h).
The thing that makes the pool different is what it keeps. When Zhou announced the project, he described it as collecting rollouts "suboptimal successes and failures alike" across robots and labs, starting with 17 contributing labs before opening to everyone. Most public robot datasets are curated down to the clean runs, where the gripper closes on the right object and the task ends the way the script says it should. The failed grasps, the near misses and the awkward recoveries usually never leave the lab.
suboptimal successes and failures alike
Why the messy runs matter
A policy trained only on perfect demonstrations has never seen what going wrong looks like, so it has no idea what to do when it gets there. Failure data is also the raw material for things like reward models and evaluation, which need examples of bad behavior to learn what bad means. Nobody has to stage anything to produce it either, which is the appeal of the multi-lab format. The data already exists on somebody's drive.
Shown versus claimed
This is a dataset count, not a result. No model trained on OopsieData has been published in the sources, and there is no breakdown yet of which robots or which labs the 15,000 episodes come from. The scale is also modest next to the fleets the large humanoid companies run internally, where 140 hours is not a lot of robot time.
What the number does show is that the contributors are still logging. Going from 10,000 to 15,000 episodes in a week is the part worth watching, because shared data efforts usually stall after the launch post.