Подтвердите e-mail

Для публикаций, комментариев, реакций и сообщений подтвердите адрес.

←

Публикация

Claas Voelcker

This is a robot failing to grasp a ball. Almost every robot lab produces clips like this daily… and almost all of them get thrown away. This is the most abundant but underused resource in robot learning. We’re collecting all of it now as “OopsieData”, please join us at oopsie-data.com! (1/13)

Обсуждение

1 прямых ответов · 11 сообщений

Claas Voelcker

Successful demonstrations are the fuel of robot learning today. Labs and companies spend millions collecting diverse and high-quality demos for policy training, and have produced capable policies. So why isn’t this enough? Why do we need failure data? (2/13)

Ответ для Claas Voelcker
Claas Voelcker

Behavioral cloning on successes gives you a policy with no notion of what a bad grasp looks like and how to avoid it, and nothing to learn from once it drifts off-distribution. To learn those, you need failure and suboptimal data. (3/13)

Ответ для Claas Voelcker
Claas Voelcker

Anything that reasons about the quality of behavior — RL, reward modeling, failure detection, world models — benefits from seeing behavior that isn't optimal. However, there is no good existing mixed-quality real-world dataset for researchers to use. (4/13)

Ответ для Claas Voelcker
Claas Voelcker

All of us produce this data constantly but just throw it away. Every policy eval, every play-data collection run, every online RL session generates suboptimal behaviors and failures. The most useful data in robotics is routinely deleted! (5/13)

Ответ для Claas Voelcker
Claas Voelcker

So we're pooling it. We are announcing “OopsieData”, a multi-lab effort to collect real-world manipulation rollouts — suboptimal successes and failures alike — across robots and labs. 17 labs have already contributed, and now it's open to everyone. (6/13)

Ответ для Claas Voelcker
Claas Voelcker

You are already producing this data, so contributing costs almost nothing. Next time you roll out a policy, keep the rollouts and send them our way. Our toolkit records, formats, and uploads in a few lines – and we provide an agent skill so Codex/CC can wire it up for you 😉 (7/13)

Ответ для Claas Voelcker
Claas Voelcker

Significant contributors can qualify for authorship on the dataset release 😉 — but more importantly, you get access to the pooled data from every other lab. You contribute your failures, you get everyone's. No single lab can build this alone. (8/13)

Ответ для Claas Voelcker
Claas Voelcker

Contribution details and the toolkit are all here: oopsie-data.com. We're registering labs now, accepting data contributions till Oct 5! You can either (1) convert existing rollouts to our format or (2) get started in minutes to log new robot rollouts. We have tooling for both! (9/13)

Ответ для Claas Voelcker
Claas Voelcker

This took a lot of effort to come together. Huge credit to my wonderful co-leads Arpit Bahety, Zhiyuan (Paul) Zhou, and @renwang435.bsky.social, the organizing team @maxbrudolph.bsky.social, Maria Attarian, Carl Qi, Sriniket Ambatipudi, Jiahui (Karen) Chen, @marcelhussing.bsky.social, (10/13)

Ответ для Claas Voelcker
Claas Voelcker

and Siddhant Agarwal. And to our faculty advisors Roberto Martín-Martín, Peter Stone, Amy Zhang, and Sergey Levine. Finally, thanks to the contributors who already started submitting data: Jagdeep Bhatia, Andrew Wagenmaker, Kensuke Nakamura, Andrew Zou Li, Yilin Wu, Junwon Seo, Albert Zhan, (11/13)

Ответ для Claas Voelcker
Alex Ene 🦀

was compelling until the ai generated art :)