← sources

Built an open source ecosystem to label factory videos for robot training. Now takes me 5 mins per clip now instead of 40-50mins raw

source: reddit (reddit:robotics)published 2026-10-20 07:28ingested 2026-10-21 02:40

retained text

Hi everyone,

I'm an AI engineer. Recently I have been working on labelling egocentric (head camera) videos for robot learning and training. I took Hf Egocentric-10K dataset and picked one worker who packs shirts into polybags. To use these videos for imitation learning, every step has to be marked: take the shirt, pick up the polybag, put the shirt inside, fold, tape, and place it on the stack. Each step needs a start and an end.



Doing this by hand was very slow. One 5 min clip was taking me 40-50 mins.

So I built roboreview - a whole open source ecosystem to do this all. A small model first predicts all the steps, and then I just check them in a review app (the video above). Most steps I approve with one key. For the wrong ones I move the start or end a bit. The approved steps get exported as training data, and the model is retrained on my reviews, so it keeps getting better.

On a clip the model had never seen, it got almost 93% of the frames right and found all 12 shirts. Reviewing that clip took me 5 mins. The model is trained on only 30 mins of video , which also includes a lot of not-working time. 

To be fair, it is one worker and one task for now, and I am the only one who reviewed the labels. So the numbers need more testing.

Everything runs on my laptop (M1 Pro) and the code is open source: [https://github.com/abhishek0093/roboreview](https://github.com/abhishek0093/roboreview)

If you work with robot data, I would really like to know:

\-  what format you need these labels in (LeRobot or something else)

\- which task I should try next



Do check it out and let me know your feedback. Thanks!

document fetched 2026-10-21 00:37 · canonical https://reddit.com/r/robotics/comments/1xalost/built_an_open_source_ecosystem_to_label_factory