Label your robot data
Upload demonstrations and we run them through the pipeline we used to audit the open robotics datasets on our published board. You get your own board, the same viewer as our published boards: a dense timeline of every episode, whether it met its task, and each data problem and operator mistake we find, with the labels to download.
01Recorded with
What was recorded
This sets how the footage is read: which cameras ride on which arm or hand, and what the recorded state means.
02Footage
Your files
A LeRobot dataset folder, MCAP files, or plain videos. We read them here first and show you what will be labelled before anything is sent.
Drop a folder or files
or choose or
What you can upload
- LeRobot dataset
- The dataset folder (meta, data, videos), versions 2.0, 2.1 and 3.0, or a folder of datasets, one per task. Missing metadata is fine: without info.json we read the file layout, and without the episode list we recover the episodes from the data files or label each video whole. Recorded state is used when it holds 6 joints and a gripper per arm, or a 6D pose and opening per gripper.
- MCAP
- One file per episode. The ABC-130k and RealOmin layouts are read with their robot state. Any other layout is read as video from its compressed image or video channels.
- Video files
- MP4, MOV, MKV or WebM. One file per episode, or one folder per episode with up to three cameras; put left or right in the wrist cameras' file names. A file is never split: a long recording is labelled in parts cut at still moments and comes back as one timeline.
- Your annotations
- Optional. A .txt or .json next to a video with the same name, in any format. The model reads it as a claim and tells you where the footage disagrees.
03Send
Label it
Choose what was recorded and add your files.