6 · Review and upload
Goal: know what is in the dataset before you spend GPU hours on it.
scripts/status.sh # episodes per sentence
scripts/review.sh 12 # replay episode 12: both cameras and all joints, in Rerun
scripts/review.sh delete 12 40 # remove bad episodes; the previous version is kept as so101-qc_old
scripts/review.sh push # upload after deleting: training reads the Hub copy, not your disk
Online, without installing anything: open
https://huggingface.co/spaces/lerobot/visualize_dataset and enter fport/so101-qc (you must be logged in for a
private dataset).
What to look for
Watch at least two episodes of every block, and every episode you were unsure about.
- Both videos play, are not black, and show the same moment (the gripper closes in both at once).
- The
actioncurve (leader) and theobservation.statecurve (follower) overlap with a small lag. A flat line is a servo that was not read. - The episode starts moving within about half a second and ends at the start pose.
- No hand, no leader arm in frame.
What is in the folder
outputs/datasets/so101-qc/
meta/info.json fps, features (6 joints in, 6 out, two cameras), episode count
meta/tasks.parquet the 12 sentences
meta/episodes/ which frames and which sentence belong to which episode
meta/stats.json mean/std per feature: used to normalise during training
data/ joint angles and actions, 30 rows per second (parquet)
videos/ one mp4 stream per camera
One frame = observation.images.top, observation.images.wrist, observation.state (where the follower is),
action (where the leader told it to go), and the task sentence. Training is: given the first three and the sentence,
predict the next 50 actions.