Datasets:
`meta/info.json` says 0 episodes, and the camera moves after episode 19
Hi! Two things I noticed while checking this dataset:
meta/info.jsonhastotal_episodes: 0, so the LeRobot dataset visualizer says there are no recorded episodes, even though the videos and data are there. Updating the totals should fix it.- The front camera's position changes after episode 19: episodes 0-19 are filmed closer and from a different angle than the later ones. If that's intentional, great; if not, a policy trained on one framing may struggle on the other.
The checker I used (rdq, open source) also reports that for episodes 0-19 the video segments don't line up with the length of the episode data, which may be related. Happy to share the full report.
A correction to my note above, after looking at both cameras side by side: in episodes 0-19 the two camera streams seem to be swapped. observation.images.front shows the gripper camera and observation.images.back the overhead view, while from episode 20 on it's the other way round. That's what made the "front" view look closer. (Camera devices changing order between recording sessions is a common cause.) Separately, the overhead camera's position changes after episode 39.
If it helps, swapping the two video keys for episodes 0-19 would make all 60 episodes consistent. rdq now checks for this too: rdq check EPITECH-LILLE/eval_training --video-full reports it as video.cameras_swapped.