Episodes 409 and 829 have no hand_image video; episode 364 repeats 326
Hi, while comparing two dataset checkers on this dataset (at revision 692eff9f) I found two things that look worth fixing:
1. Episodes 409 and 829 have no hand_image footage. Each is the last episode in its video file, and each file ends exactly where that episode should start:
| file | file length | episode starts at | |
|---|---|---|---|
| episode 409 | videos/observation.images.hand_image/chunk-000/file-000.mp4 |
6063.1 s | 6063.1 s |
| episode 829 | videos/observation.images.hand_image/chunk-000/file-001.mp4 |
5876.4 s | 5876.4 s |
So any frame requested for those two episodes falls past the end of the file.
2. Episode 364 repeats episode 326. The camera shows the same bin layout and the same grasp at the same moments, and observation.state matches after a 7-frame offset (mean absolute difference 0.07, versus 0.77-0.80 between 326 and its neighbours). It looks like the same recording cut twice, probably inherited from the original release.
Dropping episodes 364, 409 and 829 (or re-encoding the two missing segments, if the source videos have them) would fix both.
Found with rdq (open source) and checked by hand.