Leaderboard / stanford hydra dataset
stanford hydra dataset
Camera dropoutShare of frames from live episode cameras that are almost entirely one flat black or white value. A camera blank for a whole episode is counted as absent, shown underneath, not as a dropout. Method →
<0.1%
1 frames
BlurShare of camera frames inside an interval at least 0.5 s long where the image is much blurrier than that camera's typical frame. Method →
2.8%
20,199 frames
DarknessShare of camera frames inside an interval at least 0.5 s long where 60% or more of the pixels are near-black. Method →
<0.1%
6 frames
FreezeShare of camera frames inside an interval at least 1 s long where the video repeats the exact same image. Method →
0.6%
3,982 frames
Duration mismatchShare of assessed episode-camera pairs whose metadata video length disagrees with the Parquet frame count by more than 1 frame. A hatched cell with * means only some pairs could be assessed (the small number shows how many); the rest are unknown, not clean. –* means none could be assessed. Method →
–*
0 of 1,140 assessed
Flagged episodesEpisodes flagged by any episode check: actions that never change, a robot that barely moves, non-finite values, broken row indexes or timestamp gaps. The count is on top, its share of assessed episodes underneath. Method →
0
of 570 assessed episodes
Video offsets in this dataset are inconsistent, so video intervals can’t be assigned to episodes. Counts still cover every decoded frame.
All checks
| Check | Flagged | Share | Coverage |
|---|---|---|---|
| Files 7 | |||
| Video decoding | 0 | 0 | complete |
| Episodes 570 | |||
| Episode boundaries | 0 | 0 | complete |
| Finite action and state | 0 | 0 | complete |
| Timestamp gaps | 0 | 0 | complete |
| Static actions | 0 | 0 | complete |
| Idle episodes | 0 | 0 | complete |
| Episode-camera pairs 1,140 | |||
| Duration mismatch | 0 | – | 0 assessed |
| Video/action sync | – | – | not measured |
| Frames 716,468 | |||
| Camera absent | 0 | 0 | complete |
| Camera dropout | 1 | <0.1% | complete |
| Blur | 20,1991,118 intervals | 2.8% | complete · episodes can’t be placed |
| Darkness | 61 interval | <0.1% | complete · episodes can’t be placed |
| Freeze | 3,9826 intervals | 0.6% | complete · episodes can’t be placed |
By camera
| Camera | Frames | Absent | Dropout | Blur | Darkness | Freeze |
|---|---|---|---|---|---|---|
| image | 358,234 | 0 | 0 | 0 | 0 | 0 |
| wrist_image | 358,234 | 0 | <0.1% | 5.6%1,118 intervals | <0.1%1 interval | 1.1%6 intervals |
Evidence
Longest video intervals up to 4 per signal · times are within the video file
| Signal | Camera | Episode | File | Start | End | Duration | |
|---|---|---|---|---|---|---|---|
| Blur | wrist_image | – | file-000.mp4 | 147:12.6 | 147:32.5 | 19.9 s | Inspect → |
| Blur | wrist_image | – | file-000.mp4 | 107:08.2 | 107:25.1 | 16.9 s | Inspect → |
| Blur | wrist_image | – | file-000.mp4 | 180:13.8 | 180:30.2 | 16.4 s | Inspect → |
| Blur | wrist_image | – | file-001.mp4 | 140:32.4 | 140:47.9 | 15.5 s | Inspect → |
| Darkness | wrist_image | – | file-000.mp4 | 186:54.5 | 186:55.0 | 0.5 s | Inspect → |
| Freeze | wrist_image | – | file-000.mp4 | 38:00.7 | 39:15.7 | 1:15.0 | Inspect → |
| Freeze | wrist_image | – | file-002.mp4 | 3:31.4 | 4:43.8 | 1:12.4 | Inspect → |
| Freeze | wrist_image | – | file-002.mp4 | 62:30.0 | 63:41.6 | 1:11.6 | Inspect → |
| Freeze | wrist_image | – | file-000.mp4 | 230:10.5 | 231:19.7 | 1:09.2 | Inspect → |
| Camera dropout | wrist_image | – | file-001.mp4 | 69:05.7 | 69:05.7 | 0.0 s | Inspect → |
Provenance
- Revision
- ff06383708a1df7f38c21ec960a286ca6d0b7e7e (main)
- Scanned
- 2026-10-10 00:34 UTC · NVIDIA L4
- Flag method
- rule-based-video-events-v1
- Episode method
- episode-motion-v1
- Code
- 9f54cd5fa0437e2c lock 282482d5a82a
Pinned dataset on Hugging Face ↗Result JSONMethodologyReport an issue with this result →