mahathi_bottle_pick

The pick phase of firefly1709/mahathi_bottle, segmented into its own LeRobot v3.0 dataset. Each episode covers the grasp of the bottle (episode start through gripper close).

Task	pick up the bottle
Episodes	17
Frames	920
FPS	15
Robot	franka
Cameras	observation.images.primary, observation.images.wrist (256x256, AV1)

How it was produced

Segment boundaries were marked by hand per episode against the primary and wrist video, guided by the gripper signal and by bottle tilt derived from the unwrapped wrist roll (observation.state.cartesian[3], Euler).

Two corrections were applied relative to the source:

• Video/data frame offset. The source videos carry 2751 leading frames from discarded takes, and source episode 0 declares from_timestamp = 0.000 where the true value is 183.400. The correct mapping is video_frame = dataset_index + 2751, confirmed by correlating wrist-camera motion against end-effector velocity. Frames here are cut on that basis.
• Image statistics. Per-episode image std is 0.0 for every episode in the source, which is a divide-by-zero hazard for mean/std image normalisation. Statistics here are recomputed from the decoded frames.

Source episode 0 is excluded. Episodes are renumbered contiguously; meta/source_mapping.json records the provenance of every episode, reproduced below.

Video is re-encoded (SVT-AV1, CRF 20) because segment boundaries do not fall on keyframes; non-video columns are bit-identical to the corresponding source rows.

episode	source episode	source frames	length
0	1	0–58	59
1	2	0–54	55
2	3	0–59	60
3	4	0–59	60
4	5	0–53	54
5	6	0–53	54
6	7	0–56	57
7	8	0–48	49
8	9	0–52	53
9	10	0–55	56
10	11	0–47	48
11	12	0–56	57
12	13	0–47	48
13	14	0–62	63
14	15	0–44	45
15	16	0–50	51
16	17	0–50	51