Datasets:

nepyope
/

can_clean_final_30fps

This is can_clean_final resampled from 50 fps to 30 fps. Nothing else was changed: same 105 episodes, same task, same review, same observation.state and action values on the frames that were kept.


episodes	105
frames	127,395 (was 212,290)
fps	30 (was 50)
duration	70.8 min (unchanged)
codebase version	v3.0
task	`Bring the can to the white table`

Schema

feature	dtype	shape	contents
`observation.state`	float32	[31]	29 body joints in `G1_29_JointIndex` order, then left/right gripper
`action`	float32	[66]	64-D SONIC motion token, then left/right gripper
`observation.images.ego_view`	video	[480, 640, 3]	head camera
`observation.images.left_wrist`	video	[480, 640, 3]	left wrist camera
`observation.images.right_wrist`	video	[480, 640, 3]	right wrist camera

How the resampling was done

Decimation, not interpolation. For each episode independently, source frame k was kept when

k = round(i * 50 / 30)    for i = 0, 1, 2, ...

which is exactly k % 5 in {0, 2, 3} — three of every five frames, so every episode keeps 60% of its length and wall-clock duration is preserved. Because the pattern restarts at every episode, frame 0 of each episode is always kept and timestamp == frame_index / 30 holds exactly.

observation.state and action are copied bit-for-bit from the kept frames; no averaging or filtering was applied. timestamp, frame_index and index were renumbered for 30 fps.

Videos were re-encoded (dropping frames requires it) with h264 / yuv420p / crf=23 / g=30. Measured against the exact source frames they replace, the result is 43.2 dB mean PSNR (33.9 dB worst case over a 216-frame sample), i.e. visually transparent relative to an already compressed source.

One incidental cleanup: in the source, each episode's video held length + 1 frames because the recording ran one frame longer than the action sequence. Here each episode's video holds exactly length frames, so videos and parquet rows now line up one-to-one.

Caveat: `action` no longer chains into the next state

In the source, action[t] is the command that produces observation.state[t+1]. That relationship does not survive decimation — the next kept frame is 2 or 3 source steps ahead, so action[t] here produces a state 1/50 s later that is no longer the row that follows it. The actions are still the correct commands for the observations they sit next to, which is what an observation-to-action policy consumes, but do not treat consecutive rows as a 30 Hz dynamics model. If you need that, resample from the source with a proper action-chunk aggregation instead.

Review

Inherited from the source dataset. All 124 episodes of can_to_martino_2 and can_to_martino_3 were watched and marked pass/fail with a 1-10 quality rating. 19 failed and were removed:

cause	episodes
failed to pick up the can	6
dropped the can	5
camera blocked the view or fell	5
never attempted the pick	2
took over a minute	1

Of the 105 that remain, 82 carry a rating: mean 7.2, with 3 rated 4 or below. The camera failures are a rig problem rather than a teleoperation one — the head camera physically obstructed the grasp or fell mid-episode — so they cluster in time rather than being spread evenly.

Known issue: the left gripper is dead

Carried over from the source. observation.state[29] and action[64] are identically 0.0 in every frame — the episodes were effectively recorded one-handed. Their q01/q99 in meta/stats.json are set by hand to 0/1 so quantile normalisation does not divide by zero. A policy trained on this will never open or close the left hand. The right gripper is healthy, closed in 41% of frames. No other channel is degenerate: all 64 token dimensions and all 29 joints vary.