NONHUMAN-RESEARCH/tblock4, measured from the public release on 13 September 2026. frame level, every row read
Nobody opens a robot dataset and watches all of it. We read every frame of every episode and report what is in there.
The release's own footage for one of the 56 episodes we flagged. The red block is carried onto the blue board and the task is completed, and left_gripper_width holds one value from the first frame to the last while the right gripper works. This is a hand-held UMI rig rather than arms, so the left device is carried through the episode without ever being squeezed. Whether that is intended is the question worth asking, because 56 of the 122 episodes are effectively one-handed.
episode 3 of NONHUMAN-RESEARCH/tblock4, cut from the release's own video, 8 seconds. Nothing is re-rendered or sped up.
We read every frame of NONHUMAN-RESEARCH/tblock4, 122 episodes and 32,817 rows, and in 56 of them left_gripper_width holds one value from the first frame to the last while the right gripper actuates in every episode. On a hand-held UMI rig that is exactly what a one-handed take looks like rather than a fault, so the question is whether 46% of the release is meant to read as bimanual training data.
| property | value |
|---|---|
| episodes | 122 |
| frames | 32,817 |
| frame rate | 30 fps |
| robot | handumi_raw |
| action dimensions | 16 |
| cameras | 3 |
| distinct task strings | 1 |
| shortest episode | 41 frames |
| median episode | 254 frames |
| longest episode | 737 frames |
A trailing hold is the number of frames at the end during which the commanded action does not change by more than 1e-4 on any component. Seconds come from the frame rate the release declares.
| trailing hold at least | episodes | share |
|---|---|---|
| 1 frames , 0.03 seconds | 0 | 0.0% |
| 5 frames , 0.17 seconds | 0 | 0.0% |
| 10 frames , 0.33 seconds | 0 | 0.0% |
| 30 frames , 1.0 seconds | 0 | 0.0% |
| 60 frames , 2.0 seconds | 0 | 0.0% |
The release is a meta directory plus one or more parquet files. Read data/chunk-000/file-000.parquet, group rows by the episode ranges in meta/episodes, and compute the same columns. No credentials.
This page is what a published release will tell you on its own. The full grading removes the part that scores worst and retrains on what is left, then reports whether the smaller corpus scores higher, with rollout counts and intervals. The one result measured on physical hardware is Stera-10M, two arms on a four arm SO-ARM101 bench, 21.9% against 36.9% task success across 1,270 rollouts.