Gantry · dataset evaluation · 13 September 2026

one action dimension is constant in 424 of 821 episodes

BeingBeyond/PND_Adam-U_pick-bottle, measured from the public release on 13 September 2026. frame level, every row read

We read every frame of BeingBeyond/PND_Adam-U_pick-bottle, 821 episodes and 252,503 rows. In 424 of them action dimension 29 of 31 holds one value for the whole episode. The release does not name its action columns, so we are not guessing which joint that is. For that share of episodes the policy is learning a constant on that channel.

The release, as it actually is

propertyvalue
episodes821
frames252,503
frame rate20 fps
robotMyDexHand
action dimensions31
cameras
distinct task strings
shortest episodenot declared
median episodenot declared
longest episodenot declared

How each episode ends

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 leastepisodesshare
1 frames , 0.05 seconds00.0%
5 frames , 0.25 seconds00.0%
10 frames , 0.5 seconds00.0%
30 frames , 1.5 seconds00.0%
60 frames , 3.0 seconds00.0%

How to check this without taking our word

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.

What a full grading adds

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.