Grably/usa-egocentric, measured from the public release on
16 September 2026.
Every row is read out of the release. Nothing here is estimated and nothing is rounded up.
| segments in the release | 663 |
| segments whose end frame is not after their start frame | 561 |
| the same segments by the release's own duration_sec column | 561 |
| rows where the two measures disagree | 0 |
| share of the corpus | 84.6% |
| longest segment in the whole release | 0.84 seconds |
| most common non zero duration | 0.067 seconds, two frames |
| segments carrying confidence 0.333 | 561, the same rows |
| segments with the literal string nan in hazards | 498, 75.1% |
| columns null in every row | category, domain, video_context |
| episodes in the release | 3 |
A VLA training pair built from a zero length segment has an instruction and one frame under it. That is a caption. The model learns to name the action rather than to produce it, and the loss will look fine while it does.
Read data/episodes/<episode>/annotations/vla_action_segments.parquet and compare start_frame with end_frame, or just read duration_sec. No credentials.
This page is what the release tells you on its own. A full grading removes the part that scores worst, retrains on what is left, and 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. The Stera-10M report.