FlywheelAI/excavator-dataset, 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.
| tracking csv files | 176 |
| csv files whose filename column names a published mp4 | 0 |
| csv files whose filename column names a missing mp4 | 176 |
| distinct missing videos referenced | 51 |
| camera angles actually published | D01, D02, D03, D04 |
| camera the tracking belongs to | D05, not published |
| rows of tracking read | 4,444,628 |
| recordings where every marker is frozen for the whole file | 22 |
| longest fully frozen recording | 8 minutes 45 seconds, 26,278 rows |
| rows inside fully frozen recordings | 187,848 |
The marker tracks are the only machine readable signal in the release, and there is no published video they can be aligned to. Anyone training on this has to either drop the tracking or guess at the correspondence, and a wrong guess is silent.
List the mp4 files in the repo, then read the filename column of any csv under a recording directory and look for that name. 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.