ExylosAi/bimanual-table-cleanup-cross-embodiment-rich-modality-sample, measured from the public release on
16 September 2026.
episode 3 of the widowxai split, the release's own top camera, all 101 seconds of it. The arms work and the spill goes away. The action column for these frames is zero throughout. Nothing here is sped up or re-rendered.
Every row is read out of the release. Nothing here is estimated and nothing is rounded up.
| frames in episode 3 | 3,037, which is 101.2 seconds at 30 fps |
| value of all twelve arm joints, frame 11 to frame 3036 | 0.0 |
| total range of any joint across the whole episode | 0.0036 radians, about a fifth of a degree |
| the same figure on episodes 0, 1, 2 and 4 | 2.41, 2.54, 2.56 and 3.04 radians |
| observation.state range across the episode | 0.0039 radians, so the state agrees with the action |
| next.success on episode 3 | True |
| video published for episode 3 | six cameras, each 101.23 seconds |
| what the video shows | both arms working and the spill being cleaned, start to finish |
The video and the trajectory cannot both be describing this episode. The six camera video runs the full 101 seconds and shows the task being done. The recorded joint command is the zero vector from frame 11 onward, and observation.state agrees with it, so the log is internally consistent and simply does not contain the take that was filmed. It is still flagged a success, so it passes any filter keyed on that flag and lands in training as 101 seconds of real manipulation video paired with a constant zero action.
Read robots/widowxai/data/chunk-000/episode_000003.parquet and take the min and max of the action column. Then open robots/widowxai/videos/chunk-000/observation.images.top_cam/episode_000003.mp4 and watch it. 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.