nvidia/PhysicalAI-Robotics-Open-H-Embodiment, 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.
| LeRobot roots in the repo | 23 |
| episodes across the release | 40,548 |
| episodes whose task field is a uuid and a frame offset | 29,839 |
| share of the release | 73.6% |
| roots affected | 4 of 23, all under Surgical/cmr_surgical |
| roots carrying real instructions | 19 |
| distinct task strings inside the affected roots | 29,508 |
| distinct task strings per episode there | 0.99 |
| an actual task string | 49c39290-911e-4aec-8178-ee2409ff7854_3600 |
| a task string from a root that is fine | Tie a surgical knot in the suture. |
| source recordings behind the prostatectomy root | 423 |
| chunks cut from each recording | median 18, longest 122 |
| modal episode length in the affected roots | 3,600 frames |
| declared frame rate there | 60, so exactly 60 seconds a chunk |
| offsets that are an exact multiple of 3,600, prostatectomy | 80.4% |
The same data provider's other two roots, dry_box and peg_transfer, carry ordinary labels such as cutting. So this is not a house style and not a convention of the corpus. It is four roots out of twenty three.
In a LeRobot release the task field is the natural language instruction, and it is the string a language conditioned policy reads. For 73.6% of this release that string is a unique identifier that describes nothing and identifies the episode exactly. A model cannot generalise from it, and because it is unique per episode it is available as a memorisation shortcut instead. The loss will look healthy while that happens.
Open meta/episodes.jsonl under any of the four Surgical/cmr_surgical roots and read the tasks field, then open the same file under Surgical/hamlyn/knot_tying and compare. No credentials, no download of the parquet.
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.