Hillbot, haosulab/ManiSkill_Demonstrations. Measured for this page from the public release.
33,304 episodes across 16 environments, read from the success and fail flags in the release's own trajectory sidecars. 33,279 are labelled a success, which is 99.9%, and two are labelled a fail. That is a cleaner demonstration corpus than most. The part worth knowing is that 12,731 episodes, 38.2%, run to the step cap and are still labelled a success, so a curation rule that trims long episodes would cut working demonstrations.
| environment | source | episodes | success | at the cap |
|---|---|---|---|---|
| PegInsertionSide-v1 | rl | 1,000 | 97.5% | 100.0% |
| AnymalC-Reach-v1 | rl | 1,022 | 100.0% | 99.7% |
| PickCube-v1 | rl | 1,022 | 100.0% | 99.7% |
| PickCube-v1 | rl | 1,013 | 100.0% | 99.0% |
| TwoRobotStackCube-v1 | rl | 1,007 | 100.0% | 52.8% |
| LiftPegUpright-v1 | rl | 1,015 | 100.0% | 50.0% |
| DrawTriangle-v1 | motionplanning | 1,000 | 100.0% | 0.0% |
| all 36 files | 33,304 | 99.9% | 38.2% |
By source, the reinforcement learning demonstrations are 24,294 episodes at 100.0% success, motion planning is 8,000 at 100.0%, and the one unset file is 1,000 at 97.5%. Teleoperation is 10 episodes.
PegInsertionSide-v1 under reinforcement learning is the one set where both things are true at once. All 1,000 episodes reach the step cap and 25 of them are not successes, so for that environment the cap is doing the terminating and the flag is doing the judging.
Fetch any demos/
Every report on the results page takes a public robot dataset, measures one property of it across every episode, removes the part that scores worst, and retrains on what is left. The corpus figures, the defect rates and the cuts are computed from the releases themselves. The lift column on that page is read from each report's fine tune table rather than from a training run, and the page says so.
The one result that comes off physical hardware is Stera-10M. Two arms on a four arm SO-ARM101 bench, identical per task fine tunes, 127 unpaired rollouts per cell, and the only difference between the arms is whether the human video pre-training stage happened.
This page is the part that can be computed from a published manifest and the labels a release ships. The full grading retrains on the corpus with the worst scoring part removed and reports whether the smaller version scores higher, with the rollout counts and the intervals. That needs the data itself and a conversation about which property is worth measuring on your corpus rather than ours.