LightwheelAI/leisaac-pick-orange is 60 teleoperated demonstrations of picking three oranges. Every episode was scored on how much of the task it completes, how directly it moves, how long it sits still, and how often the gripper closes on nothing. 24 episodes fail at least one of those. Every clip below is on this page, front and wrist stacked, with the measured numbers under it.
| Tag | Rule | Episodes |
|---|---|---|
| INCOMPLETE | Fewer than 3 oranges carried. Cannot teach the scored task. | 5 |
| WANDERS | Path efficiency under 0.50 against a corpus median of 0.739. | 3 |
| IDLE | Over 29% of frames below 2 cm/s. Operator pausing. | 5 |
| REGRIPS | 3 or more gripper closures that carried nothing. | 8 |
| DROPS | 5 or more carries for a 3 orange task, so 2 or more re-picks. | 10 |
An episode can carry more than one tag, so the column sums past 24.
One orange of three, and a path efficiency of 0.056 against a median of 0.739. The clearest failure in the corpus.
35.1% of frames idle and six carries for a three orange task. Two different defects inside one demonstration.
The important one. Path efficiency 0.879, the best in the flagged set, and it looks completely fine. It still carries three regrips and a fourth carry. This is the shot that argues watching demonstrations does not find this.
Ordered by oranges carried, then by path efficiency, so the most damaged demonstration is first.
| # | Ep | Defect | Measured | Path eff | Idle | Carries | Regrips | Sec |
|---|---|---|---|---|---|---|---|---|
| 01 | 28 | INCOMPLETEWANDERS | 1 of 3 oranges carried . path efficiency 0.056 vs median 0.739 | 0.056 | 23.9% | 1 | 1 | 17.6 |
| 02 | 31 | INCOMPLETEWANDERS | 2 of 3 oranges carried . path efficiency 0.199 vs median 0.739 | 0.199 | 18.7% | 2 | 0 | 15.2 |
| 03 | 4 | INCOMPLETEREGRIPS | 2 of 3 oranges carried . 3 closures that carried nothing | 0.505 | 24.2% | 2 | 3 | 20.4 |
| 04 | 5 | INCOMPLETE | 2 of 3 oranges carried | 0.708 | 17.8% | 2 | 2 | 18.4 |
| 05 | 38 | INCOMPLETEREGRIPS | 2 of 3 oranges carried . 3 closures that carried nothing | 0.772 | 26.2% | 2 | 3 | 17.3 |
| 06 | 27 | WANDERS | path efficiency 0.446 vs median 0.739 | 0.446 | 19.8% | 3 | 0 | 19.7 |
| 07 | 20 | IDLE | 29.0% of frames below 2 cm/s | 0.703 | 29.0% | 3 | 1 | 22.9 |
| 08 | 37 | REGRIPS | 3 closures that carried nothing | 0.746 | 27.8% | 3 | 3 | 23.3 |
| 09 | 13 | IDLE | 30.1% of frames below 2 cm/s | 0.772 | 30.1% | 3 | 2 | 20.9 |
| 10 | 16 | IDLE | 32.4% of frames below 2 cm/s | 0.784 | 32.4% | 3 | 2 | 27.4 |
| 11 | 19 | REGRIPS | 3 closures that carried nothing | 0.642 | 27.0% | 4 | 3 | 29.8 |
| 12 | 45 | REGRIPS | 3 closures that carried nothing | 0.663 | 25.4% | 4 | 3 | 30.8 |
| 13 | 49 | IDLE | 30.1% of frames below 2 cm/s | 0.663 | 30.1% | 4 | 2 | 19.6 |
| 14 | 18 | REGRIPS | 3 closures that carried nothing | 0.879 | 28.0% | 4 | 3 | 27.2 |
| 15 | 23 | REGRIPSDROPS | 3 closures that carried nothing . 5 carries for a 3-orange task, so 2 re-picks | 0.587 | 26.4% | 5 | 3 | 27.9 |
| 16 | 29 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.590 | 22.0% | 5 | 0 | 22.7 |
| 17 | 24 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.594 | 26.7% | 5 | 2 | 25.9 |
| 18 | 25 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.634 | 22.4% | 5 | 1 | 17.8 |
| 19 | 47 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.731 | 21.2% | 5 | 0 | 17.8 |
| 20 | 21 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.750 | 25.7% | 5 | 0 | 25.8 |
| 21 | 43 | DROPS | 5 carries for a 3-orange task, so 2 re-picks | 0.760 | 20.2% | 5 | 1 | 23.1 |
| 22 | 46 | REGRIPSDROPS | 3 closures that carried nothing . 5 carries for a 3-orange task, so 2 re-picks | 0.824 | 25.8% | 5 | 3 | 25.7 |
| 23 | 26 | IDLEDROPS | 35.1% of frames below 2 cm/s . 6 carries for a 3-orange task, so 3 re-picks | 0.562 | 35.1% | 6 | 0 | 35.1 |
| 24 | 30 | DROPS | 6 carries for a 3-orange task, so 3 re-picks | 0.658 | 18.3% | 6 | 1 | 28.5 |