Trained in the real world. Running in ours.
A robot policy that learned only from real robot data, dropped into our simulation of the same stations. No fine-tuning, same scoring. Real runs on the left, sim on the right.
Throw six plastic bottles into the bin
| Real | Sim | |
|---|---|---|
| Per run | ||
| Task progress | 93% | 85% |
| Success (6/6) | 78% | 60% |
Task progress: points out of 6 per run (one for each bottle in the bin), averaged over runs.
Load two plates into the dish rack
| Real | Sim | |
|---|---|---|
| Per run | ||
| Task progress | 80% | 50% |
| Per plate | ||
| Picked | 99% | 79% |
| In rack | 94% | 45% |
| Correct | 46% | 27% |
Task progress: points out of 6 per run (each plate: picked, in rack, correct), averaged over runs.
Turn four upside-down mugs right side up
| Real | Sim | |
|---|---|---|
| Per run | ||
| Task progress | 41% | 14% |
| Success (4/4) | 2% | 0% |
Task progress: points out of 4 per run (one for each mug right side up), averaged over runs.