Stage alignment
Each stage-level environment uses the matching physical success predicate from the long task.
Open the middle drawer, retrieve a charger, place it on the table, and align both prongs into a socket.
Stage 01 / 03DrawerPlug-v1Canonical instructions, handedness, action width, horizon, and success conditions are drawn from the current task registry and success-contract documentation.
| # | Stage / environment | Hand | Action | Max steps | Success summary |
|---|---|---|---|---|---|
| 01 | Open drawerDrawerPlugOpen-v1 | Left | 29 | 600 | Move the middle drawer by more than 0.18 m and maintain the condition for five steps. |
| 02 | Place chargerDrawerPlugPickPlace-v1 | Right | 29 | 900 | Remove the charger from the drawer and place it stably near the designated table region. |
| 03 | Insert plugDrawerPlugInsert-v1 | Right | 29 | 900 | Insert both prongs 18–31 mm with under 9 mm lateral offset and aligned plug orientation. |
Stage success in the training scene (%, 50 episodes). Best scores in each row are highlighted.
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| OpenDrawer | 98.0 | 100.0 | 98.0 | 98.0 | 98.0 | 98.0 |
| PickCharger | 0.0 | 62.0 | 92.0 | 0.0 | 78.0 | 78.0 |
| InsertPlug | 0.0 | 0.0 | 16.0 | 12.0 | 6.0 | 12.0 |
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| OpenDrawer | 97.3 ± 4.6 | 95.3 ± 4.2 | 82.7 ± 10.1 | 4.0 ± 5.3 | 90.0 ± 10.6 | 100.0 ± 0.0 |
| PickCharger | 0.0 ± 0.0 | 5.3 ± 9.2 | 88.0 ± 5.3 | 0.0 ± 0.0 | 64.7 ± 19.6 | 88.0 ± 0.0 |
| InsertPlug | 0.0 ± 0.0 | 0.0 ± 0.0 | 0.7 ± 1.2 | 0.0 ± 0.0 | 2.0 ± 3.5 | 1.3 ± 1.2 |
| Task | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| Drawer and plug | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
Source: manuscript experiment tables. Compare all methods and tasks →
Each stage-level environment uses the matching physical success predicate from the long task.
Hold counters or latches prevent unstable contact from being counted as task completion.
State and action traces can be restored and replayed for validation and offline rendering.