Stage alignment
Each stage-level environment uses the matching physical success predicate from the long task.
Household · 4 stages
Cabinet & Pouring
Open a cabinet, retrieve and place a bottle, unscrew and place its cap, then pour the contents into a bowl.
Stage 01 / 04Long environment
CabinetPour-v1EmbodimentBimanual Sharpa
Observation178 state dims
Action58 targets
Horizon3400 steps
Stage definitions
Exact slices of the parent sequence.
Canonical 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 cabinetCabinetPourDoorOpen-v1 | Left | 29 | 600 | Open the cabinet door to at least 80° and hold for five control steps. |
| 02 | Place bottleCabinetPourBottlePickPlace-v1 | Right | 29 | 900 | Remove the bottle, place it upright inside the target region, release it, and hold stable for five steps. |
| 03 | Remove capCabinetPourCapRemovePlace-v1 | Right | 29 | 1400 | Rotate the cap by at least 180°, detach it, and place it flat and released on the table. |
| 04 | Pour contentsCabinetPourPour-v1 | Right | 29 | 700 | Place at least 12 of 16 particles in the bowl, spill at most two, and hold for ten steps. |
Paper results
Cabinet & Pouring in the benchmark.
Stage success in the training scene (%, 50 episodes). Best scores in each row are highlighted.
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| OpenCabinet | 62.0 | 100.0 | 94.0 | 88.0 | 92.0 | 98.0 |
| PickBottle | 0.0 | 14.0 | 68.0 | 18.0 | 78.0 | 54.0 |
| UnscrewCap | 0.0 | 2.0 | 62.0 | 50.0 | 62.0 | 46.0 |
| PourBeads | 0.0 | 96.0 | 100.0 | 32.0 | 78.0 | 52.0 |
Cross-scene results · mean ± scene standard deviation
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| OpenCabinet | 8.7 ± 9.0 | 63.3 ± 54.9 | 93.3 ± 6.4 | 84.0 ± 11.1 | 78.7 ± 17.9 | 97.3 ± 2.3 |
| PickBottle | 0.0 ± 0.0 | 0.0 ± 0.0 | 76.0 ± 8.0 | 4.0 ± 5.3 | 46.0 ± 36.7 | 69.3 ± 6.4 |
| UnscrewCap | 0.0 ± 0.0 | 0.0 ± 0.0 | 4.0 ± 3.5 | 1.3 ± 2.3 | 9.3 ± 14.5 | 33.3 ± 11.0 |
| PourBeads | 0.7 ± 1.2 | 0.0 ± 0.0 | 42.0 ± 22.3 | 6.0 ± 10.4 | 16.7 ± 17.5 | 36.0 ± 5.3 |
Complete sequence · end-to-end success (%)
| Task | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| Cabinet and pouring | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 2.0 |
Source: manuscript experiment tables. Compare all methods and tasks →
Shared contractSame workstation.
Same workstation.
Same predicates.
Persistence
Hold counters or latches prevent unstable contact from being counted as task completion.
Replay
State and action traces can be restored and replayed for validation and offline rendering.