Stage alignment
Each stage-level environment uses the matching physical success predicate from the long task.
Household · 3 stages
Meal Preparation
Open a lunch box, place bread on a plate, then use a spoon to scoop and transfer a tomato.
Stage 01 / 03Long environment
DualSharpaMealPreparation-v1EmbodimentBimanual Sharpa
Observation213 state dims
Action58 targets
Horizon1900 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 box lidAngledMobileBaseSharpaOpenCloseBox-v1 | Left | 29 | 300 | Open the lid to at least 90° with verified left-hand interaction and hold for 20 control steps. |
| 02 | Place breadAngledMobileBaseSharpaPickAndPlaceBread-v1 | Right | 29 | 400 | Set the bread within 0.129 m horizontally and 0–0.07 m above the plate; hold for 20 steps. |
| 03 | Scoop tomatoAngledMobileBaseSharpaScoopTomato-v1 | Right | 29 | 1200 | Use the spoon to scoop and transport the tomato, then settle it on the plate for 20 steps. |
Paper results
Meal Preparation 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 |
|---|---|---|---|---|---|---|
| OpenLunchBox | 74.0 | 100.0 | 100.0 | 96.0 | 100.0 | 100.0 |
| PlaceBread | 12.0 | 86.0 | 88.0 | 20.0 | 60.0 | 70.0 |
| ScoopTomato | 0.0 | 0.0 | 44.0 | 0.0 | 12.0 | 32.0 |
Cross-scene results · mean ± scene standard deviation
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| OpenLunchBox | 35.3 ± 14.5 | 33.3 ± 57.7 | 100.0 ± 0.0 | 35.3 ± 31.6 | 100.0 ± 0.0 | 99.3 ± 1.2 |
| PlaceBread | 10.0 ± 3.5 | 0.0 ± 0.0 | 40.0 ± 30.0 | 8.7 ± 4.6 | 24.0 ± 26.2 | 71.3 ± 7.0 |
| ScoopTomato | 0.0 ± 0.0 | 0.0 ± 0.0 | 23.3 ± 9.0 | 0.0 ± 0.0 | 6.7 ± 6.1 | 18.7 ± 3.1 |
Complete sequence · end-to-end success (%)
| Task | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| Meal preparation | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
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.