Stage alignment
Each stage-level environment uses the matching physical success predicate from the long task.
Industrial · 3 stages
Furniture Assembly
Lay a panel flat, slide a rail into its groove, and fasten it by turning a wing screw by hand.
Stage 01 / 03Long environment
DualSharpaFurnitureAssemblyLong-v1EmbodimentBimanual Sharpa
Observation177 state dims
Action58 targets
Horizon2400 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 | Lay panel flatSharpaAngledMobilePanelLayFlatLeft-v1 | Left | 29 | 600 | Place the panel groove-up, inside the target footprint, released and stable for five steps. |
| 02 | Insert railSharpaAngledMobileRailInsert-v1 | Right | 29 | 700 | Seat the rail within 2 mm depth, 4 mm lateral and 8 mm wall tolerances; hold for five steps. |
| 03 | Fasten wing screwSharpaAngledMobileScrewFasten-v1 | Right | 29 | 900 | Turn the wing screw clockwise to at least 87.1°; no screwdriver is used. |
Paper results
Furniture Assembly 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 |
|---|---|---|---|---|---|---|
| LayFlatPanel | 0.0 | 96.0 | 78.0 | 24.0 | 64.0 | 0.0 |
| InsertRail | 0.0 | 0.0 | 22.0 | 2.0 | 16.0 | 8.0 |
| FastenScrew | 100.0 | 100.0 | 0.0 | 0.0 | 84.0 | 94.0 |
Cross-scene results · mean ± scene standard deviation
| Stage | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| LayFlatPanel | 0.0 ± 0.0 | 0.0 ± 0.0 | 54.0 ± 27.5 | 0.0 ± 0.0 | 34.7 ± 28.4 | 0.0 ± 0.0 |
| InsertRail | 0.0 ± 0.0 | 0.0 ± 0.0 | 0.0 ± 0.0 | 0.0 ± 0.0 | 6.0 ± 10.4 | 9.3 ± 3.1 |
| FastenScrew | 86.0 ± 24.2 | 70.7 ± 50.8 | 85.3 ± 11.5 | 0.0 ± 0.0 | 20.0 ± 26.2 | 94.0 ± 3.5 |
Complete sequence · end-to-end success (%)
| Task | PPO | RLPD | ACT | Diffusion Policy | GR00T | π0.5 |
|---|---|---|---|---|---|---|
| Furniture assembly | 0.0 | 0.0 | 8.0 | 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.