DexArenaDexterous intelligence
Tasks/Package Opening
Household · 2 stages

Package Opening

Open both package flaps with two hands, then extract the apple and place it stably on the plate.

Bimanual openingGrasp / extractStable placement
Package OpeningStage 01 / 02
Long environmentSharpaPackageOpeningStandalone-v1
EmbodimentBimanual Sharpa
Observation134 state dims
Action58 targets
Horizon800 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 / environmentHandActionMax stepsSuccess summary
01Open packageSharpaPackageOpeningOpenPackage-v1Bimanual58500Rotate both lid panels more than 90° from their closed poses.
02Place appleSharpaPackageOpeningExtractAndPlaceApple-v1Right29700Lift the apple clear of the box and settle it on the plate, released, for 40 steps.
Paper results

Package Opening in the benchmark.

Stage success in the training scene (%, 50 episodes). Best scores in each row are highlighted.

Stage benchmark: success rates in percent
StagePPORLPDACTDiffusion PolicyGR00Tπ0.5
OpenPackage2.092.044.02.064.058.0
ExtractApple0.042.088.022.0100.082.0
Cross-scene results · mean ± scene standard deviation
Cross-scene: success rates in percent
StagePPORLPDACTDiffusion PolicyGR00Tπ0.5
OpenPackage2.0 ± 0.02.7 ± 4.636.7 ± 22.11.3 ± 2.335.3 ± 22.130.7 ± 24.8
ExtractApple0.0 ± 0.00.0 ± 0.078.7 ± 3.13.3 ± 3.154.7 ± 39.570.0 ± 26.5
Complete sequence · end-to-end success (%)
Long-horizon: success rates in percent
TaskPPORLPDACTDiffusion PolicyGR00Tπ0.5
Package opening0.00.04.00.00.014.0

Source: manuscript experiment tables. Compare all methods and tasks →

Shared contract

Same workstation.
Same predicates.

01

Stage alignment

Each stage-level environment uses the matching physical success predicate from the long task.

02

Persistence

Hold counters or latches prevent unstable contact from being counted as task completion.

03

Replay

State and action traces can be restored and replayed for validation and offline rendering.