Office Desk

Open a laptop, type PAIR followed by Enter, and complete the sequence with a left mouse click.

Open / closeSequential pressingPrecision click

Stage 01 / 03

Long environmentDualSharpaOfficeDeskSequence-v1

EmbodimentBimanual Sharpa

Observation382 state dims

Action58 targets

Horizon1400 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 laptop`SharpaAngledOpenLaptop-v1`	Right	29	600	Open the lid beyond 90°, then maintain at least 85° for four control steps.
02	Type PAIR`SharpaAngledTypeKeyboardAllKeys-v1`	Left	29	700	Press and release P, A, I, R and Enter in order using complete key cycles.
03	Click mouse`SharpaAngledClickMouse-v1`	Right	29	500	Depress the left button at least 60% while keeping the right button below 35% for three steps.

Paper results

Office Desk in the benchmark.

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

Stage benchmark: success rates in percent
Stage	PPO	RLPD	ACT	Diffusion Policy	GR00T	π0.5
OpenLaptop	100.0	100.0	100.0	88.0	100.0	82.0
TypeKeys	0.0	0.0	0.0	0.0	32.0	0.0
ClickMouse	96.0	100.0	14.0	92.0	98.0	62.0

Cross-scene results · mean ± scene standard deviation

Cross-scene: success rates in percent
Stage	PPO	RLPD	ACT	Diffusion Policy	GR00T	π0.5
OpenLaptop	99.3 ± 1.2	48.0 ± 50.1	84.7 ± 18.1	88.7 ± 3.1	95.3 ± 8.1	66.7 ± 5.0
TypeKeys	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	17.3 ± 17.0	0.0 ± 0.0
ClickMouse	98.0 ± 3.5	6.7 ± 11.5	12.7 ± 4.6	68.7 ± 15.0	100.0 ± 0.0	64.7 ± 3.1

Complete sequence · end-to-end success (%)

Long-horizon: success rates in percent
Task	PPO	RLPD	ACT	Diffusion Policy	GR00T	π0.5
Office desk	0.0	0.0	0.0	0.0	0.0	0.0

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

Shared contract

Same workstation.
Same predicates.

Stage alignment

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

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.