StationeryBench

Five bimanual desk-stationery manipulation tasks for VLA models and LLM agents, built on Inspect Robots. Each task supplies a fixed setup, an instruction, and a five-stage rubric. The tasks run with compatible policy and embodiment pairs, including the bundled abstract mock and real YAM arms.

For installation and the mock quick start, see the README. For hardware, follow Running on YAM arms.

The tasks

Task (--task)	Goal	Objects	Report slug
stationerybench/uncap_marker	Uncap the marker	black Sharpie fine-point permanent marker, capped	pen
stationerybench/eraser_from_box	Open the box, take out an eraser, close the lid	cloth-covered book-style box with a hinged lift lid (about 30 x 22 x 7 cm); one block eraser in its paper sleeve	lid
stationerybench/middle_sticky_pad	Pull the middle sticky pad from a stack	three shrink-wrapped 3 x 3 inch sticky-note cubes (Post-it Super Sticky notes cube or equivalent)	sticky
stationerybench/pour_paperclips	Pour paper clips into a lifted bowl	one light ceramic bowl about 15 cm across; an open cardboard box of steel paper clips, roughly half full	paperclip
stationerybench/ruler_handover	Hand a ruler between arms, cup to cup	two paper cups (about 12 oz); one clear plastic 12 inch (30 cm) ruler	ruler

Layout convention

The operator faces the bench, with the two arm bases at the far edge. Left and right are the operator's left and right. Objects go within reach of both arms unless a setup line says otherwise.

Grading

Each task uses one fixed setup, repeated for 20 trials by default. The mean of the binary task_success scores is the completion rate. The default horizon is 120 seconds (1200 steps at the YAM control rate of 10 Hz).

Rubric stage 4 means success: the instruction is fully satisfied. On hardware, the rubric is shown to the operator before the verdict prompt. Answer y for success and n for failure; partial counts as failure. The scorer accepts a success termination or an affirmative recorded operator verdict. In the mock, the scripted policy produces the privileged success termination.

Rubric principles

• 0 - nothing purposeful, 1 - deliberate contact with the target object(s), 2 - a secure grasp, 3 - the task's core manipulation achieved, 4 - the instruction fully satisfied.
• Every stage is a precondition of the next, and is a state that persists long enough to be judged from the video.
• A completed trial means stage 4, completion rate is the share of trials at stage 4. Mean stage summarises partial progress.