September 10, 2026StationeryBench is five bimanual manipulation tasks built from everyday desk stationery: a marker, a ruler, a lidded box, a stack of sticky-note pads, and a box of paper clips. We gave OpenAI's GPT-6 Astra, driven by the Inspect Robots agent policy, and Ai2's MolmoAct2, a vision-language-action model, control of the same YAM arms on all 5 tasks, 200 trials in total.Uncap the markerPour paper clips into a lifted bowlHand a ruler between arms, cup to cupPull the middle sticky note from a stackOpen the box, take out an eraser, close the lidUncap the marker: GPT-6 Astra (trial 8, completed) above MolmoAct2 (trial 2, progress 50 of 100), seen from the robot's cameras: left wrist, top, right wrist. Both runs play at 2.4× real time (GPT-6 Astra with its thinking pauses cut out) until 1 s after GPT-6 Astra finishes; MolmoAct2's remaining 108 s then fast-forward at 27×, marked by the centre chip. Timers show real time since each run's first camera frame, and a jump forward is a thinking pause that was removed.Astra completed 7 of 100 trials across StationeryBench, MolmoAct2 completed none
2026-09-11T00:45:25.806908
image/svg+xml
Matplotlib v3.11.1, https://matplotlib.org/
MolmoAct2
