Droidwright drives Android apps through reusable app skills — parameterised routines for one app's capability, invoked with arguments rather than replayed. For this submission a model receives only the task goal and selects which skill to run and what to pass it; the skills act on the device through the accessibility tree. Every episode below ships its screen recording and its full action trace, so each claim on this page can be checked against what actually ran.
Changes made to the upstream AndroidWorld environment, in full. Task definitions and reward checks were not modified.
null means the bundle did not record the field; it is left as-is rather than filled in.
Every episode ships with its screen recording and its action trace. Click any step to seek the video to that moment; where the target was captured, the element the agent acted on is outlined on the frame. Each row is addressable — the link icon copies a URL that reopens that episode directly.
The archive is the durable record. This site can go away; a hash cannot be edited after
the fact. Verify with shasum -a 256 before relying on it.