Mahal Systems AndroidControl
Fits teams training mobile action models; borrow the paired high-level and low-level instruction per step and the in-domain / out-of-domain split.
What it asks the agent to do
- predict the next action on an Android screen given a high- or low-level instruction
- generalize to apps unseen in training
What to borrow from it
case templatetaxonomy
Caveats
Offline step prediction only; label errors reported (a corrected AndroidControl-Curated exists); recorded on real apps so screens may contain personal data.
Details
| Workloads | mobile |
|---|---|
| Interaction | single turn |
| Environment | static |
| End state checked | No |
| Tools | screenshot and accessibility tree in, action out |
| Grading | exact match |
| Size | 15,283 |
| Data | real scraped |
| Personal data | possible |
| License | unknown |
| Harness | none |
| Status | active |