| --- |
| task_categories: |
| - reinforcement-learning |
| language: |
| - en |
| tags: |
| - embodied-ai |
| - world-models |
| - navigation |
| - openapps |
| - computer-vision |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # OpenApps Scripted Navigation (3,024 episodes, 20 routes) |
|
|
| Inter-app navigation trajectories in OpenApps, collected with a scripted |
| policy using only real UI actions (click "Return to List of Apps", click the |
| target app icon) — no `goto()` teleports, so transitions are learnable and |
| plannable from the action space alone. |
|
|
| - 20 routes: 5 source apps (todo, calendar, messages, codeeditor, map) × 4 targets |
| - 3,024 episodes after filtering (from 4,000 collected), 20 steps each |
| - Episode structure: ~10 random scrolls in the source app, click Return (home screen), click the target app icon, then random scrolls in the target app |
| - `pixels`: 1024×640 RGB (JPEG) |
| - `action`: `[type, grid_x, grid_y]`, where type 0 = click, 1 = scroll-down, 2 = scroll-up on a 32×20 grid |
| - `task`: `navigate_from_{src}_to_{tgt}` |
| - No reward column (scripted policy, not agent-collected) |
|
|
| Collector: `tools/nav_policy.py` in the OpenApps data collection codebase. |
|
|
| Note: content uses OpenApps' canonical fixed app state with theme/language variation. For randomized underlying app state with ground-truth labels, see the companion `openapps-todo-proprio-300ep` dataset. |
|
|
