---
dataset_info:
  features:
  - name: episode_idx
    dtype: int32
  - name: step_idx
    dtype: int32
  - name: task
    dtype: string
  - name: pixels
    dtype: image
  - name: action
    list: int32
    length: 3
  - name: reward
    dtype: float32
  - name: terminated
    dtype: bool
  - name: truncated
    dtype: bool
  - name: id
    dtype: int64
  splits:
  - name: train
    num_bytes: 48689465
    num_examples: 564
  download_size: 48666802
  dataset_size: 48689465
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- reinforcement-learning
pretty_name: OpenApps UI-TARS Navigation (150 episodes)
size_categories:
- n<1K
tags:
- gui-agents
- world-models
- ui-navigation
- openapps
- ui-tars
---
# OpenApps UI-TARS Navigation (150 episodes)

UI-TARS-1.5-7B navigation trajectories on the [OpenApps](https://github.com/facebookresearch/OpenApps) synthetic web-app suite, for world-model / GUI-agent research. Each episode is a short cross-app navigation: from a starting app, click **Return to List of Apps** to reach the launcher, then click the target app's tile.

## Schema (one row per step)
- `pixels` — 1024x640 RGB screenshot the agent observed.
- `action` — discrete `[type, grid_x, grid_y]`: `type` 0=click / 1=scroll-down / 2=scroll-up; `(grid_x, grid_y)` is a cell on a 32x20 grid over the viewport (cell centre = the click pixel). This is the `MultiDiscrete([3, 32, 20])` action space.
- `reward`, `terminated`, `truncated` — per step; `terminated` (reward>=1) marks reaching the target app.
- `task` — the navigation route (e.g. `navigate_from_map_to_calendar`).
- `episode_idx`, `step_idx`, `id`.

## Routes & success
6 routes x 25 episodes; 4 start apps (map, messages, codeeditor, calendar) and 3 targets (calendar, todo, messages). Overall **87.3%** reached the target.

| Route | Success |
|---|---|
| messages -> calendar | 25/25 |
| codeeditor -> todo | 25/25 |
| map -> todo | 25/25 |
| map -> calendar | 25/25 |
| calendar -> todo | 18/25 |
| calendar -> messages | 13/25 |

Mean ~3.8 steps/episode. `manifest.json` records the episode-index range per route. Derived from OpenApps — check its license for downstream use.