UI-MOPD/AndroidControl-Star

AndroidControl*

A curated step-level evaluation subset extracted from AndroidControl, used for static mobile GUI understanding evaluation in UI-MOPD (Multi-platform On-Policy Distillation for Continual GUI Agent Learning).

Overview

AndroidControl* contains 4,260 step-level records from 781 Android trajectories. Each record includes the trajectory identifier, step index, high-level task goal, per-step instruction, normalized action, screenshot path, screenshot resolution, and optional UI grounding metadata.

AndroidControl* preserves the same normalized mobile action space used in UI-MOPD training, including click, scroll, input_text, open_app, wait, navigate_back, long_press, and navigate_home. For actions that can be matched to a UI element, grounding metadata (target bounding boxes, widget class, text, resource id, package name, ancestor information, and matching node count) is included. This subset is mainly used to evaluate static mobile GUI understanding, verify action-screen alignment, and illustrate the format of mobile data after normalization.

Dataset Statistics

Metric	Value
Step Records	4,260
Trajectories	781
Platform	Android Mobile (1080x2400)
Format	JSONL + PNG screenshots
Coordinate System	Absolute pixel (actions normalized to [0, 1000] for evaluation)

Action Space

Action	Description
click	Tap at coordinate
long_press	Long press at coordinate
scroll	Scroll in direction (up/down/left/right)
input_text	Type text into active field
open_app	Launch application by name
navigate_back	Return to previous interface
navigate_home	Return to home screen
wait	Wait for UI response

Grounding Metadata

For actions that can be matched to a UI element, each record includes grounding metadata:

Field	Description
target_bbox	Bounding box of the target element [left, top, right, bottom] in pixels
target_class	Android widget class (e.g., android.widget.TextView)
target_text	Visible text on the element
target_content_desc	Accessibility content description
target_resource_id	Android resource ID
target_package	Application package name
ancestor_bbox	Bounding box of the nearest ancestor with text
ancestor_text	Text on the ancestor element
n_matching_nodes	Number of UI nodes matching the action target

Steps without a directly groundable target (e.g., waiting or app-level operations) have grounding: null.

Data Format

AndroidControl-Star/
  steps.jsonl              # All 4,260 step-level records
  images/
    episode_0/
      step_0.png
      step_1.png
      ...
    episode_100/
      ...

JSONL Record Structure

{
  "episode_id": 0,
  "step_idx": 0,
  "total_steps": 3,
  "goal": "Open the Zoho Meet app, view the scheduled meetings.",
  "instruction": "Open the Zoho Meet app",
  "action": {
    "action_type": "open_app",
    "app_name": "Zoho Meeting"
  },
  "screenshot": "images/episode_0/step_0.png",
  "screenshot_width": 1080,
  "screenshot_height": 2400,
  "grounding": {
    "target_bbox": [494, 365, 586, 416],
    "target_class": "android.widget.TextView",
    "target_text": "",
    "target_content_desc": "",
    "target_resource_id": "",
    "target_package": "com.zoho.meeting",
    "ancestor_bbox": [360, 317, 720, 464],
    "ancestor_text": "Past",
    "n_matching_nodes": 14
  }
}

Evaluation Protocol

Coordinates are normalized to [0, 1000] for both model input and evaluation:

# Normalize pixel coordinate to [0, 1000]
norm_x = round(x / width * 1000)
norm_y = round(y / height * 1000)

# Denormalize model output back to pixels
pixel_x = norm_x / 1000 * width
pixel_y = norm_y / 1000 * height

Evaluation Metrics

Metric	Description
Action Type Accuracy	Predicted action type matches ground truth
Grounding (target)	Predicted coordinate falls within target_bbox
Grounding (ancestor)	Predicted coordinate falls within ancestor_bbox (more lenient)
Overall Accuracy	Joint accuracy of action type and grounding

Prompt Template

The evaluation uses the mobile_use tool interface with the following action set:

click, long_press, swipe, type, system_button, open_app, wait, terminate

Each step is prompted with the episode-level goal and the current step instruction. The model outputs structured reasoning (Thought + Action) followed by a tool call.

Usage

import json

with open("steps.jsonl", "r") as f:
    steps = [json.loads(line) for line in f]

# Group by episode
from collections import defaultdict
episodes = defaultdict(list)
for step in steps:
    episodes[step["episode_id"]].append(step)

print(f"Total steps: {len(steps)}, Episodes: {len(episodes)}")

Or clone directly:

git clone https://huggingface.co/datasets/UI-MOPD/AndroidControl-Star

Citation

@article{ui-mopd,
  title={UI-MOPD: Multi-platform On-Policy Distillation for Continual GUI Agent Learning},
  year={2025}
}

License

Apache 2.0