AndroidLens: Long-latency Evaluation with Nested Sub-targets for Android GUI Agents

AndroidLens is a challenging benchmark for mobile GUI agents, featuring 571 real-world, long-horizon tasks in both Chinese and English, with an average of 26.1 steps per task. It supports evaluation of critical capabilities:

• Long-horizon planning under multi-constraint & multi-goal scenarios
• 298 cross-app tasks and 273 single-app tasks, covering 74 real-world applications (e.g., WeChat, Google Drive, Taobao, Maps).
• Robustness to real anomalies: ads, permission pop-ups, login redirects
• Multi-trajectory ground truth to reduce path bias
• Milestone-based progress tracking via nested sub-targets

This dataset is designed for both static (step-wise prediction) and dynamic (real-device execution) evaluation.

📄 Paper: AndroidLens: Long-latency Evaluation with Nested Sub-targets for Android GUI Agents
💾 GitHub: https://github.com/alibaba/AndroidLens
🤗 Hugging Face: yuecao0119/AndroidLens

---

🗂️ Dataset Structure

Your data is organized as:

test/
├─ en/                        # English tasks
│  └─ <episode_id>/
│     ├─ <episode_id>.json   # Full episode trajectory (list of steps)
│     ├─ <episode_id>_0.png  # Screenshot at step 0
│     ├─ <episode_id>_1.png
│     └─ ...
└─ zh/                        # Chinese tasks
   └─ <episode_id>/
      ├─ <episode_id>.json
      ├─ <episode_id>_0.png
      └─ ...

Each <episode_id>.json file contains a list of step objects, with one object per interaction step.

---

🏷️ Task Category Codes (types)

The types field uses a hierarchical two-digit code system to classify task complexity and structure. These categories align with AndroidLens’s taxonomy of Multi-goal (1-X), Multi-constraint (2-X), and Domain-specific (3-X) tasks, enabling fine-grained analysis of agent performance across different challenge dimensions.

Code	Category	Description
1-1	Single-app Unrelated Tasks	Multiple independent subtasks within one app, with no logical dependency.
1-2	Single-app Related Tasks	Multiple dependent subtasks within one app (e.g., “Search product → add to cart → checkout”)
1-3	Cross-app Unrelated Tasks	Independent actions across multiple apps (e.g., “Send message on WeChat, then play music on QQ Music”)
1-4	Cross-app Related Tasks	Interdependent operations across multiple apps (e.g., copy link from Chrome → paste in Drive → upload)
2-1	Operation-Level Constraints	Tasks requiring precise widget-level control, such as exact text input, time/date pickers, or multi-condition filtering
2-2	Page-Level Constraints	Tasks with navigation constraints like specific tab/category selection, sort/filter order, or view state
3-1	Batch Operation Tasks	Repeated actions on multiple items (e.g., “Empty shopping carts on Taobao, JD, and Pinduoduo”)
3-2	Combine with VLM Capabilities	Tasks that leverage the agent’s built-in multimodal abilities, such as translation, comparison, summarization, or OCR-to-action

---

📑 Step-level Data Format

Each step in the JSON list includes:

Field	Type	Description
episode_id	str	Unique task ID (UUID)
language	str	"en" or "zh"
app	List[str]	Sequence of involved apps (e.g., ["Google Chrome", "Google Drive"])
episode_length	int	Total steps in the full trajectory
step_id	int	Current step index (0-based)
instruction	str	High-level user goal (same for all steps in the episode)
image_path	str	Relative path to screenshot (e.g., en/.../0.png)
image_width, image_height	int	Original resolution of the screenshot
result_action_type	List[int]	Action code ([4] = Click; see note below)
result_touch_yx	List[str]	Normalized touch coordinates as string: "[y, x]" in range [0, 1]
result_lift_yx	List[str]	End point for swipe (same as touch for tap)
result_action_text	List[str]	Text to input (empty if none)
duration	List[null/float]	Action hold time (null for tap)
low_instruction	str	Step-specific guidance (for low-level evaluation)
milestone	dict	Nested sub-goal info (see below)
types	List[str]	Task category code (e.g., "1-2"); see full mapping below.

🔍 Coordinate Note:

• result_touch_yx uses relative coordinates in [0, 1], with format "[y, x]" (note: y first).
• To convert to absolute pixel:

  y_abs = float(y_rel) * image_height
  x_abs = float(x_rel) * image_width
  

---

🎯 Milestone Format

The milestone field enables fine-grained progress evaluation:

{
  "sub-target": "Open Google Chrome and search for 'panda'",
  "idx": 1,
  "bbox": [0.023, 0.121, 0.976, 0.189],  // [x1, y1, x2, y2] in normalized coords
  "text": "panda",
  "state": ["selected"]
}

• idx: milestone index (ordered)
• bbox: bounding box of key UI element (normalized, xy format)
• text / state: expected content or widget state

Milestones support ordered and unordered sub-goals for complex tasks.

---

📊 Action Type Mapping

Although the original AgentCPM-GUI defines actions via names, your data uses numeric codes in result_action_type. Based on AndroidLens annotation practice, the common mapping is:

Code	Action	Required Fields
1	Wait	duration
3	Type	result_action_text
4	Click	result_touch_yx
4	LongPress	result_touch_yx, duration
4	Swipe	result_touch_yx, result_lift_yx
5	PressBack	—
6	PressHome	—
10	Terminate	—

Confirm exact mapping from your annotation code if needed. AndroidLens uses ADB-based actions with explicit start/end for swipe.

---

📜 License

AndroidLens is released under the Apache-2.0 License.
Screenshots are derived from real app usage for research purposes only. Comply with app store policies and local regulations.

---

✏️ Citation

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{cao2025androidlens,
    title={AndroidLens: Long-latency Evaluation with Nested Sub-targets for Android GUI Agents}, 
    author={Yue Cao and Yingyao Wang and Pi Bu and Jingxuan Xing and Wei Jiang and Zekun Zhu and Junpeng Ma and Sashuai Zhou and Tong Lu and Jun Song and Yu Cheng and Yuning Jiang and Bo Zheng},
    year={2025},
    journal={arXiv preprint arXiv:2512.21302},
}