TVWorld

• Repository: https://github.com/Lqf-HFNJU/TVTheseus
• Paper: https://arxiv.org/pdf/2601.13142

---

⭐️ Introduction

📺 TVWorld is an offline graph-based abstraction of real-world TV navigation that enables reproducible and deployment-free evaluation. On this basis, TVWorld comprises two complementary benchmarks that comprehensively assess TV-use capabilities: 🕹️ TVWorld-N for topology-aware navigation and 🎯 TVWorld-G for focus-aware grounding.

• 🕹️ TVWorld-N is an offline interactive TV navigation environment for evaluating agents’ topology-aware planning under focus-based remote-control, supporting both textual and visual goals. Operating purely on static graph assets, it is fully replayable and deployment-free (e.g., no VMs/emulators), and enables millisecond-level interaction, avoiding the instability and overhead of online GUI benchmarks.
• 🎯 Complementarily, TVWorld-G evaluates focus-aware grounding by requiring the agent to localize the currently highlighted element within the global screen layout using bounding-box annotations, directly reflecting the focus-based nature of TV control.

---

⚙️ Data Structure

The dataset is structured as follows:

TVWorld
|-- eval_data
|   |-- TVWorld_N.json
|   |-- TVWorld_G.json
|   `-- graph
|       |-- graph_*
|       |   |-- graph.json
|       |   `-- tasks.json
|       |-- image
|       |   `-- screenshot_*.png
|       `-- xml
|           `-- ui_check_*.xml
`-- training_data
    |-- graph.json
    |-- image
    |   `-- screenshot_*.png
    |-- stage_1.json
    |-- stage_2.parquet
    `-- xml
        `-- ui_check_*.xml

Test graphs collected from Google TV: eval_data

• TVWorld_N.json: task file for topology-aware navigation in TVWorld-N
• TVWorld_G.json: task file for focus-aware grounding in TVWorld-G
• graph/: offline navigation graphs collected from Google TV:
  • it contains 5 subdirectories, each corresponding to a specific category of UI navigation subgraphs:
    1. graph_config/ — Config: system-level configuration and global settings;
    2. graph_display/ — Display: display and rendering settings;
    3. graph_audio/ — Audio: audio-related capabilities;
    4. graph_app/ — Apps: application entry points and privacy-related settings;
    5. graph_channels/ — Channels: hardware-interface components (e.g., TV channels).
  • Each subdirectory contains the following files:
    • graph.json defines the structure and metadata of a UI navigation graph:
      • nodes: a list of all nodes in the graph.
      • edges: a list of directed edges representing feasible navigation transitions. Each edge includes:
        • source: source node ID
        • target: target node ID
        • key: navigation action used for the transition (e.g., "UP", "DOWN", "LEFT", "RIGHT", "OK", "SETTING", "HOME", "EXIT")
      • node_mapping: a mapping from node IDs to detailed node-level metadata. Key fields include:
        • text: the name of the node
        • up_focus_text: the name of the UI element currently in focus in the screenshot
        • bounds: the bounding box of the focused UI element, formatted as [x1,y1][x2,y2], where (x1, y1) and (x2, y2) denote the top-left and bottom-right coordinates, respectively.
        • screenshot_url: the screenshot corresponding to the node
        • viewtree_path: the file path to the view-tree associated with the node
        • parent_layout_info: metadata describing the parent layout of the focused UI element
        • unique_id: the unique identifier for the node
        • neighbors: a dictionary that enumerates reachable neighboring nodes under different navigation actions. Each key corresponds to a navigation action (e.g., "OK"), and the value is a list of possible transitions.
    • tasks.json defines navigation tasks associated with the corresponding graph, each task entry includes the following fields:
      • start: the unique ID of the start node
      • end: the unique ID of the target (goal) node
      • start_name: the name of the start node
      • end_name: the name of the target (goal) node, used for text-based instructions
      • end_image: the filename of the screenshot corresponding to the target (goal) node, used for vision-based instructions
  • image/: screenshots corresponding to UI nodes in the navigation graphs
  • xml/: view-tree files (in XML format) associated with each UI nodes

Training graph collected from TCL TV: training_data

Please refer to the paper for details of the training data.

• graph.json: the offline UI navigation graph collected from TCL TV, used to define state transitions during training. The field format is identical to that of the test-time graph.json files described above
• image/: screenshots corresponding to UI states in the graph.
• xml/: View-tree files associated with UI states.
• stage_1.json: training data for Stage I, used for Topology-Priming SFT
• stage_2.parquet: training data for Stage II, used for Topology-Augmented RL

---

📝 Licensing Information

This work is licensed under a Creative Commons Attribution 4.0 International License.

---

📢 Disclaimer

This dataset is intended primarily for research purposes. We strongly oppose any harmful use of the data or technology.

---

📄 Citation

@misc{ma2026tvworldfoundationsremotecontroltv,
      title={TVWorld: Foundations for Remote-Control TV Agents}, 
      author={Zhantao Ma and Quanfeng Lu and Shuai Zhong and Dahai Yu and Ping Luo and Michael K. Ng},
      year={2026},
      eprint={2601.13142},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.13142}, 
}