DroidBot-GPT is an easy-to-use UI agent framework for Android. It can control a smartphone app automatically based on a natural language task description. It is built upon DroidBot.

Implementing a domain-specific mobile agent is very simple with DroidBot-GPT. You only need to modify two files: droidbot/input_policy.py which specifies how to call LLM based on the task and UI state and droidbot/device_state.py which specifies how to convert the UI state to something that LLM (or other multi-modal models) can understand.

The policy implemented in this repo is a naive baseline for mobile task automation. It simply converts the view list to text, and sends the text to GPT and ask for the next action to perform. A more advanced version (with app-specific memory injection and other optimizations) can be found at AutoDroid.

Technical report of this repo:

Hao Wen, Hongming Wang, Jiaxuan Liu, Yuanchun Li. "DroidBot-GPT: GPT-powered UI Automation for Android"

Make sure you have:

1. Python (both 2 and 3 are supported)
2. Java
3. Android SDK
4. Added platform_tools directory in Android SDK to PATH
5. Access to an LLM/LVM/LMM (e.g. ChatGPT API)

Then clone this repo and install with pip:

git clone https://github.com/MobileLLM/DroidBot-GPT.git
cd DroidBot-GPT/
pip install -e .

If successfully installed, you should be able to execute python start.py -h.

1. Prepare:

   • An app to use. Download the .apk file to your host machine.
   • A device or an emulator connected to your host machine via adb.
   • Set the OpenAI URL and API KEY with environment variables GPT_URL and GPT_KEY.

2. Start DroidBot-GPT:

   python start.py -a <path_to_apk> -o output_dir -task <your_task>
   

   That's it! The options are mostly the same as DroidBot except for the -task option, where you can specify any task you want to complete. For example,

   • Create a contact named Alice with phone number 1234567.
   • Book a table for 4 people on Saterday.
   • Send a message to Sam to have a chat tonight.
   • ...

After execution, the action trace and screenshots can be found in output_dir/index.html.

1. Decide which LLM (or other foundation models) to use. Modify how to query the model in the _query_llm function in droidbot/input_policy.py.

2. Decide how you want to represent the app UI in droidbot/device_state.py. Each DeviceState instance is a snapshot of current smartphone UI, which include the view hierarchy (the views field), the screenshot (the screenshot_path field), etc.

3. Design how to query the foundation model with the UI representation. Modify the code in the _get_action_with_LLM function in droidbot/input_policy.py. The function takes the current state (current_state parameter), the previous actions (action_history parameter) and the task description (self.task field) as input, and its job is to decide which action to take in the next step.

To further improve the agent performance, you may consider optimize how you invoke the LLM(s) and/or train/fine-tune the LLM(s).

Enjoy!