Skip to content

DroidAgent: Intent-Driven Mobile GUI Testing with Autonomous LLM Agents

Notifications You must be signed in to change notification settings

coinse/droidagent

Repository files navigation

DroidAgent: Intent-Driven Android GUI Testing with LLM Agents

Test with artificial user intents!

DroidAgent produces high-level testing scenarios with executable scripts (currently in the format of UIAutomator2) by autonomously exploring a given application under test (AUT). It is built as a modular framework driven by multiple LLM instances.

droidagent_overview

Table of Contents

  1. Setup
  2. Running DroidAgent
  3. Replicate Evaluation Results
  4. Troubleshooting

Setup

DroidAgent requires ADB (Android debugging bridge) and several Python libraries installed.

Prerequisite

  • >=16GB RAM

  • >=Python 3.10

  • Android SDK & ADB installed and configured

    1. Install Android Studio
    2. Install command line tools: Android Studio > Tools > SDK Manager > SDK Tools > Android SDK Command-line Tools (latest)
    3. Setup Environment Variables: Check official document
      export ANDROID_HOME="[YOUR_HOME_DIRECTORY]/Android/sdk"
      export PATH="$PATH:$ANDROID_HOME/tools:$ANDROID_HOME/platform-tools:$ANDROID_HOME/cmdline-tools/latest/bin"
  • Android device connected or emulator running

  • DroidAgent uses the slightly modified version of DroidBot (included as submodule)

  • OpenAI key: rename .env.example to .env and add your own OpenAI API key

Clone & Install Dependencies

$ git clone --recurse-submodule https://github.com/coinse/droidagent.git
$ cd droidbot
$ pip install -e . # install droidbot
$ cd ..
$ pip install -r requirements.txt
$ pip install -e . # install droidagent

Run DroidAgent

  • Make sure that your Android device is connected or emulator is running.
  • Locate the APK file of the application under test (AUT) in the target_apps directory, with the name [APP_NAME].apk.
  • Run the following command in the scripts directory (add --is_emulator option if you are using an emulator):
$ cd script
$ python run_droidagent.py --app [APP_NAME] --output_dir [OUTPUT_DIR] --is_emulator
# example: python run_droidagent.py --app AnkiDroid --output_dir ../evaluation/data_new/AnkiDroid --is_emulator

Custom Persona & User Goals

According to your needs, use DroidAgent with a persona (a set of user characteristics) and a set of intents (a set of user goals) to guide the testing process.

  • You can add a custom persona by adding a new .txt file in the resources/persona directory, and giving the name of the file as an argument to --profile_id option to run_droidagent.py.
  • You can adjust the goal by modifying the following line in run_droidagent.py (we plan to make this easily configurable in the future):
    persona.update({
        'ultimate_goal': 'visit as many pages as possible while trying their core functionalities',
        # 'ultimate_goal': 'check whether the app supports interactions between multiple users', # for QuickChat case study
        'initial_knowledge': initial_knowledge_map(args.app, persona_name, app_name),
    })

Generate UIAutomator Scripts

DroidAgent supports generating corresponding UIAutomator2 scripts from the exploration history. In the script directory, run the following command:

$ python make_script.py --result_dir [RESULT_DIR] --project [PROJECT_NAME] --package [PACKAGE_NAME]
# example: python make_script.py --project AnkiDroid --package_name com.ichi2.anki --result_dir ../evaluation/data/AnkiDroid

[RESULT_DIR] should point to the [OUTPUT_DIR] you used in the previous step. Refer to the evaluation/package_name_map.json to find the package name of the AUT used in our evaluation. The replay script will be generated in the gen_tests directory. You can check the generation example in the gen_test_examples directory.

Note that the scripts are not guaranteed to be fully reproducible due to the possible flakiness of the AUT, but they can further be processed (e.g., manual exception handling) for constructing robust regression testing suite.

Generate Reports

We provide a script to generate a markdown report consisting of the generated tasks during the testing process. Run the following command in the script directory:

$ python make_report.py --result_dir [RESULT_DIR] --project [PROJECT_NAME]
# example: python make_report.py --project AnkiDroid --result_dir ../evaluation/data/AnkiDroid

[RESULT_DIR] should point to the [OUTPUT_DIR] as well. The task-by-task markdown report will be generated in the reports/[PROJECT_NAME] directory. Each report contains the task description, performed GUI actions, and the observation of the application state after the task execution with screenshots.

Reports on the evaluation data - published on website (TBD)

Example of a report: report

Replicate Evaluation Results

To replicate evaluation results for the paper "Intent-Driven Android GUI Testing with Autonomous Large Language Model Agents", follow the steps:

  • Download evaluation data from the below links (these are linked to Google drive uploads linked to an anonymised account).
  • Locate the DroidAgent data in the evaluation/data directory.
  • Locate the baseline data in the evaluation/baselines directory.
  • Locate the ablation data in the evaluation/ablation directory.
  • Run the notebooks in the evaluation/notebooks/ directory to replicate the results for each of our research questions.

Troubleshooting

SQLite (ChromaDB dependency) troubleshooting

  • If you get a sqlite3 version-related error message from chromaDB dependency, follow the instruction about building pysqlite3 from source: ChromaDB documentation

  • In short: pip install pysqlite3-binary

  • Add following lines into ~/<your_python_directory>/lib/python3.10/site-packages/chromadb/__init__.py:

__import__('pysqlite3')
import sys
sys.modules['sqlite3'] = sys.modules.pop('pysqlite3')

About

DroidAgent: Intent-Driven Mobile GUI Testing with Autonomous LLM Agents

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published