Skip to content

instructions.md

Latest commit

 

History

History
115 lines (94 loc) · 5.46 KB

instructions.md

File metadata and controls

115 lines (94 loc) · 5.46 KB

AndroidEnv - Running the environment

In order to create an AndroidEnv instance you will need to provide two main components: a simulator and a task. In the following sections you will learn how you can create them.

The simulator

First, you will need to provide your Android virtual device (AVD) that the environment (and through it, the agent) can communicate with. While this can also be a physical device, in most cases you will need a virtual emulated device. There are many ways to emulate an AVD - in our examples, we will use Android Studio to create one.

  1. In Android Studio, create a virtual device by following this step-by-step guide.
  2. Follow the steps below to attach the AVD to your environment.

The task and examples with games and other apps

A task is a particular definition of an RL problem that the agent will be interacting with. A task may include critical RL information, such as what the rewards are, when the episodes are supposed to terminate, and what the reset procedures are that the environment should perform upon episode termination (e.g. start or relaunch an app, clear cache, etc.). This information is packaged into a Task() proto message, which gets passed passed to AndroidEnv.

  • For ready-made example tasks provided with AndroidEnv, check out the Available tasks, featuring Vokram (with Markov Decision Process (MDP)), Pong, DroidFish (a chess clone), Blockinger (a tetris clone), and more.

  • See the Tasks guide for details on features and capabilities of tasks, as well as how to create custom ones.

Create the environment

After setting up the simulator and creating a task, you may find the loader.load() function handy for creating an environment instance by providing relevant arguments, such as:

  • task_path: the path pointing to the .textproto file describing the desired task.
  • avd_name: the name of the AVD specified when your created it in Android Studio.
  • android_avd_home (Optional): the path to where the AVD is installed. (default value: ~/.android/avd).
  • android_sdk_root (Optional): the root directory of the Android SDK. (default value: ~/Android/Sdk).
  • emulator_path (Optional): the path to the emulator binary. (default: ~/Android/Sdk/emulator/emulator).
  • adb_path (Optional): the path to the ADB (Android Debug Bridge). (default value: ~/Android/Sdk/platform-tools/adb).

For the AVD name and path, in Android Studio, go to Tools > AVD Manager, right click on your virtual device, and select View Details, where you will find the avd_name and its path.

For the Android SDK location, in Android Studio, go to Preferences > Appearance & Behavior > System Settings > Android SDKs and note the Android SDK Location. In the SDK folder, you will find /emulator/emulator as well as the ADB path (/platform-tools/adb).

Your example configuration may look like this, depending on how you set up your emulator:

from android_env import loader

env = loader.load(
    avd_name='my_avd',
    android_avd_home='/Users/username/.android/avd',
    android_sdk_root='/Users/username/Library/Android/sdk',
    emulator_path='/Users/username/Library/Android/sdk/emulator/emulator',
    adb_path='/Users/username/Library/Android/sdk/platform-tools/adb',
    task_path='/Users/username/android_env/my_tasks/my_task.textproto',
)

Example RL agent scripts

The examples directory contains a few simple example agent setups, such as:

  • run_random_agent.py: Runs a simple loop performing randomly selected actions in the environment.
  • run_acme_agent.py: Runs a training loop with an Acme DQN agent, implemented in the popular DeepMind RL framework. This will require to install the acme dependency.
  • run_human_agent.py: Creates a pygame instance that lets a human user interact with the environment and observe environment mechanics, such as rewards or task extras. You will need to install the [PyGame] dependency.

For instance, here is how you can run run_random_agent.py in a folder where you have your APK file, such as Apple Flinger from Example tasks. (The downloaded TAR file contains the APK file and .textproto definitions.)

python3 run_random_agent.py \
--avd_name='my_avd' \
--android_avd_home=/Users/username/.android/avd \
--android_sdk_root=/Users/username/Library/Android/sdk \
--emulator_path=/Users/username/Library/Android/sdk/emulator/emulator \
--adb_path=/Users/username/Library/Android/sdk/platform-tools/adb \
--num_steps=1000 \
--task_path=/Users/username/<PATH-TO-APP-TASK-FILES>/apple_flinger_M_1_1.textproto