Voyager Documentation

To start the lifelong learning process, instantiate the Voyager class with your OpenAI API key and Azure login configuration.

Note on Azure Login: While you can use mc_port instead of azure_login, azure_login is highly recommended. When running azure_login for the first time, follow the CLI instructions to generate a config file.

Minecraft World Setup: After running voyager.learn(), you must manually configure the world:

1. Select Singleplayer -> Create New World.
2. Set Game Mode to Creative and Difficulty to Peaceful.
3. Once created, press Esc and select Open to LAN.
4. Set Allow cheats: ON and press Start LAN World.

from voyager import Voyager

# Configuration for Azure Login
azure_login = {
    "client_id": "YOUR_CLIENT_ID",
    "redirect_url": "https://127.0.0.1/auth-response",
    "secret_value": "[OPTIONAL] YOUR_SECRET_VALUE",
    "version": "fabric-loader-0.14.18-1.19", # the version Voyager is tested on
}
openai_api_key = "YOUR_API_KEY"

voyager = Voyager(
    azure_login=azure_login,
    openai_api_key=openai_api_key,
)

# Start the learning process
voyager.learn()

To execute specific tasks using previously learned skills rather than starting a new learning session, follow this pattern:

1. Instantiate with Skill Library: Pass skill_library_dir to load existing skills. Set resume=False because this is an inference task, not a learning task. Use a separate ckpt_dir for recording new events.
2. Decompose Task: Use decompose_task() to break the high-level goal into sub-goals. If sub-goals appear illogical, rerun the decomposition.
3. Inference: Use inference() to execute the sub-goals using the skill library.

from voyager import Voyager

# 1. Instantiate with skill library
voyager = Voyager(
    azure_login=azure_login,
    openai_api_key=openai_api_key,
    skill_library_dir="./skill_library/trial1",
    ckpt_dir="YOUR_CKPT_DIR",
    resume=False,
)

# 2. Run task decomposition
task = "Craft a diamond pickaxe"
sub_goals = voyager.decompose_task(task=task)

# 3. Run inference
voyager.inference(sub_goals=sub_goals)

This method allows Voyager to automatically resume if a request timeout occurs (e.g., when GPT-4 generates an infinite loop). This setup uses the minecraft-launcher-lib library.

Setup Steps:

1. Sign in to the Azure Portal.
2. Navigate to Azure Active Directory > App Registrations.
3. Click New registration and configure:
   • Name: YOUR_APP_NAME
   • Supported account types: Accounts in any organizational directory (Any Azure AD directory - Multitenant) and personal Microsoft accounts
   • Redirect URI Type: Public client/native (mobile & desktop)
   • Redirect URI Value: https://127.0.0.1/auth-response (Note: If you encounter KeyError: 'access_token', try changing the type to Web).
4. Click Register. The Application (client) ID is your client_id.
5. [Optional] Under Certificates & Secrets, create a New client secret. The generated value is your secret_value.
6. Identify your Minecraft version by checking the folder names in your Minecraft install location: YOUR_MINECRAFT_GAME_LOCATION/versions.

Note: Voyager uses fabric-loader-0.14.18-1.19 for experiments. If you do not have this version, refer to the Fabric Mods Install guide.

Configuration Object:

Use the following structure for your azure_login configuration:

azure_login = {
    "client_id": "CLIENT_ID FROM STEP 7",
    "redirect_url": "https://127.0.0.1/auth-response",
    "secret_value": "[OPTIONAL] SECRET_KEY FROM STEP 8",
    "version": "MINECRAFT VERSION YOU WANT TO USE",
}