Description
Voyager represents a significant advancement in embodied AI, functioning as the first large language model (LLM)-powered lifelong learning agent within the Minecraft environment. This agent is designed to autonomously explore the game world, acquire a diverse range of skills, and uncover novel discoveries without requiring any human input. The architecture of Voyager is built upon three core components: an automatic curriculum engineered to maximize exploration, an expanding skill library that stores and retrieves complex behaviors as executable code, and an innovative iterative prompting mechanism. This mechanism allows the agent to refine its programs by incorporating feedback from the environment, learning from execution errors, and employing self-verification.
Read More
Voyager interacts with GPT-4 through blackbox queries, a method that circumvents the need for fine-tuning the model's parameters. The skills developed by Voyager are characterized by their temporal extension, interpretability, and composability. This design facilitates rapid compounding of the agent's abilities and effectively mitigates catastrophic forgetting, a common challenge in AI learning. Empirically, Voyager exhibits robust in-context lifelong learning capabilities and demonstrates exceptional proficiency in playing Minecraft. It has been shown to acquire 3.3 times more unique items, travel 2.3 times longer distances, and unlock key technological milestones up to 15.3 times faster than previous state-of-the-art methods.
A key feature of Voyager is its ability to leverage its learned skill library in new Minecraft worlds to solve novel tasks from scratch, a capability where other techniques often struggle to generalize. The repository provides the Voyager code under an MIT License. Installation requires Python version 3.9 or higher and Node.js version 16.13.0 or higher, with tested compatibility on Ubuntu 20.04, Windows 11, and macOS. The setup involves cloning the repository, installing Python dependencies, and installing specific Node.js packages. A Minecraft instance and necessary Fabric mods are also required. To run Voyager, an OpenAI API key is necessary, as it utilizes GPT-4 for its language model. The agent can be initiated for lifelong learning or configured to resume from checkpoints and run specific tasks using learned skill libraries.
- CategoryAI Agent Builders
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsWindows, macOS, Linux
Voyager Embodied Agent's Core Features
Open-ended embodied agent for Minecraft
Utilizes large language models (LLM) for decision-making
Lifelong learning capabilities
Automatic curriculum for exploration maximization
Executable skill library for storing and retrieving behaviors
Iterative prompting mechanism with environment feedback
Self-verification for program improvement
Interacts with GPT-4 via blackbox queries
Temporally extended, interpretable, and compositional skills
Mitigates catastrophic forgetting
Demonstrates strong in-context learning
Achieves state-of-the-art performance in Minecraft tasks
Generalizes learned skills to new environments
Getting Started with Voyager Embodied Agent
Clone: Clone the Voyager GitHub repository.
Install Python dependencies: Run `pip install -e .` in the project directory.
Install Node.js dependencies: Navigate to `voyager/env/mineflayer` and run `npm install -g npx` and `npm install`.
Configure Minecraft: Set up a Minecraft instance and install required Fabric mods.
Obtain API Key: Get an OpenAI API key for GPT-4 access.
Initialize Voyager: Instantiate the Voyager agent with Azure login details and API key.
Start Learning: Execute `voyager.learn()` to begin the lifelong learning process.
Resume Learning: Instantiate Voyager with `ckpt_dir` and `resume=True` to continue from a checkpoint.
Voyager Embodied Agent's Use Cases
- Autonomous Minecraft Exploration
- Skill Acquisition and Composition
- Lifelong Learning in Embodied AI
- Task Decomposition and Execution
- AI Research in Game Environments
- Novel Discovery and Innovation
- Generalization of AI Capabilities
