Voyager is an open-ended, lifelong learning agent for Minecraft, designed to explore, acquire skills, and make discoveries autonomously. It targets researchers and developers in embodied AI and reinforcement learning, offering a framework for agents that can continuously improve and generalize their capabilities without human intervention.

How It Works

Voyager employs a modular architecture comprising an automatic curriculum for exploration, a dynamic skill library for storing and retrieving executable code (skills), and an iterative prompting mechanism. This mechanism leverages GPT-4 for program generation, incorporating environment feedback, execution errors, and self-verification to refine skills. The "blackbox" interaction with GPT-4 avoids the need for model fine-tuning, enabling rapid skill compounding and mitigating catastrophic forgetting.

Quick Start & Requirements

• Install: pip install -e . after cloning the repository.
• Prerequisites: Python ≥ 3.9, Node.js ≥ 16.13.0, OpenAI API key (GPT-4), Minecraft Java Edition, Fabric mods, and specific Node.js packages (npm install in voyager/env/mineflayer and mineflayer-collectblock).
• Setup: Requires manual Minecraft setup, including launching a LAN world with cheats enabled.
• Docs: Website, Arxiv.

Highlighted Details

• Achieves 3.3x more unique items, 2.3x longer travel distances, and up to 15.3x faster tech tree milestone unlocks than prior state-of-the-art.
• Demonstrates lifelong learning and generalization by utilizing learned skills in new Minecraft worlds.
• Skills are temporally extended, interpretable, and compositional.
• Interacts with GPT-4 via blackbox queries, bypassing fine-tuning.

Maintenance & Community

• Developed by MineDojo.
• Research project, not an official NVIDIA product.

Licensing & Compatibility

• MIT License. Permissive for commercial use and closed-source linking.

Limitations & Caveats

The project is strictly for research purposes. Task decomposition may occasionally require reruns if sub-goals are flawed. Initial setup involves several manual steps across Python, Node.js, and Minecraft configurations.