AI research framework for embodied agents in Minecraft
Top 22.5% on sourcepulse
MineDojo is an AI research framework for developing open-ended, generally capable embodied agents within a Minecraft simulation. It offers a vast benchmark suite with over 3,000 diverse tasks, ranging from programmatic challenges to creative building and full game progression, leveraging internet-scale knowledge bases derived from YouTube, Wikipedia, and Reddit to enhance agent learning.
How It Works
MineDojo provides a Gym-style interface for agent interaction, featuring multimodal observations (RGB, compass, voxels) and a compound action space. Agents can learn from a massive dataset of 730K YouTube videos, 7K Wiki pages, and 340K Reddit posts, enabling them to acquire skills and strategies from human player data. The framework supports programmatic task scoring and offers a unique "playthrough" task to defeat the Ender Dragon.
Quick Start & Requirements
pip install minedojo
python minedojo/scripts/validate_install.py
(use xvfb-run
or MINEDOJO_HEADLESS=1
for headless environments).Highlighted Details
Maintenance & Community
Licensing & Compatibility
Limitations & Caveats
The Wiki database has a non-commercial use restriction. While the framework is extensive, achieving complex tasks may require significant computational resources and agent training time.
1 year ago
Inactive