Odyssey empowers Large Language Model (LLM)-based agents with open-world skills for complex tasks in Minecraft. It offers a comprehensive framework for researchers and developers aiming to advance autonomous agent capabilities beyond basic programmatic goals, enabling exploration and long-horizon strategy learning.

How It Works

Odyssey utilizes a skill library with 40 primitive and 183 compositional skills. It fine-tunes LLaMA-3 models on a large Minecraft Wiki-derived question-answering dataset (390k+ entries). The framework supports various agent roles, including action execution, curriculum planning, and task critique, facilitating complex behaviors like combat and autonomous exploration.

Quick Start & Requirements

• Installation: Python ≥ 3.9, Node.js ≥ 16.13.0. Install via pip install -e . and pip install -r requirements.txt for Python, and yarn install for Node.js components.
• Dependencies: Requires a Minecraft server (Docker deployment recommended) and an embedding model (e.g., paraphrase-multilingual-MiniLM-L12-v2 from Hugging Face, requires git-lfs).
• Configuration: A config.json file is needed to specify server details, LLM backend, and embedding model paths.
• Running: Execute python main.py after setup. Task-specific examples are provided in the README.
• Resources: The project provides links to demo videos and a comprehensive skill library.

Highlighted Details

• Introduces a new benchmark for evaluating LLM-based agents across long-term planning, dynamic-immediate planning, and autonomous exploration tasks.
• Fine-tuned LLaMA-3 models are trained on a substantial Minecraft Wiki dataset for enhanced knowledge and instruction following.
• Features an interactive agent with an extensive library of primitive and compositional skills.
• Open-sourced components include the agent code, a web crawler for data collection, and the skill library.

Maintenance & Community

The project is developed by VIPA Lab at Zhejiang University. Recent updates include open-sourcing a multi-agent framework and a web crawler. Contact is available via email for inquiries.

Licensing & Compatibility

The codebase is licensed under the MIT License. The Minecraft Q&A Dataset is licensed under CC BY-NC-SA 3.0, which may restrict commercial use or derivative works without similar licensing.

Limitations & Caveats

The CC BY-NC-SA 3.0 license for the dataset imposes non-commercial restrictions. While tested on Ubuntu, Windows, and macOS, specific compatibility nuances for all environments are not detailed.