MineStudio provides a comprehensive toolkit for developing AI agents in Minecraft, targeting researchers and developers in reinforcement learning and embodied AI. It streamlines agent creation through a customizable simulator, efficient data handling, pre-trained models, and robust training and inference pipelines.

How It Works

MineStudio leverages MineRL for its customizable Minecraft simulator, enabling efficient RL experimentation. Its data component uses a trajectory structure for optimized storage and retrieval, while the models section offers policy templates and baseline implementations. For training and inference, it integrates PyTorch Lightning and Ray for distributed capabilities, including memory-based policies and simulator crash recovery.

Quick Start & Requirements

• Installation: pip install MineStudio
• Prerequisites: Python 3.10+, JDK 8. Requires rendering tools (VirtualGL for NVIDIA, Xvfb for CPU).
• Setup: Conda environment setup and JDK installation recommended.
• Documentation: https://github.com/CraftJarvis/MineStudio

Highlighted Details

• Offers a simulator based on MineRL with customizable data structures.
• Provides a gallery of pre-trained models (VPT, GROOT, STEVE) on Hugging Face.
• Supports offline and online RL training with distributed inference via Ray.
• Includes automated benchmarking for diverse Minecraft tasks.

Maintenance & Community

The project is under active development, welcoming contributions. Version 1.1.4 (May 2025) introduced significant refactoring for data handling and improved documentation.

Licensing & Compatibility

The repository does not explicitly state a license. It acknowledges dependencies on MineRL, Project Malmo, Ray, and PyTorch Lightning.

Limitations & Caveats

The project is still under development, and users should expect ongoing changes. Minecraft simulation requires specific rendering tools, and compatibility with non-NVIDIA GPUs may involve performance trade-offs with Xvfb.