Microsoft MineWorld: Real-Time Open-Source Interactive World Model Built on Minecraft 247

What is MineWorld?

MineWorld is a real-time interactive world model based on Minecraft, open-sourced by Microsoft on April 14, 2025. Its goal is to build a dynamic environment for intelligent agents that can perceive, reason, and act.

MineWorld Technical Features

• Vision-Action Autoregressive Transformer: MineWorld leverages a vision-action autoregressive Transformer driving model, converting game scenes and actions into discrete token IDs, then trained through next-token prediction.

• Parallel Decoding Algorithm: A novel parallel decoding algorithm has been developed to simultaneously predict spatially redundant tokens in each frame, enabling models of different scales to generate 4 to 7 frames per second, achieving real-time interaction with game players.

• Input and Output: The model takes an initial image of the game scene and user-selected actions as input, and generates the corresponding next scene.

• Evaluation Metrics: New evaluation metrics are proposed that not only assess visual quality but also evaluate action-following capability when generating new scenes.

MineWorld Applications

• Embodied AI Research: Provides a complex and dynamic virtual testing platform for embodied AI research.

• Reinforcement Learning Training: Can be used for reinforcement learning training, helping agents learn and adapt in simulated environments.

• Interactive AI Agent Testing: Provides a real-time, interactive environment for testing interactive AI agents.

How to Use MineWorld

• Web Demo: Users can launch a web game by running specific commands and access it via a local or public URL. Users can select the initial frame, control camera movement, and execute game actions such as jumping and moving.

• Local Execution: Users can also run the model locally to reproduce the evaluation results from the paper.

MineWorld Scope and Limitations

Scope: MineWorld is specifically trained for the Minecraft game domain and is only applicable to generating game-related scenes.

Limitations:

• Cannot generate results in other video domains (such as internet videos).

• The generated video resolution is limited, which may lead to loss of detail.

• The model may inherit biases, errors, or omissions from the training data.

• Further testing and validation are required before considering application in commercial or real-world scenarios.

MineWorld Installation and Usage

1. Environment Configuration:

• Clone the repository:

git clone https://github.com/microsoft/mineworld.git
cd mineworld

• Create and activate the Conda environment:

conda create -n mineworld python=3.10
conda activate mineworld
pip3 install -r requirements.txt

• A high-performance GPU (such as A100 or H100) is recommended for inference.

2. Download Pre-trained Models:

• Download the pre-trained model and related configuration files, and place them in the checkpoints folder.

• The data structure is as follows:

├── checkpoints
│   ├── 300M_16f.ckpt
│   ├── 700M_16f.ckpt
│   ├── 700M_32f.ckpt
│   ├── 1200M_16f.ckpt
│   └── 1200M_32f.ckpt
│   └── vae
│       ├── config.json
│       └── vae.ckpt
├── validation
│   └── validation.zip
└── gradio_scene
├── scene.mp4
└── scene.jsonl

3. Run the Web Demo:

• Launch the Web Demo with the following command:

python mineworld.py --scene "path/to/scene.mp4" --model_ckpt "path/to/ckpt" --config "path/to/config"

• Access the local or public URL and interact through the web page.

• You can select the initial frame, control camera movement, and execute game actions (such as jumping, moving, etc.).

4. Local Inference:

• Use the following command to perform local inference:

python inference.py --data_root "/path/to/validation/dataset" --model_ckpt "path/to/ckpt" --config "path/to/config" --demo_num 1 --frames 15 --accelerate-algo 'naive' --top_p 0.8 --output_dir "path/to/output"

• Compute evaluation metrics:

bash scripts/setup_metrics.sh
bash scripts/compute_metrics.sh

GitHub:https://github.com/microsoft/MineWorld