Microsoft has unveiled Mineworld, an open-source world model for enhancing the Minecraft experience, which allows for real-time interactions and high controllability. Mineworld improves upon previous models, particularly the closed-source Oasis, by addressing computational inefficiencies. Its key components include:
- Real-time interactivity for fast, dynamic gameplay.
- A parallel decoding algorithm that accelerates frame generation.
- A novel evaluation metric for assessing controllability.
Mineworld processes video game footage and player actions through unique tokenization methods. Each frame in a video clip consists of quantized tokens representing distinct pixel sets. Player actions are tokenized into 11 distinct tokens, including 7 for action groups, 2 for camera angles, and 2 for action sequence markers.
Mineworld was released on April 11, 2025.