Nvidia AI “MineDojo” wins award for learning to play Minecraft

Robert Collins

Nvidia announced on its blog on Monday that a research paper describing its generalist AI agent, MineDojo, has received industry recognition. Specifically, the paper won an Outstanding Datasets and Benchmarks Paper Award at the 2022 Neural Information Processing Systems (NeurIPS) conference.

The paper described the process of how MineDojo was trained to play Minecraft. Researchers fed the artificial intelligence 730,000 YouTube Minecraft videos (with more than 2.2 billion words transcribed), scraped 7,000 webpages from the Minecraft Wiki and 340,000 Reddit posts with over 6.6 million comments discussing Minecraft’s gameplay.

Utilizing all this data, MineDojo can process commands like “find a desert pyramid” or “play fireball with a ghast” and execute the steps necessary to complete the task. MineDojo accomplishes this via a custom transformer model called MineCLIP that associates video clips with specific in-game activities.

Examples of tasks that MineDojo can perform.
Via MineDojo.org

As the Nvidia blog post puts it,

While researchers have long trained autonomous AI agents in video-game environments such as Starcraft, Dota and Go, these agents are usually specialists in only a few tasks. So NVIDIA researchers turned to Minecraft, the world’s most popular game, to develop a scalable training framework for a generalist agent — one that can successfully execute a wide variety of open-ended tasks.

You may recall that earlier this year another AI achieved a similar feat, as we reported back in June. And this isn’t the first time NeurIPS has acknowledged Nvidia for its published research materials. A full list of accepted research papers from Nvidia—covering robotics, machine learning, simulation, self-driving cars and more—can be viewed here. And if you’d like to read the complete MineDojo paper (in PDF form) you can find it here.

Featured image via ArsTechnica.