Newsroom News Breaking Comics Tags RSS
News video games 27 June 2022, 19:29

author: destroytheardor

AI Learned to Play Minecraft by Watching YouTube

OpenAI's artificial intelligence appears to be a visual learner - it learned to play Minecraft from YouTube videos. The researchers had to resort to a few tricks to achieve satisfactory results, but that doesn't detract from their impressive achievement.

KEY INFORMATION:
  • OpenAI's AI learned to hunt animals and build simple structures in Minecraft.
  • The algorithm gained knowledge by watching videos posted on YouTube.

OpenAI, a company specializing in the development of neural networks (and known, for example, for the increasingly popular AI Dall-E), has developed two algorithms, which together learned to play Minecraft after watching 70,000 hours of YouTube videos. The first algorithm learned the actions taken by the players, while the second algorithm described the actions viewed by assigning labels with the keys used to perform each action.

Artificial intelligence puts up blocks in Minecraft

From the very beginning, the AI algorithm was able to learn how to cut down trees, make planks out of them, and put up simple structures made of blocks. It also learned how to swim, hunt, cook, and do everything that any novice user of Minecraft does. After a few grinds here and there, the researchers even managed to make the AI create a diamond pickaxe, which takes an efficient player about 20 minutes and requires roughly 24,000 actions.

AI Learned to Play Minecraft by Watching YouTube - picture #1
source: openai.com

The AI learned from videos using the VPT (Video PreTraining) method. However, there was a problem - video clips show what can be done in the game, but do not demonstrate how to do it specifically. The screen does not show the keys pressed, and it was from these that the algorithm gained knowledge and skills. Help was needed.

The second neural network was therefore delegated to label the videos. The clever algorithm accurately judged what keys on the keyboard the person playing the video was pressing and what mouse movements the person was making, and then assigned specific button combinations to the actions seen on the screen.

Thanks to this procedure, the first AI was able to explore the secrets of more advanced actions and as a result - to master Minecraft without having to try everything at random. The entire process is also described in detail on the OpenAI website.

Researchers believe that their method can work in other games, and also be useful in everyday computer use.. We, on the other hand, got proof that watching Minecraft for hours on YouTube can actually be useful.

  1. OpenAI - official website
  2. Minecraft - official website