OpenAI built a neural network that can play Minecraft like a person


The importance of the fact that AI can now play Minecraft as well as you do

(Image credit: Microsoft)

A neural network has been taught by OpenAI experts to play Minecraft at a level that is comparable to that of human gamers.

The neural network was trained using 70,000 hours of different video game footage, together with a tiny database of films of contractors carrying out specified in-game tasks and recording their keyboard and mouse actions.


After some tweaking, OpenAI discovered that the model could carry out a wide range of complicated tasks, including swimming and searching for and eating flesh from various species. It also understood the “pillar leap,” which is a maneuver in which the player lowers a piece of material beneath oneself mid-jump to boost height.

Most impressively, the AI created diamond tools (which required a complex series of steps to be carried out in order), which OpenAI called a “unprecedented” feat for a computer agent.

An AI breakthrough?

The Minecraft experiment is significant because it shows the effectiveness of Video PreTraining (VPT), a novel approach used by OpenAI to train AI models that might hasten the creation of “universal computer-using agents,” according to the firm.


In the past, utilizing raw video as a source for training AI models has been challenging since it is easy to comprehend what happened, but not always how. The intended outputs would effectively be absorbed by the AI model, but it would be unable to comprehend the input combinations needed to get there.

However, to create the fundamental model for VPT, OpenAI combines a sizable video dataset culled from public online sources with a pool of film that has been carefully selected and annotated with the pertinent keyboard and mouse actions.

The team adds in smaller datasets created to teach certain activities to further fine-tune the underlying model. In this case, OpenAI used video of players taking part in early-game behaviors like felling trees and creating crafting stations, which is stated to have resulted in a “huge increase” in the model’s capacity to complete these tasks reliably.


Reinforcement learning is a different method that involves “rewarding” the AI model for completing each task in a series of tasks. The neural network was able to gather all the components for a diamond pickaxe using this procedure, with a success rate comparable to that of a person.

“VPT clears the way for allowing agents to pick up acting skills by viewing the many videos on the internet. In contrast to contrastive techniques like generative video modeling, which would only provide representational priors, “VPT offers the intriguing potential of directly learning large-scale behavioral priors in more domains than simply language,” wrote OpenAI in a blog post (opens in new tab).

“We only explore in Minecraft, but the game is quite open-ended and the natural human interface (mouse and keyboard) is fairly general, so we feel our results are promising for other related domains, including computer usage,” the authors write.


The MineRL NeurIPS competition has teamed with OpenAI to encourage more research in the field by providing its contractor data and model code to competitors using AI to complete challenging Minecraft challenges. The top reward is $100,000.

Leave a Comment