OpenAI trains a bot to play Minecraft using 70,000 hours of gameplay videos

(Image Courtesy: Microsoft)

OpenAI, the artificial intelligence research and deployment company, has managed to successfully train a bot to play Minecraft using 70,000 hours of gameplay footage.

It is the first bot that can create the much coveted Diamond Tools within the game and while that is impressive, this marks an important advancement in the tools used to train AI.

Researchers are calling this a breakthrough in Imitation Learning, a method of teaching AI by showing them how humans do tasks. The AI then learns by mimicking those tasks, and doing it as efficiently as possible.

By tapping into online videos, the researchers behind the project hope to achieve something similar to GPT-3 for large scale language models. GPT-3  was trained on thousands of texts and books, and can imitate human-written text.

“In the last few years we’ve seen the rise of this GPT-3 paradigm where we see amazing capabilities come from big models trained on enormous swathes of the internet,” says Bowen Baker at OpenAI, in an interview with Technology Review.

“A large part of that is because we’re modeling what humans do when they go online,” Baker added.

The current limitation to Imitation Learning is that videos that are used to train the AI, have to be individually labeled with each step shown. This is to make the AI understand the context of the actions shown and the results.

The problem with this approach is that it takes a lot of time and manpower to maintain these videos. The OpenAI team got around this by using a technique called Video Pre-Training (VPT), which trains another neural network to decipher actions and label videos.

To capture gameplay, the OpenAI team hired people to play Minecraft and then recorded their keyboard and mouse clicks along with the gameplay. With about 2000 hours of annotated video, the researchers trained a neural network model to recognize actions and label them.