Using a new technology, an AI even learns to make diamond pickaxes in Minecraft. But the way to get there is quite complicated.
You like playing Minecraft, but don’t really have the time to mine thousands of blocks of stone and wood for your next monumental structure? Then you might be in luck, because in the future, a well-trained AI could do it for you OpenAI, a research company that studies artificial intelligence, is responsible for this. OpenAI is known for developing (DALL E 2), an AI that creates images based on text input.
A (free variant of DALL E) By the way, you can try it out for yourself, the results can be funny, but also quite disturbing. But back to the topic: The researchers’ goal was to teach the AI actions not through speech input, but through video footage. Find out how they managed to do this and what the artificial intelligence is capable of here.
Fed with 70,000 hours of gameplay
To teach the AI Minecraft, the researchers took several steps – you can find the full report in their (blog post). First, they had to make sure that the AI could learn from video material from YouTube at all. This is because the actions performed there are not specifically described: The AI cannot know which buttons, mouse clicks and movements players use to move, open their inventory and dismantle blocks.
That’s why we first recorded our own gameplay with the respective mouse and keyboard inputs. With this data, a Inverse Dynamics Model
was trained to predict the next action in the video. It was able to access both past and future frames of the video. The model was then in turn used to match 70,000 hours of unscripted video footage with the appropriate mouse and keyboard input.
KI even builds diamond pickaxes
Now you could feed the AI the processed video footage and have it mimic the behaviour demonstrated in the gameplay. For example, it managed to chop down trees, make boards out of them and eventually a workbench. This may not sound impressive at first, but it requires about 1,000 inputs from the player, all of which have to be mimicked by the AI.
The artificial intelligence learned even more human
skills, such as swimming, hunting and eating animals and even so-called pillar jumping. But this was far from the end for the researchers.
The AI now had some basic abilities, but was to learn how to use them more specifically by means of fine-tuning. So it was presented with gameplay from newly launched Minecraft worlds in order to strengthen skills from the early game. As a result, the AI even learned to make wooden and stone tools, to build (really very) rudimentary shelters and to loot chests in NPC villages.
With the help of another method of fine-tuning, reinforcement learning, it was even possible for the AI to craft a diamond pickaxe at the end! And that, although it carelessly dug straight down! For comparison: While a workbench required about 1,000 actions, humans need about 24,000 actions to craft a diamond tool, all of which were carried out by the artificial intelligence.
What’s the point of all this?
As you can probably guess, the research is not about practical AI companions for stressed Minecraft players. Rather, artificial intelligences are supposed to learn actions based on the numerous videos available on the internet. The researchers write in their blog entry that VPT (their technology) offers the exciting possibility of learning large-scale behavioural patterns directly in areas other than language.
Minecraft is ideally suited for these experiments because it has a very generic interface with mouse and keyboard, and offers a very open world with many possibilities. The results achieved there can be easily transferred to similar areas, such as the general operation of computers.
What do you think of the AI researchers’ Minecraft experiment? Do you find artificial intelligences boring and don’t expect much from them or do you believe in their future importance? Feel free to write us your opinion in the comments!
