Article Lead Image

Photo via (CC BY 2.0)

Robots are learning things in Minecraft, the end is nigh

And yet, I still can't build anything cool in Minecraft.


AJ Dellinger


Posted on Jul 14, 2015   Updated on May 28, 2021, 8:59 am CDT

There is no shortage of incredible things being built in Minecraft, but perhaps the most impressive creation that can be attributed to the game is taking place away from the screen. Professor Stefanie Tellex at Brown University is using the open-ended crafting game to teach computers to solve problems quickly and more efficiently.

The artificial intelligence we interact with today is still fairly limited. Think of the way you use Siri or your virtual assistant of choice: You ask it a question or prompt it to complete a single task, and that task must be concluded before tasking it with anything else. Generally, those tasks are pre-defined things that the A.I. is capable of. Give it a task outside of its wheelhouse and it’ll return the digital equivalent of a blank stare.

As computer learning advances and robots and machines are given more more open-ended commands, they will need a way to process less specific directions. Without contextual understanding, the robotic mind will run through every possible option before performing, lagging and using up considerable processing power.

That’s where Minecraft comes in. 

Tellex and her team of researchers used the blocky game to help create an algorithm that would help a robot strip away unnecessary paths that don’t match the criteria of commands. The researchers set up a scenario in Minecraft in which the on-screen character had to play a gold block in a furnace while avoiding a lava pit. 

Over time, the algorithm learned to eliminate certain possible actions that would not work toward accomplishing a goal. For example, setting down the gold block would not advance the mission. Nor would jumping in the lava, though it surely must have been tempting after playing through the same scenario thousands of times.

Utilizing the understanding of the context of its tasks, the algorithm was able to complete more complicated goals more efficiently. 

The team applied the algorithm to a Baxter robot from Rethink Robotics and tasked it with helping a person bake brownies. The bot was programmed to remove recipes that wouldn’t meet the requirements (i.e.: spaghetti is not brownies, so don’t put tomato sauce in the bowl) and was able to respond more efficiently because of the reduced load of information processing.

H/T Technology Review | Photo via (CC BY 2.0)

Share this article
*First Published: Jul 14, 2015, 3:45 pm CDT