Taking artificial intelligence to the next level.
Google wants to teach a computer to master the sci-fi video game StarCraft.
Such a win would be a big breakthrough for artificial intelligence, but it’s likely years away because of the complexity required to master the game. Although DeepMind has had success creating AI software that has performed well at playing classic Atari games, StarCraft poses a much more difficult challenge.
StarCraft 2, released in 2010, is one of the most popular video games in the burgeoning e-sports market, in which people compete against one another in live video game competitions. The video game and its predecessor, StartCraft, are widely considered by gaming enthusiasts to be among the best and most challenging strategy games of all time.
DeepMind, which Google goog bought in 2014 for $650 million, has gained notoriety for creating software that defeated the world’s leading players of the Chinese board game Go. Demis Hassabis, the CEO of DeepMind, said in May 2016 that his team considered StarCraft as a next big test to spur advancements in artificial intelligence.
Get Data Sheet, Fortune’s technology newsletter.
Video games are important testing grounds for researchers to examine how AI software reacts and learns from virtual environments. For example, Apple’s aapl director of AI research, Ruslan Salakhutdinov, recently detailed how Carnegie Mellon researchers have improved their AI-infused software using the 1990’s first-person-shooter video game Doom.
As a trio of DeepMind researchers explained in a blog post, it’s important to test AI software in games “where humans play well” so they have a good baseline for comparing the capabilities of their technology.
StarCraft 2 poses significant challenges for AI. Winning requires doing tasks like gathering resources, building structures, maintaining armies with different kinds of troops, and scouting unfamiliar terrain in other planets. Winning in the game typically involves amassing a big military and defeating opponents while surviving their attacks.
That’s a big challenge for AI, considering it’s hard enough for researchers to create technology that can excel in just one task, like Google’s use of deep learning to identify pictures of dogs in photos. So far, DeepMind’s researchers said their software has learned to master a few individual tasks like collecting minerals, but then fails when playing the full game.
To help create more powerful AI software, DeepMind and Blizzard are releasing AI software tools to the public so that outside researchers can develop their own AI software that can be used in the computer game. The two companies also co-authored a research paper that outlines the project and describes the AI techniques they’re using.
The hope is to enlist outside help and get data from the public that DeepMind can use to improve its software.