The blockish and slightly dorky computer game Minecraft may turn out to be a great place for humans and AI to learn how to work together.
An experimental new version of the game, released by Microsoft researchers this month, can be used to train an AI to perform all sorts of tasks, from crossing bridges to building complex objects. The new platform, called Project Malmo, makes it possible for a learning algorithm to control a Minecraft character that’s normally operated by a human player. But it also provides ways for human players and AI agents to work together, and a chat window through which a person can talk with a nascent AI.
“In the long run I want to work toward AI that can be taught by any user to help them achieve their goals,” says Katja Hofmann, a researcher at Microsoft Cambridge in the U.K. who leads the project.
Hofmann, who gave a demo of the software to AI researchers at an academic conference in New York last week, says that human-AI collaboration is a key goal for the project: “We’ve built in all the capabilities that a researcher would need in order to work toward collaborative AI.”
Malmo is geared toward testing reinforcement-learning algorithms, a way of training a computer to perform a task by providing simulated rewards. It is possible, for instance, to use reinforcement learning to train an AI controlling a Minecraft character to navigate through a room filled with obstacles by providing a reward when it reaches the other side. A human Minecraft player could take part in that process by offering helpful instructions, which the AI would also learn to recognize over time.
Increasingly clever machine-learning algorithms have the potential to make people more productive and efficient, an idea that Microsoft’s CEO Satya Nadella has emphasized as especially important for his company. However, relatively little research emphasis has been placed on getting humans and AI to team up.
Hofmann believes the work might eventually feed into the regular game. “I could imagine that in the future you train your AI agent to take over the boring parts,” she says.
She says the development of AI that helps Minecraft players may be the first step toward computer software that augments people’s abilities in everyday settings. “What we really should be developing is something that is genuinely useful and empowers people,” Hofmann says.
AI researchers have turned to computer games as a proving ground for some of the most sophisticated machine-learning approaches around. Early last year, researchers at Google DeepMind made a splash by demonstrating an algorithm that could learn to play various 2-D Atari games to a superhuman level. Their algorithm used a combination of deep learning—a very large neural network trained to respond to an input—together with reinforcement learning. The company’s researchers have since shown video clips of algorithms learning to navigate more complex, 3-D game environments, but they have yet to publish any work on this.
Minecraft, an indie game acquired by Microsoft in 2014, became hugely popular despite its no-fuss graphics and its lack of any obvious objective. Fans of the game have used it to build an incredible array of complex structures and machines (see “The Secret to a Video-Game Phenomenon”).
Although Minecraft is relatively simple, it can provide a useful testing ground for robot algorithms (see “Minecraft Shows Robots How to Stop Dithering”).
Stefanie Tellex, a professor at Brown University who has used Minecraft for this purpose, says Malmo will be very useful because it will make it easier for researchers to compare their approaches. “The idea of a having Minecraft as a kind of AI Olympics is really great,” she says. She adds that the platform will offer an efficient way to gather large amounts of data, for example concerning human-AI interaction, that is needed for modern machine learning.
While Malmo is primarily aimed at those working on artificial intelligence, machine learning, and robotics, anyone with sufficient technical savvy can download it and experiment with an in-game AI. The software comes with several machine-learning packages and example AIs. Indeed, some of those who have so far downloaded the game are hobbyists and software engineers who do not specialize in AI.
The big new idea for making self-driving cars that can go anywhere
The mainstream approach to driverless cars is slow and difficult. These startups think going all-in on AI will get there faster.
Inside Charm Industrial’s big bet on corn stalks for carbon removal
The startup used plant matter and bio-oil to sequester thousands of tons of carbon. The question now is how reliable, scalable, and economical this approach will prove.
The hype around DeepMind’s new AI model misses what’s actually cool about it
Some worry that the chatter about these tools is doing the whole field a disservice.
The dark secret behind those cute AI-generated animal images
Google Brain has revealed its own image-making AI, called Imagen. But don't expect to see anything that isn't wholesome.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.