Facebook Researchers Turn to Minecraft to Develop AI System that Can Handle Multiple Tasks

In a white paper published in July, a team of Facebook researchers led by Arthur Szlam laid out why and how they are using the popular game to train an AI system that can handle multiple requests. 

In the report, the researchers said Minecraft is an ideal choice to improve upon AI systems because it's easy to use and its sandbox nature make it easy to create AIassistants that perform multiple tasks.  Minecraft has emerged as one of the best selling video games ever released with more than 170 million copies sold. More than 90 million people play the game each month that Microsoft bought for $2.5 billion in 2014.  "Minecraft players can build things, gather resources, craft (combine resources), attack other players and mobs (nonplayer characters), and chat," wrote the researchers. "Even focusing only on building, the set of things a player could possibly do in the game is enormous." 

The Facebook researchers said they want to build a in-game assistant powered by AI that can perform whatever tasks players want to do in the game. It could be tasked with building simple or complex structures, creating entire cityscapes, breaking down structures, dancing and chatting with Minecraft creatures. The AI system would receive its commands via natual language using Minecraft chat. 

The goal with the AI system is to explore and evaluate different approaches to building a complex assistant, learning how to build a system that understands what a human wants and how to get the system to self improve.  A big knock against the current AI systems is that they can't learn new tasks beyond what they are trained for.  Minecraft could be the ideal environment to work on an AI system that improves on its own, the researchers said. 

Minecraft isn't the first game researchers have used to test AI systems with scientists turning to Starcraft, Atari Games, Doom and Text adventure games among others. But the Facebook researchers said many are built to study the development of reinforcement learnings based on algorithms and don't study languages with some, not involving humans at all.

"One crucial difference between all of these environments and Minecraft is that Minecraft is an engaging game that already has a huge player base, hopefully enabling the study of learning agents that interact with humans via natural language at scale," wrote the scientists. “Since we work in a game environment, players may enjoy interacting with the assistants as they are developed, yielding a rich resource for human-in-the-loop research."