In the future, the team hopes to expand Minecraft worlds to help robots develop even more real-world skills. In their paper, the team reported that robots showed “improvements in planning.” They found that once they had trained in Minecraft, robots were able to anticipate human needs and do things like hand them whisks so they could beat eggs. Next, the team tested the Minecraft-honed robot brains in the real world, asking robots to help people make brownies. When the robot algorithms played the game, they had to figure out how to do things like build a bridge or find buried gold - activities that helped the algorithms learn enough to move on to new domains to try out their new skills. To put their robots’ algorithms to the test, Tellex and her team created small Minecraft domains and gave characters simple tasks to solve. She notes that since it’s cheap and open-ended, the game was the perfect way for her team to test problem-solving algorithms and collect plenty of data in the process. “Minecraft is a really good model of a lot of these robot problems,” computer scientist Stefanie Tellex said in a release about the research.

The game, which has been purchased (by humans) nearly 20 million times and gained over 100 million users since its release in 2009, uses sandbox-style play that lets users build their own worlds to explore. That can be tough, the authors write, because of the “exponential number of ways a robot can affect the world.” New research by computer scientists at Brown’s Humans to Robots Laboratory is focusing on how to help robots better plan complex actions, even in a quickly-changing environment. That’s changing, though, and scientists at Brown University are using the virtual world of Minecraft to help train robots to “think” in real-world environments. Humans are really good at deciding what to do next.

Robot herobrine