Minecraft is known as one of the most popular games on the market right now. The open world is bigger than Neptune (Neptune can fit 57 Earths in it and still have room left over). Players explore, create extravagant buildings, and use the game without limit. The newest use is training robots how to prioritize.
Robots can’t ignore irrelevant objects and actions when it comes to basic action planning.
As Futurity described it, “if someone asked you to empty the trashcan in the kitchen, you would know there’s no need to turn on the oven or open the refrigerator. You’d go right to the trashcan.” Robots consider everything around them before doing their plan, and would turn on the stove and open the fridge.
When it comes to complex environments, robots can suffer from when computer scientists refer to as “state-space explosion”. That’s when there are too many objects within a robot’s space that it confuses them.
To fix that, Stefanie Tellex, assistant professor of computer science and her team of researchers at Brown University in Providence, Rhode Island, are working on an algorithm to help properly train robots to prioritize tasks. The algorithm uses “goal-based action priors” which are sets of objects and actions that are most likely to help complete tasks.
The priors can be given by an operator or learned through the algorithm itself by trial and error. Tellex told Futurity that Minecraft is great for solving these problems. “There’s a huge space of possible actions somebody playing this game can do, and it’s really cheap and easy to collect a ton of training data. It’s much harder to do that in the real world.”
Tellex and her team made small domains within their own Minecraft mock-up and put a character in to solve problems using the algorithm like mining gold or building a bridge. After successfully completing the task, the character was put into a different domain they hadn’t seen before to solve another problem and test if it would apply what it learned. The researchers found that using the right priors, the tasks would get done quicker than with the regular algorithm. Using the real Minecraft, the researchers tested the algorithm in larger spaces, and eventually the entirety of the game.
After successfully using the best actions to get the job done, the robot was put in the real world to help a human make brownies using the same algorithm. Tasks included things like understanding that eggs need to be beaten, so when the robot saw eggs, it would anticipate handing the cook a whisk.
Tellex sees the goal-based action priors as a method to help robots solve problems outside of a controlled environment.
Until then, you’ll have to make your own brownies.