Brown U Uses Minecraft To Train Robots
Researchers at Brown University are
developing a new planning algorithm for use in robots and they are using the
videogame Minecraft to test and refine it.
According to information from the university, basic action planning such as
taking out the trash is particularly difficult for robots because they have
trouble distinguishing objects or actions relevant to the task at hand from
those that are irrelevant, causing them to do things like turn on the stove in
the process of taking out the trash. The more complex the environment, the
greater number of irrelevant objects or actions, resulting in what computer
scientists call "state-space explosion."
Stefanie Tellex, an assistant professor of computer science at Brown, and
several members of Tellex's lab at
Brown are developing an algorithm to help robots filter out irrelevant
information in order to complete a task efficiently. The algorithm uses
"goal-based action priors" which are "sets of objects and actions in a given
space that are most likely to help an agent achieve a given goal," according to
information from the university. People can program robots with the priors for
a given task, but the algorithm Tellex and her team are developing is intended
to help robots learn the priors for a task through trial and error and then
apply them in a novel situation.
The researchers discovered that they could use the videogame Minecraft, a
virtual world made up of three-dimensional blocks where players can gather
resources and build or destroy structures to accomplish goals, to test the
algorithm. The researchers created a tiny space in Minecraft and a character
controlled by the algorithm, and then let the algorithm complete a task in the
game through the process of trial and error. Once the algorithm had discovered
the priors, the researchers placed it in a novel Minecraft space. "Indeed, the
researchers showed that, armed with priors, their Minecraft agents could solve
problems in unfamiliar domains much faster than agents powered by standard
planning algorithms," stated a news release from the university.
According to Tellex, the test demonstrates that goal-based action priors are
a viable strategy to help robots function in unstructured environments, and
that virtual worlds like Minecraft are useful tools for testing and refining
these types of robotics algorithms.
Eventually, the researchers said they would like the algorithm to be capable of
completing tasks in larger Minecraft spaces, and eventually the entirety of the
virtual world. "The whole of Minecraft is what we refer to as 'A.I.
complete,'" said Tellex in a prepared statement. "If you can
do all of Minecraft you could solve anything. That's pretty far off, but
there are lots of interesting research objectives along the way."
David Abel, a graduate student in Tellex's lab will present a paper
about the team's work on goal-based action priors at the International
Conference on Automated Planning and Scheduling. Other authors of the paper
were David Ellis Hershkowitz, Gabriel Barth-Maron, Stephen Brawner, Kevin
O’Farrell and James MacGlashan.
About the Author
Leila Meyer is a technology writer based in British Columbia. She can be reached at [email protected].