A team of engineers from MIT and the University of Toronto released a paper last week detailing their findings on a new virtual program that could help teach machines to perform household chores. Dubbed “VirtualHome,” the system itself is a Sims-style world that features at least eight distinct home environments, such as a kitchen, bathroom and living area. Within this universe, the team’s artificial agent can be instructed to perform over 1,000 basic tasks like “get the newspaper” or “take out the garbage.” Each task is laid out for the agent individually, raising the possibility that VirtualHome could someday be used to teach robots individualized chores.
Spelling It Out
While humans intuitively understand how to complete the micro-tasks that go into accomplishing more complex activities, artificial forms of intelligence lack this natural reasoning ability. For a machine to learn how to go get a cup of coffee, for example, it really needs to learn something like 10 individual steps—grabbing a coffee cup from the cabinet and turning on the coffee maker won’t come naturally, so they would each need to be spelled out.
To facilitate this type of learning, the researchers put together verbal descriptions of thousands of subtasks. They then coded the tasks to play a part in each simulated sequence that went into a larger activity. The VirtualHome synthesized these basic programs into videos in which the artificial agent actsout commands on the screen. The collection of coded verbal descriptions of subtasks amassed by the researchers can be thought of as the “raw material” that future robots could learn in order to string together complete activities.
Robot Butlers: Coming Soon?
VirtualHome is important in that it demonstrates that artificial agents can put together an entirely new task from scratch. In this way, simple building blocks of verbal instructions can be pieced together by artificial intelligence to form an understanding of a more complex task.
This development opens the door to the mass production of robots as blank slates—assistants that can be taught on a more personalized basis rather than programmed by a manufacturer. The ability to customize tasks could be the key that makes widespread domestic implementation of robots feasible. Going forward, the team speculates that its simulation may not even be necessary. This system of learning small actions could be applied visually, with robots simply watching a human perform a task and reasoning through the mini-steps that are needed to get there. That day may still be far off, but VirtualHome is an important theoretical building block toward a world of robot helpers.
For more on the developing trend of robots learning and retaining for themselves, check out this article.