This robot learns to pick up mugs by first learning a theory of mugness

For all of the recent progress in machine intelligence, robots still struggle to adapt relatively simple tasks to new situations. Take, for example, picking up a mug and hanging it on a mug rack; even small changes in a mug’s shape, size, color, and orientation can throw a robot off.

In a new paper, researchers at MIT are now proposing a new technique for helping robots generalize their learning with relatively little data. They do so by training a neural network to extract just a few key points from a mug or other object that needs to be picked up and placed, giving the robot a visual road map for how to grasp and orient it. During testing, they found that the bot only needed three key points for a mug—one on the center of its side, one on the bottom, and one on the handle—and six key points for a shoe.

Unlike previous techniques that require hundreds or even thousands of examples for a robot to learn to pick up a mug it has never seen before, this approach requires only a few dozen. The researchers were able to train the neural network on 60 scenes of mugs and 60 scenes of shoes to reach a similar level of performance. When the system initially failed to pick up high heels because there were none in the data set, they quickly fixed the problem by adding a few labeled scenes of high heels to the data.