|By Kramer Phillips | 2 years ago|
University of California researchers have developed a robot that can figure out on the spot how to pick up an unfamiliar object with no pre-programming required. According to the researchers, this marks a vital step toward building robots that are truly intelligent and capable of self-navigation.
“It doesn’t actually even know anything about that the object is, ” roboticist Ken Goldberg, who developed the robot with post-doc Jeff Mahler, told Wired. “It just says it’s a bunch of points in space, here’s where I would grasp that bunch of points. So it doesn’t matter if it’s a crumpled up ball of tissue or almost anything.”
Picking up an unfamiliar object is a simple task for us humans because our brains can assess the object’s shape and dimensions and instantly determine how our hands should hold it. But a typical robot’s brain can only operate with pre-loaded instructions. When it sees an object it has been programmed to pick up, it can do so. An unfamiliar object will leave it stumped.
This University of California robot’s brain is not so limited. It has an installed library of “point clouds” that represent objects in three-dimensional space and can reference these fields of points to get its bearings. It also has a Microsoft-built sensor that views its surrounding environment and the object in front of it. The sensor and the cloud library work in tandem to analyze what is in front of the robot and calculate what it should pick up and how.
“The way it’s trained is on all those samples of point clouds, and then grasps,” Goldberg told Wired. “So now when we show it a new point cloud, it says, ‘This here is the grasp, and it’s robust.’”