While businesses like Amazon are using robots to move things around their factory floors, getting them to identify and then actually pick up awkwardly-shaped objects “remains a difficult challenge” because most everyday objects aren’t uniform.
Berkeley scientists however, have invented a robot that does just that. Described by its inventors as “the most nimble-fingered robot ever built,” DexNet 2.0 is able to pick up any object, regardless of its shape, even those it’s never seen before. To help DexNet do that, scientists created a database of thousands of irregular 3D objects and included data on where best to pick them up. That created 6.7 million data points for the robot to draw on. When looking at a new object, the robot rifles through the database and makes a guess at how best to pick it up. If it’s less than 50% sure it can pick the object up, it sort of nudges it around until it gets a better angle.
In a test with a number of objects the robot had never seen before, DexNet 2.0 managed a pick-up 99% of the time, surprising even the scientists.