The publication notes that the system as a person determines how to lift objects. The robot sees the object, understand its form and compares it with other objects raised in the past, and uses this information to choose the best way to capture things.
The developers have made the system more than six million 3D images of the objects, indicate in TechCrunch.
Dex-Net presented tens of objects that the robot has never seen. The machine was not able to select only one object. “The system is quite reliable, despite the fact that it was trained on artificial data”, — notes the edition.
TechCrunch adds that Dex-Net requires less than a second to choose how to choose the subject.