We all know that robots can pick up objects. What’s perhaps not so obvious is that these objects normally have to be in a well-defined or predictable orientation. Some robots can compensate for position changes via a vision system, but for the most part they need to be in exactly the same spot. This means that enabling a robot to work properly involves complicated auxiliary feeding equipment, and inevitable human interactions as well as lost production when something doesn’t fit into a robot’s parameters.
A new system, however, under development at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has been able to turn this concept on its head, allowing a robot to inspect objects from several angles in order to understand what they are seeing. Based on inspection, objects are resolved into collections of data points, known as Dense Object Nets, or DONs. The system then uses these points to determine how to grip the identified object, which may be presented in a variety of different orientations.
Experiments shown below use a typical industrial-style robot for manipulation, and could certainly have applications in a manufacturing setting. In addition to this, with a generalized object inspection/pickup algorithm, it would even be possible to program a robot for more general tasks, such as putting dishes away or cleaning up a desk.