Building a “universal” robot is tricky. One part of the challenge is software: making the robot understand a range of objects and interaction scenarios. Another is physical: giving the robot the ability to actually manipulate both delicate and heavy objects with different textures. The usual solution is to try and design a very complex end effector that can handle everything. But, a far more practical solution is to just give an ambidextrous robot two types of end effectors.
This research comes from a team of engineers at the University of California, Berkeley, who were working on way for robots to efficiently handle a large variety of objects types. In a warehouse setting, robots need to be able to sort through and place a lot of different kinds of objects. A shrink-wrapped box, for example, needs to be handled different than a loose t-shirt. Normally, that would require a complex, and therefore error-prone, end effector. But, it turns out a seemingly-obvious tactic is much better: give the robot two arms.
Dex-Net 4.0 has two arms, and one arm could have a suction cup end effector for sealed objects, and another arm could have a gripper mechanism for loose objects. That could also be expanded to include any number of arms, each with their own end effector. Using a training data set, the robot learns to identify which end effector should be used for a certain kind of object. During testing in a warehouse pick-and-place setting, the ambidextrous robot was able to perform more than 300 picks per hour, and could do so with 95% reliability.