An extraordinarily clever gripping robot which independently learns how best to deal with particular objects has been built by Professor Ken Goldberg and Jeff Mahler from UC Berkeley . What makes it special are not the grippers but the Dex-Net algorithms behind them, that the robot uses to calculate the best angle for gripping the object. It also moves and shifts products to allow it to recognize them better. Ultimately it learns through trial and error, and once it has identified an object it is able to abstract a class of similar objects from it. The robot stores the functioning solution for the next piece and thus steadily becomes faster and more efficient.
The current prototype consists of two arms that are independent of one another: a conventional gripper and a vacuum suction arm. The 3D camera identifies at lightning speed the object to be gripped, and the algorithm determines which is the best arm to use. On average the Dex-Net bot already operates at a rate of 200 to 300 grip operations per hour. Conventional automation solutions currently manage around 100 picks. For chaotic sorting tasks humans still far outclass the machines, at over 400 grip operations per hour.