While robots are excellent at doing things they have been programmed for, they are horrible at improvising a solution to an unfamiliar issue. DexNet is a special AI system that has been designed to augment a robotic arm’s capacity to handle objects without prior training. It is the hope of DexNet’s programmers that their system will allow robots to solve a physics-oriented problem with full competence, without a trace of hesitation and faster, if not as fast, than a human could ever hope to achieve.
The underlying framework of DexNet’s artificial intelligence, referred to as “deep learning,” is that it learns to grip objects in the same manner that humans learn to do so; we notice an item, get an understanding of its particular shape, compare and contrast that object to objects we remember from past life experience and then combine those details in order to hypothesize the best approach to manipulating the item. Unlike humans, who have the benefit of vision and memory, DexNet’s creators have uploaded over six million computer simulated objects into their system, complete with the best ways that those objects can best be manipulated. The execution of this system results in the robot analyzing an object, comparing it to the database of objects in DexNet’s system and then proceeding with the most recommended approach.
In their test work, the researchers, a group of roboticists at Berkeley University, presented the robotic arm with a variety of objects it had never encountered before, such as rubber ducks, spray bottles and shoes. Despite being presented with a wide array of materials, the DexNet-connected arm only failed to correctly grip a single object. This rate of failure would seem to indicate that the system is quite robust for one built on synthetic data. Additionally, the arm was able to decide on a proper grip in less than a second, an average speed nearly three times as fast as the initial version of DexNet.
DexNet’s research team is set to present their latest iteration of the system, DexNet 2.0, at a July conference. They also have plans to publicize their collection of object data and point clouds. Their intended application for this pursuit is in the field of industry, particularly within supply chains and manufacturing.