UC Berkeley’s Laboratory for Automation Science and Engineering, also known as AUTOLAB, has developed Dexterity Network, or Dex-Net, a software program that is trained to detect and pick up objects.
AUTOLAB, directed by professor Ken Goldberg, department chair of UC Berkeley’s Industrial Engineering and Operations Research department, or IEOR, is a center for research in robotics and is known for projects such as “the Grabber,” “the Picker” and “the Bed Maker.”
The Picker, however, is the only robot to use Dex-Net, which allows it to pick up and place objects — even rigid objects such as tools, household items, packaged goods and industrial parts, according to the Dex-Net website.
“It’s got a 3D sensor and is using that to build a model to analyze that image to collect its grab point — the best place it thinks it can grab the object,” Goldberg said. “It then uses one of its arms to grab and place the object.”
The researchers also created a new way of measuring the performance of Dex-Net — “mean picks per hour,” also known as MPPH. MPPH calculates the time per pick and multiplies it by the probability of success for a set of objects, according to an article by MIT Technology Review.
Dex-Net is far ahead of the competition, according to Goldberg. During the 2017 Amazon Robotics Challenge, the top robots achieved 70 to 95 MPPH, according to an article by MIT Technology Review. Dex-Net has more than double the MPPH of those robots, but is still far from the speed and efficiency of humans.
“Humans can be 400 to 600 (MPPH) and (Dex-Net) is 200 to 300 (MPPH),” Goldberg said. “It’s quite a gap and that’s what we’re working on, to improve Dex-Net to have better MPPH.”
Goldberg said Dex-Net is “state of the art” in the way that it pushes the boundaries of what a robot can do. In 2,192 trials with a data set of 50 novel objects, Dex-Net had a success rate of 94 percent and could place 10 objects in less than three minutes, according to Dex-Net’s website.
Goldberg said he believes Dex-Net will mainly be used to complete shipping orders as well as the maintenance and housekeeping of warehouses. Researchers are concerned about objects being placed incorrectly and too slowly, according to Keith McAleer, spokesperson for UC Berkeley’s IEOR department.
McAleer added that with robots being able to efficiently grab objects, they can potentially grab, place and package orders, making supply chains fully automated. Robots such as these can also help children clean up and help the elderly, according to an article by Fast Company.
“It’s a step toward making robots more capable,” Goldberg said. “It should have an impact on e-commerce. We think that robots should be working with humans, not replacing them.”