Researchers at UC Berkeley have developed and unveiled a first in robotics: a robot that matches two highly versatile limbs with the ability to reason and simulate different outcomes via two onboard neural networks.
The university team claims that this combination of technologies makes it the world’s most dextrous robot.
Professor Ken Goldberg and one of his graduate students, Jeff Mahler, displayed the results of their work at EmTech Digital, an AI event in San Francisco organised by MIT Technology Review.
The two-armed robot relies on software called Dex-Net, which gives it the ability to reason and make decisions with as much dexterity as its arms are capable of moving. Via this onboard system, the robot is capable of quickly determining how best to grasp objects, based on simulations that take place within two separate, deep neural networks.
The robot can be seen in action here.
Read more: Ocado bots offer safe pair of hands for packing shopping
Mean picks per hour
The potential for dexterous robots in industrial, commercial, and other settings is obvious. That’s one of the reasons why companies such as Ocado are pouring resources into developing sophisticated warehouse robots.
To move on from looking at dexterity in a simple or binary sense, the Berkeley researchers use the preferred metric of ‘mean picks per hour’. This is calculated by multiplying the average time per pick with the average probability of success for a consistent set of objects.
The Dex-Net system can determine how to grasp an object based on what it has seen before. It can even nudge an item to gain more insight into how to handle it.
The system has gone through several iterations over the past year. The latest version combines a high-resolution 3D sensor with two arms – one with a robot gripper, and the other with a suction system. Each is controlled by its own neural network. The Dex-Net software scans an object and uses both networks to determine which approach is best for that particular object.
The new machine is already capable of between 200 and 300 mean picks per hour, according to Goldberg. That’s substantially more than the winner of a recent Amazon robotics contest. Humans are capable of between 400 and 600 mean picks per hour, so the machines are fast catching up.
Internet of Business says
While the mass-media has grown more hostile to robots and AI in recent months – and arguably more hostile to technology overall – robots’ ability to work quickly, safely, and non-stop among human beings is fast developing. ‘Cobots’ are rapidly becoming smart, programmable platforms, rather than dumb, single-task machines – following the smartphone model of the apps being the most important factors, not the hardware.
And while many people may be misinformed about the threat to humanity from intelligent, dextrous machines, most will enjoy the lower costs, faster services, and cheaper products that automated production will bring.
However, buy-side organisations need to consider the application of robotics carefully from both a strategic and an operational angle. Flippy, the burger-flipping robot in the US, was removed from service after less than one week because it was working so fast that human employees were unable to keep up with it.