Machine Learning Helps Robots Adapt

Robots are faster, stronger, and more precise than human workers. Industrial machines do more work in less time, and they can work far longer than people can. Robots don’t need breaks, they don’t get food poisoning, and they don’t take time off for holidays. Of course, there’s still one area where robots can’t compete with humans. For now, that is.

As great as robots are, they’re still terrible at adapting. They don’t like variables and changes. If something doesn’t fit within a strictly defined set of parameters, they can’t handle it. This is why sorting miscellaneous items in a bin remains one of the big hurdles for robots.

Researchers have developed a robot that can successfully perform back flips, but we still don’t have a robot that can sort our recycling. This could soon change, however, and machine learning could make it possible.

Brains over brawn

According to MIT Technology Review, researchers from UC Berkley have given us the world’s most dexterous robot ever.

So what makes this robot so good at picking things up?

The robot uses a suction mechanism and a fairly standard gripper to pick up items. It’s not how the robot grasps items that makes it so special, however.

It’s the software that makes this robot so good at sorting. We have machine learning to thank for giving us a super sorting robot.

Learning from past mistakes

The robot uses software called Dex-Net that uses a virtual environment to train neural networks. It simulates picking up different objects and learns through trial and error.

Dex-Net also allows the robot to pick up new items it’s never seen before by associating with an object it has already seen

The robot uses a high-res 3D sensor along with Dex-Net, and each of the robot’s two arms is controlled by a different neural network. It can tell whether an item should be picked up with its gripper or its suction mechanism.

So how good is it?

The researchers developed a metric called “mean picks per hour”. This measures the time it takes to pick up an item and the likelihood of success.

According to the researchers, humans can manage between 400 and 600 mean picks per hour, Amazon’s sorting robots manage 70 to 95 mean picks per hour, and the Dex-Net robot can swing 200-300 mean picks per hour.

This means that the new robot isn’t quite as capable as a human, but it’s as much as three times are capable as other robots. The lead researcher also predicts that robots will be as good as human, if not better, in the next 5 years.

A capable sorting robot would certainly shake up a number of fields including manufacturing and packaging industries.