Get best of the week

Warehouse robot learns to sort non-standard things

In a warehouse near Berlin, a new robot automates tasks that were only recently inaccessible to machines.




In a warehouse in the backyard of Berlin, a long row of blue boxes with switches, sockets and other electrical goods advanced along a conveyor belt. After they stopped, five workers sorted out these small items, and laid them out in cardboard boxes.

At Obeta, a manufacturer of electrical goods, which opened in 1901, workers have been doing this kind of monotonous job for years.

However, a few years ago a new employee joined the team. The robot behind the protective glass, using three vacuum suction cups at the end of a long arm, does the same job, sorting goods with amazing speed and accuracy.

You may not be impressed, but such a robot sorting components is a major breakthrough in areas such as AI and the human labor that machines are capable of.

Millions of goods pass through the warehouses of retail stores such as Amazon, Walmart, etc., and their low-wage workers have to sift box by box, containing all sorts of things - from clothes and shoes to electronics - so that every product can be packaged and send as directed. And until today, cars could not handle it.


More than 80,000 such boxes are stored in an Obeta warehouse in a suburb of Berlin.

“I have been working in logistics for over 16 years, and have not seen anything like it before,” said Peter Pachwein, vice president of Austrian warehouse automation company Knapp.

California engineers who made this robot photographed the surrounding environment on smartphones, standing next to the Obeta warehouse. They spent more than two years developing the system at their Covariant.AI startup, based on previous research from the University of California at Berkeley.

Their technology demonstrates that in the near future there will be very few tasks in warehouses that are too trifling or difficult for robots. And the greater the number of tasks that are traditionally performed by people, machines take possession of, the more there are reasons to worry about warehouse workers losing their jobs due to automation.

Economists believe that due to the rapid growth of online trading - despite the fact that most companies are unlikely to master the latest automation technologies very quickly - all these technologies are unlikely to lead to a sharp drop in the number of jobs in logistics in the near future. However, the engineers creating these technologies acknowledge that the day will come when machines will perform most of the tasks in the warehouses. Living workers will have to do something else.

Covariant engineers specialize in such a sub-area of ​​AI as reinforcement learning. Machines are trained to perform new tasks independently, through a huge number of trial and error. And the best place to teach them is the real world.

“If you want to promote AI while sitting in the lab, you can’t do this,” said Peter Chen, director and co-founder of Covariant. "There is a big gap between the ideal and the real world."


A factory worker works with a robot. So far, the robot only automates the packaging station.

Warehouses are already highly automated. In this warehouse near Berlin, inside a fenced-in room larger than a football field, other robots have long been used to remove large boxes from high shelves.

However, this task for the car is relatively easy. Engineers can program the robot to repeat the same movement. All boxes are the same. The robot can take them each time making the same movement.

Sorting baskets with random items is another matter. Their shape and surface vary. Some switches may lie face down, while others may be reversed. Another product can be packed in a plastic bag that reflects light in a way that the robot has never encountered before. This required human participation.

It is impossible to program the robot arm to handle any situation by writing many rules into the program. For many years, Knapp Pachwein and partners have been trying to create a robot with the right dexterity and flexibility, and failed.


Pachwain Knapp has been trying for many years to invent a robot sorter

Covariant, working with Knapp, has created a program that can learn by trial and error. At first, the system was trained using a digital simulation of a task - a virtually recreated basket with random objects. Then, when Chen and colleagues transferred the program to a robot, he was able to pick things up in the real world.

The robot is able to continue learning while sorting things never seen before. A robot from a German warehouse is able to select and sort over 10,000 different items, and with 99% accuracy, according to Covariant.

And this is a sign of significant changes in areas such as online retail and logistics.

At the end of last year, the international robot manufacturer ABB held a competition. He invited 20 companies to develop software for his robotic manipulators, capable of sorting random objects, from cubes to plastic bags containing other objects.


Robots on rails are looking for the necessary cargo to send it for packaging.

Ten companies were from Europe and the other half from the USA. Most and did not close the task. Some were able to handle almost all the tasks, but could not cope with the most cunning examples. The only company that was able to cope with all the tasks as quickly and efficiently as people was Covariant.

“We tried to find weaknesses,” said Mark Segura, managing director of the service robot division at ABB. “Reaching a certain level in these tests is quite easy, but it’s very difficult not to demonstrate a single weak point.”

Knapp, which helped to implement the system near Berlin, and ABB, believe that this technology can be used in other similar warehouses.

Engineers from Covariant believe that their robots, constantly practicing, will better cope with the tasks. While the robot in one of the warehouses is learning more convenient ways to lift certain objects, this information enters the central brain controlled by Covariant, and this will allow the machines to work even better.


Dirk Jandura, Managing Director of Obeta, said that such companies are very active in improving efficiency. Automation is the key to reducing waste.

Like many warehouse operators, Obeta had problems finding workers who want to do monotonous work. Each sorter processes about 170 orders per hour, about three per minute, eight hours a day. In summer, the temperature in the warehouse exceeds 38 degrees. It is difficult to keep workers longer than six months.

For Obeta, the new robot is the perfect solution. The work of three people is performed by one robot that does not know fatigue.

“He does not go to smoke, is always healthy, does not chat with neighbors, does not take breaks on the toilet,” said Zhandura. “He's more effective.”

Knapp is also considering warehouse projects where robots work in place of people, which will allow for denser placement of packages that robots will then pick up.

“New warehouses will be built with an eye on AI robots, not humans,” said Pachwain.

Knapp plans to make it difficult for companies to refuse to replace people with robots. Pachwain said they would take from companies the amount that would always be less than the salary of a worker. If the company paid the worker $ 40,000 a year, then Knapp will take $ 30,000, he said.

“We’ll just go down,” he said. - This is our business model. And it will be easy for the client to make a decision. ”


Peter Chen and Peter Abbeel, founders of Covariant.AI

Beth Gutelius, the first assistant director of the Center for Urban Economic Development at the University of Illinois in Chicago, who studied the impact of automation on work, said that such technology is unlikely to cause changes in the labor market in the near future.

She said that a more serious problem would be that when people start working together with robots, they will be judged differently. “After we begin to compare the speed and effectiveness of people with robots, a whole new set of health and safety problems will appear,” she said.

Peter Abbil, a professor from Berkeley and co-founder, president and chief scientist of Covariant, said that people will continue to work together with machines in such warehouses. However, he acknowledged that the labor market will move significantly with improved machine learning.


Truck loading at Obeta's warehouse in Germany

“If this happens in 50 years, the education system will have plenty of time to pull itself up to the state of the labor market,” he said.

In a German warehouse, a woman in a baggy T-shirt diligently sorts goods in boxes, occasionally glancing at English-speaking visitors, taking photos of the robot and admiring its effectiveness.

An engineer from Covariant approached the group to share information about how the robot completed more than 200 orders in the last hour - if it were a human, it would receive a bonus.

More articles:


All Articles