Innovation

Amazon's robot worker challenge won by AI-powered suction arm

Amazon's annual contest to build bots that can pick items from shelves in its fulfillment centers sees the robots' capabilities come on leaps and bounds.

the-winning-robot-of-team-delft-at-apc-2016.png
The Delft team robot that won this year's Amazon Picking Challenge.
Image: Amazon
While humans are undoubtedly better than robots at all sorts of tasks, slowly but surely the bots are catching up.

Helping narrow that gap is retail giant Amazon, which this past weekend funded a contest in Leipzig, Germany, where robots tried their hand at picking items from shelves in a mock Amazon warehouse.

Compared with human pickers, the performance of the robot arms in this year's competition looks sub-par. They are slower to spot items with their cameras and more likely to drop them with their grippers and suction cups. But compared with the robots that competed in last year's event they excelled.

"It was inspiring to see 16 top teams with so many different approaches to the same problem. And we also saw the advancements robotic technology has made since last year," said Tye Brady, chief technologist at Amazon Robotics.

Even though this year's competition was tougher than 2015's, the new champion bot picked items roughly three times as quickly as last year's winner. And while half of the 26 robots competing in the inaugural Amazon Picking Challenge (APC) in 2015 failed to score a single point, only four crashed out of this year's contest.

Like the previous year, bots were scored on their ability to correctly select individual items from shelves, with more points awarded for picking out items mixed in with other objects. These items represented a cross section of products commonly found in Amazon's warehouses, and ranged in shape and size from clothing to a toothbrush.

Bots in this year's contest performed more effectively, with just under half scoring more than 40 points, a feat that would have landed them third place in last year's challenge.

This leap forward occurred despite this year's challenge being more difficult. Bots were scored less generously and had less time to rack up points, as well as having to pick from 40 different items, compared to 27 last year. Given the harsher contest, the winning bot this year scored fewer points than the first-placed machine in 2015 — managing a still impressive 105 compared to 148 the year before.

The bots' growing prowess will be of interest to the warehouse staff that make up a large proportion of Amazon's 230,000-strong global workforce.

However, Amazon is quick to point out the competition is not about replacing these workers — it's about augmenting the workforce and allowing them to do more.

As proof that bots will not supplant people, the company points to the growth in employee numbers since it began using Kiva robots to carry shelves of products about its warehouses, known as fulfilment centers.

"Robotics enhance the job for employees but does not replace them," said an Amazon spokesman.

"In fact, we continue to hire. Many of those roles are being created in buildings where employees are working alongside Amazon robotic drive units."

Amazon uses robots in 13 of its 123 fulfilment centers (FCs) worldwide, with plans to expand their use to more, and so far "the data shows that the more robots we put into our FCs, the more jobs we create", the spokesman said.

"In a robotics-enabled FC, you would see a massively parallel process, a symphony of humans and technology," he said, giving the example of how automation in one area creates demand for more people in other areas such as sorting or packing goods.

Without automation, Amazon would be unable to ship items to millions of people each day and as the retail giant moves towards its goal of using drones to deliver packages within 30 minutes, it needs to continue to streamline delivery.

Deep learning in the warehouse

The winning bot in this year's APC was made by a Team Delft, a partnership between researchers from the TU Delft Robotics Institute and the company Delft Robotics in The Netherlands.

The team trained its bot to recognise the 40 Amazon warehouse items by feeding a deep learning system a database of 3D scans of the items. Doing the grasping was a robot arm with seven degrees of freedom — a term that reflects its ability to move in any direction in 3D space, as well as yaw, pitch and roll — together with high-quality 3D cameras for vision and with a combined gripper/suction cup designed in-house.

Crucially, said Carlos Hernández Corbato from TU Delft Robotics Institute, the bot demonstrated its ability to operate in an uncontrolled environment such as a warehouse, where items can be in different positions every time the bot comes to grab them.

'The robot needs to be able to handle variety and operate in an unstructured environment. We are really happy that we have been able to develop this successful system', he said, while his colleague stressed the arm's ability to accurately pick out the right items and not drop them thanks to its combined gripper/suction cup.

The Delft team's robotic arm not only won the APC's picking contest but also the stow task, a new challenge where bots were tasked with taking an item from a tray and placing it on a shelf.

Despite the progress, this year's bots still have some way to go before they match humans. A human can pick about 400 items from shelves per hour with minimal errors, whereas the Delft team's winning bot can select about 100 items per hour with a 16.7 percent failure rate. Nevertheless, it is still far faster than last year's APC winner, which could pick about 30 items per hour with a similar failure rate.

The speed and accuracy of the Delft robot is such that Corbato suggested that similar systems could be employed to augment a workforce.

"A SME [small and medium-sized enterprise] providing these systems to their operators to work with, let's say a human operator could handle four of these robotized workcells, could handle the delivery rate required, for example, for online shopping."

The complexity of getting a robot to pick an item from a shelf is perhaps more obvious when the task is broken down into the many areas it requires a robot to master: object and pose recognition, grasp planning, compliant manipulation, motion planning, task planning, task execution, and error detection and recovery.

Amazon's Brady believes the APC will identify ways to tackle this complexity, by encouraging science and industry to "share approaches to the unresolved challenges of unstructured automation projects typically found in the logistics industry".

About Nick Heath

Nick Heath is chief reporter for TechRepublic. He writes about the technology that IT decision makers need to know about, and the latest happenings in the European tech scene.

Editor's Picks