For anyone wondering whether robots are gaining ground on people when it comes to performing repetitive tasks, you need look no further than the Amazon Picking Challenge. Having long ago automated the movement of goods in its warehouses through its acquisition of Kiva Robotics, Amazon is now looking at technology to reduce the number of people needed to pack boxes. It recently hosted a DARPA-challenge-style contest, featuring over a dozen teams from the around the world, whose robots competed for the title of best autonomous box packer. This contest, its second one, took place in Leipzig, Germany, and featured a more complex “Pick” task than the first challenge Amazon held last year in Seattle, and a new “Stow” task for unloading.
Team Delft’s robot, featuring both a two-fingered gripper and a suction device, achieved top scores and times in both the box picking-and-packing, and the reverse un-packing and restocking, to take home the $50,000 first place prize. Amazon kept things interesting by using a dozen differently shaped objects in each task, and 40 items overall in the contest. This meant the robot had to adapt its picking strategy to each specific item. Teams were given a JSON file with an item list and work order five minutes before the challenge began.
The flexibility to handle the wide-variety of objects found in an Amazon warehouse is perhaps the largest breakthrough needed to make general-purpose warehouse robots a reality. Currently most package handling robots are designed with a single, or small number of related, package sizes and shapes in mind — or their cargo is prepackaged or paletted for easy automated handling. Many of the robots, including the one from Team Delft, included a depth-sensing camera to help identify the objects and their exact size and location.
In addition to timing the robots, Amazon deducted points for damaging an item, dropping it more than a foot, or misplacing it on the shelf. Objects ranged from a T-shirt to a dumbbell. Delft’s point score was matched by Japan’s Team PFN, but the Dutch team performed its task about 30 seconds faster, giving it first place.
Team leaders and event organizers provided some cheery words about a future where people and robots would work side-by-side in the warehouse — especially since current-technology robots are only expected to be able to handle about 50% of the variety of products — but the long-term trajectory is the replacement of workers performing repetitive tasks with machines.
A large part of what makes these new applications possible is deep learning. Using the same type of software tools that have enabled facial recognition and early-stage autonomous vehicles, these prototype warehouse robots identified objects and pattern-matched their shapes and attributes with the appropriate picking and packing strategies.
Now read: Artificial neural networks are changing the world. What are they? and Robots are coming for your job — but maybe that’s okay