TOKYO Picking stuff up and putting it away sounds like child's play, but to online retailer Amazon.com, it is a high-tech, high-stakes game.
Always keen on leveraging new technologies, the e-commerce giant hosted its Amazon Robotics Challenge last month, a tournament for warehouse robots. Leading corporations and research institutions brought their next-generation robots to the contest, showing off some surprising advances in artificial intelligence and other technologies, both to other competitors and to Amazon, which hopes to continue automating its logistics operations with more robots.
IDEA INCUBATOR The tournament has been a cradle of future technologies. Last month's third annual challenge was held in Japan, with 16 teams from 10 countries and territories showing up in Nagoya to compete at the four-day event.
This year, contestants competed in timed stowing and picking operations -- simple enough for humans yet nearly impossible for robots to perform reliably. For stowing, each robot was rated on how efficiently it could pick up items and put them on a shelf. In the picking phase, the robots had to put specified items into the proper boxes. The top eight advanced to the final round.
Half of the 32 items used were not known to the teams until the contest began. Unfamiliar items are particularly difficult for robots to handle, hence good recognition skills are required. By not revealing some of the items, tournament organizers forced the robots to deal with unknown situations that humans regularly encounter.
Tye Brady, Amazon Robotics' chief technologist, said that tournaments like this help advance machine learning, object recognition and other technologies, which in turn drive wider use in warehouse robots.
"I learned so much by participating in the competition," said Masaki Yamamoto, chief engineer at the robotics solution department of Panasonic. The Japanese company teamed up with the Nara Institute of Science and Technology. Their robot finished sixth in what was their first ARC. "It changed our values and [the way we think about] nurturing talent," Yamamoto noted.
Kazuyuki Ikeda, an executive officer of Askul, a Japanese office supply company that employs robots at its logistics centers, said: "Unlike places like factories, warehouses are filled with merchandise that greatly vary in material, shape, size and weight."
The Panasonic-Nara Institute entry was imbued with democratic voting system technology, which improves a robot's ability to recognize unknown items.
The system consists of several subsystems, such as AI, deep learning, high-performance cameras and weight sensors. Based on the data it gleans, the system "votes" on what the most plausible action should be. This way, even if one subsystem fails, the robot can carry on.
TECH GAPS Yamamoto's team struggled after reaching the final round: A roll of adhesive tape being removed from a shelf got stuck on the robot's arm. After numerous attempts to deal with the problem, the AI-equipped entrant finally figured things out and put the tape away. "In terms of creating a reliable robot that can function continuously, we attained what we aimed for," Yamamoto reasoned.
Another Japanese team -- comprised of Mitsubishi Electric, Chubu University and Chukyo University -- also did fairly well. It was Mitsubishi's third time at the event.
"I am interested in how robots can be used in the world of logistics," said Makito Seki of the company's Advanced Technology R&D Center. "The competition offers a chance to develop technology and learn how engineers from around the world are approaching [the same issues we are confronting]."
Mitsubishi's entrant had two arms, each able to pick and stow separate items independently. Past versions of the robot required the two arms to work together, handling one item at a time.
The team scored high in the preliminaries, handling unknown objects well by weighing them when picking them up. But the contest revealed huge technological gaps between Japanese entrants and the top finishers.
GOOD PICKS Japanese experts were especially surprised by the team from Singapore's prestigious Nanyang Technological University. Their robot was nearly flawless in recognition, despite the venue's varied lighting, which made it difficult to discern items.
First place, however, went to the Australian Centre for Robotic Vision. Their entrant picked items from above rather than reaching out, and also featured unique AI that allowed it to deftly handle unknown items.
The ACRV robot figured out how to handle items by taking and analyzing 200 to 300 pictures of each. Deep learning like this typically requires 10,000 or more pictures to achieve acceptable results. The team got around this by pre-installing image data covering a range of items. This required the robot to shoot fewer photos before it could successfully recognize an item by comparing it to those in memory.
Only last year did Amazon Japan install robots at its logistics centers. But these are relatively primitive machines, like conveyor shelves. Receiving products from suppliers then getting them packaged for customers requires a different breed of robot, one with the smarts to efficiently pick and stow.