Intelligent Machines

Help Wanted: Robot to Fulfill Amazon Orders

Amazon awarded $20,000 to the creators of a robot that can fill orders most like a human, but it’s not going to replace employees yet.

May 28, 2015

Amazon announced plans this week to hire 6,000 new warehouse workers to meet growing demand. They will join over 50,000 people currently manning dozens of Amazon fulfillment centers across the U.S., picking, packing, and shipping orders. But if the company has its way, some of that work will soon be done by robots.

At the IEEE’s International Conference on Robotics and Automation in Seattle this week, Amazon sponsored a competition to find a robot that can autonomously pick a selection of goods from shelves and pack them safely for shipping. No industrial robot can recognize even a tiny fraction of the millions of items Amazon sells, let alone pick, grasp, and move them safely.

Among the 25 diverse items the robots were asked to retrieve was a duck bath toy, a box of Oreo cookies, safety glasses, a bag of balls, and a plush frog-shaped dog toy. The teams had 20 minutes to pick and pack as many items as possible from a standard Amazon shelving unit, itself designed to be moved around a warehouse by small autonomous robots. Teams earned points by correctly moving items, and lost them by damaging, dropping, or misplacing anything along the way.

Even the simplest products seemed to fool some of the 31 competing robots. “Thin paperback books cause us the biggest problem,” says Camilo Alfonso Perez from the University of Alberta, of his dexterous three-fingered robotic arm with a human-like range of motion. “They’re too thin for our robot to see them, and its fingers are too big to grasp them.”

The Alberta team used lidar (laser ranging) and video to recognize products, and has augmented the arm with what it calls a tongue. This small flap of plastic is attached to a motorized, extendable tape measure that drags items to the front of the shelf, where they are easier to grab.

The University of Texas Austin also used a hand-like gripper, attached to the right arm of a Yaskawa Motoman SDA5F robot. Its left arm holds a depth camera made by Primesense (recently acquired by Apple) that peers into each shelf cubby to capture the positions of objects within. “This setup is something we’re familiar with,” says UT’s Jack Thompson. “Our primary sponsor is Los Alamos National Labs, so we do handling of special nuclear materials. But the rubber duck in the plastic bag is proving difficult for our depth sensor to get a handle on.”

Kentaro Wada of the University of Tokyo was unsure whether his system, based
on Rethink Robotics’s Baxter, would make it through. The Tokyo team’s robot, like several others, gently picks up objects using suction, in this case generated by a pair of commercial vacuum cleaners strapped to the side of the robot.

In the end, Wada’s vacuum completed the test, although his Baxter robot managed to retrieve just one object. The winning team, from the Technical University of Berlin, snagged the top $20,000 prize by successfully picking and packing 10 out of 12 items during its 20-minute test.

A good performance, then, but not one that Amazon spokesperson Kelly Cheeseman envisions threatening any human positions. “We already have over 15,000 robots in our fulfillment centers working alongside our employees,” she says. “These technologies enhance jobs for employees, making them in many cases more efficient. Certainly the role for employees is still vital.”

At least until a robot can figure out how to pick up those fiddly rubber ducks.