They call me the Master of Robots—or at least they should. I grab a flat package, hold its barcode under a red laser dot, and place it on a tiny conveyor belt which is, in fact, a robot. I hit a button to my left, and off zips the robot to do my bidding, bound for one of more than 300 rectangular holes in the floor corresponding to zip codes. When it gets there, the robot engages its conveyor belt, and the package slides off its back and down a chute to the floor below, where it can be loaded onto a truck for delivery.
This is not an experimental system in a robotics lab. These are real packages going to real people with the help of real robots in Amazon’s sorting facility of tomorrow, not far from the Denver airport. With any luck, my robot friend and I just successfully ferried a parcel to someone in Colorado. If not—well, blame the technology, not the user.
Seen from above, the scale of the system is dizzying. My robot, a little orange slab known as a “drive” (or more formally and mythically, Pegasus), is just one of hundreds of its kind swarming a 125,000-square-foot “field” pockmarked with chutes. It’s a symphony of electric whirring, with robots pausing for one another at intersections and delivering their packages to the slides. After each “mission,” they form a neat queue at stations along the periphery, waiting for humans to scan a new package, load the robots once again, and dispatch them on another mission.
Matt Simon covers cannabis, robots, and climate science for WIRED.
You don’t have to look far to see what a massive shakeup this is for the unseen logistics behind your Amazon deliveries. On the other side of the building are four humans doing things the old way, standing at the base of a slide flowing with packages. Frenetically they pick up the parcels, eyeball the label on each, and walk them over to the appropriate chutes. At the bottom of the chutes, yet more humans grab the packages and stack them on pallets for delivery. It’s all extremely labor-intensive and, in a word, chaotic.
Amazon needs this robotic system to supercharge its order fulfillment process and make same-day delivery a widespread reality. But the implications strike at the very nature of modern labor: Humans and robots are fusing into a cohesive workforce, one that promises to harness the unique skills of both parties. With that comes a familiar anxiety—an existential conundrum, even—that as robots grow ever more advanced, they’re bound to push more and more people out of work. But in reality, it’s not nearly as simple as all that.
If only the Luddites could see us fulfilling online orders now.
This Colorado warehouse is, in a way, a monument to robots. It’s not one of the Amazon fulfillment centers you’ve probably heard of by now, in which humans grab all the items in your order and pack them into a box. This is a sorting facility, which receives all those boxes and puts them on trucks to your neighborhood. The distinction is important: These squat, wheeled drives aren’t tasked with finely manipulating your shampoos and books and T-shirts. They’re mules.
Very, very finely tuned mules. A system in the cloud, sort of like air traffic control, coordinates the route of every robot across the floor, with an eye to potential interference from other drives on other routes. That coordination system also decides when a robot should peel off to the side and dock in a charger, and when it should return to work. Sometimes the route selection can get even more complicated, because particularly populous zip codes have more than one chute, so the system needs to factor in traffic patterns in deciding which portal a robot should visit.
“It's basically a very large sudoku puzzle,” says Ryan Clarke, senior manager of Special Amazon Robotics Technology Applications. “You want every column and every row to have an equal amount of drops. How do we make sure that every row and every column looks exactly equal to each other?” The end goal is to minimize congestion through an even distribution of traffic across the field. So on top of tweaking the robots’ routes, the system can actually switch the chute assignments around to match demand, so that neither the robots nor the human sorters they work with hit any bottlenecks.
To map out all this madness, Amazon runs simulations. Those in turn informed how the drives themselves should be performing. What’s the optimal speed? What’s the optimal acceleration and deceleration, given you want the deliveries to be as efficient as possible while keeping the robots from smashing into one another? After all, a bump might toss a package to the ground, which other robots would spot with their vision sensors and route around, adding yet another layer of complexity to the field. (The robots have sensors on either end of their conveyor belt, by the way, so if a package starts to slip off the side, the belt automatically engages to pull the package back on.)
The temptation might be to get these machines moving as quick as possible. “But it would be like having a Ferrari in downtown San Francisco—all you're doing is stop and go,” says Clarke. “We looked at tuning it to many different parameters and found that more speed and more acceleration actually had a reverse effect. They were just bumping into each other and causing more pileups.”
Ready for more complexity? Amazon had to tweak the built space itself to keep the machines happy. Humans doing things the old way on the other side of the building, for instance, enjoy basking in the photons that pour through skylights. Above the robot’s field, though, the skylights are covered, because the glare might throw off the machine’s sensors: To navigate, they’re using a camera on their bellies that reads QR codes on the ground. Even the air conditioning units hanging from the roof are modified. On the human side, they blow air straight downward, but above the robots they blow out to the side, because gusts of air could blow light packages off the machines’ conveyor belts.
The WIRED Guide to Robots
Worse yet, precarious packages like liquids could send the system into chaos. So although the system is automated, humans still monitor the robots on flatscreens below the field, where the packages come down the chutes, and respond to crises. “Think about if I had a package and it had a gallon of paint in it, and that gallon of paint was damaged and it leaked down one of these chutes,” says Steve McDonnell, general manager of the sorting center. “Within minutes I'm able to shut that chute off, redirect drives to another chute, and I'm done.”
The key here is flexibility—not a word that first comes to mind when you think of robots. Flexibility in the robots’ pathways, in their destinations, in the number of robots on the field at once. You might, for example, think the more machines out there, the better. Amazon could deploy up to 800 drives simultaneously, but that could jam up the floor like traffic in a city. Instead, they’re typically operating 400 or 500, with others parked off to the side and waiting to be circulated in.
Beyond coordinating the robots themselves, there’s the question of how to make them good coworkers for the human employees. The humans’ job is to place packages in 6-foot-tall boxes below the field, taking care not to toss in heavy packages first. To make that work manageable, the robots have to distribute packages between the multiple chutes for a particular zip code, so a given chute doesn’t overflow. At the same time, the system considers how to best group packages downstairs by their departure time, so workers don’t have to run around hunting for them.
“The interaction between the associate and the drives is almost like a 3D chess set,” says McDonnell, “because you can optimize the drive field, but then you can make the associate's job harder below the field.”
Across the field from the human workers distributing packages to the drives, a prototype robotic arm, named Robin, sits at the end of a conveyor belt. Its “hand” is a vacuum manipulator, designed to snag boxes and flat packages.
This robotic arm is a test of what it might look like to further automate the work of shuffling packages around. The idea is that the conveyor will deliver packages to the arm, which would then load the drives. “We're going to feed it a little bit differently than we do with humans,” says Rob Whitten, senior technical program manager. “We're not going to just give it a pile coming down a chute—we're going to kind of toss it softballs. We're going to give it a little more structure so it can handle it.” For parcels it can't manipulate, like if they're too heavy or weirdly shaped, humans would step in to help.
As I walk down the line of human robot-loaders, I come across a worker who’s set aside a broken box, which has spilled out bottles and other entrails. That uniquely capable human could do two things here: Use his problem-solving skills to say, Something is wrong, I need to set these aside, and then manipulate those objects with exceedingly fine motor skills.
This robot arm has neither problem-solving prowess nor fine motor skills. Imagine if clear laundry liquid had broken inside a package and soaked the bottom of the box. A human might smell the detergent or feel its stickiness before they see it. A robot arm relying on sight alone would miss the problem, loading the package on a drive robot that then snail-slimes the floor of the field.
Even if they had some semblance of judgment, robots are still awful at manipulating complex objects like bottles. That’s why Amazon is keeping it simple here, with a suction arm meant to stick to flat surfaces, as opposed to an analog of the human hand . For quite some time, humans will need to (nearly) literally hold these robots’ hands.
The bottom line is this: We humans have to adapt to the machines as much as the machines have to adapt to us. Our careers depend on it.
Amazon runs simulations to figure out how to keep their human workers comfortable when loading robots with packages. This includes their range of movement from an ergonomics standpoint, and their safety. Or such questions as how best for a human to grab a parcel, scan it, place it, and reach over to hit the button that sends the robot on its way. “There's an art to making it feel seamless between what the robot is doing and what the humans are doing,” says Brad Porter, VP of robotics at Amazon.
It’s the kind of dynamic environment that’s perfect for the development of Amazon’s next iteration of its system. The company is working on a new modular robot called Xanthus with different attachments, say to hold containers instead of using a conveyor belt. This machine will in a sense bridge the divide between fulfillment centers, where humans are loading products into boxes by hand, and sorting centers, where they’re mostly working with those assembled boxes.
“You can see how combined with maybe the addition of a sensor platform, you could have an autonomous drive that's driving totes around,” says Porter. But you can also take that same thin sled and replace the tote-carrying unit with a conveyor top, and deploy it in the sorting center.
Herein lies Amazon’s huge advantage: It’s got the funds and the talent to develop robots in-house, tailoring each to solve problems specific to Amazon. Other warehouses are starting to go robotic, but they’re working with other companies’ machines. For instance, Boston Dynamics—maker of the hypnotically impressive SpotMini and Atlas —will soon offer a box-lifting robot called Handle . But it’s a generalist machine, not developed exclusively for one client.
Amazon, on the other hand, can iterate on a robot until it's perfectly adapted for a specific task. “They're building it for themselves and they're building it for their environment and circumstances,” says John Santagate, research director of service robotics at IDC, which does market research. “It's hard to build any one product that suits all of it.”
And every worker they hire into a machine-facing role is doing something no other human has ever done before—lower-level workers in this facility have been promoted to help oversee the massive system whirring around them, as well as the humans intimately integrated with it. “The fully automated or highly automated fulfillment center isn't a North Star we're trying to hit,” says Porter. “Do we see additional levels of automation, at higher and higher levels? Yeah, I think that will increase as the capabilities of our systems increase.”
Here’s the big question, though. Is this kind of automation bound to replace human jobs entirely, or replace parts of those jobs? “Most of the research seems to suggest that the direction that automation is moving in is the displacement of skills, not jobs,” says R. David Edelman, formerly President Obama's special assistant on the digital economy, and now the director of MIT’s Project on Technology, Economy, and National Security. “That suggests those individuals can, by Amazon, be reskilled or leverage other skills they already have in the same job.”
These days, industries that are short human labor need automation to survive. Consumers still want fresh produce, but California’s farms are facing a labor shortage of 20 percent, and are increasingly turning to agricultural robotics . Amazon’s business is booming, yet America is enjoying historically low unemployment, so laborers have lots of options for work. “The demand on that company is increasing, but the availability of resources to fill that demand isn't necessarily increasing,” says Santagate. “In fact it's probably contracting.” Robots are filling the void.
Here in this sorting center of tomorrow, I walk along the edge of the field and hear the morning break for humans, called out on loudspeakers. The drive robots continue to shuffle around for a few minutes, with their incessant electric white noise, until suddenly the place falls almost silent. Having delivered their packages to chutes, the robots have run out of work. They park off to the side of the field, some of them in charging stations. Only when the loudspeakers call the end of break do the machines start up again, ready for their humans to feed them more packages.
If only the Luddites could see our codependency now.
- My glorious, boring, almost-disconnected walk in Japan
- What do Amazon's star ratings really mean?
- Drugs that boost circadian rhythms could save our lives
- The 4 best password managers to secure your digital life
- What tech companies pay employees in 2019
- 🏃🏽♀️ Want the best tools to get healthy? Check out our Gear team's picks for the best fitness trackers , running gear (including shoes and socks ), and best headphones .
- 📩 Get even more of our inside scoops with our weekly Backchannel newsletter
“This is all about letting the robot supervise itself, rather than humans going in and doing annotations,” says coauthor Lucas Manuelli, also of MIT CSAIL.“I can see how this is very useful in industrial applications where the hard part is finding a good point to grasp,” says Matthias Plappert, an engineer at OpenAI who has developed a system for a robot hand to teach itself how to manipulate, but who wasn’t involved in this work.