Today I had the opportunity of touring an Amazon fulfillment center (FC). This was my local MIA1 robotic FC in Opa Locka, FL.
Our group assembled at 9am and was led to a classroom-style room, where we were given water and a headset and radio, so that we could follow along the tour.
Our team leader, Patrick, was wearing a microphone. The group was small—maybe 20 or so—but there’s a lot of ambient noise from the constant whirr of conveyor belts (all 15 miles of them).
Patrick began the tour by pointing out the sheer size of the building we were in: 800,000 square feet, 10x the size of Amazon’s very first warehouse, and the equivalent of 14 football fields.
The center is split between the ground floor, which has the receiving and shipping bays for trucks, and three identical floors above that. Those three floors have armies of Kiva robots (which Amazon acquired in 2012) lifting, moving and lowering stacks of bins. Photos and videos aren’t allowed so I couldn’t take any, but this short video gives you a good idea of what I’m describing:
The flow of goods under 25 lbs (those that fit in these yellow bins) is from the truck in the receiving bay to an unpacking station. In the unpacking station, workers manually open the cardboard boxes that arrived from the trucks and start placing items in the yellow bins.
Every time an item is picked up though, it is first scanned with a handheld gun. After the worker places it in the bin, the worker scans the bin location. Amazon uses a random process so each bin might contain completely unrelated products. Now, the computer system knows exactly in which bin each item resides, and the item is immediately available for sale on Amazon. A purple light shone by a Ricoh projector illuminates the bins the worker should not use, so as to not confuse the system.
This is what the packed bins look like:
The bins are then lifted and routed to a VBI (visual bin inspection) station, where they are photographed and filmed. This information is fed into machine learning models that verify the contents of each bin against what the computer thinks is in that bin. Error checking in action!
When an order is placed, the Kiva robot picks up the bin and moves it to a picking station. The Ricoh projector shines a white light on the correct bin—where the product that was ordered resides—and the worker manually picks out the item, holds it so an overhead Cognex camera can scan it, then places it in a yellow tote box. A monitor confirms by showing a picture of the picked item. After placing the item in the tote box, the worker must tap a round button indicating the process was completed. The tote is then routed along the conveyor belts, down to the packing area.
We then toured the packing area and the automated labeling system. But this was just for single items packaged in an envelope, so we didn’t get to see more complex orders of multiple items in boxes.
Last month, I attended the Amazon re:Mars conference in Las Vegas. Mars actually stands for machine learning, automation, robotics and space. Here’s a picture of a bin and a robotic arm at the conference:
There was a lot of focus at the re:Mars conference on robotic arms, specifically, arms that can handle exactly the types of objects in those bins. Robots still have trouble both with the grasping part and also with understanding what they’re seeing. But they’re getting better.
Having finished this fulfillment center tour, Amazon’s interest in robotics and machine learning—as applied to logistics—makes so much sense now. Although I am no FC expert, it’s probably fair to assume that this Amazon robotic FC is reasonably state-of-the-art. Despite that, it’s striking how manual so many of the processes still are.
There are dozens of teams and startups around the world working on this problem. At re:Mars, UC Berkeley professor Ken Goldberg described how Amazon has been running a challenge: how many picks per hour can a robotic arm perform over objects they provide?
Professor Goldberg’s team created a robotic arm that could do 25 picks per hour in 2015 (very good humans can do 400-600). Today, the robot can already do 400 picks per hour on objects it’s never seen before. (Interestingly, many of the companies working in the field run their machine learning algorithms and simulations on Amazon AWS, so Amazon makes money while others figure out the technology that will enable Amazon to save money in the future.)
Given these advances, Amazon will soon be able to apply automation to the unpacking, picking, and packing processes. This will massively improve productivity—doing it faster and with fewer mistakes than humans—and will also allow these workers to move from repetitive, low-skill jobs to higher-skilled jobs.
This is why Amazon is investing so much in upskilling its workers. They’ll pick up tuition and textbook costs, whether an FC worker wants to move onto a career that benefits Amazon or not.
Another company I saw at re:Mars and later visited in San Francisco, Lightform, is using projectors and cameras to superimpose information around us in, well, light form, which we can interact with using gestures and voice (through Amazon Alexa). The demo I saw had exciting consumer applications, but it’s also easy to imagine a use case for Lightform to replace the Ricoh projectors in FCs with something more useful and interactive. Unless, of course, the robotic arms take over first. (Amazon is a Lightform investor.)
Overall, the tour was useful. There were areas we didn’t see, but it’s hard to argue that it was a show put on for visitors: on an average day, they host three tours, and the FC is pretty wide open. Looking in any direction, you can look around and see what’s going on.
As we have been reminded in the past couple of days of tech earnings, Google, Facebook and Amazon have enabled the rise of millions of small and medium size businesses through their advertising, discovery, and merchant programs. Tyler Cowen, in this excellent debate with Tim Wu, has pointed out that Google and Facebook in particular are the biggest anti-monopoly institutions in the American economy.
More companies—in particular Google and Facebook—should do some version of Amazon’s FC tours. The vast majority of consumers love these companies, and a tiny number of academics have made a career out of being upset about them, but their voice is loud. The risk is that the government listens to them, not the people, and passes laws that end up having the unintended consequence of hurting the American consumer. You can book your FC tour here.