A Brief Primer on Amazon Robots

Not all robots are made equal

Most of the recognizable innovations of the last fifty years or so are bound within the devices we carry around with us. For everything else, housing, cars, appliances, etc., we live in a world not so different from the 1970s, and in some ways visibly more degraded. I can only hope that in twenty years time, we look back on how the richest country on earth could have such obviously dilapidated infrastructure, from grimy public buildings to potholed roads.

There is something undeniably attractive about actual physical innovations, things you can see in action interacting with the physical world. It’s great to have AI-powered search on our phones, but to see it changing our physical environments is magnetic. We turn into children looking on in awe and terror when such things are trotted out.

Amazon knows this, and it’s why it makes great hay of new robotics deployments. This primer is intended to demystify their robotics game, first by reviewing all of the robots that they have developed and put to use, and second by spelling out their importance to Amazon’s operation. Local newscasts unhesitatingly cover every new development like its wizardry, but not all robots are made equal. Some have allowed true systemic innovations. Some are going to add very little to their operation other than partially replacing human workers. And some are novelties that aren’t likely going anywhere. 

Kiva Developments

Undoubtedly the biggest win for Amazon has been the widespread adoption of Kiva technology. The workhorses of AR Fulfillment Centers are the Hercules robots, originally developed by Kiva Systems, where it was called the DU 3000. Amazon acquired Kiva in 2012, at which point its robotics game took off. The Hercules looks like a giant Roomba, and it moves “pods”—giant stacks of inventory weighing up to 2,000 pounds—to and from human stowers and pickers. Hercules robots move around on floors with QR codes every couple feet, which is how they recognize their position and know where to go next. They are generally physically separated from workers by a chain link fence, and only special technicians wearing tech vests that deactivate Hercules robots nearby are allowed behind that fence.

The Hercules

Amazon has a lot of cool robots. None are as close to important as Hercules. It’s what allows randomized storage and saves a human being from having to walk a few football fields to sift through items for a particular order. According to Brad Stone in Amazon Unbound, the average pick rate in a Fulfillment Center jumped from 100 items/hour to 300-400 after the introduction of Kiva technology. 

Other robots more or less replace individual human tasks. The Hercules robots have made possible a complete overhaul of inventory management. Here’s the Hercules in action:

The importance of the Hercules is demonstrated further by the fact that the second most important robot in Amazon facilities is essentially a Hercules robot with a mini-conveyor belt on top: the Pegasus. Unveiled in 2019, the Pegasus robot is intended to replace conveyor belt sorters: instead of being fed into a singulator and then kicked via a shoe sorter or something similar to different chutes, broken down by zip code, inbound packages are scanned and individually placed on top of a Pegasus, which takes them directly to a chute.

I personally have a difficult time wrapping my mind around how precisely sortation robots make more sense than conveyor belts. If people only traveled between a set number of terminals, trains would make way more sense than cars, right? So how does it make sense to give each package its own little vehicle? 

A few possibilities: 1) Sortation robots are more scalable than belts. Need to process more packages? Add more Pegasuses to the floor. 2) Reduced labor costs. One analyst estimates that sortation robots save 40% on labor costs, but prepping packages for a singulator v. tossing them on a Pegasus don’t seem so different, so I’m not sold yet. 3) Low sorting error. Here I’m more convinced, as a Pegasus is going to much more reliably deliver a package to the right chute as compared to any shoe/push/tilt sorter on a conveyor belt. Amazon claims that it reduces sorting errors by 50%. To my knowledge, most Amazon sortation facilities still rely on conveyor belt sorting, but perhaps that’s changing quickly. In any event, here’s the Pegasus in action:

The newest Hercules offshoot is the Proteus, just unveiled in 2022. This is the first fully autonomous robot used in Amazon facilities, and a new, bigger version of the previously unveiled Bert robot. Unlike the Hercules and the Pegasus, it travels amongst humans. For years, human “water spiders” moved storage totes around facilities, and human workers were also primarily responsible for inbound and outbound GoCart (the big blue storage cart found throughout Amazon facilities) movement. The Proteus potentially changes all that, and Amazon explicitly said the goal is to automate all outbound GoCart handling with this robot. Here’s the Proteus moving one of those GoCarts:

There’s also the Xanthus, which is simply a modular version of the Kiva, so it could do both Hercules and Pegasus tasks. 

The Xanthus, showing off its modularity

And while most robotics people wouldn’t include these in the same category, I might also mention here Scooter and Kermit, which are larger autonomously guided vehicles meant to move storage totes and carts around. They’re bigger than the Kivas, but they serve a similar function, and it may be that that function is better served by the smaller Proteus. 

The Scooter robot

The Kermit robot

Robotic Arms

Amazon has used the massive, six-axis Robo-stow (developed by Fanuc) since 2014. It can lift pallets of up to 3,000 pounds up to 24 feet in the air. It’s a regular feature of Amazon warehouses, but it’s not a core Amazon Robotics technology.

The robotic arm people at Amazon seem to like birds, and so the key ones have names like Cardinal, Robin, and Sparrow. Both Cardinal and Robin pick up packages using vacuum suckers; the main difference appears to be that Cardinal can pick up heavier packages (up to 50 pounds), while Robin is a speedier sorter. Superficially, Sparrow is just another one of these vacuum-sucking robotic arms, but it represents a leap in sophistication, as it is intended to automate large parts of the “picking” role. 

The Sparrow robot

Again, the Hercules made possible a form of inventory management that would have been difficult to imagine fifteen years ago; its utility today goes far beyond the individual task of moving a thing from point a to point b. The robotic arms, by contrast, really are just about the individual task. They’re meant to replace human hands, that’s all.

Sequoia: Next Generation Stowing and Picking

That being said, the robotic arm program has been key to the development of Sequoia, which is not an individual robot but more of a robotics system being tested out in newer Fulfillment Centers. Instead of the pods with mesh bands that hold in the inventory, and which human hands can easily navigate, Sequoia uses pods composed of plastic totes that a large gantry arm moves on and off the pods. The tote removal/storage looks very similar to what was accomplished by Amazon’s Ernie robot, but since in Sequoia, the pods are brought directly to the picking station by Kivas, the simpler gantry arm suffices.

Of course, Amazon touts the safety of this new system, as it does with any automation technology. (I would very much like some diligent reporter to ask them why they think unemployment is “safe” at one of these robot unveilings for the press.) Instead of human stowers and pickers reaching up and down the mesh-band pods, with Sequoia, individual totes are delivered to human pickers within the ergonomic “power zone,” thus preventing any reaching up high or bending down low. 

What this also makes possible is the automation of large parts of the stowing and picking operations. Sparrow can supposedly handle 65% of the products in Amazon’s system. That means that in any Fulfillment Center using Sequoia, 65% of the stowing/picking work in any FC using the mesh-band pods is redundant. For a large FC, that could be 600-800 warehouse jobs. That’s a huge development. MWPVL’s Marc Wulfraat estimates that the full rollout of Sequoia would mean 25,000 jobs lost. Here’s the Sequoia unveiling:

Rolling out Sequoia in existing FCs is quite a process, however. It means replacing all of those mesh-band pods with plastic-tote pods, and it also means putting these large gantry robots at every picking station. If they want to take advantage of the automation, it also means incorporating Sparrow. This would all be of huge cost to Amazon, and Sequoia is going to have to prove its worth in the Shreveport, LA Gen 12 facility before its rolled out system-wide. My guess is that human pickers are actually a tad more efficient than Sparrow, as genius as that particular robot is. The average human pick rate is about 350 items an hour, which is actually astoundingly fast. But maybe the reduced labor costs are enough to make up for the investment. If pursued, again, it would entail a drastic reduction of Amazon’s warehousing workforce.

Automated Packing: “The Holy Grail”

In an interview, Wulfraat referred to automated packing as the “holy grail” of e-commerce automation. Human pickers are pretty darn fast; human packers, by contrast, are still rather slow. At some Fulfillment Centers, the computer will tell you what size box to use, and automatically dispense the amount of tape you need. Even then it takes a bit of time to work through multiple steps. 

The CartonWrap 1000, developed by the Italian firm CMC Srl, is used in some Fulfillment Centers and can box 700 items an hour, about 4-5 times faster than a human packer. It was piloted in 2019, and it’s not clear how widely Amazon has adopted it. Reportedly it’s being used in the much touted Gen 12 FC in Shreveport.

Novelties

Very smart people have been made to look like fools by saying that certain elements of Amazon’s robotics program aren’t going to amount to much. Still, I’m willing to take a shot here.

First, there’s Digit, which at one glance indicates that it’s just meant to replace flesh-and-blood human beings. But the very fact that it’s humanoid in look and movement tells me that it’s more of a fun story than a serious development. Where will they use Digit? Sparrow, as deployed within the Sequoia system, is going to be much more efficient at picking than Digit could ever be. The same story is true of the CW1000 and packing. They’re saying that Digit can move totes around, but that’s what Proteus and Kermit are for, and they’re on wheels! For every conceivable Digit function, Amazon already has a more efficient robot. 

The Digit robot

Then there’s last-mile technologies. Scout, the autonomously guided sidewalk roamer, was decommissioned in 2022, the same year that Amazon began experimenting with drone delivery. Perhaps the air affords advantages in last-mile delivery that the ground does not, and Amazon is piloting drone delivery in Tolleson, AZ right now with its MK30 drones. 

The MK30 drone

All I’ll say about the drone program is that it’s a very complicated affair compared to paying a guy to drive around a neighborhood and drop off packages—so comparatively costly and complicated that one has to wonder whether this kind of thing is more about projecting a certain mystique than actually delivering packages. I also think Americans have enough Ron Swanson in them not to abide.