The Box Does Half the Picking - by Jaimin

Atoms to Algorithms

SubscribeSign in

The Box Does Half the Picking<br>Monday, June 8, 2026 · Logistics & Warehousing

Jaimin<br>Jun 08, 2026

Share

On Friday we ended with ask, why a company that built a superhuman picking arm would spend billions redesigning its shelves. Here is the answer up front. Amazon’s older storage unit is a fabric pod divided into compartments about a foot square, each stuffed with up to ten items leaning on each other. Getting a robot to pick from that took years of work on force sensors and a literal sense of touch, and the robot still hands off about a quarter of items to a human. Amazon’s newer system stores the same inventory in rigid plastic totes, open on top, identical by design. In a tote, a camera and a suction cup are enough. The breakthrough is not a better arm. It is a better box.<br>This week we leave the research lab and walk into the building where robot economics get decided. Amazon has deployed more than a million robots since 2012, the largest industrial fleet on Earth, so Monday is theirs. Tomorrow, a British grocer that took the box idea to its logical extreme. Later this week, the company automating Walmart, the robots on your sidewalk, and the question that decides everything: when does the machine pay for itself?<br>How it actually works

Picking in a warehouse is really four problems wearing one name. The robot faces an endless variety of products, including ones that did not exist last week. The products sit in clutter, leaning on and hiding behind each other. The robot must choose exactly where to attach its suction cups and at what angle. And it must do all of this in about a second, because a warehouse is a throughput machine.<br>Amazon runs two opposite solutions side by side, and the contrast is the story.<br>The first solution is to make the robot smarter. For its classic fabric pods, where items press against each other in cramped compartments, Amazon built Vulcan, a robot that can feel. Its tool looks, in Amazon’s own words, like a ruler stuck onto a hair straightener: the ruler nudges neighboring items aside, the paddles grip the target, and force sensors tell the robot how hard it is pushing so it stops short of crushing anything. Vulcan can handle roughly three quarters of the items Amazon stores, and it knows to call a human colleague for the rest. That is the price of being clever in a messy scene.

The second solution is to make the scene simpler. Sequoia, the storage system Amazon has been rolling out since 2023, repacks inventory into rigid totes carried by mobile robots. A tote is everything a fabric pod is not: a known shape, a bounded depth, an open top, no fabric in the way. Amazon says the system stores incoming inventory up to 75 percent faster and speeds order processing by up to 25 percent. And inside that tidy world works Sparrow, the suction-cup arm from Friday’s cliffhanger. Sparrow does not need to feel, because the tote has already done the hard part. The clutter problem was not solved. It was designed out.

What is left for Sparrow is still a real problem: pick up something it has never seen, quickly. Amazon’s published approach works like a short audition. The system proposes many candidate grasps, each a specific spot, angle, and combination of suction cups. A learned model then scores each candidate’s chance of success and the best one wins. The clever part is where the model’s education comes from: the recorded successes and failures of millions of real picks made by the previous generation of software, running in live warehouses. The fleet teaches its own successor. A million robots generate a study guide nobody else can buy.

The image to keep is bowling with bumpers. Vulcan is a bowler trained to throw hooks on a warped lane. Sequoia just installs bumpers, and suddenly an ordinary throw scores. The world’s most experienced robot operator keeps deciding that reshaping the environment is often cheaper than perfecting the robot, which is worth remembering every time a demo video promises a robot that can do anything anywhere.<br>New this week

At its Delivering the Future event in London on June 4, Amazon announced a next-generation Proteus, its free-roaming warehouse robot. The new version can work anywhere in a building rather than only loading docks, and workers can now assign it tasks in plain conversational language; the robot figures out priority, route, and timing itself. It arrives in European buildings in 2027, part of a commitment of more than 10 billion euros to modernize the network, alongside 25,000 new jobs. (About Amazon)<br>The same announcement quietly confirmed today’s thesis: STARK, a robot that lifts and carries full totes, expands to 15 European sites by 2027, and Vulcan rolls out to more European buildings. Stowing, picking, and now carrying are all being standardized around one humble object, the plastic tote. (About Amazon)<br>What to notice

The visualization shows the two...