Since then, Amazon has categorised all its operations into six types of automation: movement, manipulation, sorting, storage, identification, and packing, the chief technologist at Amazon Robotics, Tye Brady, said. “We want to have a world-class capability in each of those.”

This approach tries to address a central tension in developing robotics, between creating a system that can do many things but is harder to develop and creating one that has a narrower focus but is more likely to succeed.

The company has developed robots that tackle particular challenges.

That includes Greek-named updates to Kiva, like Hercules, which moves heavy carts, and Pegasus, which shuttles and sorts packed orders.

There is also a flock of robotic arms, including Robin and Sparrow, that manipulate items and packages.

The robots Amazon is using now

Several years ago, Amazon began rethinking how its primary warehouses operated. The single biggest change was overhauling how Amazon stored and moved items.

In the old system, Amazon stored products in towers of cubbies that had a fabric front; workers put a hand into the cubby and fished around for the desired product.

In the new system, called Sequoia, those cubbies have been replaced with plastic bins that robotically slide in and out of a frame. Products can move around the warehouse in those bins, and using computer vision, Amazon can look into the bins from overhead to identify items. Then robotic arms move the items with suction cups.

“We thought that change could be a simplifier in allowing the robots around the system to be more effective and allow us to take away a part of the process in a way that could actually lead to safer outcomes and to more efficient output,” said Udit Madan, Amazon’s head of operations.

At Amazon’s most advanced warehouse, in Shreveport, Louisiana, employees touch products at just a few stages, such as taking them out of shipping boxes and placing them in bins.

After that, the Sparrow robotic arm looks into a bin of items, picks the one it wants and puts it in another bin.

The robotic arm called Robin places packed packages on a small robot called Pegasus, which shuttles packages to drop down specific chutes depending on where they will be shipped.

Beneath that chute, a beefy, tall robotic arm called Cardinal grabs sealed boxes and stacks them into carts.

“As the packages come down into the chutes, Cardinal is able to pick those boxes, lift it, and play Tetris very nicely” to fit the boxes in the carts, said Abhishek Gowrishankar, who runs the Shreveport facility.

A tortoise-looking robot named Proteus slides under those carts and autonomously carries them to shipping docks. When it navigates around workers, its lights form a smile.

There are other smaller advancements.

Different machines pack items into boxes and envelopes, depending on the customer order.

One blows air to keep the sides of a paper envelope apart so a worker can easily slide in an item.

And the first new address labeller in two decades has an arm that moves in different directions to place the labels on 3000 packages an hour.

What’s ahead for Amazon’s robots

For now, the Sparrow arm is used for consolidating inventory between the bins. But internal documents viewed by the New York Times show that Amazon has tested arms to pick inventory for individual customer orders, one of the key tasks done by workers.

A process known as “decanting” — cutting open boxes, unpacking products and getting them into bins — has remained stubbornly manual.

So far, robotic decanting prototypes have not kept pace with the rest of the automated systems.

People are needed to ensure that the inventory isn’t damaged, or that it matches the expected shipment.

For now, workers at the Shreveport facility stand at an updated station that uses computer vision to detect which bin a product is placed in, requiring fewer steps.

In the smaller facilities that Amazon uses for same-day deliveries, it has experimented with a system called Jupiter to store and robotically retrieve a lot of inventory. Amazon is still years away from what internal documents describe as the goal of “near lights-out automation” at those buildings.

Executives have been focusing on everyday products, like deodorant and groceries, which customers buy more of when they are delivered quickly.

Those products are housed at ultrafast facilities that are already highly efficient because they are geographically close to customers. Because these are low-margin items, cutting costs to fulfil the orders and being able to store more inventory in the buildings are critical.

And Amazon has only begun to integrate this next generation of artificial intelligence into its systems. The company paid US$400 million a year ago to hire the founding team and license the technology of Covariant, a start-up developing AI systems that act like a robot’s “brain”.

Madan said the team had already improved the vision models of the Sparrow system, letting the arm better understand what is in a bin, which item to grasp and where to best place it in another bin.

Those advanced systems are part of Amazon’s experiments to create a new generation of robotic arms, called Bluejay and Starling, that can manipulate items and packages in a broader range of tasks and in different types of buildings.

This article originally appeared in The New York Times.

Written by: Karen Weise

Photographs by: Emily Kask

©2025 THE NEW YORK TIMES