Amazon Has 1,000,000 Robots. Who Directs the Traffic?

A robotic arm lifts a parcel, turns and feeds the next part of the line. One task, done. But what if the next step is not ready?

Amazon says it has deployed more than a million robots across more than 300 facilities. At that scale, even waiting becomes a business problem, and the answer is not always another robot.

The $775 million Kiva agreement

In March 2012, Amazon agreed to acquire Kiva Systems for approximately $775 million in cash. Amazon described Kiva’s technology as a way to bring products directly to employees to pick, pack and stow. Instead of sending a person walking through the warehouse to find a product, the shelf comes to the person.

That changes the shape of the problem. The worker stays at a station and the inventory travels. A job that used to be about walking becomes a problem of movement, timing and software.

Warehouse rush hour

Then scale it up. Hundreds of robots want to arrive at stations while others need to leave, all sharing the same space, intersections and destinations. Automating the walking does not abolish rush hour.

In 2022, Amazon’s robotics researchers described the catch. At first, adding robots increases throughput. At a certain point, their numbers start to cause congestion: the robots interfere with each other and reduce the efficiency of the whole system. The shortest route for one machine can be a terrible route for everybody else.

DeepFleet and the traffic problem

That is the business problem behind DeepFleet, the generative AI foundation model Amazon introduced in June 2025 alongside the news that its millionth robot had reached a fulfillment center in Japan. DeepFleet learns from robot movement data to predict how traffic is likely to unfold, and Amazon’s scientists say those predictions can help assign tasks to robots and route them around potential congestion.

Picture two routes to the same station. One is shorter, but everybody wants it. The other takes a little longer but keeps moving. The clever decision can be choosing the second route before the first one backs up. What matters is what gets through the building, not which machine wins the race to the intersection.

What the 10% claim actually means

Amazon put a number on the opportunity: a 10% improvement in robot fleet travel efficiency. That is Amazon’s claim, and it describes how the robots travel. It is not a 10% saving on everything Amazon spends delivering an order. Still, less wasted travel means more useful movement from equipment and floor space the company has already invested in.

Follow the real handoffs

The footage in the video was filmed at Amazon’s BRS2 fulfillment center in Swindon, UK, in March 2025. It shows different stations, not one customer’s order followed from start to finish: a robotic arm placing a package onto the line, a measuring gate where parcels pass through sensors, and a station that applies shipping labels to moving boxes.

Each successful handoff raises a new question: is the next stage ready for what is arriving? If the robot can feed boxes faster than the next station can process them, speeding up the robot again just builds a longer queue. An impressive demonstration is the beginning of the business case, not the end of it. The same lesson runs through FedEx’s overnight network, where fast sorting only matters if the rest of the chain keeps up.

The cost around the machine

The $775 million Kiva figure was an acquisition price, not the total cost of Amazon’s robotics program. Machines, maintenance, software, training and the rest of the operation around them all have to be paid for. A faster piece of equipment does not automatically lower the cost of a completed order. The gain has to survive the next handoff, and the one after that.

People and the ordinary shift

There are people in this system. Bringing a shelf to a worker can reduce walking, but on its own it does not answer questions about pace, repetitive work or safety. In December 2024, the U.S. Department of Labor announced a corporate-wide settlement with Amazon requiring ergonomic measures at covered facilities, including fulfillment centers, sortation centers and delivery stations. Amazon, for its part, reports that its recordable incident rate improved 34% over five years.

A robot count is not a safety score. For anyone evaluating automation, the useful question is what happens across an ordinary shift, including the interruptions.

The real test is consistency

Once a delivery experience works well, it becomes the expectation. Customers never see the warehouse choreography. They see whether the promise was kept. The million-robot milestone gets the headline. The less visible challenge is getting the machines, the software and the people to work as one system, again and again.

Next time you see an impressive robot demo, look just beyond it. What is waiting on the other side?