In large warehouses, AI mostly works as software that routes and schedules fleets of mobile robots, plus computer vision for picking and sorting. Amazon announced on June 30, 2025 that it had deployed its millionth robot, delivered to a fulfillment center in Japan, and introduced DeepFleet, a foundation model for coordinating robot movement across more than 300 facilities. Amazon says the model improves fleet travel time by 10 percent, a self-reported figure with no published baseline.
Key takeaways
- DeepFleet predicts how mobile robots interact on a fulfillment-center floor, so tasks can be assigned and congestion avoided.
- Amazon Science tested four model designs; one of them, which treats the floor as an image, did not work well.
- The 10 percent efficiency figure is Amazon's own and the company does not explain how it was measured.
- Federal occupational projections show 4 percent growth for hand laborers and material movers from 2025 to 2035, and says some machine-feeding tasks are likely to be automated.
- Most public detail comes from one operator, so results at other warehouses cannot be inferred from it.
What the announcement said
Amazon states that it has deployed more than one million robots across its operations network since it bought Kiva Systems in 2012, the deal that led to Amazon Robotics. In the June 2025 post by Scott Dresser, vice president of Amazon Robotics, DeepFleet is described as a generative AI foundation model that works like an intelligent traffic management system for the robots. It was built on inventory-movement data from Amazon sites using AWS tools, including Amazon SageMaker, and the post says it is already improving the travel time of the fleet by 10 percent.
Amazon Science gave more detail on August 11, 2025. The blog says development is focused first on predicting congestion, with assigning robots to retrieval tasks and target locations as a longer-term goal. Independent verification of the 10 percent result was not found.
How the model is built
The floor is represented as a grid of cells that hold robots, obstacles, storage or drop-off points, and travel lanes. Every design uses Transformer attention, the same family of techniques behind language models, applied to robot positions instead of words. Amazon Science describes four versions.
| Design | How it represents the floor | Reported outcome |
|---|---|---|
| Robot-centric | Looks at one robot and its nearby robots, cells, and objects | Best overall; 97 million parameters |
| Robot-floor | Separate encoders for robots and floor cells, linked by cross-attention | Best timing estimate; 840 million parameters |
| Graph-floor | Floor as a graph of cells and movement edges | Slightly behind but strong; 13 million parameters |
| Image-floor | Floor grid treated as an image | Did not work well |
The author of the blog says much remains to be figured out. A motive for a learned model is that simulating a couple of thousand robots faster than real time is, in the blog's words, prohibitively resource intensive. Performance improved as training data grew, but scaling tests covered only two of the four designs.
The machines around the model
DeepFleet only coordinates; other systems move and handle goods. Amazon's own descriptions of its robots, which are company statements, include the following.
- Hercules: brings pods of items to workers, uses a forward 3D camera to tell people, pods, and objects apart, and lifts up to 1,250 pounds (567 kilograms) according to the announcement post.
- Proteus: described as Amazon's first fully autonomous mobile robot, working outside restricted robot-only areas.
- Vulcan: described as Amazon's first robot with a sense of touch, which can ask a human partner to take over an item it cannot handle.
- Cardinal: a robotic arm that reads labels and places packages of up to 50 pounds (23 kilograms) into carts.
Amazon says its Shreveport, Louisiana center, launched in 2024, uses eight robotics systems. Claims on speed, such as one robot storing inventory up to 75 percent faster, are the company's and were not independently checked.
Jobs and workforce data
Federal data list 6,917,800 hand laborer and material mover jobs in 2025, paid a median USD 38,220 per year, and project 4 percent growth to 2035, about as fast as average. It says that machine feeders and offbearers' tasks are likely to continue to be automated. Amazon says its Shreveport center needs 30 percent more employees in reliability, maintenance, and engineering roles, and that more than 700,000 employees have been upskilled since 2019. Public data does not yet show a net effect of fleet-coordination software on warehouse employment.
Questions for evaluating warehouse AI claims
- What is the baseline for any percentage gain, and over what period?
- Does the gain come from the AI model, from layout changes, or from new hardware introduced at the same time?
- Was the result measured by the vendor, the operator, or an outside party?
- What happens when the model errs, and who can override it on the floor?
- Which tasks still need people, and how is injury risk tracked?
An illustrative example: if a robot spends 100 minutes per shift traveling, a 10 percent travel-time cut would save 10 minutes per robot per shift. Whether that saving turns into more packages depends on the other steps in the process, which this arithmetic ignores.
FAQ
Is DeepFleet a chatbot?
No. Amazon calls it a generative AI foundation model, but it predicts robot movement on a warehouse floor. It uses the Transformer technique common to language models, trained on movement data instead of text.
Does the 10 percent figure apply to every warehouse?
The figure is Amazon's statement about its own robot fleet. It was not presented as a result for other companies, and the baseline is not published.
Do warehouse robots replace workers?
Public sources do not settle this. Amazon says robots take on heavy and repetitive tasks and that some technical roles grew; federal projections show moderate growth in material-moving jobs with some tasks automated.
How many robots does Amazon operate?
The company said in June 2025 that it had deployed its millionth robot. That count covers deployments since 2012 and the post gives no current active total.
Related reading
- AI in warehouse management
- AI and human resource management
- What an AI engine does
- Automation in accounting
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Amazon builds first foundation model for multirobot coordination, Joey Durham, Amazon Science, August 11, 2025
- Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot, Scott Dresser, Amazon, June 2025
- Amazon uses robots that sort, lift, and carry packages, Amazon, 2025
- BLS Occupational Outlook Handbook entry on hand laborers and material movers, page modified August 27, 2026
- Amazon deploys its 1 millionth robot in a sign of more job automation, Dylan Butts, CNBC, July 2, 2025