Amazon puts its millionth robot to work in Japan and launches DeepFleet, an AI model it says cuts fleet travel time by 10%

Amazon has passed one million robots in its operations network. In a post on its corporate newsroom dated 1 July 2025, Scott Dresser, vice president of Amazon Robotics, wrote that the millionth unit had been delivered to a fulfillment centre in Japan and joined a fleet that works across more than 300 facilities worldwide. In the same post the company introduced DeepFleet, a generative AI foundation model that coordinates how its robots move through a building, and said the model would improve the travel time of its robotic fleet by 10%.

The milestone caps a steep climb. On 18 October 2023 Amazon said it had more than 750,000 robots working alongside its employees. Adding 250,000 machines between those two disclosures implies fleet growth of about 33% in roughly 20 months, by MEKAWIRE calculation. Amazon did not say how the fleet splits by robot type, how many units sit in each building or what the programme has cost. Dividing one million robots by the 300 facilities it cites gives an average of at most about 3,300 robots per site, again by MEKAWIRE calculation, because the company gives the site count only as more than 300.

Amazon describes DeepFleet as a traffic management system for its warehouse floors. The model forecasts how robots will move and where congestion will build, so that routes can be planned around bottlenecks. Amazon said it built the model with Amazon SageMaker and other AWS tools, trained it on its own inventory movement data, and that it keeps learning as it runs. Faster robot travel, in the company's account, means orders move through a building faster and the network costs less to operate. Amazon repeated the 10% figure in its second quarter results on 31 July, where it described DeepFleet as making its one million plus robots work smarter by finding better paths and reducing bottlenecks.

The technical detail arrived on 11 August, when Joey Durham, a senior manager of applied science at Amazon Robotics, described the work on the Amazon Science blog alongside a paper on arXiv. The team trained on millions of hours of data from fulfillment and sortation centres and compared four designs. A robot-centric model with 97 million parameters, which looks at each robot together with its nearest 30 robots, 100 grid cells and 100 objects, performed best overall. A robot-floor model with 840 million parameters scored best on a timing measure, a graph-based floor model with 13 million parameters held up well for its size, and an image-style model that treats floor cells as pixels did not work well. Results improved as training data grew, which is the scaling behaviour that underpins the foundation model label. The near-term job is predicting congestion; the stated longer-term aim is to have the models assign robots to specific retrieval tasks.

Amazon has run robots in its fulfillment network since 2012. The fleet now includes Hercules drive units, which lift up to 1,250 pounds of inventory, Pegasus robots that carry single packages on precision conveyor belts, and Proteus, which Amazon calls its first fully autonomous mobile robot able to work safely around people. In October 2024 it opened a five-storey site of more than 3 million square feet in Shreveport, Louisiana, with ten times the robotics of earlier designs. It said that building had cut fulfillment processing times by up to 25% and that it expected a 25% improvement in cost to serve there during peak periods. In its second quarter release Amazon also highlighted Vulcan, its first robot that combines sight and touch.

No outside supplier operates at anything near that scale. Geek+, which listed in Hong Kong under code 2590 in July 2025, said in its interim results on 29 August that it had deployed more than 66,000 robots in more than 40 countries and regions, with first half revenue of RMB1.025 billion, up 31.0%. Symbotic reported fiscal third quarter revenue of $592 million, up 26%, on 6 August. AutoStore said on 14 August that about 1,750 of its systems were installed for about 1,250 customers, while second quarter revenue fell 13.1% to $133.9 million. Teradyne's robotics segment, which houses Universal Robots and MiR, reported second quarter revenue of $75 million.

The workforce question follows every Amazon robotics announcement, and the company addressed it with training figures rather than headcount. Amazon said it has upskilled more than 700,000 employees since 2019 and that its next-generation sites need 30% more people in reliability, maintenance and engineering roles. It did not say how the million robots have changed staffing per building, and in October 2023 it had pointed to lower recordable and lost-time incident rates at robotics sites as the main worker benefit.

The open questions are about measurement. The 10% gain is Amazon's own figure, and the company has not published the baseline, the sites where it was measured or whether it holds at peak. The arXiv paper says the models were trained on movement data from hundreds of thousands of robots, which suggests DeepFleet does not yet cover the full fleet. Amazon also has not said whether DeepFleet will ever be offered to other warehouse operators through AWS.

The next markers are Amazon's third quarter results in late October 2025, the fourth quarter peak season, when any travel time gain should show up in throughput, and evidence that DeepFleet has moved from forecasting congestion to assigning tasks.