Warehouse automation

Warehouse automation is the use of machinery, control software, and increasingly machine learning to move, store, sort, and pick goods inside distribution centers and fulfillment centers. The field predates modern AI by decades: the first computer-controlled high-bay storage systems went in during the early 1960s, and the first driverless industrial vehicles in the 1950s ^[3][5]. What changed in the 2010s and 2020s is that perception and decision-making moved from fixed rules to learned models, which let robots handle the messy parts of the job that conveyors never could: recognizing an unfamiliar product, deciding where to grip it, and navigating a floor full of people.

The clearest measure of the shift is Amazon's fleet. The company acquired Kiva Systems in 2012 and had deployed more than 520,000 mobile drive units by mid-2022, over 750,000 robots by late 2023, and its one millionth robot by July 2025 ^{[1][8][10][12]}. No other operator is close, but the same technology stack, mobile robots under a fleet manager, dense automated storage, and vision-guided arms, now appears in grocery, apparel, pharmaceutical, and third-party logistics operations worldwide.

Automation in warehouses is not uniformly successful. Vendors have gone bankrupt or been wound down, forecasts have been cut, and the hardest task in the building, picking an arbitrary item out of a cluttered bin, remains only partly solved. This article covers the system types, the AI methods behind them, the companies building them, and the labor and safety questions that follow.

What the term covers

"Warehouse automation" spans a wide range of equipment, and the categories matter because they have very different economics.

Fixed automation is bolted to the floor: conveyor lines, chutes, and high-throughput sorters that divert cartons or parcels into destinations. It is fast and cheap per unit handled, but the layout is baked in.

Automated storage and retrieval systems (AS/RS) hold inventory in a dense structure and use machines rather than people to put things away and get them back. Variants include unit-load systems for pallets, mini-load systems for small parts, shuttle systems where a shuttle runs on each rack level and a lift handles vertical movement, horizontal carousels, and vertical lift modules ^[4].

Mobile robots move goods across the floor. Automated guided vehicles (AGVs) follow infrastructure such as buried wire, magnetic tape, or reflective targets; autonomous mobile robots (AMRs) localize against the building itself and replan around obstacles ^[5].

The organizing distinction on the picking side is goods-to-person versus person-to-goods. In a person-to-goods warehouse, a worker walks or rides to the inventory. In goods-to-person, the inventory comes to a fixed station, which removes most of the walking time from the pick, whether the carrier is a shelf-lifting robot, a bin on a grid, or a tote fetched by a case-handling robot ^[43]. Both patterns are in active use. Goods-to-person dominates purpose-built automated buildings; person-to-goods assist robots are cheaper to retrofit into an existing site, and Interact Analysis has projected that segment to grow around 30 percent annually to 2030 ^[40].

From stacker cranes to mobile robots

The first automated guided vehicle came to market in the 1950s from Barrett Electronics of Northbrook, Illinois. It was, in the plainest description, a tow truck that followed a wire in the floor instead of a rail ^[5]. Navigation later moved through floor markers, laser triangulation against reflectors, inertial systems, and finally natural-feature methods related to SLAM, which removed the need to modify the building at all ^[5].

Automated storage arrived a decade later. In 1962 Demag, a predecessor of Dematic, installed the first stacker crane of its kind for the German bookseller Bertelsmann. It stored 7 million books on pallets at heights of up to 20 meters, which established the high-bay warehouse as a building type ^[3].

In the mid-1990s a Norwegian route ran in a different direction. AutoStore, founded by Jakob Hatteland after a new warehouse filled up within a month, replaced racking and aisles entirely with a cube: bins stacked directly on top of each other, with robots driving on a grid across the top to dig out whatever is needed. The company says the approach raises storage density by up to 400 percent, and reports an installed base of more than 1,950 systems in over 65 countries. AutoStore listed on the Oslo Stock Exchange in October 2021 at a valuation of $12.4 billion ^[6][48].

Ocado built a comparable grid for grocery and has iterated its bots through several generations. Its current 600 Series bot is substantially lighter than its predecessors, using a chassis produced in part by additive manufacturing, and the bots work as a swarm to lift totes off the grid and deliver them to pick stations and arms ^[26].

Kiva Systems and the goods-to-person turn

Kiva Systems was founded in 2003 by Mick Mountz, with Peter Wurman and Raffaello D'Andrea as co-founders. Mountz had worked at the failed online grocer Webvan and concluded that the problem there was the inflexibility of conventional material handling systems ^[2]. Kiva's answer was a squat robot that drove under a mobile shelving pod, lifted it, and carried the whole pod to a human at a station, so the shelves moved and the picker did not.

Amazon announced on 19 March 2012 that it would acquire Kiva, based in North Reading, Massachusetts, for approximately $775 million in cash. Dave Clark, then Amazon's vice president of global customer fulfillment, framed it in operational terms: "Amazon has long used automation in its fulfillment centers, and Kiva's technology is another way to improve productivity by bringing the products directly to employees to pick, pack and stow" ^[1]. The subsidiary was renamed Amazon Robotics LLC in August 2015 ^[2].

System types at a glance

Scroll sideways for more →

AGVs, AMRs, and fleet coordination

The AGV/AMR line is about how much the building has to change. An AGV depends on something installed for its benefit, which makes routes rigid but behavior predictable. An AMR builds and localizes against a map of the space as it finds it, so a rack can move without a reprogramming project ^[5]. In practice many deployments blend the two, and vendors apply the labels loosely.

Fleet coordination is where the software gets interesting. Hundreds or thousands of robots share aisles, charging stations, and pick faces, and the assignment problem, deciding which robot takes which task and by which route, has a large effect on throughput. Amazon's DeepFleet, announced in July 2025, is a generative AI foundation model trained to coordinate robot movement across the network; the company says it improves the travel time of its robotic fleet by 10 percent, and that fleet spans more than 300 facilities worldwide ^[9][10]. Geek+ says its multi-robot hybrid scheduling system supports over 5,000 devices working simultaneously in a single warehouse ^[30].

Safety for these machines is governed by the ANSI/A3 R15.08 series for industrial mobile robots. Part 1 sets out the definitions and common terminology, including the first definition of safety considerations for mobile manipulators; Part 2, released in October 2023, covers the adaptation of those systems for particular applications and their deployment at specific sites; Part 3 sets out the user requirements for safe operation ^[42]. Robot safety constraints also shape where autonomous units are allowed to work: Amazon's original Proteus was confined to dock areas, and only the next-generation version shown in 2026 is designed to operate anywhere items need moving across a site ^[16].

Amazon Robotics

Amazon's fulfillment network is the largest single deployment of warehouse robots, and its product names have become shorthand for the categories.

Scroll sideways for more →

Proteus and Cardinal were shown in June 2022, when Amazon Robotics had deployed more than 520,000 drive units ^[12]. Sequoia went live at a Houston fulfillment center in October 2023, alongside the start of testing with Agility Robotics' bipedal Digit robot for tote recycling ^[8]. Vulcan, described by Amazon as its first robot with a sense of touch, was introduced in 2025 with pilots in Spokane, Washington and Hamburg, Germany; it uses force feedback sensors and was trained on physical data that includes touch, so it can push aside neighboring items in a fabric bin rather than treating the bin as rigid ^[13]. Blue Jay, a workcell that coordinates several arms in one workspace, entered testing at a South Carolina site, where Amazon said it could pick, stow, and consolidate approximately 75 percent of the item types stored there; in a note dated 25 February 2026 the company said it is no longer using Blue Jay in operations, though the underlying technology continues to support its network ^[14]. Project Eluna, an agentic AI system that reasons over live facility data and recommends actions to operators, was piloted at a Tennessee fulfillment center for the 2025 holiday season ^[14][15].

Amazon announced its one millionth deployed robot in July 2025; the unit went to a facility in Japan ^[9][10]. At a June 2026 event in London, the company said a next-generation Proteus would accept plain conversational instructions from employees rather than predefined tasks, with European deployment planned for the first half of 2027, as part of a stated investment of more than 10 billion euros in its European network ^[16][17].

Perception, grasping, and bin picking

Moving totes is largely solved. Reaching into a bin of mixed goods and pulling out the right one is not, and that gap in robot manipulation is where most of the AI research sits.

Amazon put the problem to academia directly. The Amazon Picking Challenge, later renamed the Amazon Robotics Challenge, asked teams to pick and stow items from shelving units. Team Delft won both the picking and stowing competitions in 2016 with an industrial arm, 3D cameras, and a custom gripper, integrating deep learning and computer vision for object recognition and pose estimation with grasp planning and motion planning under ROS ^[20]. The third edition was held at RoboCup in Nagoya in 2017, where sixteen teams competed and the Australian Centre for Robotic Vision at QUT won first place and US$80,000 with a robot called Cartman, out of $270,000 in total prizes ^[21][22].

On the research side, Dex-Net 2.0 (2017) showed that grasp planning could be learned largely from synthetic data. The authors trained a Grasp Quality Convolutional Neural Network on 6.7 million synthetic point clouds with analytic grasp metrics, then reported 93 percent success on eight known objects and 99 percent precision on 40 novel household objects, with grasp planning in 0.8 seconds ^[23].

Commercially, the direction of travel is toward robot foundation models that generalize across items instead of per-SKU engineering. In August 2024 Amazon hired Covariant founders Pieter Abbeel, Peter Chen, and Rocky Duan along with roughly a quarter of the startup's staff, and took a non-exclusive license to Covariant's robotic foundation models; Covariant continued under new leadership ^[18][19].

Ocado's On-Grid Robotic Pick (OGRP) is one of the more documented production systems. Arms mounted over the storage grid pick items directly from bins, using machine vision to choose grasp points, reinforcement learning to adapt to unfamiliar items, and behavior cloning, a form of imitation learning, for harder cases; errors are shared back across the fleet so one robot's mistake updates the others ^[24]. The published specification lists a UR10e arm with a bespoke end effector, a 1.5 kg payload, a burst rate of 630 units per hour, uptime above 98 percent, and a minimum item dimension of 2.5 cm. Exceptions fall through to a human remote pilot, a form of teleoperation ^[25]. Ocado said it picked over 30 million items with OGRP in 2024, and reported more than 50 robotic pick arms in active operation ^[24][25].

Amazon's Vulcan takes the complementary approach of adding a sense the arm never had. Rather than treating a bin as a rigid scene to be solved by vision alone, it uses force feedback to tell how hard it is pressing, which is what makes stowing into an already-full fabric pocket tractable ^[13].

Vendors and market structure

Scroll sideways for more →

The Symbotic and Walmart relationship shows how deep the buyer side has gone. On 28 January 2025 Symbotic completed the acquisition of Walmart's Advanced Systems and Robotics business; Walmart agreed to pay Symbotic a total of $520 million, including $230 million at closing, and committed to purchasing and deploying systems for 400 accelerated pickup and delivery centers at stores over a multi-year period ^[27]. Symbotic's own marketing claims for its systems, which should be read as vendor figures rather than independent measurements, include a 60 to 80 percent reduction in warehouse labor cost, a 9x throughput improvement, and accuracy above 99.99 percent ^[28].

Consolidation and failure are as much a part of the story as growth. Ocado agreed in May 2023 to buy 6 River Systems from Shopify; its Chuck collaborative robot was then deployed in over 100 warehouses with more than 70 customers ^[37]. Zebra Technologies, which had bought Fetch Robotics for $290 million in 2021, announced in December 2025 that it was winding down the resulting AMR division, terminating most robotics staff by the end of 2025 and keeping about a quarter through March 2026 to support existing deployments ^[36].

There has also been litigation over the underlying patents. Ocado and AutoStore settled their global dispute on 22 July 2023, with AutoStore paying £200 million to Ocado in installments over two years and both sides entering a cross-license of their pre-2020 patents ^[7].

Humanoid pilots

Bipedal robots are the newest and least proven category. The argument for them is that warehouses are already built for human bodies, so a machine with a human footprint can use existing stairs, totes, and conveyors without a retrofit. The open questions are cost and demonstrated uptime, since wheeled robots and fixed arms already handle most transport and picking work in production.

The most substantiated deployment is Agility Robotics' Digit at GXO Logistics. In June 2024 the two companies signed a multi-year robots-as-a-service agreement, which Agility described as the first formal commercial deployment of humanoid robots and the first humanoid RaaS arrangement; Digit moves totes from collaborative robots onto conveyors under Agility's Arc fleet platform ^[34]. In November 2025 Agility reported that Digit had moved over 100,000 totes at GXO's Flowery Branch facility in Georgia ^[35]. Amazon began testing Digit for tote recycling in October 2023 ^[8].

Published throughput and cost-per-unit figures for humanoids in logistics remain scarce, so pilot announcements are better read as engineering milestones than as evidence of unit economics.

Market conditions

The sector went through a correction after the pandemic-era buildout. In 2025 Interact Analysis cut its mobile robot forecast for the year by $800 million and reduced its 2025-2030 compound growth rate from 26 percent to 21 percent, putting 2030 mobile robot revenue at $15.6 billion; it also reduced its estimate of the market's historical size for 2024 and earlier by 8 percent, and forecast new warehouse construction at -2.0 percent year on year through 2030 ^[40].

Warehouse automation as a whole held up better. Interact Analysis reported that order intake grew 7 percent in 2025, but that the growth was concentrated in a handful of large customers including Amazon, Walmart, Tesco, and Ahold Delhaize. Vendor results diverged sharply: Dematic order intake grew 50 percent over the first three quarters of 2025 and Toyota Industries Logistics Systems 65 percent, while AutoStore saw a 5 percent decline ^[39]. For 2026 the firm expects slightly slower but more evenly distributed order growth across vendors ^[41].

Robots-as-a-service has become a common commercial model, particularly for mobile robots, because it converts a capital project into an operating expense and shifts uptime risk to the vendor. Locus Robotics reported 5 billion cumulative picks in April 2025, 24 weeks after passing 4 billion, then 6 billion in October 2025 after another 24 weeks, across over 350 sites and more than 150 brands ^[32][33].

Labor and safety

Warehouse automation is one of the most direct AI-and-jobs stories in industry, and the evidence points in more than one direction.

A 2020 investigation by Reveal from the Center for Investigative Reporting, based on internal Amazon records, reported more than 14,000 serious injuries at Amazon fulfillment centers in 2019, a rate of 7.7 serious injuries per 100 employees, and found that Amazon's own data showed serious injury rates were higher at warehouses with robots than at those without, contrary to the company's public safety claims ^[44].

Amazon disputes that picture and publishes its own figures. The company states that in 2022, recordable and lost-time incident rates were lower at Amazon Robotics sites than at non-robotics sites ^[45], and reports that it has invested more than $2.5 billion in safety since 2019, with its global recordable incident rate improved 43 percent and lost time incident rate improved 70 percent over six years ^[46].

On employment, The New York Times reported in October 2025 that internal Amazon documents described a plan to roughly double product sales by 2033 without adding to the US workforce, implying more than 600,000 jobs the company would not need to fill, and a robotics team goal of automating 75 percent of operations. Amazon spokesperson Kelly Nantel responded that "leaked documents often paint an incomplete and misleading picture of our plans," that the materials "appear to reflect the perspective of just one team and don't represent our overall hiring strategy," and noted the company's plan to fill 250,000 seasonal positions ^[47]. TechCrunch reported at the time of the one-millionth-robot announcement that Amazon says roughly 75 percent of its global deliveries are now assisted in some way by a robot ^[10].

Limits

The long tail of items is the binding constraint. Amazon has described both Vulcan and Blue Jay as handling approximately 75 percent of item types, which leaves roughly a quarter to a person ^[13][14]. Ocado routes OGRP exceptions to a human remote pilot ^[25]. Deformable packaging, transparent and reflective surfaces, heavy or oddly balanced goods, and anything that has shifted inside its bin remain difficult for both vision and grasp planning.

Capital intensity is the second limit. A dense AS/RS or cube system is a multi-year building project, and its throughput profile is fixed once built, which is a poor match for a retailer whose order mix changes. That is much of the appeal of mobile robots, which can be added a unit at a time, and much of the reason vendors pushed service-based pricing.

Third, integration is the part that fails quietly. A warehouse control system has to reconcile the robot fleet manager, the warehouse management system, sortation, and labor scheduling. Zebra's wind-down of its Fetch-derived AMR business after a $290 million acquisition is a reminder that having competent hardware is not the same as having a business ^[36].

See also

• Amazon Robotics
• Warehouse robot
• Autonomous mobile robot
• Robotics as a service
• Humanoid robot deployments
• Industrial robot

References

1. ^Amazon, "Amazon.com to Acquire Kiva Systems, Inc.", press release, 19 March 2012. press.aboutamazon.com/...-acquire-kiva-systems-inc
2. ^Wikipedia, "Amazon Robotics". en.wikipedia.org/...Amazon_Robotics
3. ^Dematic, "Why the Best Automation Decisions Are Informed by Experience". dematic.com/...ecisions-are-informed-by-experience
4. ^Wikipedia, "Automated storage and retrieval system". en.wikipedia.org/...d_storage_and_retrieval_system
5. ^Wikipedia, "Automated guided vehicle". en.wikipedia.org/...Automated_guided_vehicle
6. ^Wikipedia, "AutoStore". en.wikipedia.org/...AutoStore
7. ^Ocado Group, "AutoStore and Ocado settle all patent litigation", 22 July 2023. ocadogroup.com/...ado-settle-all-patent-litigation
8. ^About Amazon, "Amazon announces new fulfillment center robots, Sequoia and Digit", October 2023. aboutamazon.com/...troduces-new-robotics-solutions
9. ^About Amazon, "Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot". aboutamazon.com/...lion-robots-ai-foundation-model
10. ^TechCrunch, "Amazon deploys its 1 millionth robot, releases generative AI model", 1 July 2025. techcrunch.com/...bot-releases-generative-ai-model
11. ^About Amazon, "Amazon uses robots that sort, lift, and carry packages: see them in action". aboutamazon.com/...otics-robots-fulfillment-center
12. ^TechCrunch, "Amazon debuts a fully autonomous warehouse robot", 22 June 2022. techcrunch.com/...fully-autonomous-warehouse-robot
13. ^About Amazon, "Introducing Vulcan: Amazon's first robot with a sense of touch". aboutamazon.com/...on-vulcan-robot-pick-stow-touch
14. ^About Amazon, "Introducing Blue Jay and Project Eluna, Amazon's latest robotics and AI technology for its operations". aboutamazon.com/...s-amazon-fulfillment-agentic-ai
15. ^About Amazon, "Amazon's newest AI and robotics systems are empowering employees and speeding up delivery", October 2025. aboutamazon.com/...-online-shopping-speed-delivery
16. ^About Amazon, "Amazon unveils next-gen Proteus robot as part of 10 billion euro European investment in its fulfillment network". aboutamazon.com/...ope-investment-employee-support
17. ^TechStartups, "Amazon unveils Proteus, an AI warehouse robot that understands spoken commands", 4 June 2026. techstartups.com/...-amid-11-6b-european-expansion
18. ^About Amazon, "Amazon hires from AI robotics startup Covariant, licenses technology". aboutamazon.com/...amazon-covariant-ai-robots
19. ^TechCrunch, "Amazon hires the founders of AI robotics startup Covariant", 31 August 2024. techcrunch.com/...of-robotics-ai-startup-covariant
20. ^Hernandez et al., "Team Delft's Robot Winner of the Amazon Picking Challenge 2016", arXiv:1610.05514. arxiv.org/...1610.05514
21. ^IEEE Spectrum, "Aussies Win Amazon Robotics Challenge", 2 August 2017. spectrum.ieee.org/...win-amazon-robotics-challenge
22. ^About Amazon, "Amazon Robotics Challenge winners announced". aboutamazon.com/...ics-challenge-winners-announced
23. ^Mahler et al., "Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics", arXiv:1703.09312. arxiv.org/...1703.09312
24. ^Ocado Group, "Ocado's AI-powered robotic arms". ocadogroup.com/...ocado-robotic-arms
25. ^Ocado Intelligent Automation, "On-Grid Robotic Pick". ocadointelligentautomation.com/...rid-robotic-pick
26. ^Ocado Group, "Engineering the future: the evolution of Ocado's fulfilment robots". ocadogroup.com/...tion-of-ocados-fulfilment-robots
27. ^Symbotic, "Symbotic Completes Acquisition of Walmart's Advanced Systems and Robotics Business and Signs Related Commercial Agreement", 28 January 2025. symbotic.com/...signs-related-commercial-agreement
28. ^Symbotic, corporate website (vendor performance claims). symbotic.com
29. ^StockTitan, "Symbotic Reports Fourth Quarter and Fiscal Year 2025 Results", 24 November 2025. stocktitan.net/...nd-fiscal-year-2025-su0q6gxaemjc
30. ^Geekplus, "Geekplus Lists on HKEX Main Board", 9 July 2025. geekplus.com/...stics-transformation-with-robotics
31. ^Geekplus, "Geekplus Maintains the Largest AMR Market Share for the 7th Consecutive Year", 10 December 2025. geekplus.com/...ership-for-seven-consecutive-years
32. ^PR Newswire, "Locus Robotics Surpasses 5 Billion Pick Milestone", 24 April 2025. prnewswire.com/...e-warehouse-automation-302436658
33. ^Locus Robotics, "Locus Robotics Reports Record Growth, Achieving 6 Billion Picks in Fastest Time Yet", 22 October 2025. via.tt.se/...g-6-billion-picks-in-fastest-time-yet
34. ^Agility Robotics, "GXO Signs Industry-First Multi-Year Agreement with Agility Robotics", 27 June 2024. agilityrobotics.com/...ement-with-agility-robotics
35. ^Agility Robotics, "Digit Moves Over 100,000 Totes in Commercial Deployment", 20 November 2025. agilityrobotics.com/...digit-moves-over-100k-totes
36. ^The Robot Report, "Zebra Technologies winding down Fetch-based mobile robot group", 12 December 2025. therobotreport.com/...tch-based-mobile-robot-group
37. ^Ocado Group, "Ocado Group announces agreement to acquire 6 River Systems", 4 May 2023. ocadogroup.com/...ement-to-acquire-6-river-systems
38. ^Berkshire Grey, "Berkshire Grey Enters into Definitive Merger Agreement with SoftBank Group for Go-Private Transaction", 24 March 2023. berkshiregrey.com/...up-for-go-private-transaction
39. ^Interact Analysis, "Warehouse automation in 2025: a market defined by concentrated growth", February 2026. interactanalysis.com/...warehouse-automation-in-2025
40. ^Interact Analysis, "Mobile robot market forecast slashed". interactanalysis.com/...ot-market-forecast-slashed
41. ^Interact Analysis, "Warehouse automation: what to expect in 2026", February 2026. interactanalysis.com/...ion-what-to-expect-in-2026
42. ^The Robot Report, "New AMR safety standard available with release of ANSI/A3 R15.08-2", 26 October 2023. therobotreport.com/...-release-of-ansi-a3-r15-08-2
43. ^Hai Robotics, "Types of Warehouse Robots, ACR Robotic Case Picking System". hairobotics.com/products
44. ^TechCrunch, "Serious injuries at Amazon fulfillment centers topped 14,000 despite the company's safety claims", 29 September 2020. techcrunch.com/...spite-the-companys-safety-claims
45. ^About Amazon, "How robotics improve safety in Amazon's operations". aboutamazon.com/...amazon-robotics-safety
46. ^About Amazon, "Amazon invests $2.5 billion in workplace safety, reduces workplace injuries". aboutamazon.com/...ce-safety-2025-injury-reduction
47. ^Fox Business, "Amazon defends ambitious AI strategy that could prevent 600,000 future hires through innovation", 22 October 2025. foxbusiness.com/...future-hires-through-innovation
48. ^AutoStore, corporate website (installed base and country count). autostoresystem.com