Last reviewed
May 2, 2026
Sources
27 citations
Review status
Source-backed
Revision
v1 · 4,552 words
Add missing citations, update stale details, or suggest a clearer explanation.
An Automated Guided Vehicle (AGV) is a driverless mobile robot used in factories, warehouses, hospitals, ports, and distribution centers to move materials along defined paths. Classical AGVs follow physical or virtual guidance (an inductive wire under the floor, a magnetic strip, painted lines, optical markers, or a map of laser reflectors mounted on walls and columns) and rely on a central controller to assign jobs and resolve conflicts. The technology dates to 1953, when Barrett Electronics of Northbrook, Illinois converted a tow tractor into a wire-guided vehicle for a grocery warehouse. Seven decades later, AGVs and their close cousins, autonomous mobile robots (AMRs), are the workhorses of e-commerce fulfillment, automotive assembly, and intralogistics, with Amazon alone running more than a million mobile units across its global fulfillment network as of 2025.
The distinction between an AGV and an AMR is a matter of degree rather than kind. A traditional AGV sticks to a fixed route and is, in effect, a moving conveyor; an AMR uses SLAM and on-board perception to navigate freely, replan around obstacles, and share unstructured space with people. The two categories now sit on a spectrum, and standards bodies treat them under the same safety umbrella. AGVs and AMRs are also a major surface where modern artificial intelligence reaches industry: SLAM, deep learning perception, multi-agent path finding, reinforcement learning for fleet routing, and (more recently) generalist policies for humanoid variants such as Apptronik's Apollo and Agility's Digit all run on AGV-class hardware.
A modern AGV is an unmanned, electrically powered ground vehicle with five core subsystems:
A group of AGVs working together with a controller and the surrounding infrastructure (charging stations, doors, lifts, RFID tags) is called an Automated Guided Vehicle System (AGVS). The German VDI 2510 guideline lays out this system view in detail.
The first commercial AGV was built in 1953 by Arthur "Mac" Barrett Jr., founder of Barrett Electronics. Barrett's "Guide-O-Matic" was a tow tractor that followed an overhead wire installed in a grocery warehouse in Mercury, Pennsylvania. Within a few years the wire was moved into a slot cut in the floor, and Barrett's design became the basic template for the next thirty years: a low-voltage signal in a buried wire generates a magnetic field, an antenna under the vehicle senses the field, and steering corrects the vehicle to keep the antenna centered. By the late 1950s and 1960s, US factories and a handful of European auto plants had begun using these tow trains for line-side delivery. Volvo's Kalmar plant, opened in 1974, was an early showcase: AGVs ferried car bodies between assembly stations, replacing the conventional moving line and giving workers more control over their pace.
The wire-guidance approach had two drawbacks that made customers grumble. Cutting trenches in concrete is expensive, and once the wire is laid, changing the route means breaking up the floor again. In the early 1980s, the Swedish firm NDC (Netzler & Dahlgren Co., later acquired by Kollmorgen) commercialized laser-guided navigation. The vehicle carries a rotating laser on top; the laser pulses out and reads back returns from retro-reflective tags placed on walls and columns at known coordinates. Triangulating from at least three reflectors gives a position fix to a few millimeters, with no floor work required. This unlocked the European intralogistics boom: paper mills, automotive plants, and pharmaceutical warehouses could redraw routes in software. The same period saw the rise of magnetic-tape and magnetic-spot guidance for cheaper, less precise applications, and the spread of inertial navigation (gyroscopes and odometry) as a backup or as the primary system in environments where reflectors were not practical.
The deepest change came not from a German auto plant but from a Boston startup. Kiva Systems, founded in 2003 by Mick Mountz, Peter Wurman, and Raffaello D'Andrea, threw out the conveyor-and-tote layout that had defined warehousing for decades. Instead of moving items past pickers on a belt, Kiva's drive units (small, orange, four-wheeled robots about the size of a footstool) lifted entire shelving pods and brought them to a stationary picker. Pickers stood still; the inventory came to them. The architecture worked because Kiva's central planner solved a dense multi-agent path finding problem in real time, routing hundreds of robots across a grid floor without collisions. In March 2012, Amazon acquired Kiva for $775 million in cash. Kiva was renamed Amazon Robotics in 2015, and Amazon stopped selling Kiva drives to outside customers.
That last fact created a vacuum, and a wave of new companies rushed in: Locus Robotics (spun out of the Quiet Logistics 3PL in 2014), 6 River Systems (founded in 2015 by ex-Kiva engineers, sold to Shopify in 2019 and again to Ocado in 2023), Fetch Robotics (Melonee Wise, 2014, sold to Zebra in 2021), and MiR (Mobile Industrial Robots, Denmark, 2013, sold to Teradyne in 2018). Around the same time, China saw the rise of Geek+, Hai Robotics, and Quicktron.
The wave that came after Kiva did not call itself "AGV." The new term, popularized by MiR in early 2017, was Autonomous Mobile Robot. The shift was real: instead of following fixed paths on a sealed grid, AMRs navigated open floors using SLAM, swerved around boxes and humans, and could be rerouted with a fleet-management UI rather than a contractor with a tape gun. AMRs were also pitched at small and mid-sized warehouses, not just billion-dollar fulfillment centers. By 2020 the global installed base of AMRs and AGVs combined had crossed several hundred thousand units. By 2025 Amazon alone passed one million.
The latest twist is humanoids on the warehouse floor. In 2022 Amazon unveiled Proteus, its first fully autonomous AMR, which works in shared space with people. In 2023 GXO Logistics announced a paid deployment of Agility Robotics' Digit, a bipedal robot that picks up and walks totes between racks and conveyors. Amazon began testing Digit at its Sumner, Washington research facility around the same time. Apptronik's Apollo entered a proof-of-concept with GXO and a manufacturing pilot with Mercedes-Benz. None of these humanoids will replace classical AGVs, which are cheaper, faster, and more reliable for the wheeled-cart job. But they extend the AGV/AMR category into tasks (unloading trailers, depalletizing mixed cases, stuffing totes into shelves) that wheeled platforms cannot do.
The industry uses the two terms inconsistently, but the practical distinctions look like this.
| Dimension | Classical AGV | Modern AMR |
|---|---|---|
| Path | Fixed, defined by wire, tape, or reflector map | Free; planned on the fly |
| Primary navigation | Inductive guide wire, magnetic tape, laser triangulation | SLAM (LiDAR or visual), natural-feature mapping |
| Obstacle response | Stop and wait | Replan and detour |
| Layout changes | Hours to days; physical infrastructure | Minutes; software map edit |
| Traffic management | Centralized blocking and zoning | Mix of centralized and decentralized |
| Typical sectors | Auto plants, paper mills, semiconductor fabs | E-commerce fulfillment, hospitals, mid-size 3PLs |
| Throughput per robot | High and predictable | Lower per unit, more flexible at the fleet level |
| Capex per unit | Higher (often $80K to $250K) | Lower (often $30K to $80K) |
A semiconductor fab still picks classical AGVs because it values determinism over flexibility: the same wafer carrier always takes the same path on the same cycle. An e-commerce 3PL leans toward AMRs because its product mix and floor plan churn every quarter. Many sites run hybrid fleets, which is exactly the use case that VDA 5050 was written to support.
| Method | Year introduced | How it works | Strengths | Weaknesses |
|---|---|---|---|---|
| Inductive guide wire | 1953 | Buried wire emits low-frequency signal; antenna under vehicle steers to null the signal | Cheap to operate; immune to dirt and lighting | Floor cutting; inflexible routes |
| Magnetic tape | 1980s | Thin magnetic strip glued to floor; Hall-effect sensor detects it | No floor cutting; cheap | Tape wears out; cannot handle complex intersections |
| Magnetic spots | 1990s | Discrete magnets buried in floor as a coarse grid | Durable; no surface marks | Requires drilling many small holes |
| Optical / painted line | 1980s | Camera follows a colored stripe | Easy to relay routes | Sensitive to lighting; lines get dirty |
| Laser triangulation | early 1990s | Rotating laser hits retro-reflectors at known coordinates; vehicle solves for its pose | Millimeter precision; route flexibility | Reflectors must be installed and maintained; line-of-sight required |
| Vision marker (QR, AprilTag) | 2010s | Cameras read fiducials on floor or shelf | Cheap markers; works in cluttered space | Needs clean tags; lighting sensitive |
| Natural-feature SLAM | 2010s | LiDAR or stereo cameras build a map of fixed features (walls, racks); robot localizes against it | No infrastructure; rapid layout change | Compute-heavy; struggles in featureless or highly dynamic spaces |
| Inertial / odometry | 1970s | Wheel encoders plus gyroscope integrate position from a known start | Works in open spaces; fast | Drifts; usually a complement, not a primary |
| GPS / GNSS | 1990s | Outdoor only; satellite fix | Works on yards and ports | Useless indoors; insufficient precision indoors |
In practice, most modern AMRs run sensor fusion. A typical stack might combine a 2D LiDAR for SLAM, a 3D depth camera for obstacle perception, wheel odometry for short-horizon dead reckoning, and a Kalman filter or particle filter to merge the streams into a single pose estimate.
Three standards matter for anyone deploying AGVs at scale.
ISO 3691-4:2023. "Industrial trucks, safety requirements and verification, Part 4: Driverless industrial trucks and their systems." The global safety baseline. Covers obstacle detection, bumper design, ground clearance to prevent foot entrapment, E-stop placement, performance levels per ISO 13849-1, and the responsibilities of OEM, integrator, and end user. The 2023 revision is now the reference text in most jurisdictions, with ANSI/RIA R15.08 covering the same ground in the US for industrial mobile robots.
VDA 5050. A standardized communication interface between mobile robots and a master fleet controller, written by the German Association of the Automotive Industry (VDA) and the VDMA Materials Handling and Intralogistics Association, with input from the Karlsruhe Institute of Technology. The interface uses MQTT as the transport and JSON as the message schema. The point of VDA 5050 is to let a single fleet manager dispatch, route, and monitor AGVs and AMRs from different vendors on the same site. Before VDA 5050, every brand needed its own master controller, which made mixed fleets a nightmare. Adoption is voluntary, but tenders in the German auto industry now routinely require it.
VDI 2510. A multi-part guideline from the Verein Deutscher Ingenieure (Association of German Engineers) covering AGV systems. Sheet 1 deals with infrastructure, Sheet 2 with safety, Sheet 3 with peripheral equipment such as fire doors and lifts, Sheet 4 with power supply and charging.
Other relevant texts include EN 1525 (the older European safety standard for driverless trucks, largely superseded by ISO 3691-4), ANSI/ITSDF B56.5 (the US analogue for driverless industrial trucks), and IEC 62998 series for safety-related sensors.
The AGV and AMR market is fragmented at the bottom and concentrated at the top. The list below is not exhaustive, but it covers most of the players a procurement team will see in a serious tender.
| Company | Headquarters | Specialty | Notes |
|---|---|---|---|
| Daifuku | Osaka, Japan | Conveyors, AS/RS, AGVs, airport baggage | One of the world's largest material-handling firms; trailing 12-month revenue around $4.4 billion as of late 2025; over 20,000 vehicles delivered |
| Toyota Industries | Kariya, Japan | Forklifts and AGVs (Toyota, Raymond, CESAB), plus Vanderlande and Bastian Solutions for systems | Largest forklift manufacturer in the world by revenue |
| KION Group (Linde, STILL, Dematic) | Frankfurt, Germany | Forklifts, AGVs, AS/RS, software | Second-largest forklift manufacturer; Dematic supplies large e-commerce systems |
| Jungheinrich | Hamburg, Germany | Forklifts, AGV-converted forklifts, intralogistics | Strong in European auto and retail DCs |
| Kuka Swisslog | Augsburg, Germany / Buchs, Switzerland | AGVs, AS/RS, healthcare logistics | Owned by China's Midea since 2017 |
| Murata Machinery (Muratec) | Kyoto, Japan | Semiconductor AMHS, AGVs for fabs | Dominant in 300mm wafer fabs |
| JBT (now Oshkosh AeroTech and others after 2025 spinoffs) | Chicago, US | AGVs for food processing and airports | Long history with Volvo and General Mills |
| Locus Robotics | Wilmington, Massachusetts | AMRs for goods-to-person picking | Around $400M raised; thousands of robots deployed at DHL, FedEx, GXO |
| Amazon Robotics | North Reading, Massachusetts | Captive fleet for Amazon | Not sold to third parties; over 1 million robots deployed |
| Symbotic | Wilmington, Massachusetts | End-to-end case-handling robotics | Public on NASDAQ as SYM; deploys across all 42 Walmart US regional DCs; acquired Walmart's Advanced Systems and Robotics business in January 2025 |
| Geek+ | Beijing, China | AMRs (goods-to-person, sortation, forklift) | Listed on Hong Kong Stock Exchange in 2025 |
| Hai Robotics | Shenzhen, China | Autonomous case-handling robots (ACR) | Pioneer of box-level dynamic storage |
| Quicktron | Shanghai, China | AMRs and goods-to-person | Founded 2014; client list includes DHL, Nike |
| MiR (Mobile Industrial Robots) | Odense, Denmark | Indoor AMRs | Owned by Teradyne |
| OTTO Motors | Kitchener, Canada | Heavy-payload AMRs | Owned by Rockwell Automation since 2023 |
| Boston Dynamics | Waltham, Massachusetts | Humanoids and case-handling | Stretch is its warehouse robot for trailer unloading |
| Agility Robotics | Salem, Oregon | Bipedal humanoids for logistics | Digit deployed at GXO and tested at Amazon |
| Apptronik | Austin, Texas | Bipedal humanoid Apollo | Pilots with GXO, Mercedes-Benz, Jabil |
Amazon Robotics deserves its own section because of its scale and influence on the rest of the industry. The story starts with the $775 million acquisition of Kiva Systems in March 2012. By 2015 the unit had been renamed Amazon Robotics. By July 2025 Amazon had crossed one million mobile robots across more than 300 facilities, with the milestone unit going to a Japanese fulfillment center.
The fleet has evolved through several generations.
| Robot | Year introduced | Function |
|---|---|---|
| Hercules (and earlier Kiva drive units) | 2012 (Kiva), upgraded as Hercules | Lifts and moves shelving pods to stationary pickers |
| Pegasus | 2018 | Sorts finished parcels by zip code or delivery route |
| Xanthus | 2019 | Modular drive platform; can swap top modules for different tasks |
| Robin | 2021 | Robotic arm with vacuum end-effector that picks parcels off conveyors |
| Cardinal | 2022 | Larger arm that handles heavier packages (up to ~50 lb) |
| Proteus | 2022 | Amazon's first fully autonomous AMR; navigates open floor in mixed human spaces |
| Sequoia | 2023 | Storage and retrieval system using mobile robots and containerized pods |
| Sparrow | 2022 | Arm with computer-vision sorting that picks individual items from totes |
| Vulcan | 2025 | Tactile robot for shelf stowing |
| Titan | 2025 | Heavy-lift mobile drive (up to 2,500 lb) |
On top of the fleet, Amazon runs a software stack called DeepFleet, announced in 2025, which is a generative AI foundation model for routing the entire mobile fleet. According to Amazon's own figures, DeepFleet improves average travel time across the fleet by about 10 percent. The internal name conceals what is, in practice, a large-scale reinforcement learning and sequence-modeling system trained on the operational data from a million robots, which is more than any academic group has ever had access to.
The AGV started as a controls problem. The AMR is, increasingly, an AI problem. Five subfields are doing most of the work.
Simultaneous Localization and Mapping is the foundational capability for any AMR that does not rely on fixed beacons. The robot builds a map of its environment while estimating its own pose within that map, and updates both at every sensor frame. Two algorithm families dominate. 2D LiDAR SLAM uses a horizontally mounted laser scanner to build an occupancy grid; classic implementations include GMapping (a Rao-Blackwellized particle filter) and Cartographer (Google's pose-graph optimizer). Visual SLAM uses cameras and runs feature extraction or direct photometric optimization; ORB-SLAM, RTAB-Map, and VINS-Mono are common research baselines. Industrial AMRs typically run 2D LiDAR SLAM in production because it is cheap, fast, and works in featureless aisles. Visual SLAM is more common on lighter platforms and on humanoids.
A single robot navigating a static map is a solved problem. Use A* on a graph, RRT on a continuous space, or a sampling-based optimizer like Hybrid A*. The interesting problem is the fleet. With hundreds of robots in a Kiva-style grid, the planner must assign a path to each robot such that no two robots occupy the same cell at the same time, no robot deadlocks waiting for another, and the makespan or sum-of-costs is small. This is multi-agent path finding, and it has been an active research field since at least the 2010s. Conflict-Based Search (CBS), introduced by Sharon, Stern, Felner and Sturtevant in 2015, is the canonical optimal algorithm: a high-level search over conflicts spawns single-agent A* searches with extra constraints, and the high-level tree expands until the leaf is conflict-free. Variants include ICBS, EECBS, and CBS with priorities for suboptimal but faster solutions. Real warehouses use heuristic priority planners and rolling-horizon replanning rather than provably optimal MAPF, because robots arrive and leave at different times and the problem is never quite the textbook formulation. Amazon's DeepFleet, Geek+'s fleet manager, and Symbotic's coordinator all sit somewhere on the spectrum between classical MAPF and learned policies.
When MAPF instances grow past a few hundred agents on a few thousand cells, exact methods slow down and engineers reach for reinforcement learning and learned heuristics. The literature on multi-agent deep RL for warehouse routing is large and growing. Papers from 2020 onward apply DQN, MADDPG, QMIX, and centralized-training-decentralized-execution variants to warehouse-scale instances. Cooperative MARL has been used for joint task assignment and path planning, where the agent's policy outputs both a target shelf and a movement direction. The published gains are often impressive but tend to be benchmarked on simulators; real-world deployments still rely heavily on hand-engineered priority rules with RL as a tie-breaker or a recommender. Generative models that imitate dispatcher behavior across millions of historical decisions (DeepFleet's likely architecture) are the more recent direction.
AMRs use computer vision at three layers. Safety perception detects humans, pallets, and other obstacles in the path; production stacks rely on safety-rated 2D LiDAR for the regulated detection function and use 3D depth cameras and CNN-based object detection (YOLO, SSD, or proprietary networks) for the soft layer that triggers replanning. Localization aids include AprilTag and ArUco fiducial detection on aisle ends, ceiling cameras for overhead localization in tight spaces, and shelf-spine cameras for verifying that the right pod has been lifted. Pick-and-place uses convolutional neural network-based segmentation and grasp planning on arm-equipped units like Cardinal, Sparrow, and Robin. The 2024 to 2026 generation increasingly mixes traditional vision with vision-language models for unstructured situations.
The newest direction is policies that handle multiple manipulation tasks from natural-language instructions. OpenVLA (the 7B-parameter open vision-language-action model from a 2024 Stanford-led collaboration), Pi Zero (Physical Intelligence's flagship policy), RT-2 from Google DeepMind, and several proprietary stacks at Apptronik, Figure, and 1X are training on large mixtures of teleoperation and warehouse video. The pitch is that one policy can run a Digit, an Apollo, or a Stretch, with task-level prompting replacing per-task engineering. Whether this generalist approach beats specialized stacks for narrow warehouse jobs is still being argued out at the deployment level. Locus and Geek+ continue to ship more units per quarter than every humanoid company combined, and probably will for several more years.
AGVs and AMRs show up in nearly every industry that moves heavy or repetitive things, but a few sectors dominate.
E-commerce fulfillment. The largest single application by unit count. Amazon, JD.com, Cainiao (Alibaba), Walmart through Symbotic, Ocado in groceries, and a long tail of 3PLs (DHL, GXO, FedEx Supply Chain, GEODIS) run goods-to-person AMR fleets at scale. Locus is the dominant third-party choice in North America; Geek+ and Hai Robotics dominate in China.
Automotive manufacturing. The original AGV market, and still a steady one. Volkswagen, BMW, Mercedes, Tesla, and the Asian OEMs use AGVs for sub-assembly delivery, kitting, and final-assembly conveyance. The push toward VDA 5050 is largely driven by mixed-vendor auto deployments.
Semiconductor fabs. A specialized world. 300mm wafer fabs use overhead transport (OHT) systems that are a kind of inverted AGV (the vehicle hangs from rails and carries FOUPs between tools). Murata Machinery and Daifuku effectively split this market.
Hospitals. Aethon's TUG, Swisslog Healthcare's TransCar, and several others move medications, lab samples, linens, and food trays. The hospital environment is brutal on perception (crowded corridors, kids, beds in the way) and is one of the few places where AMR safety design has had to deal with truly unstructured pedestrian traffic at scale.
Ports and yards. Shipping container terminals at Rotterdam, Long Beach, Qingdao, and others use heavy-duty outdoor AGVs (Kalmar, Konecranes, ZPMC) to ferry containers between cranes and stacks. These are massive (50+ tonne payloads) and use a mix of GPS, transponders, and laser scanners.
Retail and grocery. Brain Corp's BrainOS runs autonomous floor scrubbers in big-box retailers; Walmart and Sam's Club use them at scale. Ocado Smart Platform uses small grid-based bots in a vertical structure to fulfill grocery orders.
AGVs and AMRs are not a clean solution to every materials handling problem.