Autonomous Mobile Robot
Last reviewed
Apr 27, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 4,475 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 27, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 4,475 words
Add missing citations, update stale details, or suggest a clearer explanation.
An autonomous mobile robot (AMR) is a category of [[robotics|robot]] that moves through unstructured or partially structured environments without relying on fixed physical guides such as magnetic tape, embedded wires, painted lines, or QR-code grids. Instead of following a pre-programmed track, an AMR builds and updates its own internal map of the world, decides on its own route from a current pose to a goal, and replans that route in real time when people, forklifts, pallets, doors, or other robots get in the way. The defining property of the class is local autonomy: the robot can perceive, decide, and act without continuous instructions from a human operator or a centralized track-control system, although it usually still reports to a higher-level fleet manager that assigns missions and prevents conflicts between machines.
Most commercial AMRs are wheeled platforms designed for flat indoor floors, but the same core architecture also applies to outdoor sidewalk delivery bots, four-wheel field robots, tracked inspection units, and a growing population of legged and bipedal machines. Underneath the chassis differences sit the same building blocks: a multi-modal sensor suite, a [[slam|simultaneous localization and mapping]] (SLAM) layer that fuses those sensors into a consistent map and pose estimate, a planner that produces collision-free trajectories, a controller that executes them, and a connectivity layer that talks to a fleet manager or warehouse management system.
AMRs are the workhorses of contemporary warehouse automation, but they have spread far beyond the loading dock. They sweep airport terminals, deliver linens and medications inside hospitals, ferry food across restaurant dining rooms, patrol corporate campuses, inspect substations, vacuum living rooms, and complete the last few hundred meters of e-commerce orders on city sidewalks. According to the International Federation of Robotics, around 81,800 mobile robots for intralogistics applications were sold worldwide in 2024 alone, making transportation and logistics the single largest segment of the professional service-robot market. [1]
The phrase "autonomous mobile robot" has been used in academic robotics since the 1980s, when research platforms such as Stanford's Cart, Carnegie Mellon's Navlab, and the Hilare series at LAAS in Toulouse demonstrated that a robot could move through an unmodified environment using onboard sensing alone. The hardware and algorithms needed for reliable indoor autonomy (low-cost laser rangefinders, real-time SLAM, fast onboard processors, and longer-life batteries) only matured in the 2000s.
The modern industrial usage of the term gained traction in the 2010s as a way to differentiate new sensor-driven machines from earlier [[automated guided vehicle|automated guided vehicles]] (AGVs). AGVs had been a fixture of factory and warehouse floors since the 1950s, when the first towing tractors followed wires buried in the floor at the Mercury Motor Freight terminal in South Carolina. AGV navigation evolved through magnetic tape, optical lines, reflective targets, and laser triangulation, but for most of its history the technology required some form of fixed infrastructure. The newer AMR designs dropped that requirement and instead used [[lidar|lidar]] and cameras to build their own maps. Market analysts at firms such as Interact Analysis now forecast AGV revenue to fall from around 33 percent of total mobile robot revenue in 2024 to roughly 20 percent in 2030. [2]
The boundary between AGVs and AMRs is blurrier in 2026 than it was a decade ago. Many traditional AGV vendors now ship platforms with [[slam|SLAM]] navigation and dynamic obstacle avoidance, while several AMR vendors have added route-following modes for tight aisles where deterministic behavior is preferred.
AGVs and AMRs share a goal, moving payloads from one place to another without a human driver, but they reach that goal through fundamentally different design philosophies. The table below summarizes the contrast as it is generally understood in the intralogistics market.
| Aspect | Automated Guided Vehicle (AGV) | Autonomous Mobile Robot (AMR) |
|---|---|---|
| Primary navigation | Follows fixed paths defined by magnetic tape, wires, optical lines, or reflective targets | Builds and updates its own map; plans routes dynamically using SLAM |
| Infrastructure required | Floor markings, beacons, or buried wires; changes require recommissioning | Minimal; mostly Wi-Fi coverage and chargers |
| Obstacle handling | Stops when blocked, waits for the obstacle to clear | Reroutes around the obstacle and continues |
| Sensor suite | Often a single 2D laser scanner plus magnetic or optical sensors | Multi-modal: 2D and 3D lidar, depth cameras, RGB cameras, IMU, ultrasonic |
| Best fit | High-volume, repetitive flows in stable layouts (assembly lines, automotive) | Variable flows, mixed-use floors, frequent layout changes (e-commerce, third-party logistics) |
| Typical cost | Lower per unit, higher commissioning cost | Higher per unit, lower commissioning cost |
| Software updates | Rare, tied to layout changes | Frequent; new map regions or behaviors pushed over the air |
In practice, large operators run mixed fleets. AGVs handle predictable, high-throughput trunk routes between fixed stations; AMRs handle the messy first and last legs where humans, forklifts, and changing pallet locations make rigid paths impractical. The VDA 5050 standard, discussed below, exists precisely because hybrid fleets need a common control protocol. [3]
An AMR can be decomposed into a stack of overlapping subsystems. Each layer feeds the layer above it and depends on the layers below.
The sensor suite is the AMR's window onto the world. Most warehouse-class machines combine a 2D safety-rated lidar at the base with one or more 3D lidars or depth cameras higher up the chassis, plus an inertial measurement unit (IMU) and wheel encoders for odometry. RGB cameras add semantic context such as pallet labels, AprilTags, and human-form recognition. Each sensor type has complementary strengths and weaknesses: lidar gives centimeter-level geometric accuracy and works in poor lighting, cameras add color and texture but struggle with motion blur and glare, and the IMU fills in the high-frequency gaps between slower sensor updates. Combining them is the job of [[sensor_fusion|sensor fusion]], which is now the standard approach for industrial-grade SLAM. [4]
A typical fielded AMR carries between five and fifteen sensors. Outdoor delivery robots add GNSS receivers and stereo cameras for traffic perception. Hospital robots emphasize ultrasonic and short-range cameras for reflective floors and people moving unpredictably. Security robots include thermal cameras and gunshot detectors. The sensor mix is where vendors most visibly differentiate their hardware.
The dominant technique is some flavor of [[slam|SLAM]], which solves localization and mapping simultaneously by treating new sensor frames as both observations of the existing map and constraints on the robot's pose. 2D laser SLAM, built around algorithms such as Cartographer, gmapping, and SLAM Toolbox, remains the workhorse of warehouse automation. Visual SLAM (V-SLAM), which uses cameras as the primary sensor, is now mature enough for industrial deployment and was the core technology behind the Sevensense product line that ABB acquired in 2024 to add 3D-vision smarts to its mobile robots. [5]
Newer systems fuse [[lidar|lidar]], cameras, and IMU at the front end of the SLAM pipeline. Tightly-coupled lidar-inertial-visual SLAM produces drift-free trajectories over hundreds of meters even when one sensor stream is briefly degraded, which matters in featureless aisles, glaring sunlight near loading bay doors, or under metal mezzanines that confuse lidar reflections.
Most commercial stacks layer two planners. A global planner computes a coarse route across the whole map using algorithms such as A*, Dijkstra, or Theta*. A local planner then refines that route into a smooth, dynamically feasible trajectory while reacting to sensor updates at 10 to 50 Hz. Common local planners include the Dynamic Window Approach, Timed Elastic Bands, and Model Predictive Path Integral methods.
The ROS 2 [[navigation|Nav2]] project has become a de facto reference implementation of this stack and is used in production by more than 100 companies. Nav2 provides plug-in slots for SLAM, behavior trees, costmaps, and recovery behaviors. [6]
A behavior layer, often expressed as a behavior tree or finite state machine, sequences high-level actions such as "go to dock", "pick up cart", or "return to charger". The fleet manager sits above many robots and assigns missions, balances charging schedules, prevents conflicts at intersections, and exposes APIs to a warehouse management system. In healthcare or hospitality settings, the fleet manager also integrates with elevators, badge readers, and door openers.
Open interoperability matters because customers do not want to be locked into a single vendor. The VDA 5050 standard, published jointly by the German automotive association (VDA) and the mechanical-engineering association (VDMA), defines a JSON-over-MQTT protocol that any compliant AGV or AMR can use to talk to any compliant fleet manager. The current published version is 2.1.0, released in January 2025, and supports heterogeneous fleets with vehicles from different manufacturers under one master controller. [3]
AMRs span an enormous range of physical sizes and operating environments. The table below maps the main commercial categories to representative products and use cases as of 2026.
| Category | Typical use case | Representative products and vendors |
|---|---|---|
| Warehouse goods-to-person | Bringing shelves or totes to a stationary picker | [[geek_plus |
| Warehouse person-to-goods | Following or leading a human picker through aisles | [[locus_robotics |
| Warehouse pallet movement | Moving full or partial pallets between stations | OTTO Lifter, Vecna Pallet Truck, MiR 1350 |
| Carrypick / cube storage AMR | Hauling shelf modules in a goods-to-person grid | Swisslog CarryPick, KMP600 (KUKA / Swisslog), [[autostore |
| Manufacturing line-side delivery | Bin replenishment in automotive and electronics plants | KUKA KMP, Mobile Industrial Robots MiR 600, ASTI Mobile Robotics platforms |
| Hospital service | Linen, meal, lab specimen, and medication transport | Aethon TUG, Diligent Robotics Moxi, Savioke Relay |
| Restaurant and hospitality | Food delivery, dish bussing, hotel room service | [[bear_robotics |
| Cleaning | Vacuuming, scrubbing, and sweeping | [[roomba |
| Outdoor last-mile delivery | Sidewalk and street parcel and food delivery | Starship Technologies sidewalk bot, Nuro R2/R3, Coco Robotics, Serve Robotics |
| Security and inspection | Patrol, intrusion detection, equipment monitoring | Knightscope K3, K5, K7, Cobalt Robotics, ANYbotics ANYmal |
What follows looks at each major category in more depth.
Intralogistics is where AMRs first reached commercial scale, and it is still the largest market by a wide margin. The IFR's World Robotics 2025 report found that around 102,900 robots for transportation and logistics tasks were sold worldwide in 2024, of which roughly 81,800 were intralogistics mobile robots; the Asia-Pacific region produced about 84 percent of the total. [1]
Within warehousing, AMRs split into distinct flows. Goods-to-person systems use small low-profile robots to lift mobile shelves or totes and bring them to a stationary picker. The original Kiva Systems robots, acquired by [[amazon_robotics|Amazon]] in 2012 for 775 million USD, popularized the model; today [[geek_plus|Geek+]], [[hai_robotics|Hai Robotics]], Quicktron, and Libiao operate the largest goods-to-person fleets, with Geek+ alone shipping tens of thousands of robots and going public on the Hong Kong Stock Exchange in 2025.
Person-to-goods systems work in the opposite direction: the robot moves to the picker, who walks shorter loops between consecutive picks. [[locus_robotics|Locus Robotics]] is the most visible vendor in this niche, with its Vector and Origin robots used by warehouse operators across North America and Europe. Locus reports that its AMRs make existing pickers two to three times more productive while reducing overall labor requirements by 20 to 30 percent. [7] 6 River Systems, founded by ex-Kiva engineers and acquired by Shopify for around 450 million USD in 2019, was sold on to Ocado in 2023 in one of the more notable consolidation moves of the cycle. [8]
Warehouses also deploy AMR pallet trucks, tugger trains, and case-pick robots for heavier loads. The Swisslog and KUKA partnership packages many of these into integrated cells, including the CarryPick AMR system that uses KMP600 vehicles to bring mobile racks to operator stations. Swisslog passed 300 [[autostore|AutoStore]] integrations in 2024.
Amazon, the single largest operator of AMRs in the world, runs more than 750,000 robots across its global network, including the Proteus AMR introduced in 2022 that operates in shared space with human associates without safety cages. Proteus uses fused lidar and camera perception to detect people, marking Amazon's transition from caged AGV fleets toward genuinely autonomous floor traffic. [9]
After intralogistics, the largest mobile-robot category in the IFR data is hospitality, with more than 42,000 units sold in 2024. [1] These are smaller, friendlier-looking robots that ferry plates and dirty dishes between the kitchen and the dining room.
[[bear_robotics|Bear Robotics]] Servi is one of the most widely deployed restaurant AMRs in North America. The robot uses lidar and depth cameras to navigate dining rooms and can carry up to roughly 20 kilograms across multiple trays. SoftBank originally invested in Bear, and in January 2025 LG Electronics acquired a 51 percent controlling stake in the company, building on its earlier 60 million USD investment. [[pudu_technology|Pudu Robotics]] and Keenon Robotics, both Shenzhen-based, dominate the Chinese market and have grown rapidly abroad, with Pudu's BellaBot and PuduBot 2 and Keenon's DINERBOT T8 widely used in Asian and Middle Eastern dining venues.
Hospital AMRs handle the logistical underbelly of a modern medical center: delivering medications, lab specimens, surgical packs, food trays, linen, and biohazardous waste between floors and departments. Aethon's TUG, used in more than 140 U.S. hospitals including 37 Veterans Affairs facilities, is a bin-on-wheels AMR that integrates with elevators and automatic doors so that a robot can travel from a third-floor pharmacy to a fifth-floor nursing station without human intervention.
Diligent Robotics' Moxi takes a different approach: it has a single robotic arm on top of a mobile base, which lets it handle items like supply totes rather than only towing carts. Moxi targets the "last 100 yards" of nursing logistics, the small but cumulative tasks that pull nurses away from patient care.
Cleaning robots are the highest-volume AMR category if consumer products are included. The IFR notes that the professional cleaning robot segment alone grew 34 percent in 2024 to more than 25,000 units, while the consumer side is dominated by [[roomba|iRobot Roomba]] and Roborock units that sell millions of devices a year. Modern home robots use lidar SLAM, camera-based object recognition to avoid pet waste and electrical cords, and persistent floor maps. The Roborock Saros 20, launched in 2025, includes mechanical mop arms that extend beneath couches, and the iRobot Roomba 105 added lidar navigation and an AutoEmpty dock for 75 days of unattended operation.
Professional floor scrubbers such as the Tennant T7AMR and the Avidbots Neo scale the same idea to airports, warehouses, and supermarkets, typically retrofitted from standard ride-on scrubbers with sensor packs and SLAM software.
Outside the building, AMRs face a harder problem: pedestrians, weather, curbs, intersections, and adversarial humans. Despite that, sidewalk delivery has reached real scale. Starship Technologies, founded by two of the cofounders of Skype, reported 9 million autonomous deliveries, 19 million kilometers of autonomous driving, and 2,700 robots in operation across 270 locations in 7 countries as of October 2025. Most routes are college campuses and residential neighborhoods, where robots cross controlled crosswalks at speeds below 6 km/h.
Nuro builds larger road-going pods; its R2 has no passenger compartment and operates commercial pilots with Domino's, Walmart, FedEx, and 7-Eleven, with regulatory clearance for driverless operation in California and Texas. Coco Robotics takes a hybrid approach with sidewalk bots remotely supervised by human operators, trading some autonomy for faster deployment. Serve Robotics, spun out of Postmates and now public on Nasdaq, partners with Uber Eats for last-mile food orders.
Security AMRs trade payload capacity for sensor breadth. Knightscope's K-series robots, including the K3 indoor unit, the cylindrical K5 outdoor unit, and the upcoming K7 off-road unit announced in November 2025, patrol parking lots, malls, and corporate campuses while streaming video, thermal, and audio data back to a security operations center.
Industrial inspection AMRs target oil and gas, electrical substations, and process plants. Boston Dynamics' Spot, ANYbotics' ANYmal, and Unitree's quadrupeds are legged rather than wheeled, but share the same software stack. They carry thermal cameras, gas sensors, and ultrasonic gauges to detect failing seals or rising temperatures before they cause unplanned downtime.
Most commercial AMRs share a fairly standard mechanical layout. A low chassis houses lithium-iron-phosphate or lithium-ion batteries that provide six to 12 hours of operating time. Two driven wheels in a differential-drive configuration are the most common arrangement for indoor robots up to a few hundred kilograms; larger pallet movers use Mecanum wheels or articulated steering. Outdoor robots adopt skid steering, four-wheel independent drive, or legged locomotion.
The top of the chassis carries the payload interface: a flat deck for tugging carts, a lifting plate for goods-to-person robots, a forklift mast for pallet handling, or a rotating turret for shelf-to-shelf transfers. Charging is handled by autonomous docking with a contact pad or by opportunity charging during idle periods, with wireless inductive charging becoming increasingly common.
Developing AMR software has historically been a slow, hand-crafted process: each new deployment required mapping the site, tuning costmap parameters, integrating with the customer's WMS, and manually testing edge cases. The arrival of large-scale simulation and foundation models is starting to change that pattern.
NVIDIA's Isaac platform offers a synthetic-data-and-simulation pipeline aimed at AMRs and humanoids alike. [[nvidia_isaac_sim|Isaac Sim]] runs photorealistic, physically-accurate simulations of warehouses and factories so that perception models can be trained and policies validated on tens of thousands of synthetic episodes before hardware is deployed. Isaac ROS provides CUDA-accelerated implementations of common Nav2 components. The Isaac GR00T family, including the GR00T N1.7 vision-language-action model released in 2025, brings generalist manipulation capabilities to mobile platforms. [10]
Figure AI's Helix is another visible example. Helix is a vision-language-action model that controls humanoid upper bodies at 200 hertz from a single neural network, accepts natural-language prompts, and generalizes across thousands of household and warehouse items the model never saw during training. In 2025 Figure showed the same model handling logistics package triage on a humanoid with an AMR-style mobile base, blurring the historical line between AMR and humanoid robotics. [11]
The practical effect is that AMRs are moving from pure transport ("drive there") toward integrated pick, place, and manipulation behaviors ("drive there, find the right tote, take out a specific bin, and bring it to me").
Forecasts vary, but most analysts agree the AMR market is in a phase of strong but uneven growth. Interact Analysis projects mobile-robot revenue to climb from just under 5 billion USD in 2024 to about 14 billion USD in 2030, an average annual growth rate of roughly 19 percent, although the firm trimmed its near-term forecast in 2025 to reflect tariff uncertainty and weaker capital spending in some regions. Within that envelope, AGV revenue is expected to drop from about 33 percent of total mobile-robot sales in 2024 to roughly 20 percent in 2030, with AMRs taking the rest. [2]
Grand View Research and MarketsAndMarkets, which use slightly different segmentations, both put the 2024 AMR-only market in the 4 to 6 billion USD range with compound annual growth rates between 15 and 25 percent through the early 2030s. The Robotics-as-a-Service (RaaS) model is one of the structural reasons for this growth: instead of buying robots outright, customers rent capacity by the month, which lowers the upfront barrier and lets vendors keep upgrading hardware. The IFR notes that the RaaS fleet grew 31 percent year over year in 2024. [1]
Regional dynamics matter. China alone accounted for about 58 percent of mobile-robot sales in 2024, a share that Interact Analysis expects to fall to 46 percent by 2030 as European and North American operators accelerate their own adoption. Labor-shortage data underscore the demand: Descartes' 2024 study reported that at least 76 percent of supply chain and logistics operations were experiencing notable workforce shortages, and 56 percent of warehouses specifically reported chronic understaffing. [12]
No single firm dominates more than a single-digit share of global AMR revenue. The table below highlights some of the most visible players and their focus areas.
| Vendor | Headquarters | Primary focus |
|---|---|---|
| [[locus_robotics | Locus Robotics]] | Wilmington, Massachusetts, USA |
| [[geek_plus | Geek+]] | Beijing, China |
| [[hai_robotics | Hai Robotics]] | Shenzhen, China |
| [[abb | ABB]] | Zurich, Switzerland |
| [[kuka | KUKA]] / Swisslog | Augsburg, Germany |
| Mobile Industrial Robots (MiR) | Odense, Denmark | Modular indoor AMRs (MiR100, MiR250, MiR600, MiR1350) for factories |
| [[amazon_robotics | Amazon Robotics]] | North Reading, Massachusetts, USA |
| OTTO Motors (Rockwell Automation) | Kitchener, Canada | Heavy-duty manufacturing AMRs |
| Vecna Robotics | Waltham, Massachusetts, USA | Pallet-handling and pivot-style AMRs |
| Fetch Robotics (Zebra) | San Jose, California, USA | Cloud-managed AMRs for fulfillment and manufacturing |
| [[bear_robotics | Bear Robotics]] | Redwood City, California, USA |
| [[pudu_technology | Pudu Robotics]] | Shenzhen, China |
| Keenon Robotics | Shanghai, China | Restaurant delivery, hospitality, and disinfection robots |
| Starship Technologies | San Francisco, California, USA / Tallinn, Estonia | Sidewalk delivery robots |
| Nuro | Mountain View, California, USA | Road-going autonomous delivery vehicles |
| Knightscope | Mountain View, California, USA | Outdoor and indoor autonomous security robots |
| Aethon (ST Engineering) | Pittsburgh, Pennsylvania, USA | Hospital service robots |
| Diligent Robotics | Austin, Texas, USA | Hospital service robots with manipulation arms |
| iRobot | Bedford, Massachusetts, USA | Consumer cleaning robots ([[roomba |
| Roborock | Beijing, China | Consumer vacuum and mop robots |
Consolidation has been a constant theme: ABB acquired ASTI Mobile Robotics in 2021 and Sevensense in 2024 [5], Rockwell Automation acquired OTTO Motors in 2023, Zebra Technologies bought Fetch in 2021, and Shopify sold 6 River Systems to Ocado in 2023.
VDA 5050 is the most consequential standardization effort. Version 2.0.0 launched in 2022 and version 2.1.0, published in January 2025, refined support for heterogeneous fleets. [3] Version 3.0.0 is in active development and is expected to extend the protocol toward outdoor robots. ISO 3691-4:2023 covers safety requirements for industrial trucks including AGVs and driverless vehicles, the R15.08 standard from the Association for Advancing Automation (A3) addresses industrial mobile robot safety in the U.S., and Europe's Machinery Regulation 2023/1230 takes full effect in 2027.
Most warehouse AMRs carry safety-rated 2D laser scanners certified to PL d or PL e under ISO 13849, with two protective fields: a slowdown zone that triggers a speed reduction and a stop zone that triggers an emergency stop. Software enforces speed limits in mixed-traffic zones and applies behavior trees that yield to humans by default. As shared-space deployment grows, vendors pay more attention to human factors: how the robot signals its intentions, how predictable its motion appears, and how easy it is for a worker to override or pause it. Studies cited by Locus suggest well-designed AMR rollouts increase worker retention by removing physically taxing tasks. [12]
AMRs work well inside the operating envelope they were designed for and less well outside it. Common failure modes include glare, dust, and reflective surfaces that confuse lidar and cameras; crowded narrow aisles where the local planner must choose between waiting and squeezing past a forklift; pallets left unexpectedly in routed aisles; rain and snow that degrade outdoor sensor returns; dirt or fingerprints on camera lenses that cause subtle SLAM failures; and network outages that knock fleet managers offline.
Research directions for the next several years include foundation-model perception that handles novel objects without retraining, federated SLAM that lets a fleet share a single map, and multi-agent reinforcement learning for fleet-level traffic control, an area where MIT and Symbotic recently demonstrated meaningful throughput improvements. [13]
Forward-looking trends include foundation models migrating from manipulation research into AMR products. NVIDIA Isaac, Figure Helix, Google DeepMind's Gemini Robotics, and similar systems are pushing toward generalist robots that interpret natural-language instructions and combine driving with picking. The traditional categories will blur: a walking humanoid, an [[autonomous_driving|autonomous driving]] platform on a sidewalk, and a wheeled warehouse AMR will increasingly run variants of the same vision-language-action model. Robotics-as-a-Service will keep claiming share, and AMRs will keep moving outdoors as cheaper solid-state lidar and stronger foundation-model perception make outdoor reliability achievable across a much wider set of conditions.