Catalog / Robotics, Autonomy & Physical AI / ROBO-002

Autonomous mobile robots (AMR) in intralogistics

Free-navigating transport robots that move totes, pallets and carts through warehouses and factories, coordinated by a fleet manager and integrated with WMS/MES.

Parts needed to run it yourselfOptimizationRoboticsVision

Moving material between goods-in, storage, picking, production lines and shipping is repetitive, labour-intensive and error-prone. AMRs navigate on their own using onboard sensors, replace fixed conveyors and magnetic-tape AGVs, re-route around people and obstacles, and are dispatched by a fleet manager connected to the WMS/MES. Typical benefit: flexible, scalable internal transport without rebuilding the building.

Problem & value

  • Today: forklifts, tugger trains, manual carts and conveyors. Labour shortages, walking time of pickers, forklift-pedestrian accidents and inflexible fixed installations are the main pains. Classic AGVs follow wires, tape or reflectors and stop when blocked.
  • AI value: map-based and visual localisation, dynamic obstacle handling, learned perception (pallet, rack and person detection), and optimisation of task allocation and traffic across a fleet. Robots can be added incrementally (robots-as-a-service models exist), so SMBs can start with a handful.

Variants

  • Goods-to-person / picker-assist AMRs: robots follow or meet pickers in e-commerce and retail fulfilment (e.g. Locus, Geek+, Zebra/Fetch).
  • Tote and cart transport: line-side supply in manufacturing, hospital logistics (linen, meals, waste).
  • Pallet AMRs and autonomous forklifts: pallet moves, dock-to-stock (e.g. Third Wave Automation, Vecna, Toyota/BALYO, Linde, Jungheinrich, STILL).
  • Under-ride / tugger AMRs: towing carts on milk-run routes (e.g. MiR, OTTO/Rockwell).
  • Shelf-lifting AMR systems: goods-to-person storage (Geek+, Hai Robotics, Exotec – EU).

How it works

  1. transport order (WMS/MES/ERP)
  2. fleet manager: task allocation + route/traffic planning (VDA 5050 / vendor API)
  3. robot: sensor fusion (LiDAR + camera + IMU)
  4. SLAM / localisation on map
  5. perception (obstacles, people, pallets, docks)
  6. local planner
  7. motion control
  8. safety layer (certified safety LiDAR, speed/protective fields) – independent of AI
  9. handover (conveyor, lift, door, charger)
  10. status/KPI back to WMS
  • Localisation and mapping: 2D/3D LiDAR SLAM or visual SLAM (e.g. NVIDIA cuVSLAM); maps are built once by teleoperating a robot and then shared across the fleet.
  • Navigation: global planner plus local planner (ROS 2 Nav2: costmaps, behaviour trees, controllers). Learned components assist in perception and increasingly in local planning.
  • Fleet management: optimisation (assignment, vehicle routing, deadlock avoidance, charging scheduling) is largely classical solvers/heuristics; RL and learned ETA models are emerging.
  • Interoperability: VDA 5050 (VDA/VDMA) defines a MQTT/JSON interface between a master control system and AGVs/AMRs of different vendors; Open-RMF handles multi-vendor traffic in the ROS ecosystem [2][3].
  • Human-in-the-loop: a supervisor resolves stuck robots, approves map changes and layout edits; a safety officer signs off risk assessments. The safety function itself must not depend on a learned model.

Data preparation

Needed: a floor plan or CAD layout of the site, a list of stations and transport relations, material flow volumes and times, and order or transport requests from WMS/ERP. In an SMB these sit in layout drawings on file shares or paper, in the ERP and partly only in the heads of shift leads. Typical gaps: outdated floor plans, undocumented one-way rules and narrow spots, unclear station IDs, inconsistent storage-location master data and no clean transport order interface.

Steps: (1) update the layout with real measurements and mark restricted zones; (2) count transports per route and shift for two weeks, manually or via ERP/WMS export (PREP-005, PREP-008); (3) harmonise location and station codes in the master data (PREP-011); (4) let the vendor map the site with the robot's lidar; (5) define the interface for transport requests (API, PLC signal or simple button). Tools: vendor mapping software, spreadsheets, existing WMS exports. The AI-specific data effort is low, typically 10 to 15 percent of total effort; most work is layout, floor quality and interfaces. A readiness check helps: GUIDE-003.

Implementation paths

Which path for whom

  • SMB: Path A. Ready-made AMRs with fleet software are mature and available as purchase or robots-as-a-service; layout and routes are standard.
  • Enterprise: Path A. Intralogistics transport is a commodity function; buy proven fleets and integrate with WMS/MES instead of building navigation.
  • Developer: Path C. ROS 2 with Nav2 and open fleet managers (for example Open-RMF) run on HW1 edge computers for custom platforms.
  • Complete AMR systems with fleet software: MiR (Mobile Industrial Robots, Teradyne, DK), OMRON, KUKA (DE, Midea-owned), Locus Robotics, Geek+, Zebra (Fetch), OTTO Motors (Rockwell), BlueBotics/ANT (CH), Neobotix (DE), SEER, Magazino (DE, picking), Exotec (FR), Jungheinrich/STILL/Linde for autonomous trucks (EU).
  • Choose A when: standard payloads (tote, cart, pallet), a WMS/MES with a connector, no unusual environment. Insist on: VDA 5050 or open API, CE/ISO 3691-4 conformity, safety-rated sensors, EU service partner.

Path B — Build on cloud APIs

  • Fleet orchestration and analytics on cloud: VDA 5050 master control or Open-RMF running on Kubernetes (EU region), digital twin and optimisation services, MQTT brokers (EMQX, HiveMQ – DE).
  • Simulation and perception training: NVIDIA Isaac Sim / Omniverse (cloud or on-prem), AWS/Azure IoT and robotics services for telemetry.
  • LLM/VLM layer (optional): natural-language task entry and incident triage ("why is robot 7 stuck?") through general LLM APIs with tool calls to the fleet manager; never in the safety or motion loop.
  • Choose B when: you build your own robot or need a custom master control integrating several vendors.

Path C — Self-host (open source)

  • Stack: ROS 2 (Jazzy/Kilted), Nav2, slam_toolbox or RTAB-Map, Open-RMF, Isaac ROS (cuVSLAM, nvblox, Isaac Perceptor reference workflows), Gazebo for simulation, Eclipse Mosquitto/MQTT for VDA 5050.
  • Perception models: YOLO-class or RT-DETR detectors, open-vocabulary detectors and segmentation models for pallets/people; deployed with TensorRT/ONNX Runtime on-board.
  • Hardware: HW1 (Jetson Orin-class module on the robot; CPU server for fleet manager). Chassis can come from vendors such as Neobotix, Clearpath/OTTO, or Robotnik (ES).

Path D — Custom training

Worth it for unusual loads (e.g. specific racks, transparent wrap, reflective floors), site-specific person/forklift detection, or learned local planners. Needs a few thousand labelled frames from the robot's own sensors; use simulation plus real fleet data (active learning on interventions).

State of the art (2026)

  • Vision-first navigation: multi-camera plus LiDAR stacks (Isaac Perceptor-style) and 3D semantic occupancy let AMRs operate in less structured, human-shared spaces than 2D-LiDAR-only robots.
  • Foundation models in the robot stack: open-vocabulary detection, VLM-based scene understanding and natural-language tasking are appearing in products; learned policies for navigation are mostly still research or pilot stage compared to proven classical planners.
  • Interoperability maturing: VDA 5050 is widely supported by AGV/AMR and truck makers; it is being complemented by the VDMA/MassRobotics interoperability work and by standardised map exchange efforts [2][4].
  • Autonomous forklifts are moving from pilots to production, with pallet pose estimation the key AI task.
  • Limits: glass, mirrors, shiny wrap, ramps, dense crowds, dynamic layout changes, Wi-Fi/5G dropouts and localisation loss in featureless long corridors; deadlocks in narrow aisles; battery and charger capacity often limit throughput more than the AI does.

Shared resources & combinations

  • Same stack (robotics-control + 3d-vision + edge-runtime): ROBO-001 bin picking (mobile manipulation), ROBO-014 truck unloading and palletising, ROBO-015 inspection robots, ROBO-020 delivery robots.
  • Safety: ROBO-016 (AI-based person detection and speed-and-separation monitoring) and VIS-007 (PPE and workplace safety monitoring) share cameras and detectors.
  • Optimisation: OPT-006 (slotting and picking), OPT-001 (routing), OPT-005 (replenishment) feed and consume AMR transport orders.
  • Simulation: ROBO-010 digital twins for layout validation, throughput simulation and sim-to-real training; VIS-019 inventory counting with robots or drones as a by-product of transport.

Effort, cost & KPIs

  • Path A: weeks to 3 months for a first site (mapping, WMS connector, safety validation); cost drivers are robot count, software licences per robot, integration and infrastructure (Wi-Fi, charging, doors, lifts).
  • Path B/C: 6–18 months for a robust multi-vendor system with in-house robotics skills.
  • KPIs: missions completed per hour, on-time delivery of material, robot utilisation and availability, interventions per 1000 missions, mean time to recover from stops, battery/charger utilisation, near-miss and incident rate, pick-walk time saved, cost per transport.

People & change

Involve management for budget and liability, the logistics or production lead as process owner, IT and IT security (Wi-Fi, fleet server, remote access), legal for machine safety and liability, and the drivers and operators who share the floor with the robots (GUIDE-001). Works council: yes in most cases. AMRs navigate with cameras and lidar, log positions, routes and fleet events, and run in rooms where staff work, so the system is objectively suited to monitor behaviour or performance (§87(1) no. 6 BetrVG, BAG); §87(1) no. 7 on occupational safety and §90(1) no. 3 on AI-supported work processes also apply. Conclude a works agreement on data use, retention of camera and position data and a ban on individual performance evaluation (GUIDE-004). A DPO check is needed if onboard cameras record people; prefer blurring or non-recording sensors. AI literacy: the AI Act requires measures to support staff literacy, so offer short training on behaviour around robots and how to react to blocked routes (GUIDE-006). Typical concerns are collision safety, job loss and surveillance; address them with visible safety zones, a pilot on one route with volunteers and AI champions, clear manual override and transparent communication (GUIDE-008).

Risks, regulation & compliance

  • EU Machinery Regulation (EU) 2023/1230 replaces the Machinery Directive from 20 January 2027 and explicitly addresses machinery with self-evolving/learning behaviour and safety components using AI, which may need third-party conformity assessment [5]. Until then Directive 2006/42/EC applies. Driverless industrial trucks follow ISO 3691-4; mobile-robot safety also draws on ISO 13849, ISO 10218/ISO 13482 and IEC 61508 for safety-related controls.
  • EU AI Act (Regulation 2024/1689): AMRs as safety components of machinery covered by the Machinery Regulation can fall under Annex I high-risk via Art. 6(1) when third-party conformity assessment is required (2 August 2028, postponed from August 2027 by the Digital Omnibus on AI, Reg. (EU) 2026/1744); a pure transport robot with conventional certified safety functions is typically minimal risk. Classification depends on the architecture: keep AI out of the safety function, or treat it as high-risk. No Art. 5 prohibition applies; but do not use robot cameras for emotion recognition or productivity scoring of workers (Art. 5(1)(f)) [6]. Staff need AI literacy (Art. 4); NIS2/Cyber Resilience Act apply to fleet servers.
  • GDPR: cameras and LiDAR capture employees and visitors; need a legal basis, works council agreement (Germany: §87 BetrVG), data minimisation (blur/discard images, edge-only processing), short retention, DPIA where systematic monitoring of workers is plausible.

Getting started

  1. Map flows: routes, payloads, shifts, peak volumes, doors/lifts, floor quality; quantify transport share of labour.
  2. Pick one repeatable route (e.g. goods-in to buffer) as a pilot with a clear KPI baseline.
  3. Request two or three vendor demos on your floor with VDA 5050 or open API and safety documentation.
  4. Involve works council, safety officer and IT security (network, GDPR) before purchase.
  5. Run the pilot 8–12 weeks, log interventions, then scale robot by robot.

Sources

  1. International Federation of Robotics – World Robotics service robots (transport and logistics) – https://ifr.org/service-robots/
  2. VDA/VDMA – VDA 5050 interface for AGV/AMR communication – https://github.com/VDA5050/VDA5050
  3. Open-RMF project – https://www.open-rmf.org/
  4. Nav2 documentation (ROS 2 navigation) – https://docs.nav2.org/
  5. Regulation (EU) 2023/1230 on machinery – https://eur-lex.europa.eu/eli/reg/2023/1230/oj
  6. Regulation (EU) 2024/1689 (AI Act) – https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  7. NVIDIA Isaac ROS / Isaac Perceptor documentation – https://nvidia-isaac-ros.github.io/
  8. ISO 3691-4 Industrial trucks – driverless trucks and their systems – https://www.iso.org/standard/70660.html