(1) Initial mapping (commissioning): the operator joysticks the robot around the facility while it builds a 2D/3D map with SLAM. (2) Localization: during operation the robot compares current sensor scans with the map and estimates its pose (Monte Carlo Localization, NDT). (3) Global planning: for a transport task from A to B, the planner builds the optimal route on the map. (4) Local planning: at each control cycle (10-30 Hz) the local planner generates a short trajectory in velocity space, avoiding dynamic obstacles. (5) Drive control: the trajectory is converted into wheel velocity commands.
AMR solves the rigidity problem of older AGV systems. AGVs require expensive guidance infrastructure (tracks, tape) and reprogramming with every facility layout change. AMR removes this barrier โ deployment requires no changes to the existing environment, and the robot adapts to dynamic changes (new racks, people, forklifts) without a programmer.
Environmental sensor set: 2D LiDAR (Hokuyo, SICK) or 3D LiDAR (Velodyne, Ouster, RoboSense), RGB-D cameras (Intel RealSense, ZED), IMU (Bosch BMI088, MicroStrain), wheel encoders, optionally ToF and radars. Sensor fusion gives the robot a coherent real-time picture of its surroundings.
Official
Algorithm for simultaneous mapping and robot localization. Popular implementations: Cartographer (Google), LIO-SAM, FAST-LIO (LiDAR + IMU), ORB-SLAM3 (visual), Hector SLAM (2D LiDAR-only). In production, closed commercial equivalents (MiR Sync, OTTO Fleet, Locus) are typically used.
Official
Two-stage planner: global (A*, Dijkstra, RRT*) builds the optimal route from A to B on the occupancy map; local (TEB โ Timed Elastic Bands, MPPI, DWB โ Dynamic Window Approach) generates a short-horizon trajectory in velocity space, avoiding dynamic obstacles. Replanning typically at 10-30 Hz.
Official
Central system coordinating multiple robots: task assignment, coordination at collision points (intersections, narrow passages, elevators), charging queuing, fleet health monitoring, WMS/MES integration. Interoperability standard: VDA 5050.
Official
A classical AMR based on hard safety speed limits (typically 0.3-0.5 m/s near humans) drastically loses throughput in places with heavy foot traffic โ stores, hospital corridors. The robot often stops or moves at a crawl.
A robot can lose localization when: the environment changes dramatically (facility renovation), lighting changes abruptly (visual SLAM), or the robot is physically moved while powered off. Without relocalization, the AMR is useless.
Without fleet manager coordination, two AMRs meeting in a narrow corridor can deadlock โ both wait for the other, because the local planner treats the other robot as an unpredictable human.
AMRs designed with payload margins often turn out insufficient in real operations โ pallets weigh more than assumed, loads are tall or unstable. The robot exceeds permissible load and breaches safety, or cannot stop in time.
Shakey from SRI International (Nils Nilsson and team) was the first robot combining perception, planning (STRIPS) and motion control into a single autonomous system. Slow and limited as it was, it established the sense-plan-act architectural pattern that remains the foundation of all later AMRs.
AGVs (Automated Guided Vehicles) requiring magnetic tape or QR codes spread in Toyota, GM, Volkswagen factories and warehouses. This is the predecessor to AMR โ the same goal (autonomous transport), but with rigid guidance infrastructure.
Amazon acquires Kiva Systems for $775M, closes it to external customers and deploys tens of thousands of mobile robots in its fulfillment centers. While Kiva is technically an AGV (floor QR codes), the acquisition demonstrates the business scale mobile robotics can achieve โ and opens the market for AMR competitors.
Danish company Mobile Industrial Robots (MiR) introduces MiR100 โ the first widely sold AMR without guidance infrastructure. Competitors Fetch Robotics (US, 2014), OTTO Motors (Canada, 2015), Locus Robotics (US, 2014) quickly follow. The term AMR spreads as a way to distinguish from AGV.
The German automotive industry (VDA โ Verband der Automobilindustrie) introduces the VDA 5050 standard โ a communication interface between a fleet manager and a heterogeneous AMR fleet from different vendors. Allows mixing MiR, OTTO, Geek+, Fetch in one facility.
The 2024-2026 generation of mobile humanoids (Proxie Gen 2, Digit, AGIBOT G2) combines classical AMR-style navigation (omni/swerve drive, SLAM, planning) with bimanual manipulation. The AMR classification boundary expands from pure transport to mobile manipulation.
The most compute-demanding part of the AMR stack is local planning (10-30 Hz) with a dynamic obstacle map updated from LiDAR and cameras. For a 3D LiDAR (Velodyne 64-line) plus 2-4 RGB-D cameras, a mid-range x86 CPU (Intel i7) or Jetson Orin/Thor is required. Initial SLAM mapping is a brief one-time cost; runtime localization is light.
AMR runs locally on the onboard computer. NVIDIA Jetson Orin/Thor, Intel NUC, Qualcomm RB5 are standard for new deployments. Requires low latency (local planning at 10-30 Hz) and deterministic control.
Classical AMRs with 2D LiDAR and 2 cameras run well on Intel i5/i7. For a full stack with 3D LiDAR + multiple RGB-D cameras + ML inference, an additional accelerator is required.
LiDAR sensors (2D Hokuyo/SICK, 3D Velodyne/Ouster) are the de facto standard for industrial AMRs โ lighting-independent, accurate, proven in safety certification.
