Narrative Goldmine

localisation

8 min read

Properties

Type
  • Class
Resource
  • urn:ngm:class:localisation
Domain
  • robotics
Status
  • stable
Maturity
  • established
Quality
  • 0.74
Is a
Requires
Enables
Generated
  • by: process:vault-migrate/1.0 · at: 2026-09-22T12:41:39.024503178Z
Bridges to
Contrasts with
Depends on
Label
  • Localisation
Links
    Page resource
    • urn:visionflow:page:2709daa33c3a13581cd149ecf89229599e2c4c585e53dd067b6ec2493162fb1f
    Related to
    Supports
    Uses

    Localisation is the computational process by which a mobile agent—robot, autonomous vehicle, or mixed-reality device—estimates its six-degree-of-freedom pose (position and orientation) within a reference coordinate frame, using sensor observations fused through probabilistic inference algorithms. It encompasses both map-based approaches (matching live sensor data against a prior map) and map-free approaches (dead-reckoning and visual odometry), as well as the joint Simultaneous Localisation and Mapping (SLAM) problem in which the map and pose are estimated concurrently. Accurate localisation is a prerequisite for autonomous navigation, path planning, and real-time spatial anchoring in both physical and virtual environments. The field draws on Bayesian filtering, factor graph optimisation, deep metric learning, and multi-sensor fusion across domains ranging from warehouse automation and self-driving vehicles to extended-reality (XR) headsets.

    Overview

    • Localisation answers a deceptively simple question: “Where am I?” For a mobile agent operating in an unstructured or partially known environment, answering it reliably is non-trivial. The agent must fuse noisy, incomplete, and sometimes conflicting data from multiple sensors—LiDAR, Camera, Inertial Measurement Unit, Wheel Encoders, Ultra-Wideband beacons—and reconcile that data against either a pre-built Map or an incrementally built one.
    • The problem is framed probabilistically: the agent maintains a belief distribution over its possible poses and updates it using Bayesian Inference each time a new sensor observation arrives. The choice of representation for this belief—Gaussian (parametric), particle-based (non-parametric), or factor graph (sparse)—determines the algorithm family and its computational cost.
    • Localisation is “established” as a mature research area but continues to advance rapidly in deep-learning-assisted descriptor generation, uncertainty-aware fusion, and large-scale lifelong mapping with dynamic objects.

    Key Components and Mechanisms

    • Sensor Modalities
    • Probabilistic Filters
    • Scan Matching and Map Matching
      • Iterative Closest Point (ICP): aligns successive Point Cloud scans to estimate relative motion.
      • Normal Distributions Transform (NDT): represents the environment as a set of normal distributions over voxels; faster than ICP for large scans.
      • Map Matching: aligns sensor observations against a pre-built 2-D or 3-D reference Map.
    • Deep Learning Descriptors
      • NetVLAD, PointNetVLAD, SuperPoint/SuperGlue: neural network-derived place descriptors enabling robust Loop Closure Detection and cross-condition re-localisation.
      • LocNet, DNN-based end-to-end pose regression (PoseNet): direct mapping from image to 6-DoF pose.
    • Factor Graph Optimisation
      • Pose graph formulation: poses are nodes, sensor constraints are edges; optimised with g2o, GTSAM, iSAM2.
      • Enables globally consistent trajectory estimation over long runs; central to graph-based SLAM.
    • Coordinate Frames and Transforms
      • REP-105 (ROS) defines standard frames: map → odom → base_link → sensor_link.
      • Coordinate Frame management via tf2 is essential in multi-sensor, multi-robot systems.

    Applications and Use Cases

    • Autonomous Vehicles: real-time localisation against HD maps (HERE, TomTom) using LiDAR + GNSS + HD-map-matching at centimetre accuracy. Self-Driving Car pipelines integrate GNSS RTK, LiDAR scan matching, and deep learned descriptors for lane-level precision.
    • Warehouse and Logistics Robots: AMR (Autonomous Mobile Robot) fleets use SLAM and Occupancy Grid localisation to navigate dynamic warehouse environments without fixed infrastructure.
    • Surgical Robotics: sub-millimetre localisation of instruments relative to pre-operative CT/MRI-derived maps enables Robotic Surgery with spatial precision.
    • Unmanned Aerial Vehicles (UAVs/Drones): visual-inertial odometry (VIO) fusing Camera and Inertial Measurement Unit for GPS-denied environments.
    • Extended Reality (XR/AR/VR): Spatial Anchoring and world tracking in headsets (HoloLens, Apple Vision Pro, Meta Quest) use inside-out SLAM to render virtual content co-registered with the physical world.
    • Indoor Navigation: hospital, airport, and large-venue wayfinding using Ultra-Wideband or Wi-Fi fingerprinting combined with dead-reckoning on smartphones.
    • Agricultural Robotics: field robots use GNSS RTK + visual odometry in crop rows where satellite signal is partially obstructed.
    • Search and Rescue: SLAM-equipped robots map and localise in GPS-denied collapsed structures or underground environments.
    • Digital Twin Synchronisation: real-time robot pose fed to a Digital Twin model for live simulation mirroring and predictive maintenance.

    Types of Localisation

    • Position Tracking (local localisation): the robot knows its initial pose and tracks it over time; the less challenging variant.
    • Global Localisation (the “kidnapped robot” problem): the initial pose is unknown and must be determined from scratch; significantly harder, typically requiring multi-modal belief representations such as the Particle Filter.

    Localisation in Different Environments

    • Indoor: no GNSS; structured environments; LiDAR, vision, and beacon-based sensing dominate.
    • Outdoor: GNSS available but environments are unstructured and dynamic; fusion of GNSS, LiDAR, and vision.
    • Underwater: neither GNSS nor LiDAR is usable; acoustic sensors and Doppler velocity logs (DVL) provide ranging and odometry.
    • Aerial: GNSS, IMU, and vision fused under high dynamics for drones and UAVs.
    • Space: planetary rovers rely on visual odometry and landmark matching in the absence of satellite positioning.

    Challenges

    • Sensor noise and Uncertainty; dynamic environments with moving objects; perceptual aliasing (similar-looking places); computational complexity; ambiguous situations; limited sensing range; GNSS-denied environments.

    Performance Metrics

    • Localisation accuracy (pose error); precision (repeatability); convergence time to the correct pose; computational efficiency; robustness to failures; coverage (area over which localisation succeeds).

    Standards and Context

    • ROS / ROS 2: de facto middleware for robotics localisation; REP-105 defines canonical coordinate frame conventions; the nav_stack (Navigation Stack) bundles AMCL, costmap2d, move_base for 2-D mobile robot localisation.
    • IEEE 1873-2015: IEEE Standard for Robot Map Data Representation for Navigation; specifies exchange formats for occupancy grids and metric maps.
    • OpenCV: open-source library providing camera calibration, feature detection, and optical flow primitives essential for visual localisation pipelines.
    • Open3D / PCL (Point Cloud Library): standard libraries for processing Point Cloud data used in LiDAR-based localisation.
    • OSRM / OpenStreetMap: vector map data used as a localisation reference for road-level and pedestrian navigation.
    • ISO 3691-4: safety requirements for driverless industrial trucks; indirectly mandates reliable localisation for autonomous operation.
    • IEEEE 802.15.4z (UWB): radio standard underpinning high-precision Ultra-Wideband ranging used in Indoor Positioning System.

    Current Landscape (2026)

    • Visual foundation models have become the dominant lever for robust localisation: Voedisch and Scaramuzza’s “LiDAR Registration with Visual Foundation Models” (RSS 2025) used DINOv2 surround-view descriptors to lift 6-DoF registration recall by +24.8 and +17.3 points on the NCLT and Oxford Radar RobotCar datasets, and feed-forward 3D models (DUSt3R, MASt3R, VGGT) now bootstrap dense SLAM front-ends directly from RGB.
    • End-to-end learned dense SLAM matured through 2025-2026: FoundationSLAM (AAAI 2026 Oral) couples optical flow with foundation depth priors to run monocular dense tracking and mapping in real time at 18 FPS, while DINO-VO and ZeroVO push zero-shot, cross-camera visual odometry.
    • NVIDIA consolidated a GPU-accelerated localisation stack in Isaac ROS: cuVSLAM (visual-inertial SLAM, ROS 2 Jazzy support landed October 2025), cuVGL for global re-localisation in prebuilt maps, FoundationStereo depth, and nvblox reconstruction; the January 2026 GR00T N1.6 humanoid release makes this vision-centric SLAM its core localisation layer, and PyCuVSLAM (2025) opened the library to Python users.
    • 3D Gaussian Splatting overtook NeRF as the mapping substrate for photorealistic SLAM, with SplaTAM, MonoGS, DROID-Splat (ICCV 2025 workshop) and Gaussian-LIC2 (LiDAR-inertial-camera) as reference systems; a dedicated SLAM&Render benchmark (2025) and an October 2025 survey on collaborative multi-robot 3DGS SLAM signal the field’s consolidation.
    • Localisation is shifting from purely geometric to “AI-native” semantic SLAM: 2026 Chinese patents (Hefei Keda) integrate Vision-Language-Action models into the SLAM back-end for keyframe selection, dynamic-object filtering and loop closure, building hierarchical situational maps that support natural-language navigation goals.
    • Radio positioning standards advanced sharply: 3GPP Release 18 added carrier-phase measurement and PRS/SRS bandwidth aggregation (up to 400 MHz) for centimetric accuracy, and Release 19’s Sidelink Positioning Protocol was published as ETSI TS 138 355 V19.0.0 in October 2025, defining SL-TDOA/TOA/AoA/RTT device-to-device ranging plus seven positioning service levels, including millisecond-latency tiers explicitly for AMR cooperation and collision avoidance, with AI/ML for NLOS mitigation.
    • Open frontiers as of 2026 remain robustness in degenerate, GNSS-denied and low-texture settings (planetary, subterranean, factory), long-term cross-session re-localisation against maps that are months or years old, and dynamic-scene consistency — motivating multimodal loop-closure pipelines such as MPRF (under review for ICRA 2026) that fuse DINOv2/SALAD visual retrieval with LiDAR geometric verification.

    References

    Provenance

    Graph View

    Backlinks