PaperlayerView exact revision on arXiv

Source-linked AI summary

Localization as a key enabler of 6G wireless systems: A comprehensive survey and an outlook

Stylianos E. Trevlakis, Alexandros-Apostolos A. Boulogeorgos, Dimitrios Pliatsios, Konstantinos Ntontin, Panagiotis Sarigiannidis, Symeon Chatzinotas, Marco Di Renzo

arXiv:2302.02107v1eess.SP

TL;DR

6G is envisioned to combine ubiquitous communication with high-accuracy localization while sharing network resources. This survey reviews applications, technologies, channel models, localization techniques, and system design, and uses its analysis to highlight future directions for 6G localization.

  • Problem

    The central question is how 6G networks can provide high-accuracy localization alongside ubiquitous communication while sharing resources.

    Source: §X · p. 47

  • Method

    The survey reviews 6G localization applications and enabling technologies, mmWave/THz and visible-light system models, KPIs, localization techniques, and wireless-system design.

    Source: §I · p. 2, §X · p. 47

  • Results

    The survey synthesizes its analysis into future directions for localization in 6G wireless systems.

    Source: §X · p. 47

  • Takeaways & Limitations

    The paper frames localization as a 6G service that coexists with communication while sharing available resources.

    Source: §X · p. 47

Abstract

from arXiv · show

When fully implemented, sixth generation (6G) wireless systems will constitute intelligent wireless networks that enable not only ubiquitous communication but also high-accuracy localization services. They will be the driving force behind this transformation by introducing a new set of characteristics and service capabilities in which location will coexist with communication while sharing available resources. To that purpose, this survey investigates the envisioned applications and use cases of localization in future 6G wireless systems, while analyzing the impact of the major technology enablers. Afterwards, system models for millimeter wave, terahertz and visible light positioning that take into account both line-of-sight (LOS) and non-LOS channels are presented, while localization key performance indicators are revisited alongside mathematical definitions. Moreover, a detailed review of the state of the art conventional and learning-based localization techniques is conducted. Furthermore, the localization problem is formulated, the wireless system design is considered and the optimization of both is investigated. Finally, insights that arise from the presented analysis are summarized and used to highlight the most important future directions for localization in 6G wireless systems.

I. INTRODUCTION … 2) Earth monitoring:

The paper argues that localization is vital to high-frequency 6G networks and surveys its applications, enabling technologies, system models, KPIs, methods, and design directions. It also describes how localization supports sustainable-development use cases, including supply chains and earth monitoring.

I. INTRODUCTION

  • High-frequency wireless systems need transmitter and receiver position and orientation knowledge to meet performance goals over acceptable distances, making localization vital.The survey connects this need to mmWave, THz, and optical bands.Source: §I · p. 1
  • Existing surveys cover differing subsets of applications, technologies, models, KPIs, and methods, leaving a comprehensive review focused on future 6G networks missing.This survey aims to address that gap and identify research gaps and directions.Source: §I · p. 2
  • The survey covers 6G localization applications and enabling technologies, LOS/NLOS mmWave/THz and VLP models with reflectors and RISs, and mathematically expressed performance KPIs.These topics are considered as contributions to advancing wireless-system state of the art.Source: §I · p. 2
  • It reviews conventional and learning-based methods, then discusses localization problem formulation, system design, optimization, and future 6G research directions.Conceptual figures are included to explain the methods’ basic principles.Source: §I · p. 2

II. APPLICATIONS OF LOCALIZATION IN 6G WIRELESS SYSTEMS

  • 6G localization is envisioned to combine radio and visible-light technologies for cm- to mm-level accuracy indoors and outdoors, including 3D data and device orientation.The paper situates these capabilities across applications such as immersive telepresence, trust zones, twinning, and robotics.Source: §II · p. 5

A. Sustainable development

  • Localization is a common requirement across sustainable-development use cases, whose performance must satisfy criteria set by major KPIs.The use cases include autonomous supply chains, earth monitoring, e-health, and personalized context-aware networks.Source: A. Sustainable development · p. 5

1) Autonomous supply chain:

  • 6G micro-tags can track products’ whereabouts and condition, support worldwide asset monitoring, and connect products to the internet for more adaptable supply chains.The passage also identifies UAVs and robot delivery workers as elements of next-generation delivery systems.Source: 1) Autonomous supply chain · p. 5

2) Earth monitoring:

  • Earth monitoring calls for low-cost biodegradable sensors and autonomous placement to map environmental measurements from global deployments to where data are used.Green localization can help reduce future networks’ energy footprint amid data growth and climate pressures.Source: 2) Earth monitoring · pp. 5–6

3) e-Health: … 4) Nanoscale localization:

The passages present 6G localization as an enabler of healthcare, context-aware services, immersive interaction, distributed sensing, and in-body nanonetworks. They also identify application-specific precision, availability, energy, propagation, privacy, and standardization requirements.

3) e-Health; 3) Precision healthcare:

  • 6G localization supports remote healthcare and precision medicine, from drone sample delivery and patient monitoring to robotic telesurgery.Telemedicine and remote monitoring use,; drone sample collection requires localization precision of 0.1−0.3 m vertically and 0.1 −0.5 m horizontally, with up to 99.99% availability. Precision healthcare also includes in-body robots and equipment placement, while telesurgery requires precise localization of tissues and surgical tools.Source: 3) e-Health · pp. 6–7, 3) Precision healthcare · p. 9

4) Personalized context-aware networks:

  • Context-aware 6G can reduce energy use, improve on-demand throughput, enable multimodal localization, and support personalized data processing.Healthcare sensor fusion can reveal patterns and deviations,, while devices can switch channels and technologies according to location. Temporal context also raises privacy and security risks, and distributed context awareness depends on standardized parameters and interpretation rules.Source: 4) Personalized context-aware networks · p. 7

B. Immersive telepresence

  • Immersive telepresence uses localization to support near-real-time holographic and mixed-reality interaction, including gesture and motion recognition.Applications include remote work and entertainment; localization enables gesture and motion recognition and XR features.Source: B. Immersive telepresence · pp. 7–8

1) Gesture and motion recognition:

  • 6G localization and sensing can enable device-free gesture recognition, smart-environment monitoring, and location-appropriate XR experiences.Virtual-piano interaction requires 3 mm accuracy and recognition probability above 99%. XR design also needs a quality-of-physical-experience metric combining QoS, QoE, and users’ physical and perceptual factors,.Source: 1) Gesture and motion recognition · p. 8

C. Local trust zones

  • Localization is required across local trust-zone use cases, where adaptable architectures keep sensitive enterprise and service data within their own networks.The identified use cases include infrastructure-less network extensions, sensor infrastructure webs, precision healthcare, and security and privacy.Source: C. Local trust zones · p. 8

1) Infrastructure-less network extensions:

  • Infrastructure-less extensions use cooperating localization to coordinate moving nodes, including agricultural vehicle platoons that must maintain safe spacing and coverage-edge operation.The update rate should scale with vehicle speed, and the network must rapidly solve geometric consensus problems [79].Source: 1) Infrastructure-less network… · p. 8

2) Sensor infrastructure web:

  • A sensor infrastructure web combines sensing and communication, while 6G localization can guide poorly instrumented devices and support disaster monitoring.Examples include guiding vehicles with inadequate onboard sensors and combining in-situ and space sensors to monitor volcano hazards, flooding, and wildfire hotspots.Source: 2) Sensor infrastructure web · pp. 8–9

4) Nanoscale localization:

  • In-body nanoscale localization can regulate energy-harvesting nanonodes, but THz systems must address limited range, energy and complexity constraints, and scalability.A proposed architecture combines software-defined metamaterials with location-aware and wake-up radio paradigms; other work uses Markov-chain motion modeling and ML trajectory prediction to identify and localize vascular anomalies,.Source: 4) Nanoscale localization · p. 9

5) Security and privacy: … A. Indoor

The passages describe localization needs and methods across security and privacy, digital twins, agriculture, cities, manufacturing, robotics, mapping, and indoor services. They highlight threats and defenses as well as the role of accurate positioning in enabling these applications.

5) Security and privacy:

  • When jamming signals are not correctly identified and their consequences are not mitigated, GNSS jamming reduces positioning accuracy or causes complete signal loss [88]; THz remote sensing and see-through imaging may compromise privacy,.Radio-frequency fingerprinting may identify devices without providing a device identifier,.Source: 5) Security and privacy · pp. 9–10
  • Security-aware localization methods can be computationally demanding or rely on restrictive assumptions, while proposed alternatives detect malicious nodes through signal noise or clustering.A lightweight ToA method uses noise from external distance attacks to resist impersonation. Clustering and consistency evaluation, followed by a sequential probability ratio test, detected malicious nodes more accurately and efficiently than existing approaches in simulations, but labeled anchors cannot regain trusted status. A distributed, lightweight range-free approach filters malicious or malfunctioning anchors without external resources or a single point of failure.Source: 5) Security and privacy · p. 10

D. Massive twinning

  • 6G tags and localization can connect real-world objects to digital twins, including in GPS-denied locations, while radio sensing can add untagged objects and their positions.Tags can gather sensory data with timestamps, and 6G WANs may support localization where devices cannot use GPS,.Source: D. Massive twinning · p. 10

1) Agriculture:

  • Agricultural robots need localization because GNSS can be blocked or reflected; SLAM maps surroundings while estimating robot state, and visual odometry offers an image-based alternative.SLAM can incorporate position, velocity, calibration factors, and sensor offsets; its map can be updated and reused. Visual odometry predicts onboard-camera motion from image data,,,,,.Source: 1) Agriculture · pp. 10–11

2) Smart cities:

  • Smart-city 4D maps require localized sensors to track infrastructure and services, while roadside units can also monitor traffic.Radio reflections from buildings and nearby trees can help infer local topography; digital twinning can also support precise placement of building fixtures,,.Source: 2) Smart cities · p. 11

3) Manufacturing:

  • In Industry 5.0, 6G positioning can support precise production and automate industrial operations by helping robots, terminals, and networks allocate resources and direct movement.The passage identifies connection density of 107/km2, latency of hundreds of microseconds, and data rates on the order of Tbps as industrial requirements; THz networks are presented as a route to high data rates, with positioning offering further potential,,,.Source: 3) Manufacturing · p. 11

E. Robot coordination and interaction; 1) Interactions:

  • Robots use localization both to interact with people and to map their surroundings; cobots need accurate, high-update-rate 3D positioning for smooth collaboration.Probabilistic methods such as Bayes filters and extended Kalman filters are established options. In a long-term 1,000-km cobot challenge, over 90 percent of deployments did not require human intervention to reset localization.Source: E. Robot coordination and interaction · p. 11, 1) Interactions · p. 11

2) Mapping; III. LOCALIZATION USE CASES; A. Indoor

  • SLAM supports localization and mapping when GPS is unavailable, using visual, lidar, or sensor-fusion methods across applications including indoor robots, unmanned vehicles, and AR/VR.Indoor localization identifies people or devices within buildings and supports applications ranging from healthcare and supply chains to XR and robotics [115], [116], [117], [118], [119],,,,,.Source: 2) Mapping · p. 12, §III · p. 12, A. Indoor · p. 12

B. Outdoor … D. RIS

These sections connect localization use cases across terrestrial, satellite, underwater, and nanoscale settings with 6G technology enablers, emphasizing the opportunities and constraints of high-frequency systems and reconfigurable surfaces.

B. Outdoor

  • Outdoor mmWave and future sub-THz networks may provide localization where GNSS coverage fails, using propagation measurements or location fingerprinting.Relevant measurements include channel impulse response, propagation delay, AoA, and RSS; the passage notes adjustable 5G bandwidths up to 400 MHz and future allocations of several gigahertz above 100 GHz.Source: B. Outdoor · p. 12

C. Non-terrestrial; D. Underwater

  • Satellite localization, especially GNSS, supports non-terrestrial applications such as supply chains and earth monitoring, while underwater localization serves navigation, surveillance, disaster alerts, and oceanographic collection.Underwater sensors can also monitor offshore structures, and the listed applications include mapping, security, privacy, and infrastructure-less network expansion.Source: C. Non-terrestrial · pp. 12–13, D. Underwater · pp. 13–14

E. Nanoscale

  • Nanoscale localization must balance high precision with tight energy budgets; multi-hop range or hop-count methods lose accuracy as accumulated error grows with hop count.Open challenges include non-free-space propagation modeling, hardware imperfections, and frequency- and angle-dependent nanonode responses.Source: E. Nanoscale · p. 14

A. mmWave/THz localization

  • mmWave and THz localization must address narrow-beam alignment costs alongside near-field effects, beam-split, synchronization, hardware, energy-efficiency, and coverage challenges.Although wider bandwidths and higher frequencies can support robust delay estimation, the section identifies complexity, overhead, area coverage, and hardware architecture as difficulties.Source: A. mmWave/THz localization · p. 14

IV. TECHNOLOGY ENABLERS FOR 6G WIRELESS SYSTEMS; B. Infrastructure

  • The technology-enabler review covers hardware, beamforming, RIS, AI, channel charting, radar, sensors, and visible-light methods; THz infrastructure options trade power, integration, and range differently.Silicon devices face power limits, photonic systems offer higher carrier frequencies but larger form factors, and plasmonic devices are compact and reconfigurable but short-range.Source: §IV · p. 14, B. Infrastructure · p. 15

C. Beamforming

  • Beamforming uses channel-derived angular and delay information to steer beams and infer user or scatterer locations; NLOS user positioning requires both AoD and AoA estimates.For passive targets, angular-delay beamspace profiles can be compared with references or prior samples for detection and tracking, though target identification grows more complex as target counts increase.Source: C. Beamforming · p. 15

D. RIS

  • RISs can add curvature-of-arrival information and passive-anchor diversity, enabling localization and mapping with one base station while reducing infrastructure and energy costs.Their lower energy consumption can make them preferable to active nodes in hard-to-power outdoor locations, including when supplied by solar panels.Source: D. RIS · pp. 15–16

E. AI … A. mmWave/THz

The sections describe AI-enabled localization methods and applications alongside mmWave/THz and optical system models. They emphasize learning from complex or limited data, radar and sensor use, and channel models that account for LOS and NLOS paths.

E. AI

  • AI can learn complex radio-signal properties and fuse noisy sensors for localization where conventional methods struggle, but data and labeling demands call for semi-supervised and adaptive approaches,.The passages also identify hierarchical models for restricted, structured or noisy data and cite cooperative localization with RL and crowdsourcing with FL as possible approaches.Source: E. AI · pp. 16–17

F. Channel charting

  • Channel charting uses unsupervised dimensionality reduction on CSI to create stable pseudo-locations without labeled fingerprints, supporting network functions such as beam tracking and resource management.Pseudo-locations can support self-configuration and privacy-preserving contact tracking, but cannot replace physical positions; extending charting to LOS high-frequency settings and lifelong learning remains unresolved.Source: F. Channel charting · p. 17

G. Radars

  • Radar sensory data can support AI-based beam prediction in realistic vehicular scenarios, addressing beam-selection sensitivity to transmitter, receiver and environmental changes.Narrow-beam training overhead makes highly mobile mmWave/THz applications challenging; reliable large-scale FMCW-radar SLAM in harsh weather remains an open question.Source: G. Radars · p. 17

H. Sensors

  • Sensor-network localization supplies the measurement locations needed for applications such as environmental monitoring, while GPS/GNSS limitations and indoor multipath make accurate positioning challenging [178], [179].GPS works only where satellite reception is adequate, and indoor or hostile settings can frustrate it; predetermined static anchors or GPS-equipped seed nodes can guide other nodes.Source: H. Sensors · pp. 17–18

I. VLP

  • Visible light positioning can reduce multipath effects because optical energy is concentrated on LOS links, while optical RISs or mirrors may restore link quality under blockage.Dense ceiling LEDs are expected to support accurate indoor positioning where RF systems are restricted, including sensitive environments.Source: §I · p. 18

V. SYSTEM MODELS

  • The system-model discussion distinguishes mmWave/THz localization with LOS and RIS-aided NLOS scenarios from VLC localization with LOS and optical RIS-aided NLOS models.These are the two model families outlined for the paper’s localization systems.Source: §V · p. 18

A. mmWave/THz

  • The mmWave/THz model considers a deterministic far-field MIMO channel with planar waves, in contrast to near-field models with spherical waves, and separates H into LOS and NLOS submatrices.The passage identifies the channel model as necessary for configuring and analyzing system performance.Source: A. mmWave/THz · pp. 18–19

1) LOS: · 2) NLOS:

The section presents LOS and NLOS channel models for mmWave/THz systems, describing path attenuation and the effects of random reflectors and RIS-aided reflections.

1) LOS:

  • LOS channel coefficients apply when the BS and UE share an unobstructed path, with steering vectors and gains, delay, Doppler shift, frequency, and path gain represented in the channel matrix [194].The passage identifies the channel-matrix form but does not provide its equation here.Source: 1) LOS · p. 19
  • LOS path gain depends on propagation distance and attenuation; in THz, attenuation reflects molecular absorption by atmospheric water vapor and gases, while in mmWave it denotes atmospheric attenuation.The attenuation coefficient is a function of distance and transmission frequency.Source: 1) LOS · p. 19
  • Assuming negligible Doppler effect and identical attenuation across subcarriers makes the LOS channel frequency-flat; the model also defines clock offset, subcarrier frequencies, and BS–UE distance.The passage defines B as the clock synchronization offset, c as the speed of light, and c_k as the ratio between central and k-th subcarrier frequency.Source: 1) LOS · p. 19

2) NLOS:

  • Without RIS, blocked LOS signals may reach the UE through random or deliberately placed reflectors; THz NLOS has increased losses and sparsity, while rough surfaces introduce scattering and diffraction.THz NLOS channels are commonly modeled with ray-tracing and stochastic techniques; mmWave NLOS uses stochastic reflection-loss and geometrical modeling.Source: 2) NLOS · p. 19
  • With M reflectors, the NLOS channel is modeled using a channel matrix for each reflector, with BS-to-reflector and reflector-to-UE distances, reflector steering, and RIS–UE attenuation specified.The passage defines the reflector steering vector and attenuation coefficient for the RIS–UE channel.Source: 2) NLOS · p. 19
  • In RIS-aided systems, RIS-reflected signal coefficients augment the model, and RIS reflection modifies signal phase and amplitude.The model includes separate attenuation coefficients for the BS–RIS and RIS–UE channels.Source: 2) NLOS · p. 19
  • The RIS-aided model also specifies BS–RIS and RIS–UE distances and the RIS steering vector.The section identifies the distances for both channel segments and the RIS steering vector.Source: 2) NLOS · pp. 19–20
  • The complete NLOS channel matrix is obtained by summing all individual reflection components, including both random and RIS reflections.This combines the reflection components described for NLOS channels.Source: 2) NLOS · p. 20

3) Localization geometry: · B. VLP · 1) LOS:

The section describes how localization geometry maps between global and local coordinates, then outlines visible-light positioning (VLP) and its line-of-sight (LOS) channel model. It covers optical RIS use when the direct path is obstructed and the angle- and device-dependent parameters of the LOS gain.

3) Localization geometry:

  • The geometry uses a local coordinate system translated to the global frame by an Euler-angle vector and an xyz rotation sequence, with array center and element positions defined across the frames.The system includes the base station, user device, and RIS.Source: 3) Localization geometry · p. 20
  • Because AoA/AoD angles are measured locally, azimuth and elevation are mapped to and from direction vectors to recover global angles for LOS and NLOS channels.Azimuth is measured from the y axis to the direction-vector projection on the xy plane; elevation is measured from the xy plane. The inverse mapping uses four-quadrant arctan2.Source: 3) Localization geometry · p. 20

B. VLP

  • VLP sends signals from an LED to a photodiode through a wireless optical channel, with the LED often modeled as a Lambertian source.The analysis allows obstruction between transmitter and receiver and uses an optical RIS to extend coverage.Source: B. VLP · p. 20, Figure 6 · p. 20

1) LOS:

  • Mirror-array RIS can redirect the optical path and enable beamforming through multi-mirror reflection links.Under the Lambertian LOS model, reflected light is treated as negligible after reflection from an object such as a wall.Source: 1) LOS · p. 21
  • The LOS signal model expresses received signal y using transmitted signal x, AWGN n, and channel gain h [189].The LOS channel gain depends on irradiance angle φ, incidence angle ψ, optical concentrator gain G(ψ), and optical filter gain To.Source: 1) LOS · p. 21
  • The LOS gain also uses Lambertian index m, LED half-power angle Φ1/2, receiver field of view Ψ, photodiode refractive index α, and area A.These parameters specify the optical source and receiver contributions to the gain.Source: 1) LOS · p. 21

2) NLOS: · 3) Localization geometry:

The NLOS model describes RIS-assisted first-order optical reflections and the geometry used to represent RIS elements. VLP localization geometry includes RSS-, ToA/TDoA-, and AoA-based approaches, with trilateration and phase-difference processing detailed for the first two.

2) NLOS:

  • The RIS comprises identical rectangular mirrors, with two mirrors at right angles required to determine roll and yaw angles.The mirrors are mounted with their optical centres in a plane perpendicular to the y axis; roll and yaw rotate clockwise about positive x and counterclockwise about positive z, respectively.Source: 2) NLOS · p. 21
  • RIS-enabled NLOS paths model the RIS as a receiver and then a point source, while restricting analysis to first-order reflections.The LED is assumed to illuminate K meta-atoms of area D at their geometric centres.Source: 2) NLOS · p. 21
  • Each RIS element’s channel gain depends on its LED and photodetector distances and reflection coefficient, contributing to the complete VLP channel gain.The model also specifies transmitter, receiver, and element positions and uses element roll and yaw to evaluate orientation terms.Source: 2) NLOS · p. 21

3) Localization geometry:

  • VLP geometrical methods include RSS-, ToA/TDoA-, and AoA-based approaches; RSS trilateration estimates distances from signal strengths and solves overlapping-circle measurements by least squares.Measurement errors can make the measured distances differ from the true distances, causing the circles to overlap.Source: 3) Localization geometry · p. 21
  • ToA/TDoA methods also use trilateration and require at least three transmitters with distinct frequency identifiers to obtain phase differences.The method relies on light travelling at a fixed speed over a given distance, producing phase delay at the receiver.Source: 3) Localization geometry · p. 21
  • The exposition defines transmitter signal power and received photodetector power using modulation and continuous-signal powers, initial phase, responsivity, proportionality constant, and channel impulse response.The transmitter modulation frequencies are indexed for i = 1, 2, 3.Source: 3) Localization geometry · pp. 21–22
  • The receiver filters each modulated frequency, down-converts the signals, and applies a Hilbert transform to obtain two phase differences from three transmitters.An additional phase difference is needed for accurate positioning and can be obtained by altering the frequencies of two transmitters.Source: 3) Localization geometry · p. 22

VI. LOCALIZATION KPIS … D. Coverage

The section defines localization KPIs spanning accuracy, precision, latency, coverage, and scalability. It explains how these measures characterize estimation quality, timing, and the ability to provide reliable service across users and regions.

VI. LOCALIZATION KPIS

  • Accuracy and precision are leading localization goals, while complexity, coverage, and mobility also affect system success; latency captures a separate time-related dimension.Source: §VI · p. 22, C. Latency · p. 23

A. Accuracy

  • Accuracy is assessed through location error, often using MSE or a threshold-bounded CDF; the Cramér-Rao bound gives a worst-case estimate assuming no bias and known noise and signal models.Bias and variance contribute to location-estimate error, while realistic skew from NLoS propagation can make performance fall short of the Cramér-Rao bound; tighter bounds include Ziv-Zakai, Weiss-Weinstein, and Bayesian Cramér-Rao bounds.Source: A. Accuracy · p. 22

B. Precision

  • Precision characterizes variation across localization attempts: GDoP reflects error variation and node geometry, while error CDF and outage indicate the likelihood of meeting a specified precision.For equally accurate methods, the one with smaller outage and/or larger CDF values is preferred.Source: B. Precision · pp. 22–23

C. Latency

  • Latency is the delay from a location request to its response, determined by method processing speed and reference-signal length; the use-case-dependent update rate cannot exceed latency^-1.Source: C. Latency · p. 23, Figure 7 · p. 23

D. Coverage

  • Coverage is the largest geographic region offering reliable localization at specified KPI levels; its probability KPI measures whether a user’s SINR exceeds a threshold.Coverage categories are indoor, local, and global.Source: D. Coverage · p. 23
  • Scalability complements coverage by measuring capacity to accommodate more end devices as expanded localization coverage risks channel congestion and increased computation and measurement demands.The system’s maximum supported node count is the point after which residual capacity becomes negative; available capacity, efficiency, load, contention, and transit factor inform the measure.Source: D. Coverage · p. 23

E. Complexity … 4) Multidimensional Scaling:

The section relates localization complexity and stability to implementation, accuracy, and update timing, then reviews conventional algorithms and their methods. These include geometric, filtering, sparse-recovery, and distance-matrix approaches.

E. Complexity

  • Localization complexity spans hardware, signal detection, and computation, while software complexity is often treated as estimator computational cost and is linked to accuracy, precision, update rate, and latency.Source: E. Complexity · p. 23

F. Stability

  • Deafness measures power leakage from estimation error relative to half-power beamwidth; at 100%, the estimate fails and localization must restart, so stability reflects accuracy variation over time.Source: F. Stability · p. 23

VII. LOCALIZATION ALGORITHMS

  • The review distinguishes conventional from learning-based algorithms, while noting that fingerprinting is grouped with conventional methods despite its probabilistic character and overlap with machine learning.Source: §VII · p. 25

A. Conventional

  • Conventional methods reviewed include triangulation, Kalman filters, compressive sensing, multidimensional scaling, direct localization, swarm intelligence, and fingerprinting.Source: A. Conventional · p. 25

1) Triangulation:

  • Triangulation estimates location from AoA or AoD line intersections and requires at least two access points for 2D coordinates.Source: 1) Triangulation · p. 25

2) Kalman filters:

  • Kalman-filter localization predicts and updates states, fusing anchor measurements; reported examples include 0.39 m average error below 90% packet loss and 24.8 cm at 140 GHz.EKF and UKF variants also combine ToA and DoA estimates for 3D localization, with densely deployed anchors supporting sub-meter accuracy.Source: 2) Kalman filters · p. 26

3) Compressive Sensing:

  • Compressive sensing reconstructs signal estimates from fewer, under-sampled measurements, with localization applications including TDOA, RSS, AoD, and 2D visible-light positioning.A CS-based TDOA method is described as requiring limited signal samples and lowering computational complexity.Source: 3) Compressive Sensing · p. 26

4) Multidimensional Scaling:

  • MDS derives relative coordinates from inter-node distances; eigenvalue decomposition yields positions up to rotation, translation, and reflection, which known anchor locations resolve.It is robust to individual range errors when the distance matrix has redundancy and provides a closed-form estimate regardless of network size.Source: 4) Multidimensional Scaling · pp. 26–27

5) Direct Localization: · 6) Swarm Intelligence:

Direct localization estimates target positions from accumulated signals without first estimating intermediate parameters, while swarm-intelligence localization searches through interacting exploration and exploitation procedures. The sections survey direct-localization methods and swarm algorithms applied across wireless sensor, indoor, and UAV localization settings.

5) Direct Localization:

  • Direct localization estimates a source position from signals accumulated across node pairs, avoiding intermediate parameter estimates and using signal-level fusion for lower computational overhead.Signals or functions of them are sent to a fusion centre; C-RAN can provide a centralized processing backbone for cellular networks.Source: 5) Direct Localization · pp. 27–28
  • Direct-localization methods combine AoA, ToA, TDOA, Doppler, or near- and far-field measurements, with approaches addressing multiple nodes, unknown source counts, and independent clocks –.The multiple-stationary-node algorithm in [246] is reported to attain the corresponding Cramer-Rao constraint; avoids prior knowledge of source-node count.Source: 5) Direct Localization · p. 28
  • Other direct-localization work uses beamspace design, spatial-spectrum suppression and clustering, message passing in distributed MIMO radar, and noise-subspace orthogonality for multipath OFDM.The beamspace framework includes a performance bound and an algorithm described as having low computational complexity and communication overhead.Source: 5) Direct Localization · p. 28

6) Swarm Intelligence:

  • Swarm-intelligence methods model coordinated group behavior and balance exploration of uncovered search space against exploitation that improves previously found solutions.The surveyed approaches draw inspiration from animal social behaviors, including those of bees, birds, and ants.Source: 6) Swarm Intelligence · p. 28
  • Swarm localization studies apply dragonfly, fireworks, artificial bee colony, particle swarm, firefly, monarch butterfly, and moth-search algorithms to node-location estimation and wireless sensor networks –.Examples include a multi-stage artificial bee colony method compared with particle swarm optimization and a two-stage firefly method that first uses four edge anchors, then the closest anchors.Source: 6) Swarm Intelligence · p. 28
  • Further swarm methods combine fuzzy logic with particle swarm optimization, compare firefly and bee-colony distance estimates, and study other WSN, UAV, and SLAM localization algorithms [265]–[268],.For UAV-node localization, a boundary-box technique reduces the initial search space to increase particle-swarm convergence speed; a multi-objective variant jointly optimizes localized-node count, energy consumption, and computation time.Source: 6) Swarm Intelligence · p. 28

7) Fingerprinting: … 1) K-nearest neighbors:

The section reviews fingerprint-based localization and SLAM, then surveys supervised learning with emphasis on k-nearest neighbors (kNN), including its basic operation and localization adaptations.

7) Fingerprinting:

  • Fingerprinting estimates location by matching a measured signal fingerprint against a prebuilt database, using a training phase to map signals and a matching phase to locate the node.RSS can represent reference fingerprints; matching may use Euclidean distance followed by nearest-neighbor selection.Source: 7) Fingerprinting · p. 29
  • Fingerprinting studies optimize access-point placement, update databases, incorporate context or collaboration, and develop path-loss, crowdsourced, data-rate, and CSI-based approaches –.These works also address robustness to outdated entries and interference, use signal or architectural features, define accuracy indicators, and apply entropy, maximum-likelihood, or AoA information. The passages describe the individual approaches and their aims, but do not provide a shared quantitative comparison.Source: 7) Fingerprinting · p. 29

8) Simultaneous localization and mapping:

  • SLAM jointly estimates a moving agent’s trajectory and a map of its surroundings, using data frames whose measurements are relative to the agent’s uncertain location.Maps may take forms such as occupancy grids or point clouds, and building a consistent representation requires collecting data from varied sources.Source: 8) Simultaneous localization and mapping · p. 29
  • SLAM work spans image-based 3D mapping, radar and underwater methods, uncertainty-aware estimation, and LIDAR or active-SLAM approaches –.The active approach in derives a collision-free trajectory through predictive control, convex relaxation, optimization, and rounding.Source: 8) Simultaneous localization and mapping · pp. 29–30

B. Supervised learning

  • The supervised-learning survey covers traditional methods such as kNN, SVMs, decision trees, and Gaussian processes, as well as neural-network, autoencoder, CNN, and RNN methods.Source: B. Supervised learning · p. 30

1) K-nearest neighbors:

  • kNN estimates a target’s coordinates by averaging the positions of nearby fixed nodes in a fingerprint database, with neighbors selected using criteria such as Euclidean distance.It is a non-parametric, instance-based method whose new labels are derived by comparison with training examples.Source: 1) K-nearest neighbors · p. 30
  • Weighted kNN, adaptive neighbor selection, and alternative similarity metrics address localization errors from fixed k values, node movement, or changing RSS –.Other applications combine hybrid features or use kNN as a coarse Wi-Fi location estimate to narrow image-based search,.Source: 1) K-nearest neighbors · pp. 30–31
  • An AoA-based kNN method compresses channel observations in the angle-delay domain, clusters fingerprints in two tiers, and localizes using weighted ToA and AoA features,.Its combined angle-delay similarity depends on scatterer overlap, while the sparse representation addresses the storage and sorting burden of numerous channel observations.Source: 1) K-nearest neighbors · p. 31

2) Support vector machine: … 5) Neural networks:

The section surveys support vector machines, decision trees, Gaussian processes, and neural networks for localization, describing their learning mechanisms and applications across wireless systems.

2) Support vector machine:

  • SVMs use kernels such as Gaussian, RBF, sigmoid, polynomial, and linear to identify classification hyperplanes with larger margins in feature spaces.The passage describes kernels as capturing vector similarity and enabling observed features to be represented in higher-dimensional spaces.Source: 2) Support vector machine · p. 31
  • SVM localization spans RSS- and ToA-based methods, CSI-based localization, and tracking that combines coarse RSS position estimates with accelerometer data [300],, –.Other approaches use trusted-node ranking, real-time node selection, or kernels adapted to spatial relationships and nearby and distant nodes –.Source: 2) Support vector machine · p. 31

3) Decision trees:

  • Decision-tree positioning trains a weighted-tree ensemble from fixed-node location and direction measurements, then uses online measurements to estimate mobile-node locations.Dataset size affects both localization accuracy and the time to construct and assess the ensemble.Source: 3) Decision trees · p. 32
  • Decision-tree ensembles include gradient boosting that learns a nonlinear RSS mapping before weighted kNN, and random forests used with multipath, WiFi RSS, and multi-antenna fingerprints,,,,.The cited approaches also use ray-tracing with empirical data, entropy-based classifier and time selection, or equally weighted anticipated-node positions,,.Source: 3) Decision trees · p. 32

4) Gaussian processes:

  • GPs model spatial RSS distributions with few labeled fixed nodes, reducing radio-map deployment and construction time; other methods use GP regression, Bayesian weighting, or dynamic radio-map calibration –, –.Reported approaches use Matern or RBF kernels, with location estimates obtained through maximum likelihood, weighted node positions, or weighted kNN.Source: 4) Gaussian processes · pp. 32–33
  • GP approximation techniques for massive-MIMO localization were reported to achieve RMSE near the Cramer-Rao constraint, while a UWB method used channel features for ranging,.The UWB method applies kernel PCA to RMS delay spread, RSS, and ToA features before GP regression.Source: 4) Gaussian processes · p. 33

5) Neural networks:

  • Neural-network localization uses CSI, RSS, RSRP, and WiFi RSS/ToA inputs for fingerprinting, ranging-error correction, collaborative estimation, and hierarchical environment-aware positioning,, [345]–.For MIMO-OFDM fingerprinting, one method encodes CSI magnitude in network weights and uses Bayes-rule location probabilities; extreme learning machines offer a lower-complexity alternative to backpropagation,.Source: 5) Neural networks · pp. 33–34

6) Autoencoders: · 7) Convolution neural networks: · 8) Recurrent neural networks:

The survey reviews autoencoders, CNNs, and RNNs as learning-based localization methods, describing their architectures, strengths, limitations, and representative applications. Across these methods, learned representations support localization from fingerprint, RSS, CSI, and sequential measurements.

6) Autoencoders:

  • Autoencoders reconstruct inputs to learn encoded features that can feed localization modules, classifiers, or coordinate, floor, and building prediction networks.Their encoder maps input to a hidden state and the decoder maps that state back to an output; hierarchical tuning can adjust classifier parameters before encoder parameters.Source: 6) Autoencoders · p. 34
  • Fingerprint methods use autoencoder reconstruction or output weights to assess similarity and estimate location, while device-free applications include localization, activity recognition, and device-heterogeneity transformation, [360],, –.BiLoc uses average amplitude across two antennas and AoA estimates as inputs, with output weights serving as fingerprints; other methods use reconstructed signals or transform test-device characteristics to match database-device features.Source: 6) Autoencoders · pp. 34–35

7) Convolution neural networks:

  • CNN parameter sharing and sparse connections reduce parameters and computational and data requirements, enabling efficient, translation-invariant models and mitigating overfitting.Low-dimensional filters generate outputs while reusing weights across pixels, allowing deeper structures to be trained on fewer samples.Source: 7) Convolution neural networks · p. 35
  • CNN localization methods reshape RSS measurements into images for floor, building, corridor, and coordinate prediction, including hierarchical models and Bluetooth-based pedestrian and car localization,,,,,.One approach uses a continuous wavelet transform to create a time-frequency image, then combines CNN reference-point predictions with kNN coordinate inference.Source: 7) Convolution neural networks · p. 35
  • CNNs also process CSI and channel-derived images for position likelihoods, fixed-node selection, and spatio-temporal position estimates; the survey reports CSI-based solutions outperform others in localization accuracy –, [159], –,.Applications include device-free localization, NLOS interference and range-error estimation, and massive-MIMO images representing power across delay and spatial domains.Source: 7) Convolution neural networks · pp. 35–36

8) Recurrent neural networks:

  • RNN hidden states retain information from earlier calculations, but plain RNNs can suffer vanishing or exploding gradients and lose influence over long histories, motivating gated LSTM and GRU architectures.A described localization pipeline extracts measurement features for each fixed node over a time window, uses LSTMs to learn higher-level features, and feeds them to fully connected regression layers to monitor movement.Source: 8) Recurrent neural networks · p. 36
  • Localization studies use sequential RSS or CSI with RNNs for coordinate, floor, building, and UAV-base-station position estimates, and combine CNN spatial features with LSTM trajectories or temporal-correlation models, –, –,,, [387].Reported pipelines include feature selection or dimensionality reduction before recurrent prediction, and a CNN–RNN approach followed by a mixture density network to model conditional location distributions.Source: 8) Recurrent neural networks · pp. 36–37

C. Unsupervised learning … F. Transfer learning

The section reviews unsupervised, federated, reinforcement, and transfer learning approaches for localization, including methods for learning from unlabeled data, protecting local data, adapting across domains, and improving localization.

C. Unsupervised learning / 1) Semi-supervised:

  • Unsupervised methods learn from unlabeled data, including deep belief networks that extract RSS features for location estimation and semi-supervised models that estimate access-point distances and localization precision,.A graph-based method models latent power levels and location using an RSS-likelihood Gaussian mixture model, with expectation maximization initialized by a pathloss model and fixed-node positions,.Source: C. Unsupervised learning · p. 37, 1) Semi-supervised · p. 37

2) Clustering:

  • Clustering supports unsupervised radio-map construction through RSS-based floor plans and sensor signatures, while global-local optimization with a Markov model fits unlabeled RSS traces to environmental structure,,.The methods use k-means to assign representative fingerprints or identify structural landmarks such as elevators, columns, and stairs.Source: 2) Clustering · p. 37

3) Dimensionality reduction:

  • Unsupervised classifier fusion combines location-weight estimates, while multidimensional scaling projects known wireless-node distances into a spatial map for cooperative localization,,,.Cooperative methods use signals such as ToA, AoA, and RSS to estimate device distances in cognitive radio, IoT, and RFID contexts.Source: 3) Dimensionality reduction · p. 37

D. Federated learning

  • Federated learning trains shared localization models on edge nodes while keeping training data local, with encrypted updates and centralized model aggregation supporting privacy-conscious services,,.Examples include WiFi fingerprint localization and precise IoT localization; a corridor experiment reported high precision and safety.Source: D. Federated learning · pp. 37–38

E. Reinforcement learning / 1) Markov decision process:

  • For localization, RL uses states and rewards to guide agent actions; approaches include landmark-based RSS rewards, WiFi Markov decision processes, and scheduling cooperative signal exchanges,,.The WiFi method is model-free and uses RSS, action history, and center coordinates as states; signal scheduling treats measurement choices as tasks and links as agents.Source: E. Reinforcement learning · p. 39, 1) Markov decision process · p. 39

2) Deep RL:

  • A Q-network in deep reinforcement learning improved pedestrian-position prediction accuracy by 37% over unsupervised multilateration; semi-supervised deep RL increased pedestrian-distance precision by 23% over supervised learning.A separate deep RL method for pedestrian-path prediction achieved performance similar to LSTM localization.Source: 2) Deep RL · p. 39

F. Transfer learning

  • Transfer learning addresses retraining across time, space, and hardware by extracting a signal’s semantic manifold and using it to constrain target-domain labeling.The localization system includes offline training and online localization; temporal transfer can use manifold regularization when distributions vary over time but remain consistent in low dimensions.Source: F. Transfer learning · p. 39

1) NNs: … 3) Algorithm development:

The section reviews learning-based localization methods and frames localization as a system-design and optimization problem shaped by application requirements, network architecture, cooperation, and signal choices.

1) NNs:

  • Transfer learning adapts neural localization models across antenna configurations and propagation domains, reducing target-domain data needs and training time.CNNs can be retrained in lower layers for altered antenna configurations; across two floors, 30% of the second-floor data achieved the same performance, while a pedestrian-localization approach increased training data by 45% and cut training time in half.Source: 1) NNs · p. 39

2) Conventional TL:

  • Conventional transfer learning combines source-domain fingerprints with target measurements, using learned domain-invariant kernels, transferred metrics, or fuzzy C-means clustering.Metric learning captures signal-label and feature-statistic dependence, while metric transfer selects a target-domain metric by minimizing source-target data differences; clustering can reduce external-factor effects.Source: 2) Conventional TL · p. 40

3) Manifold learning:

  • Manifold-learning localization aligns representations or graphs to construct radio maps and support supervised and semi-supervised RSS or CSI localization.Methods include Laplacian Eigenmaps that preserve local geometry, graphs using labeled RSS and unlabeled timestamped traces, modified objectives and Gaussian weights, pseudo-labeling with manifold-regularized transductive SVM, and separate WiFi/Bluetooth Laplacians. Siamese networks also compare inputs for CSI localization.Source: 3) Manifold learning · p. 40

VIII. OPTIMIZATION FRAMEWORKS FOR LOCALIZATION; A. Problem formulation

  • Localization optimization balances context-specific objectives and constraints, with accuracy, coverage, update rate, and communication KPIs potentially competing.A generic formulation uses objective f, constraint function g, and variable set X; accuracy may be either an objective or a constraint, alongside spacing, device count, and beamforming angles. System design and optimization are presented as necessary to assess achievable performance.Source: A. Problem formulation · pp. 40–41, §VIII · p. 40

B. System design; 1) Network architectures:

  • Use cases guide localization system design, including architecture choices that trade performance, privacy, scalability, coverage, and deployment flexibility.Centralized and distributed topologies suit macro settings, while clustered designs support parallel processing and are suited to low-power, short-range nanonetworks. Cell-free, RIS-enabled, 3D, and heterogeneous networks are options for large-scale scenarios; cell-free geometric base-station diversity can aid localization, while RIS can expand coverage and reshape channels.Source: B. System design · p. 41, 1) Network architectures · p. 41

2) Cooperative approaches:

  • Cooperative localization can improve precision and coverage, including when NLOS limits direct access to fixed nodes, but increases energy, computation, and time costs.Probabilistic approaches such as estimation theory and factor graphs quantify uncertainty; deterministic and probabilistic methods address the parameter-estimation problem, while some conventional methods assume Gaussian measurement errors,.Source: 2) Cooperative approaches · pp. 41–42

3) Algorithm development:

  • THz modulation choices depend on channel conditions, while bandwidth, packet length, and carrier frequency jointly affect localization and communication performance.Multicarrier schemes are favored in frequency-dependent molecular-absorption and multipath scenarios despite complexity and power-efficiency costs; DFT-spread OFDM can mitigate PAPR. More bandwidth improves delay-domain route separation, but data size and sampling rate must fit hardware limits; longer messages aid energy acquisition, whereas shorter messages reduce delay in mobile settings.Source: 3) Algorithm development · p. 42

C. Optimization … D. Heterogeneous localization system design

The paper frames localization optimization around joint design of network layout, beams, RIS settings, and resources, then outlines future research directions for high-frequency, optical, RIS-assisted, and heterogeneous systems.

C. Optimization

  • With unknown device locations, optimize node layout from surroundings and CRB; with known locations, tune beams, RIS coefficients, and bandwidth, time, and energy allocation against accuracy, coverage, overhead, and latency.Higher SNR does not necessarily yield a lower Cramér-Rao bound; in THz systems, effective bandwidth shrinks as connection distance increases.Source: C. Optimization · pp. 42–43
  • THz initial access is localization without device input, so codebook search must balance prior-aware or hierarchical discovery speed against exhaustive-search practicality and beam-strength-dependent latency.Blockage and deafness from narrow beams make initial access difficult.Source: C. Optimization · p. 42

1) Fundamental research:

  • Future mmWave/THz localization needs stochastic channel models accounting for reflectors, scatterers, and hardware impairments, plus joint UE localization and BS/RIS calibration when anchors are numerous.Existing deterministic models and known-anchor assumptions can miss practical signal variability and calibration errors.Source: A. mmWave and THz localization · p. 43, 1) Fundamental research · p. 43

2) Intelligent protocol design:

  • Intelligent mmWave/THz protocols need shared-data evaluation and privacy safeguards, while adapting UAV paths, beams, RIS coefficients, and evolving SLAM maps to dynamic connectivity and obstacles.Previous access-site distributions and local maps can support scene-aware localization and reflector use.Source: 2) Intelligent protocol design · pp. 43–44

B. Optical wireless localization

  • VLP commonly uses three or more sinusoidally driven LEDs and LOS RSS models for static positioning; overlapping-signal averaging and Kalman tracking can improve estimates, warranting evaluation across operating conditions.Future comparisons should assess positional error, computational efficiency, precision, and latency across LED frequency, noise, and user speed.Source: B. Optical wireless localization · p. 44

1) Fundamental research:

  • RIS-localization research should relax idealized assumptions by evaluating fading, discrete phases, meta-atom failures, hybrid or digital RISs, multi-RIS configurations, interference, hardware impairments, and broader KPIs.Existing studies mainly analyze analog RISs under idealized channels and hardware, often emphasizing CRLB or Fisher information rather than energy efficiency, latency, or coverage.Source: C. RIS-assisted localization · p. 44, 1) Fundamental research · p. 44

2) Intelligent protocol design:

  • RIS-assisted localization needs adaptive protocols that co-design source and RIS beam codewords, bandwidth, timeslots, and placement to maximize energy efficiency without sacrificing accuracy or stability.The protocol should account for operating band and analog, hybrid, or digital RIS type; optimization solutions can be combined with reinforcement learning.Source: 2) Intelligent protocol design · pp. 44–45

D. Heterogeneous localization system design

  • Heterogeneous localization should combine microwave, mmWave, THz, and optical bands because their range and accuracy vary with propagation conditions and design, while coordinating coexisting technologies and standards.The paper distinguishes spectrum heterogeneity from technology heterogeneity.Source: D. Heterogeneous localization system… · p. 45

1) Spectrum heterogeneity: … X. CONCLUSION

The paper surveys 6G localization applications, enabling technologies, system models, performance indicators, and conventional and learning-based methods, then identifies design challenges and future research directions for integrated localization and communication.

1) Spectrum heterogeneity:

  • Multi-band localization can use lower frequencies for wider-range coarse positioning and higher frequencies for more accurate positioning within their shorter coverage, requiring hierarchical design.An experimentally verified framework is needed to quantify performance and provide design guidelines.Source: 1) Spectrum heterogeneity · p. 45

2) Technology heterogeneity:; 6) Design of integrated localization and communication protocols for heterogeneous networks:

E. Integrated localization and communication

  • Integrated localization and communication must balance accuracy, data rate, power efficiency, reliability, and range through suitable KPIs, protocols, and system design.The cited work identifies trade-offs –; application-specific KPIs may include QoS, position integrity, and availability.Source: E. Integrated localization and… · p. 45, 1) KPIs definition · p. 45

2) Fundamental research:; 3) Advanced signal processing:

  • Future integrated systems need realistic channel models and performance frameworks, as well as co-design of signal processing and hardware.Models should capture stochastic NLOS scattering, obstacles, wideband effects, atmospheric conditions, hardware imperfections, and RISs; analysis should assess conventional and new KPIs across localizer and RIS configurations.Source: 2) Fundamental research · pp. 45–46, 3) Advanced signal processing · p. 46

4) Waveform design:

  • Localization and communication waveform design requires balancing localization accuracy against communication robustness, with adaptable power spectral density, bandwidth, and power to suit changing conditions.Spectrum power concentrated at its edges improves localization accuracy, while concentration at the center reduces inter-symbol interference; optimization is needed to find an operating balance.Source: 4) Waveform design · p. 46

5) Pro-active radio environment mapping and radio resource management:

  • Future 6G radio resource management needs accurate 3D localization for NLOS and multipath settings, alongside less complex methods that preserve UE communication performance.Existing mitigation methods can be too complex for UE self-localization, and often model only first-order reflections despite higher-order reflections in dense multipath; the relation between location precision and transmission rate also needs study.Source: 5) Pro-active radio environment… · p. 46

F. Low earth orbit (LEO) localization systems

  • Terrestrial and LEO localization can complement one another where infrastructure is absent or blocked, while LEO-PNT research targets worldwide precision through network integration, intelligent beamforming, edge computing, and satellite GNSS receivers.The paper contrasts reusing existing systems, with receiver-side optimization, and designing new systems, which requires three-segment optimization simultaneously or in stages.Source: F. Low earth orbit (LEO) localization… · p. 47

X. CONCLUSION

  • The survey presents 6G localization applications, enabling technologies, mmWave/THz and visible-light models for LOS and NLOS channels, localization KPIs, methods, system design, and future directions.Its scope also includes NLOS signals generated by traditional reflectors and RISs, and reviews both traditional and learning-based localization techniques.Source: §X · p. 47

Related papers

Loading 2302.02107v1…