PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of AI-enabled Routing Protocols for UAV Networks: Trends, Challenges, and Future Outlook

Arnau Rovira-Sugranes, Abolfazl Razi, Fatemeh Afghah, Jacob Chakareski

arXiv:2104.01283v2cs.NI

TL;DR

Highly dynamic UAV networks require routing that can accommodate changing topology, connectivity, and operating conditions. This survey reviews AI-enabled routing and related mobility, networking, simulation, dataset, and swarming topics, concluding with challenges for future UAV networking. It identifies topology prediction and learning-by-experience as central approaches while emphasizing lightweight, secure, energy-efficient, and spectrum-aware designs.

  • Problem

    Highly dynamic UAV networks challenge conventional routing because topology changes rapidly and connectivity is difficult to maintain.

  • Method

    The paper surveys AI-enabled UAV routing, including topology prediction and learning-by-experience, together with mobility models, networking protocols, tools, datasets, and UAV swarming.

  • Results

    The review reports that topology-prediction methods are more accurate for nonlinear motion paths, while learning-by-experience separates networking and topology-control layers.

  • Takeaways & Limitations

    Future UAV networking needs advanced prediction, integrated topology control, lightweight AI routing, distributed security, anti-jamming methods, and energy-efficient low-complexity designs.

Abstract

from arXiv · show

Unmanned Aerial Vehicles (UAVs), as a recently emerging technology, enabled a new breed of unprecedented applications in different domains. This technology's ongoing trend is departing from large remotely-controlled drones to networks of small autonomous drones to collectively complete intricate tasks time and cost-effectively. An important challenge is developing efficient sensing, communication, and control algorithms that can accommodate the requirements of highly dynamic UAV networks with heterogeneous mobility levels. Recently, the use of Artificial Intelligence (AI) in learning-based networking has gained momentum to harness the learning power of cognizant nodes to make more intelligent networking decisions by integrating computational intelligence into UAV networks. An important example of this trend is developing learning-powered routing protocols, where machine learning methods are used to model and predict topology evolution, channel status, traffic mobility, and environmental factors for enhanced routing. This paper reviews AI-enabled routing protocols designed primarily for aerial networks, including topology-predictive and self-adaptive learning-based routing algorithms, with an emphasis on accommodating highly-dynamic network topology. To this end, we justify the importance and adaptation of AI into UAV network communications. We also address, with an AI emphasis, the closely related topics of mobility and networking models for UAV networks, simulation tools and public datasets, and relations to UAV swarming, which serve to choose the right algorithm for each scenario. We conclude by presenting future trends, and the remaining challenges in AI-based UAV networking, for different aspects of routing, connectivity, topology control, security and privacy, energy efficiency, and spectrum sharing.

I. INTRODUCTION

UAV networks support diverse applications but face severe constraints from limited payloads, dynamic topology, and heterogeneous mobility. This survey addresses these challenges by reviewing AI-enabled routing and related mobility, networking, and evaluation topics.

  • UAVs provide low-cost, on-demand monitoring and actuation across transportation, surveillance, search and rescue, agriculture, and other domains.
  • Limited payloads constrain UAV power consumption, communication range, and computation, complicating networking, control, information processing, autonomy, and coordination.
  • The survey reviews AI-enabled UAV routing alongside mobility models, networking protocols, simulation tools, public datasets, and UAV swarming.
  • UAV networks have more dynamic topologies and lower node density than VANETs, creating connectivity and routing challenges.
  • Conventional routing relies on prior or current topology information, whereas self-adaptive and topology-predictive methods learn and anticipate network changes.
  • Earlier surveys variously focus on VANETs, security, position-based routing, or conventional protocols, leaving comprehensive AI-based aerial-network routing insufficiently covered.

III. ARTIFICIAL INTELLIGENCE IN UAV NETWORKS

AI methods are applied across UAV networking and operations to interpret observations, learn patterns, and anticipate future states. Their benefits depend on scenario complexity and available training data, while applications span communication, sensing, control, and mission domains.

  • AI integrates computational intelligence into UAV networking and control so nodes can infer network and environmental conditions from limited observations.
  • AI supports pattern interpretation and future-state anticipation, enabling learning-based decisions in UAV networks.
  • AI-based decision-making can provide continuous feedback, improve experience-based decisions, and support energy optimization and obstacle avoidance.
  • AI benefits come at higher computational cost, and their value may be negligible in straightforward scenarios with limited training data.
  • AI applications in UAV wireless networking include positioning, channel estimation, imaging, path control, scheduling, resource allocation, security, and sensing.
  • UAV applications span military, industrial, and commercial settings, including reconnaissance, agriculture, infrastructure monitoring, delivery, and disaster-related services.

A. AI features

Modern UAVs can run lightweight AI workloads onboard, but storage and networking remain less mature and require further research. UAV networking spans communication, computation, and scheduling, with edge and fog processing gaining momentum for near-real-time applications.

  • A. AI features: Onboard GPUs, TPUs, and FPGAs enable modern UAVs to run lightweight deep-learning algorithms for AI applications.
  • A. AI features: AI hardware includes computing, storage, and networking components, but storage and networking still need research for diverse UAV requirements.
  • A. AI features: UAV networking is studied through communication, computation, and scheduling requirements and constraints.
  • A. AI features: Embedded GPUs and TPUs, Mobile Edge Computing, and fog computing support accelerated and near-real-time processing as alternatives to bulk server processing.
  • A. AI features: Vast operating areas often exceed one UAV’s LoS range, making networked platforms necessary and requiring wireless technologies with sufficient capacity and acceptable QoS.

A. Wi-Fi

Wi-Fi remains a low-cost and scalable UAV communication solution, especially for ground-station links, backhauling, and local inter-drone networks. UAV connectivity can also use LTE, 5G, and future 6G systems, while AI addresses reliability, mobility, security, and networking challenges.

  • A. Wi-Fi: Commercial UAVs commonly use Wi-Fi for ground-station communication because it is low-cost, scalable, and affordable.
  • A. Wi-Fi: Wi-Fi can support wireless backhauling and inter-drone communication when one node acts as a local WLAN access point.
  • A. Wi-Fi: Wi-Fi’s drawback is difficulty handling mobility and handoff between base stations, which limits drone operation range.
  • A. Wi-Fi: LTE provides airborne connectivity beyond LoS communication and improves throughput and connectivity through hard and soft handover.
  • A. Wi-Fi: 5G supports drone communications requiring bitrates beyond 2.4 Gbps and enables drones to participate in the Internet of Things.
  • A. Wi-Fi: AI-based networking can address interference, mobility, handover, cyber-physical attacks, and authentication, including prediction of transmission success and failure.

B. Swarm architectures

UAV swarms coordinate many drones through centralized, structure-free, or cellular distributed architectures. These designs trade central control and offloaded computation against autonomy, mission-area flexibility, and greater onboard capability requirements.

  • B. Swarm architectures: A UAV swarm consists of many mostly similar-capability drones collectively performing designated monitoring or operational tasks.
  • B. Swarm architectures: Centralized swarms use a ground control station to process member information and return commands, enabling computation offloading to support low-capability UAVs.
  • B. Swarm architectures: Structure-free swarms enable direct UAV-to-UAV communication and distributed decisions without access points or routers, but require more capable UAVs and dynamic-topology routing.
  • B. Swarm architectures: Cellular architectures combine distributed decision-making with LTE or 5G connectivity for long-range missions without a central controller.
  • B. Swarm architectures: Swarm AI combines real-time inputs with algorithms to optimize overall swarm performance.
  • B. Swarm architectures: For UAV networks, swarm AI uses AI and data-driven methods to control drones toward designated goals.

2) Levy process [133]:

UAV mobility models represent different trajectory patterns, from random and smooth motion to correlated, pheromone-guided, and hybrid behaviors. These models support simulation and application-specific evaluation, but connectivity and correlated motion remain important considerations.

  • 2) Levy process [133]:: Random-step mobility uses power-law-distributed step lengths and is more appropriate for rotary drones.
  • 2) Levy process [133]:: Controlled-randomness models simulate swarm movement with smooth speed and direction changes, avoiding sudden stops and sharp turns.
  • 2) Levy process [133]:: Circular-scan mobility models curved constant-altitude movements and outperform random waypoints in connectivity and scanning coverage over a 2D disk.
  • 2) Levy process [133]:: Predetermined-path mobility gives UAVs known trajectories, randomly assigned endpoints, and specified velocity and flight time.
  • 2) Levy process [133]:: PPRZM combines motion patterns through a Markov state diagram and outperforms RWP in geometric and network performance by switching between modes.
  • 2) Levy process [133]:: Many mobility models assume independent trajectories, whereas UAV swarms may require models representing highly correlated motion.
  • 2) Levy process [133]:: Pheromone-guided mobility achieves good scanning properties but does not account for connectivity between UAVs serving different areas.
  • 2) Levy process [133]:: A hybrid Markov-chain and pheromone model shows superiority in detecting and tracking targets compared with other pheromone-based methods.

9) UAV fleet mobility model [7]:

UAV mobility models differ in how they represent and predict movement, with implications for connectivity and routing in highly dynamic networks. Mission-specific models can improve coverage and connectivity, but heterogeneous-node networks remain difficult to model accurately.

  • 9) UAV fleet mobility model [7]:: The UAV fleet mobility model uses remaining energy, area coverage, and network connectivity to determine each UAV’s next movement.Neighbor information and weighted vectors determine movement direction and speed.
  • 9) UAV fleet mobility model [7]:: This mobility model outperforms random motion methods in terms of coverage and connectivity.
  • Mobility-model selection: A proper mobility model should match the drone types, mission requirements, and path-planning method because models vary in application suitability.
  • 1) Data-driven:: Data-driven mobility methods exploit frequent motion patterns from relatively large datasets, but cannot support real-time networking decisions by network nodes.
  • 2) Model-based:: Model-based methods predict future locations online from motion-path smoothness, using models such as HMMs, Bayesian methods, Kalman filtering, and mixture models.
  • 2) Model-based:: Models customized for particular object classes are not applicable to heterogeneous-node networks and cannot balance mobility randomness with predictability.
  • Mobility and routing impact: High mobility can disrupt information exchange, connectivity, and routing efficiency, potentially fragmenting UAV networks into isolated islands.

A. Conventional routing protocols

Conventional routing protocols are organized by topology knowledge, decision randomness, route adaptability, and structural design. Many protocols developed for low-speed or stable networks struggle with UAVs’ high mobility and abrupt topology changes.

  • A. Conventional routing protocols: Conventional routing protocols introduced for low-speed ad hoc networks generally do not adapt to UAVs’ high mobility and abrupt topology changes.
  • Static routing: Static routing computes paths from the initial topology and is suitable mainly for structured networks with static or slowly varying topology.
  • Routing classifications: Routing is also classified as centralized or distributed, deterministic or probabilistic, and static or dynamic according to topology knowledge, decision policy, and route adaptability.
  • Proactive routing: Proactive routing maintains paths for source-destination pairs and can quickly reestablish links, but topology exploration and path discovery impose substantial overhead.
  • Reactive routing: Reactive routing discovers routes on demand and reduces overhead in low-traffic regimes, but route recovery after failure can take a long time.
  • Hybrid routing: Hybrid routing initially uses proactive routing and activates reactive discovery after substantial topology changes or route breaks.

6) Hierarchical routing protocols:

The review covers hierarchical, probabilistic, position-based, and AI-enabled topology-predictive routing approaches for UAV networks. Learning and prediction improve route selection by incorporating topology, mobility, and link information, although model and scenario fit remain important.

  • 6) Hierarchical routing protocols:: Hierarchical routing organizes UAVs into clusters, with cluster heads communicating to select paths while nodes maintain neighbor information.
  • Probabilistic routing: Probabilistic routing selects among multiple source-destination routes probabilistically to address congestion and security.
  • B. AI-enabled routing protocols: AI-enabled routing uses machine learning to select paths from network topology, channel status, user behavior, and traffic mobility.
  • Topology predictive routing protocols: Topology-predictive protocols estimate node trajectories and incorporate predicted movement into path selection as an approximation of network topology.
  • Topology predictive routing protocols: Predictive greedy routing uses local observations and neighbor-location estimates to adapt to dynamic topology with low complexity and overhead.
  • Topology predictive routing protocols: GPMOR uses the Gauss-Markov mobility model and mobility relationships alongside Euclidean distance, improving latency and packet delay rate while reducing intermittent connectivity.
  • Topology predictive routing protocols: RARP combines 3-D location prediction with directional transmission and hybrid unicast-geocast routing to track topology changes.
  • Topology predictive routing protocols: RARP reduces path re-establishment and service disruption time while achieving higher successful packet delivery rates.

2) Self-adaptive learning-based routing protocols:

Self-adaptive learning-based routing protocols learn routing decisions online from network conditions and their consequences on metrics such as delay, throughput, energy efficiency, and fairness. Q-Routing variants progressively address exploration, congestion, convergence, reliability, and dynamic topology concerns.

  • 2) Self-adaptive learning-based routing protocols:: Reinforcement-learning routing continuously learns from the environment and decision consequences to optimize delay, throughput, energy efficiency, and fairness.
  • RL-based routing: In the RL process, an agent selects an action from candidate neighbors using expected reward Q(s, a), then receives an immediate environmental reward.
  • Q-Routing: Q-Routing stores each neighbor’s expected travel time in Q-values and selects the next node minimizing expected travel time, updating values from observed travel time.
  • Q-Routing variants: Predictive Q-Routing improves learning speed and adaptability over Q-Routing by predicting traffic trends and reusing learned policies, but requires large memory.
  • Q-Routing variants: Dual Reinforcement Q-Routing uses forward and backward exploration and learns the optimal policy more than twice faster than standard Q-Routing.
  • Q-Routing variants: Credence-based Q-Routing methods dynamically capture congestion and adapt to current network conditions faster than conventional Q-Routing.
  • Q-Routing variants: Poisson’s probability-based Q-Routing nearly doubles delivery probability relative to Q-Routing and reduces the overhead ratio by half.
  • Q-network routing: QNGPSR uses a Q-network to estimate path quality and achieves higher packet delivery ratio and lower end-to-end delay than GPSR.

IX. TOOLS AND PUBLIC DATASETS

This section reviews tools and public datasets for simulating real UAV networking environments, emphasizing their use in testing networking solutions under different conditions.

  • The section surveys tools and public datasets available for simulating real UAV networking environments.
  • These resources support testing networking solutions, including routing protocols, for UAV networks.
  • The review emphasizes testing under different network conditions.

A. Simulation tools

Simulation and experimentation tools enable lower-cost evaluation of UAV networking algorithms before real-world testing. The reviewed ecosystem spans flight simulators, network simulators, UAV-specific platforms, remote testbeds, and AI-enabled extensions.

  • Simulation tools: UAV flight simulators model flights in near-realistic virtual environments, enabling lower-cost evaluation before real-world testing.Simulator choice depends on the testing objective and available features.
  • Simulation tools: Popular simulation tools include X-Plane, FlightGear, Gazebo, JMavSim, Microsoft AirSim, and UE4Sim.The paper compares these tools in Table V.
  • Simulation tools: Microsoft AirSim and UE4Sim are considered among the best available simulators because AI components support near-realistic algorithm development.
  • Simulation tools: Network simulators such as NS-3 and OPNET support evaluation of networking systems and protocols, including routing protocols.NS-3 is open-source, free, and discrete-event based.
  • Simulation tools: UAV-specific platforms include ROS-NetSim for perception-action-communication loops and UB-ANC for repeatable comparative evaluation across protocol layers.ROS-NetSim supports tunable communication fidelity and complexity, while UB-ANC emphasizes modularity and extensibility.
  • Simulation tools: NS3-GYM and NS3-AI connect NS-3 with AI frameworks, while AERPAW and POWDER provide AI-enabled experimental infrastructure.

X. FUTURE TRENDS AND REMAINING CHALLENGES

Future UAV networking research must address dynamic connectivity, realistic mobility, routing robustness, energy constraints, and the costs of AI computation. Proposed directions include topology control, richer learning methods, and energy-aware networking.

  • Challenges: Limited payload, processing power, and highly dynamic, structure-free platforms remain central communication challenges for UAV networks.
  • AI networking: AI can predict node locations and potential link losses, but its use may increase computation complexity and CPU power consumption on power-constrained UAVs.
  • Connectivity: Connectivity loss can cause packet drops, link re-establishment, shortened link lifetimes, delays, and mission disruption.Commercial UAV communication ranges may span only a few miles, while some monitoring areas cover hundreds of miles.
  • Connectivity: Integrating online path planning with networking algorithms is proposed for topology control with minimal connectivity issues.The approach targets scenarios with defined coverage areas and flexible UAV motion paths.
  • Routing: Routing research should address link breakages, three-dimensional mobility, vision-based topology prediction, packet prioritization, nonlinear motion, and lightweight protocols.
  • Energy efficiency: Energy-aware routing remains a future direction requiring multi-objective and constrained optimization to balance connectivity, routing efficiency, and energy consumption.The paper also identifies wireless power transfer and energy-efficient ML/DL algorithms as related research directions.

E. Spectrum management

UAV networking requires high-rate, low-latency, ultra-reliable communications, while spectrum scarcity and optical-link constraints remain important challenges. The review connects these issues with AI-based networking and future spectrum-sharing policies.

  • Communication requirements: UAV networks require high-rate, low-latency, and ultra-reliable wireless communications for future applications.Current protocols include WiFi, LTE, LoRA, and 5G for A2A and A2G communications.
  • Spectrum access: Dedicated spectrum for drone operations has expanded, including the 5030-5091 MHz radio-frequency band.
  • Alternative communications: Fiber optic, laser, and LiFi models aim to transmit larger data volumes over longer distances and alleviate spectrum scarcity.
  • Alternative communications: Optical communications still require research on interference sensitivity and adaptive antenna steering.
  • Future directions: Future AI-based networking research includes spectrum-sharing and leasing policies alongside energy-efficient, low-complexity networking.
Loading 2104.01283v2…