PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning-Aided Operations and Communications of Unmanned Aerial Vehicles: A Contemporary Survey

Harrison Kurunathan, Hailong Huang, Kai Li, Wei Ni, Ekram Hossain

arXiv:2211.04324v1cs.ROcs.CVcs.LG

TL;DR

UAV applications increasingly use ML for intelligent communications, perception, planning, and control, but integrated and trustworthy automation remains limited. This survey classifies ML techniques across four UAV-operation modules and synthesizes their applications and research gaps. It finds that different techniques dominate different modules, cross-module integration is increasing, and end-to-end ML control remains underdeveloped.

  • Problem

    The field lacks a comprehensive view of ML techniques across UAV perception, interpretation, planning, and aerodynamic control, including their research gaps and trustworthiness.

  • Method

    The survey reviews and classifies ML techniques by their applications to feature extraction, environment modeling, planning and scheduling, and control and operation.

  • Results

    Different ML techniques dominate the four UAV-operation modules; cross-module integration is increasing, but little effort has addressed end-to-end ML control from feature extraction through operation.

  • Takeaways & Limitations

    ML can support UAV communications, feature extraction, prediction, planning, control, and operations, while broader automation requires attention to reliability and trust.

  • Takeaways & Limitations

    Linear regression assumes linear relationships, is sensitive to noise and overfitting, and cannot be used when features outnumber observations.

Abstract

from arXiv · show

The ongoing amalgamation of UAV and ML techniques is creating a significant synergy and empowering UAVs with unprecedented intelligence and autonomy. This survey aims to provide a timely and comprehensive overview of ML techniques used in UAV operations and communications and identify the potential growth areas and research gaps. We emphasise the four key components of UAV operations and communications to which ML can significantly contribute, namely, perception and feature extraction, feature interpretation and regeneration, trajectory and mission planning, and aerodynamic control and operation. We classify the latest popular ML tools based on their applications to the four components and conduct gap analyses. This survey also takes a step forward by pointing out significant challenges in the upcoming realm of ML-aided automated UAV operations and communications. It is revealed that different ML techniques dominate the applications to the four key modules of UAV operations and communications. While there is an increasing trend of cross-module designs, little effort has been devoted to an end-to-end ML framework, from perception and feature extraction to aerodynamic control and operation. It is also unveiled that the reliability and trust of ML in UAV operations and applications require significant attention before full automation of UAVs and potential cooperation between UAVs and humans come to fruition.

ACRONYMS

The paper defines acronyms spanning artificial intelligence, machine learning, neural-network architectures, reinforcement learning, communications, sensing, and UAV operations.

  • AI and ML denote Artificial Intelligence and Machine Learning, the paper’s central computational concepts.
  • CNN, DNN, RNN, LSTM, MLP, GAN, GMM, R-CNN, and SNN identify neural-network or generative modeling techniques.
  • DRL, DDPG, DQN, and RL identify reinforcement-learning methods and related algorithms.
  • UAV, IoT, MEC, BS, CPU, GPU, QoS, and AoI cover aerial systems, networking, computing, hardware, service quality, and information freshness.
  • LoS, RSS, SINR, WPT, VTOL, MPC, PID, and SfM identify communication, power, flight, control, and vision terms.

1 INTRODUCTION

UAVs support communications, data collection, edge computing, and monitoring across demanding environments, while ML supplies tools for perception, prediction, planning, and control. This survey organizes those tools by application, reviews their advantages and drawbacks, and identifies gaps including limited end-to-end integration and trust assessment.

  • UAVs provide mobile data collection, forwarding, and edge-computing capabilities where terrestrial infrastructure is unavailable, unreliable, or costly.
  • UAV-assisted applications must balance data richness, freshness, representativeness, reliability, latency, and energy efficiency.
  • The survey classifies ML techniques by their roles in feature extraction, environment modeling, planning and scheduling, control, and data collection.
  • CNN predominates in UAV image processing, while low-power SNNs and continuous-action DRL methods such as DDPG support online operations and control.
  • Existing work increasingly combines ML modules, but little effort has produced an end-to-end solution from feature extraction through planning, control, and operation.
  • The survey reviews over 300 UAV-compatible ML studies and covers supervised, unsupervised, and reinforcement-learning architectures.

2 BACKGROUND TO UAV SYSTEMS

UAV systems vary by platform configuration and mission structure, while their communications support command and control, air traffic safety, and collision avoidance. UAV-assisted communications balance throughput, energy, information freshness, reliability, and resource utilization.

  • UAV System Types: UAV systems comprise single-UAV and multi-UAV configurations, with swarms enabling broader coverage and parallel task processing.Single-UAV missions rely on one aircraft, whereas multi-UAV systems can reduce mission completion time through geographic coverage and parallelism.
  • Communication Capabilities: UAV communications must support command and control, air traffic control relay, and collision avoidance.These links carry telemetry and control updates, protect manned aircraft in shared airspace, and support obstacle sensing with adequate safety distance.
  • UAV Types: UAV platforms include VTOL, fixed-wing, and hybrid VTOL designs distinguished by rotor configuration and take-off method.VTOL platforms include single-rotor and multi-rotor types; fixed-wing UAVs support long-distance cruising, while hybrid VTOLs combine vertical flight with wing-borne motion.
  • Performance Metrics: Spectral efficiency and energy efficiency are competing objectives because higher throughput must be balanced against battery consumption and mission constraints.Energy use is affected by weather, speed, trajectory alignment, and maneuvering, motivating models that maintain throughput while reducing consumption.
  • Performance Metrics: Age of Information measures the time since the latest packet update, linking data freshness to trajectory planning, sensor selection, power efficiency, and latency.Fresh information can improve control and reliability, but UAV and ground networks must compromise between freshness and resource use.
  • Performance Metrics: Communication reliability is measured using SNR, SINR, or outage probability, while resource utilization concerns allocating limited bandwidth and energy to improve data transmission.UAV communication reliability can additionally deteriorate from wind-induced jitter, and learning techniques are needed to adapt allocation to environmental conditions.

3 SURVEY OF SURVEYS

Earlier surveys address UAV communications, machine learning methods, and specialized applications, but they often organize the literature around particular functions or technologies. This survey broadens the perspective by analyzing ML techniques together with their UAV-operation applications, advantages, drawbacks, and research gaps.

  • Existing UAV Surveys: Earlier UAV communication surveys examine challenges such as cybersecurity, energy efficiency, stable communication, monitoring, channel modeling, and software-defined networking.The reviewed literature also covers fifth-generation integration, aeronautical link budgets, channel fading, and SDN-enabled cellular systems.
  • Existing ML Surveys: General ML surveys review techniques including representation learning, deep learning, distributed learning, transfer learning, active learning, and kernel-based learning across application domains.These methods support big-data processing and have been considered relevant to UAV applications.
  • ML-Aided UAV Surveys: Surveys focused on ML-assisted UAV communications address channel modeling, positioning, security, interference suppression, hand-off assistance, and authentication.Their emphasis is on functional aspects and security challenges rather than a unified application-oriented analysis.
  • Technology-Focused Surveys: Other reviews cover AI, ML, DRL, MEC, and SDN for joint UAV-system optimization, collaboration, cooperation, and changing network conditions.These surveys examine technology-specific roles in UAV-aided communications rather than the full range of UAV operational modules.
  • DRL Surveys: DRL surveys organize reinforcement-learning techniques around path planning, navigation, control, and network functions such as access control and smart caching.Reported simulation tasks include hovering, landing, random way-point movement, and target following.
  • Scope of This Survey: This survey provides a broader application view by analyzing why specific ML techniques fit UAV operations and communications, including their advantages, drawbacks, and research gaps.It differs from reviews centered on particular UAV applications or isolated ML techniques.

4 BACKGROUND OF ML FOR UAV APPLICATIONS

This section introduces supervised, unsupervised, semi-supervised, reinforcement, and federated learning approaches used across UAV applications. It also connects these techniques to trajectory planning, mission scheduling, and resource-aware operation.

  • Supervised learning uses labeled datasets to classify data or predict outcomes, adjusting model weights through a loss function.
  • CNNs extract image features through convolution, pooling, and connection layers, supporting UAV image classification and recognition.
  • RNNs model sequential data but generally require longer training and more difficult calibration than feed-forward networks.
  • Unsupervised learning discovers patterns in unlabeled data, including through GANs that train generative and discriminative networks adversarially.
  • Semi-supervised learning combines limited labeled data with larger unlabeled datasets, while K-means groups samples around repeatedly updated centroids.
  • Federated learning trains decentralized models across servers or edge devices, and multi-agent DRL enables collaborative action exploration and convergence.

5 ML FOR UAV TRAJECTORY PLANNING AND MISSION SCHEDULING

ML supports UAV trajectory planning and mission scheduling by predicting states, selecting waypoints, optimizing communication resources, and adapting decisions to dynamic environments. The surveyed approaches include supervised, unsupervised, adversarial, recurrent, and deep reinforcement learning methods.

  • RNN-based arrival-angle prediction provides UAV position information that affects beamforming in communication-service applications.
  • Autoencoders generate candidate waypoints from historical trajectories and images, while also supporting movement prediction in unknown terrain.
  • GAN-LSTM models jointly optimize transmit power, communication mode, frequency channels, and UAV selection or trajectories under partial observability.
  • DRL provides solutions for trajectory planning in dynamic environments, with DQN extending Q-learning to larger state and action spaces.
  • DQN-based methods jointly optimize trajectories, bandwidth, communication schedules, energy use, data freshness, packet loss, and MEC task offloading.

6 ML FOR AERODYNAMIC CONTROL AND OPERATION OF UAV

ML-based aerodynamic control addresses UAV modeling, trajectory tracking, and motion control when accurate mathematical models are difficult to obtain. RNN, MLP, SNN, and hybrid architectures are applied to dynamics modeling, adaptive control, and disturbance handling.

  • CNN, MLP, and RNN methods support path finding, image-based control, caching decisions, and sequential UAV control.
  • Accurate on-the-fly UAV control is difficult because unmodeled dynamics, parametric uncertainties, and disturbances reduce mathematical-model fidelity.
  • A hybrid network combines non-recurrent and recurrent blocks to model helicopter attitude and position dynamics, including roll, pitch, and yaw.
  • MLP performs well in level flight, whereas radial basis networks perform well during take-off and landing, requiring separate networks for different stages.
  • RNN-based controllers address trajectory tracking, extremum seeking, collision avoidance, feedback linearization, and fixed-wing control under disturbances.
  • MLP-based adaptive PID controllers and recurrent SNN controllers are applied to suppress control error, accelerate learning, and handle nonlinear six-degree-of-freedom control.

7 ML FOR UAV PERCEPTION AND FEATURE EX-

UAV perception and feature extraction transform aerial imagery and radio measurements into representations for classification, detection, segmentation, and communications. CNNs and related deep or adversarial models dominate imagery applications, while autoencoders and GANs support reconstruction and radio-feature learning.

  • Feature extraction reduces measured UAV data into informative, non-redundant values such as edges, shapes, and recognized objects.
  • MLP, CNN, and RNN models analyze aerial imagery for crop disease, vegetation, environmental management, and flood detection.
  • 96% accuracy was achieved for detecting flooded areas using densely connected CNN and RNN models on UAV imagery.
  • CNNs extract forest features, support video encoding under packet loss, and achieve almost 90% precision and accuracy for slope-failure detection.
  • CNN segmentation is detailed and can be time-, energy-, and resource-consuming, while deeper networks require longer training.
  • GANs and autoencoders support pixel classification, low-resolution object identification, image decomposition, missing-information recovery, and radio-channel distribution estimation.

8 ML FOR FEATURE INTERPRETATION AND RE-

ML methods support UAV feature interpretation, environment modeling, and trajectory planning through regression, clustering, and probabilistic models. These methods can improve planning and communication deployment, but their accuracy depends on sufficient prior data and individual techniques involve trade-offs.

  • Feature interpretation: Regression models interpret UAV sensory data for environmental monitoring, crop surveillance, water-quality analysis, and bathymetric mapping.Applications include linear regression, fuzzy regression, and geographically weighted regression with UAV-captured imagery or spectral data.
  • Feature interpretation: Linear regression is easy to implement but assumes linear relationships, remains vulnerable to noise and over-fitting, and fails when features outnumber observations.
  • K-means clustering: K-means clusters tasks into subsets for multi-UAV route planning and navigates UAV movement for coverage control.Subsequent route planning can use optimization methods such as simulated annealing or genetic algorithms.
  • Gaussian mixture modeling: GMM models obstacle distributions, object-detection probabilities, and complex motion patterns to generate potential fields, infer trajectories, and plan search routes.Its applications include collision avoidance, target search, and probability-guided flight planning.
  • Gaussian mixture modeling: GMM-based traffic modeling reduced UAV energy consumption by 20% for communication and 80% for mobility versus heuristic-based alternatives.
  • Cross-method observations: K-means, linear regression, and GMM support feature interpretation, spatial modeling, classification, and prediction, but are most accurate when sufficient prior data are available.

9 CHALLENGES, OPEN ISSUES, AND DISCUSSION

The survey identifies limited progress toward an end-to-end ML design integrating UAV perception, interpretation, planning, control, and operation. This gap matters as increasingly autonomous UAV teams undertake sophisticated missions and potential human-UAV collaboration expands.

  • Research gap: Considerable effort addresses four UAV ML elements, but little progress exists toward a holistic end-to-end solution integrating them for efficiency, reliability, and quality assurance.
  • Motivation: End-to-end integration is important because increasingly autonomous UAV teams perform sophisticated surveillance, environmental, detection, rescue, and herding operations.
  • Open challenges: Holistic end-to-end ML design raises challenges for sophisticated UAV operations and collaborations.
  • Open challenges: Growing UAV and IoT deployment will increase air traffic, requiring ML-assisted airspace compliance, path synchronization, safe trajectory planning, and task completion.These requirements must account for UAVs, other aircraft, pedestrians, and property.
  • Data requirements: Unexpected environmental changes and events create a lack of prior data, motivating the need for offline training data for UAV operations and communications.

9.2 Increasing Energy Requirement vs. Finite Battery

UAV ML systems face a coupled energy and computation constraint: onboard batteries and processing limits restrict feasible applications, while more accurate feature extraction can shorten mission duration.

  • Computational demand: ML operations can be computationally expensive and energy-hungry, especially for lower-level feature extraction and online DRL training.
  • Platform constraints: Limited batteries and onboard processing restrict ML applications to tasks such as object detection, depth prediction, target tracking, and localization.
  • Research needs: Energy-efficient AI/ML aerial systems require investigation, verification, efficient architectures, and suitable embedded hardware and software.The survey identifies lightweight techniques such as R-CNN and SNN as possible directions.
  • Energy-saving approaches: UAVs can collaborate with public transport vehicles by resting on roofs, turning off motors, and potentially recharging batteries to support longer-distance flight.
  • Energy-accuracy trade-off: Higher feature-extraction accuracy can increase computation, convergence latency, energy depletion, and reduce UAV mission time.

9.3 UAV Cooperation Without Persistent Connectivity

Multi-UAV learning is challenged by rapidly changing joint states, long or unstable multi-agent DRL convergence, unreliable connectivity, privacy risks, heterogeneous platforms, and differing operational capabilities.

  • Multi-agent learning: Independent UAV state observations can quickly become outdated after another UAV acts, causing multi-agent DRL to converge slowly or diverge.
  • Connectivity: Sharing actions and state observations could enable joint training, but requires consistent and reliable wireless connections that are difficult in practice.
  • Security and privacy: Coordinating collaborative UAV training requires sharing network states and rewards, exposing transmissions to eavesdropping and message modification over wireless channels.
  • Heterogeneous swarms: Heterogeneous UAV swarms differ in processing capabilities, maneuverability, communication range, and response delays, affecting ML-based coordination.

9.6 Responsible ML for UAVs Interacting With Reactive Objects

Responsible ML for UAVs interacting with reactive objects requires shared situational awareness, mutual understanding of capabilities and intent, and distributed decision-making. Current research emphasizes sensing, while higher-level supervision, planning, and interaction with humans and animals remain underexplored.

  • Human-UAV teams require shared physical-world understanding, mutual awareness of capabilities and roles, intent recognition, and distributed decision-making.
  • The science of combining complementary human and machine intelligence remains a significant knowledge and capability gap.
  • Most UAV ML methods target sensing tasks such as traffic detection and data classification rather than higher-level supervision and planning.
  • Interaction between UAVs and reactive objects, including humans and animals, has received little investigation despite the complex behaviors involved.

9.8 Theory and Implementation of RL/DRL for UAV Attitude Control

UAV autopilots primarily rely on PID control, which performs well in stable environments but is less suitable for unpredictable and harsh conditions. RL/DRL, including newer meta-learning and transfer-learning approaches, is presented as an option for more intelligent flight control and lower data or operational burdens.

  • PID-based autopilots demonstrate excellent performance in stable environments but require more sophisticated control in unpredictable and harsh environments.
  • RL/DRL techniques are being incorporated into intelligent flight-control systems to address shortcomings of conventional PID control.
  • Meta-learning and transfer learning have been used in UAV applications as alternatives to classic ML methods with higher operational costs and larger data requirements.

10 CONCLUSION

This survey reviews and classifies ML techniques for UAV operations, communications, and IoT applications across four operational modules. It finds module-specific technique dominance and growing cross-module integration, but limited end-to-end development and unresolved trust, reliability, and security challenges.

  • The survey classifies ML techniques for UAV operations and communications across feature extraction, environment interpretation, planning and scheduling, and control and operation.
  • Different ML techniques dominate the four modules; CNN is prominent in UAV image processing, while DRL is increasingly used for online control and communication scheduling.
  • Cross-module designs are increasing, including RNN-enhanced feature extraction that supplies inputs to DRL for faster UAV action exploration and exploitation.
  • Little to no effort has produced an end-to-end ML solution spanning feature extraction through UAV control and operation.
  • Security, reliability, and trustworthiness of ML remain important research gaps before full UAV automation can be realized.
Loading 2211.04324v1…