Logistics Planning of Autonomous Air Cargo Vehicles with Deep Learning Methods: A Literature Review

Abstract

Over the past decade, digitalization in the logistics sector has heightened the significance of autonomous systems and AI-based applications, while the integration of advanced deep learning technologies with air cargo carriers has ushered in a new era in the logistics industry. This study systematically addresses the current applications of these technological advances in logistics planning, the challenges faced, and perspectives for the future. These developments are transforming the role of UAVs and autonomous systems in logistics operations by improving last-mile efficiency and reducing costs. Key functions of autonomous vehicles, including environmental perception, decision-making, and route optimization, have shown notable progress through deep learning algorithms. However, major obstacles remain to their widespread adoption, particularly in terms of energy efficiency, data security, and the absence of a mature regulatory framework. Accordingly, this paper discusses these issues in detail and highlights areas for further research. This systematic literature review reveals the disruptive potential of AACV for the logistics industry and presents findings that can guide both academic inquiry and industrial practice. The results underscore that establishing a sustainable and efficient logistics ecosystem is essential in the context of these emerging technologies.

1. Introduction

In recent years, the use of autonomous air cargo vehicles (AACVs) and deep learning methods in logistics planning has become an important area of research in logistics and supply chain management. The integration of these methods offers opportunities to enhance the efficiency of logistics processes and reduce operational costs [1]. Autonomous vehicles and drones provide effective solutions to the limitations of traditional logistics approaches, particularly in last-mile deliveries and in challenging geographical conditions [2].

The application of deep learning techniques is crucial for enhancing the perceptual and decision-making capabilities of autonomous vehicles [3]. These technologies enable more effective decision-making in logistics planning by delivering results that surpass human performance in complex visual tasks [4]. For example, convolutional neural networks (CNNs) are used to process camera images for the detection of lane markings, road edges, and obstacles [3].

The application of artificial intelligence and machine learning in supply chain management has advanced considerably, particularly in demand forecasting, inventory optimization, and route planning. The deployment of these technologies enhances operational efficiency by strengthening supply chain resilience and improving performance under dynamic environmental conditions [4].

The objective of this study is to provide a comprehensive review of research on logistics planning for autonomous air cargo vehicles, with a particular emphasis on deep learning methodologies. The aim is to identify key developments, challenges, and future research directions in this field. This study is intended to serve as a valuable reference for understanding the potential applications and impacts of these technologies in logistics and supply chain management.

The growth of e-commerce and rising customer expectations for faster delivery have increased the demand for air cargo transportation [5]. In recent years, technologies such as unmanned aerial vehicles (UAVs) and autonomous delivery robots have shown strong potential to transform traditional logistics chains [6,7].

The deployment of AACVs has the potential to support the development of logistics networks extending into city centers. These vehicles can help reduce road traffic and minimize environmental impacts by establishing new logistics chains from cargo airports to urban areas [7]. Moreover, their capacity for continuous operation allows them to address the shortage of available drivers [6].

However, the integration of AACVs into supply chains presents several technical and legal challenges. Fundamental infrastructure requirements, security concerns, and regulatory frameworks remain significant barriers to their widespread adoption [6,7]. Addressing these obstacles will require the development of improved methodologies to enable more effective planning and management of AACVs.

The application of deep learning methods is becoming increasingly common in addressing complex logistics problems. These methods are distinguished by their ability to identify meaningful patterns within large datasets and to enhance the effectiveness of complex decision-making processes. Deep learning algorithms offer promising solutions to several challenges related to AACVs, including route planning, load optimization, and energy consumption minimization [8].

1.1. Scope and Aim of the Study

The purpose of this study is to conduct a comprehensive literature review on logistics planning for AACVs using deep learning methods and to identify key developments, challenges, and future research directions in this field.

This research is structured around the following main topics:

• AACVs and their applications in the logistics sector.
• Application of deep learning and artificial intelligence techniques in logistics planning.
• Integration of deep learning techniques and autonomous vehicles in supply chain management.
• Deep learning applications in logistics processes, including route optimization, demand forecasting, and inventory management.
• Use of AACVs in last-mile deliveries and their sustainability impacts.
• Application of artificial intelligence and machine learning techniques in the optimization of distribution networks.
• Technical, legal, and ethical challenges faced by autonomous vehicles and drones in logistics planning.

This study aims to examine the potential impacts and application areas of AACVs and deep learning technologies in logistics and supply chain management. A detailed analysis of these impacts and emerging trends has the potential to guide future research. The study also considers how these technologies can address current challenges and how large retail organizations adapt to them within the broader context of innovation in the sector.

1.2. The Structure of the Paper and the Path to Be Followed:

This study investigates logistics planning for AACVs using deep learning methods through a systematic and comprehensive approach. The paper begins with an introduction outlining the background, significance, and objectives of the research. The methodology section then explains the literature review strategy, inclusion and exclusion criteria, and data collection and analysis methods.

The main body of the paper first discusses AACVs and deep learning separately, presenting the fundamental principles and characteristics of these technologies. It then examines the application of deep learning in logistics planning, focusing on specific areas such as route optimization, demand forecasting, inventory management, and distribution network optimization.

The core of the study—a comprehensive review of deep learning applications in AACV logistics planning—includes analyses of application areas, success cases, and associated challenges with proposed solutions. The discussion section synthesizes the findings, identifies research gaps, and suggests future directions. It also offers recommendations for sectoral applications. The paper concludes with a summary of the main findings and a discussion of the study’s contributions and limitations.

2. Materials and Methods

2.1. Literature Review Strategy

This study conducted an extensive and systematic review of the literature on the application of deep learning methods to logistics planning for AACVs. Given the multidisciplinary nature of the topic, the review strategy was designed with a broad perspective. The process began with the identification of suitable keywords, which included “autonomous vehicles,” “drones,” “deep learning,” “artificial intelligence,” “logistics planning,” “distribution networks,” “machine learning,” “supply chain management,” and their Turkish equivalents. These keywords reflect the core components of the research area and enabled the identification of relevant studies.

The literature searches were conducted primarily in Google Scholar, Web of Science, Scopus, IEEE Xplore, and ScienceDirect. The selection of these databases reflects the multidisciplinary nature of the topic, as they encompass both engineering and management perspectives. In addition, these databases include academic publications and conference proceedings, ensuring that the most recent studies in the field are represented in the search results.

The search for relevant studies was subsequently limited to literature published within the last 10 years (2014–2024). This timeframe was considered sufficiently recent to capture the rapid advancements in autonomous vehicles and deep learning technologies, while also including foundational studies in these fields. Publications were primarily reviewed in English and Turkish, although essential works in other languages were also included when necessary.

The titles, abstracts, and keywords of the retrieved studies were analyzed, and those relevant to the topic were selected. At this stage, the methodological quality of the studies, the originality of their findings, and their contribution to the field were used as key criteria. In addition, the bibliographies of the selected studies were reviewed to identify important references that may have been overlooked during the initial screening.

Data extraction was conducted systematically, focusing on the major findings, methodologies employed, and outcomes reported in the selected studies. The extracted data were then categorized and analyzed according to predetermined themes, including the use of autonomous vehicles in logistics planning, applications of deep learning methods, supply chain optimization, and associated challenges.

Subsequently, the information obtained was synthesized to identify the current state of research, main trends, and future research directions. This synthesis process involved comparing the findings across studies, highlighting contradictions, and identifying research gaps.

This comprehensive and systematic literature review strategy seeks to provide an up-to-date and in-depth understanding of logistics planning for AACVs using deep learning methods. Its purpose is to consolidate existing knowledge in the field and to guide future research and practical applications.

2.2. Inclusion and Exclusion Criteria

In this study, a set of inclusion and exclusion criteria was established to conduct a comprehensive and systematic review of articles on AACV logistics planning using deep learning methods. These criteria were designed to keep the research focused, ensure the inclusion of current and high-quality studies, and exclude works not directly relevant to the topic.

The inclusion criteria define the scope of the review. First, studies were required to address at least three of the following topics: autonomous vehicles, drones, deep learning, artificial intelligence, logistics planning, distribution networks, machine learning, and supply chain management. This requirement ensured that the review remained focused and captured relevant literature. Regarding publication type, only articles published in peer-reviewed journals, conference proceedings, and book chapters were included, in order to guarantee academic quality.

With respect to methodology, both theoretical and empirical studies were included, ensuring coverage of the conceptual as well as applied aspects of the topic. In terms of application domain, studies addressing logistics, supply chain management, distribution networks, and last-mile delivery were incorporated, allowing for an assessment of practical applications and potential impacts.

Exclusion criteria were established to maintain the focus of the research and to filter out irrelevant studies. Specifically, logistics studies that did not involve AACVs or any aspect of artificial intelligence were excluded to ensure alignment with the research scope. In addition, newspaper articles, blogs, and non-refereed publications were excluded in order to guarantee academic quality.

Exclusion based on language was applied when translation was not possible for studies written in languages other than English or Turkish. This criterion ensured accurate understanding and evaluation of the studies. In terms of accessibility, works for which the full text was unavailable were excluded, as detailed analysis would not have been possible.

In cases of duplication, different versions or summaries of the same study were excluded, with only the most comprehensive version retained to avoid redundancy and ensure that the most complete and updated information was included. Finally, studies with weak methodology or questionable results were removed to maintain the overall quality and reliability of the literature review.

These inclusion and exclusion criteria ensured that the literature review remained systematic and comprehensive. Minor adjustments were allowed based on special circumstances or findings that emerged during the screening process. This flexibility reflects the dynamic nature of the research and helps ensure that potentially important studies are not overlooked. The study selection process is illustrated in the flow diagram (Figure 1), which shows the number of records identified, screened, excluded, and ultimately included in the review.

2.3. Data Collection and Analysis Methods

This literature review was conducted using a systematic approach to data collection and analysis, with the aim of comprehensively examining research on AACV logistics planning using deep learning methods. For data collection, articles published in peer-reviewed journals, conference proceedings, and book chapters were considered. Study selection was based on an analysis of titles, abstracts, and keywords to identify works directly related to the topic. In addition, the bibliographies of the selected studies were reviewed to capture important references that might have been overlooked during the initial screening. For data analysis, the collected studies were systematically examined and categorized. The analysis process involved the following steps:

• Content analysis: Identification of the main themes, methodologies and key findings of each study.
• Thematic categorization: Grouping of studies under major themes such as autonomous vehicles, drones, logistics planning, distribution networks, supply chain management, deep learning, machine learning and artificial intelligence.
• Comparative analysis: Comparison of the findings and approaches of different studies to identify trends and research gaps in the field.
• Synthesis: Integration of the analyzed information to determine the current state of the field, main trends, and future research directions.

Data collection and analysis were carried out through a comprehensive review of the literature, identifying key developments, challenges, and future research directions in the field. This process was designed to synthesize the existing body of knowledge and to provide guidance for future research and practical applications.

3. Results

3.1. Autonomous Air Cargo Vehicles and Deep Learning

3.1.1. Autonomous Air Cargo Vehicles and Deep Learning

AACVs have gained significant momentum in the logistics sector in recent years, driven by technological advances and developments in artificial intelligence systems. These vehicles hold the potential to increase efficiency, reduce operational costs, and expand access to hard-to-reach areas [9,10]. These vehicles employ a range of autonomous technologies. Sensor and detection systems, including lidar, radar, and optical cameras, enable obstacle detection and avoidance. Navigation and control processes rely on GPS, inertial measurement units (IMUs), and real-time kinematic positioning for precise navigation. In addition, communication systems utilize data links to support command, control, and communication with ground stations or other aircraft [11,12,13]. In terms of load capacity, small drones can transport only a few kilograms, making them suitable for applications such as parcel delivery. In contrast, heavy-lift UAVs are large vehicles designed for logistics and industrial purposes, with the capability to carry payloads of 500 kg or more [9,10,14,15]. With respect to range and endurance, lithium-ion and lithium-polymer batteries extend flight capability, while hybrid systems enable longer flight times by combining electric and fuel-powered engines [13,15,16]. Aviation authorities such as the FAA and EASA regulate the integration of autonomous vehicles into airspace and establish the necessary frameworks for their operation. Safety standards emphasize compliance with aircraft certification requirements, air traffic management procedures, and safety protocols [9,11,17]. The applications of AACVs are diverse, ranging from logistics and supply chain operations to e-commerce deliveries, which ensure rapid distribution of online purchases to customers. In addition, within just-in-time manufacturing, AACVs support the timely delivery of essential parts to production facilities [9,10,15]. In humanitarian aid and disaster response, technologies such as drones and autonomous aerial vehicles enhance the effectiveness of rescue operations by enabling fast and safe deliveries, even under extreme weather conditions. When access roads are blocked, these vehicles play a vital role by rapidly transporting food, water, and medical supplies. At the same time, they provide sustainable solutions to infrastructure challenges by offering regular service in isolated areas [14,16]. Military applications include logistics support, which enables the transport of materials and equipment in hazardous areas without endangering human life. In addition, surveillance and reconnaissance operations enhance border security and intelligence gathering through the use of autonomous vehicles [9,17]. In industrial and commercial applications, autonomous vehicles improve workforce efficiency and facilitate access in challenging geographical conditions. In agriculture, they enhance productivity, particularly across large areas. In the mining and energy sectors, they contribute to safer and more sustainable operations by reducing time and costs [9,15,17]. Despite these advantages, several challenges remain to be addressed. Technical challenges include improving energy efficiency through extended battery life and more advanced energy management systems. In addition, enhancing the decision-making capabilities and fault tolerance of artificial intelligence algorithms is essential to strengthen autonomy and reliability. These improvements are critical for achieving longer operating times and ensuring safe autonomous systems [11,12,13,16]. Regulatory barriers present another challenge. Standardization provides a framework for the safe and efficient use of autonomous aircraft by establishing common rules and standards at the international level. At the same time, airspace integration seeks to incorporate autonomous vehicles seamlessly into existing air traffic systems, thereby enhancing safety and efficiency in both commercial and military operations [9,17]. Social acceptance and security concerns must also be addressed. Public trust can be strengthened by raising awareness about the safety and reliability of autonomous aircraft. At the same time, ensuring privacy and data security is essential to protect communication and data systems against cyber threats. These measures are critical for the widespread adoption of autonomous technologies, as the confidence of both users and society in these systems must be reinforced [11,16,17,18]. AACVs have the potential to revolutionize the logistics and transportation sectors by accelerating delivery processes and reducing costs. As technology advances and regulatory frameworks mature, these vehicles are expected to play a critical role in global supply chains by enhancing operational efficiency. At the same time, they are likely to drive sectoral transformation by making significant contributions to environmental sustainability.

3.1.2. Basic Principles of Deep Learning Methods

Deep learning, a subfield of machine learning, extracts meaning from data through multiple layers of artificial neural networks. With the growth of big data and computational power in recent years, deep learning methods have achieved remarkable results in areas such as image processing, natural language processing, and speech recognition. Artificial neural networks, inspired by biological neural networks, consist of an input layer, one or more hidden layers, and an output layer [19]. They employ activation functions such as Sigmoid, ReLU (Rectified Linear Unit), and tanh to determine the output of neurons [20]. Deep neural networks (DNNs) are neural networks with more than one hidden layer, where increased depth enhances the model’s ability to learn complex functions [19]. Specialized architectures include convolutional neural networks (CNNs), which are primarily used for image data and employ convolutional layers to capture spatial relationships [21], and recurrent neural networks (RNNs), which are designed for time series and sequential data and are capable of retaining past information [21]. The learning and optimization process involves backpropagation, which updates the weights by propagating the error signal backward through the network [19]. Optimization algorithms such as stochastic gradient descent (SGD), Adam, RMSprop, Adagrad, and Nadam are used to accelerate and improve the learning process [22]. To evaluate model performance, loss functions such as cross-entropy (CE) and mean squared error (MSE) are commonly employed [22]. The schematic structure of the learning and optimization process is presented in Figure 2.

Overfitting occurs when a model fits the training data too closely but performs poorly on overall evaluation [23]. Regularization methods such as dropout, early stopping, and L1 and L2 norms are employed to improve the generalizability of the model [24].

Transfer learning is a technique widely used in machine learning and deep learning. It is based on reusing a pre-trained model for a similar task or dataset. By transferring existing knowledge to a new problem, this method enables faster and more efficient solutions, eliminating the need to train models from scratch that would otherwise require large amounts of data and computational power. Image processing plays a critical role in object recognition, face recognition, and medical image analysis. Natural language processing enables the interpretation of language data through applications such as machine translation, text classification, and sentiment analysis. In addition, speech recognition technology is applied in areas including voice assistants and automatic subtitle generation. In the domains of gaming and simulation, deep learning contributes to the development of artificial intelligence game characters and autonomous vehicle simulations [21]. The need for large amounts of labeled data and high-performance hardware for the successful training of deep learning models remains a significant challenge. Moreover, the difficulty in understanding and explaining the decision-making processes of these models raises the issue of explainability, which requires greater attention in future research. Deep learning methods form the foundation of modern artificial intelligence applications and have enabled revolutionary advances across a wide range of fields. A clear understanding of their fundamental principles is crucial for the effective and efficient application of these technologies.

3.1.3. Deep Learning Applications in Autonomous Vehicles

In recent years, autonomous vehicle technologies—particularly those based on deep learning—have advanced rapidly. These technologies enable land and air vehicles to operate safely and efficiently without human intervention. Autonomous aerial vehicles, in particular, create new opportunities in diverse fields ranging from logistics to disaster management. Deep learning, with its ability to detect complex associations and patterns in large datasets, plays a central role in perception, decision-making, and planning tasks for both land and air vehicles. As a result, these vehicles can identify obstacles in real time, assess weather and road conditions, and determine optimal routes.

Deep learning is a machine learning approach that extracts and models data features through multi-layer artificial neural networks. In the context of autonomous vehicles, its applications focus on several key areas: sensing and environmental perception, localization and mapping, decision-making and planning, and sensor fusion. Realistic perception of the surrounding environment is essential for safe navigation. Deep learning-based CNNs outperform traditional methods in feature extraction and object recognition from image data. Real-time object detection algorithms such as you only look once (YOLO) and single shot multibox detector (SSD) enable vehicles to quickly and effectively detect pedestrians, other vehicles, and obstacles [25,26]. Furthermore, deep learning architectures such as fully convolutional networks (FCNs) and U-Net are employed for semantic and instance segmentation, respectively. These techniques enable pixel-level classification of the environment, allowing the distinction between classes such as roads, pavements, and traffic signs [27,28]. PointNet and VoxelNet are deep learning models applied to the processing of 3D data from sensors such as LiDAR and radar [29,30]. These models enable vehicles to achieve a more comprehensive perception of their surroundings by detecting objects from three-dimensional point clouds. However, the two approaches differ in terms of data representation and computational efficiency. PointNet operates directly on raw point cloud data without transformation, thereby preserving greater detail and achieving higher computational efficiency. In contrast, VoxelNet converts the data into voxels to create a more structured representation. While this allows VoxelNet to benefit from convolutional operations, it also results in some loss of detail and higher computational demands. Thus, PointNet maximizes data integrity and efficiency, whereas VoxelNet provides the advantages of convolutional network processing.

The ability of vehicles to accurately localize and generate maps of their surroundings is a fundamental prerequisite for autonomous driving. Image-based localization methods, enhanced by deep learning, enable vehicles to self-localize by matching live images with previously known maps [31]. SLAM algorithms are increasingly integrated with deep learning to achieve more robust and faster mapping. In particular, automatic feature extraction through deep learning significantly enhances SLAM performance [32]. SLAM enables an autonomous vehicle operating in an unfamiliar environment to estimate its own location while simultaneously generating a map of its surroundings. Image-based SLAM, for example, relies on camera data to accomplish this task. Under dynamic traffic conditions, autonomous vehicles must make accurate decisions and generate appropriate movement plans to ensure safe and efficient operation. In the field of path planning and motion control, deep reinforcement learning (DRL) enables vehicles to learn optimal movement strategies [33]. Furthermore, recurrent neural networks (RNNs), and particularly long short-term memory (LSTM) networks, are effective in predicting the behavior of other traffic participants using time series data. By anticipating possible scenarios in advance, these models enable vehicles to make safer decisions [34]. Combining data from different sensors allows vehicles to perceive their surroundings more accurately and comprehensively. Multimodal deep learning approaches enhance detection performance by processing data from cameras, LiDAR, radar, and ultrasonic sensors using deep neural networks (DNNs) [35]. Some studies focus on end-to-end learning models that generate control commands directly from raw sensor data (Figure 3). These approaches enable deep neural networks (DNNs) to learn vehicle control signals directly from image data. For example, one model was developed to learn vehicle control by mimicking the behavior of human drivers [36].

Deep learning applications in autonomous vehicles face several challenges, one of the most critical being the need for large and diverse datasets. Collecting and labeling comprehensive datasets that accurately reflect real-world conditions is both costly and time-consuming. To address this issue, synthetic data generation and data augmentation techniques are employed in simulation environments [37]. Another major challenge concerns the reliability and explainability of deep learning models. Since their decision-making processes are often difficult to interpret, this raises both security and regulatory concerns. Explainable AI seeks to address this issue by making the inner workings of models more transparent [38]. Legal regulations and ethical considerations also play a critical role. For autonomous vehicles to operate safely, legal frameworks must be established at both international and national levels. From an ethical standpoint, ongoing discussions focus on how these vehicles should make decisions in emergency situations and on the social acceptance of such decisions [39]. The integration of 5G and vehicle-to-everything (V2X) communication technologies will be a major consideration for the future of autonomous vehicles. Through these technologies, vehicles will be able to share real-time information with other vehicles and infrastructure, thereby enabling safer and more efficient driving experiences [40]. Furthermore, shared mobility is expected to become increasingly prominent with the adoption of autonomous vehicles in public transportation and the logistics industry. Through continuous learning and online adaptation capabilities, these vehicles will be able to respond more rapidly to changing conditions and new situations [41].

Deep learning is expected to be one of the key technologies driving the rise in autonomous vehicles. To ensure safe and effective operation, it has been applied to critical areas such as sensing, localization, decision-making, and planning. As technology continues to advance and current challenges are overcome, autonomous vehicles are likely to become a routine part of daily life and play a central role in shaping the future of transportation.

3.2. Deep Learning Methods in Logistics Planning

Route optimization is a critical component of logistics operations planning, particularly in large-scale distribution networks. Compared to classical procedures, deep learning methodologies offer a more efficient and effective approach to solving route optimization problems. As deep learning can process large datasets and identify complex patterns, it serves as a powerful tool for determining optimal routes while accounting for the inherent complexity and variability of distribution networks. Techniques such as multi-agent deep reinforcement learning (DRL) have been employed to optimize delivery vehicle routes [42,43]. The available evidence indicates that deep learning-based route optimization is more efficient and robust than traditional techniques. In particular, it has been shown to offer greater statistical reliability. For example, one study demonstrated that a deep learning-based algorithm achieved superior accuracy in route optimization compared to conventional methods [44]. Deep learning-based route optimization offers greater flexibility in exploring the solution space, particularly given the inherent dynamics and uncertainties of distribution networks. Recent studies have applied deep learning-based algorithms to optimize delivery vehicle routes in real time [45]. Furthermore, deep learning-based route optimization can incorporate the environmental impacts of distribution networks. For example, one study applied a deep learning algorithm with the objective of reducing the energy consumption of delivery vehicles [46]. From this, it can be reasonably concluded that the aforementioned methods are more efficient and effective than conventional approaches. Further research in this field will support the development of more advanced and effective route optimization methodologies.

Demand forecasting is a critical task in logistics planning, as it enables more efficient use of available resources and optimization of supply chain management. In this area, deep learning methods have been shown to outperform traditional statistical models. Recently, these models have been increasingly applied to logistics demand prediction, where their ability to process complex and nonlinear data provides a significant advantage. For example, a study of 54 distribution centers belonging to a Brazilian shipping company demonstrated that LSTM networks achieved superior performance compared to autoregressive integrated moving average (ARIMA) models [47]. In addition, ensemble learning methodologies have been applied to logistics demand forecasting. For example, researchers proposed an ensemble approach that integrates the multiscale time-delay convolution model (MSTDCM) with the seasonal autoregressive integrated moving average (SARIMA) model. This method proved effective in forecasting local values and volatility using data from Singapore’s logistics sector [48]. Deep learning methodologies have also been applied in the context of small and medium-sized enterprises (SMEs). For instance, a study conducted in collaboration with an SME in Denmark reported effective results in demand forecasting and inventory management using various deep learning models, including artificial neural networks (ANNs), long short-term memory (LSTM), support vector regression (SVR), random forest (RF), wavelet-ANN (W-ANN), and wavelet-LSTM (W-LSTM) [49]. Furthermore, demand forecasting solutions using open-source platforms and Azure Machine Learning are under development. In the retail sector, a study compared the effectiveness of traditional statistical models, machine learning-based approaches, and deep neural network-based approaches for demand forecasting [50].

The application of deep learning methods in logistics planning represents a significant advancement in inventory management. A review of the literature shows that deep learning-based inventory management approaches provide greater accuracy and efficiency than traditional methods. In particular, the deep inventory management (DIM) method, which is based on LSTM theory, has been shown to achieve over 80 percent accuracy in inventory demand forecasting [51]. This approach transforms the time series problem into a supervised learning task, allowing the training process to be completed efficiently through backpropagation. Moreover, object detection methods have demonstrated high accuracy in inventory management. Studies in this area show that models such as RetinaNet, YOLO, Faster R-CNN, and Mask R-CNN achieve high accuracy in labeling stock images [52]. These models provide highly effective detection capabilities for the analysis of inventory management systems. Furthermore, deep learning methods play a crucial role in demand forecasting. Studies have demonstrated that recurrent neural networks (RNNs) and attention mechanisms can achieve high levels of accuracy in this task [53]. The implementation of these techniques has the potential to further improve the accuracy of demand forecasting and inventory management systems. Finally, an inventory management optimization framework developed on the basis of deep reinforcement learning (DRL) theory has been shown to achieve greater efficiency and adaptability compared to traditional methods [54]. This framework employs an end-to-end neural network to identify optimal inventory management strategies.

Deep learning methods are receiving increasing attention in logistics planning, particularly in distribution network optimization. Studies in this area employ deep learning techniques to address the limitations of traditional methods, with the aim of developing more efficient solutions. Examples of such applications include the control and optimization of distribution networks using methods based on deep reinforcement learning (DRL). These approaches are designed to address the inherent complexity and uncertainty of distribution networks. Their successful applications have ranged from Volt/Var control to network reconfiguration and restoration [55]. These studies demonstrate that the application of deep learning methods in logistics planning can yield more efficient and effective solutions. However, further research is still required in this area.

3.3. Deep Learning Applications in Logistics Planning of Autonomous Air Cargo Vehicles

3.3.1. Analysis of Existing Studies

One of the basic criteria applied during the review was that each article had to address at least three topics. The articles obtained through the literature review were categorized as shown in

Table 1. Distribution of the scanned studies according to the subjects.

References	Autonomous Vehicles	Drone	Logistics Planning	Distribution Networks	Supply Chain Management	Deep Learning	Machine Learning	Artificial Intelligence
[1,56]			✓		✓			✓
[2,7,14,17]		✓	✓	✓
[8,9,57,58,59]	✓		✓	✓
[10]		✓	✓					✓
[12]		✓	✓				✓
[13,60,61,62,63,64,65,66,67,68,69,70,71]	✓	✓	✓
[16]		✓	✓		✓
[19,72]						✓	✓	✓
[30,34,35,36,37,73]	✓					✓	✓
[42,43]			✓	✓		✓
[44,48,74]			✓			✓	✓
[45]	✓	✓	✓			✓
[46,47]			✓		✓	✓
[49,51,53,75]					✓	✓	✓
[54]					✓	✓		✓
[55]				✓		✓	✓
[76]		✓			✓		✓	✓
[77,78,79,80,81,82,83,84,85,86,87]					✓		✓	✓
[88]	✓		✓		✓			✓
[89]	✓		✓					✓
[90,91]	✓		✓				✓
[92,93]			✓	✓	✓
[94]	✓	✓					✓
[95]			✓		✓		✓
[96]		✓				✓	✓
[97]			✓			✓		✓
[98]	✓		✓		✓
[99]	✓	✓				✓	✓	✓
[100]	✓	✓					✓	✓
[101]		✓				✓		✓

Note: The ✓ mark in the table denotes that the subject specified in the column is covered in the study listed in that row.

.

The topics identified in the literature reviewed for the analysis of existing studies are presented in Figure 4.

One of the most important themes in the reviewed literature is deep learning, which encompasses a wide range of methods. Among these are: deep neural networks (DNNs), consisting of multiple layers that learn complex patterns from data; convolutional neural networks (CNNs), commonly applied to two-dimensional data such as images, performing feature extraction by capturing local connections; deep reinforcement learning (DRL), which integrates deep learning with reinforcement learning to optimize decision-making; recurrent neural networks (RNNs), designed for sequential data such as time series; and long short-term memory (LSTM) networks, a variant of RNNs developed to address long-term dependency problems.

Other approaches include artificial neural networks (ANNs), the most basic model often serving as a building block in AI systems; gated recurrent units (GRUs), similar to LSTMs but with a simpler structure for efficient sequential learning; generative adversarial networks (GANs), where two networks compete to generate new data; deep belief networks (DBNs), probabilistic models composed of multiple layers; and restricted Boltzmann machines (RBMs), the fundamental building block of DBNs. Additionally, transformer models (TRANs), which employ attention mechanisms primarily in natural language processing and sequential data; stacked autoencoders (SAs), which extract features by compressing and reconstructing data; and feedforward neural networks (FWNNs), which pass data forward through layers to learn, are also widely used.

All these methods fall within the domain of deep learning and are capable of automatically extracting and learning features from data. An overview of these techniques is presented in Figure 5, while the deep learning methods employed in the reviewed studies are summarized in

Table 2. Deep learning methods used in the reviewed literature.

References	DNN	CNN	DRL	RNN	LSTM	ANN	GRU	GAN	DBN	RBM	TRAN	SA	FWNN
[1,21,77,84,97,102,103,104]						✓
[3,26,28,36]		✓
[19]	✓	✓				✓
[20]									✓	✓		✓
[23]	✓	✓						✓
[24]								✓			✓
[25,27,31,45,50]	✓	✓
[29]	✓			✓
[33,37,42,43,44,49,54,55,74,96]			✓
[34,51]				✓	✓
[38]	✓	✓		✓	✓
[47]	✓			✓			✓
[48,53]				✓
[79]	✓					✓
[80]	✓												✓
[82]				✓		✓
[73,105]	✓
[100]		✓		✓		✓
[106]	✓											✓
[107]		✓				✓

Note: The ✓ mark in the table denotes that the subject specified in the column is covered in the study listed in that row.

.

Another important component of the reviewed literature is logistics, which is categorized according to thematic topics. Classifying articles by themes allows researchers and practitioners to gain deeper insights into specific domains. This approach aggregates studies in the logistics field around focal areas, thereby offering a clearer depiction of prevailing trends, research gaps, and potential directions for future development.

The field of sustainability and green logistics examines methods for redesigning logistics processes to minimize environmental impacts and improve resource efficiency. Technological innovation and digitalization address the digital transformation of logistics operations and the integration of emerging technologies such as artificial intelligence, the Internet of Things, and blockchain. Supply chain integration focuses on managing the flow of information and goods more effectively while enhancing collaboration across the various stages of the supply chain.

Cost optimization strategies emphasize models and methods designed to reduce expenses and improve efficiency in logistics processes. Quality management and standards play a key role in improving the quality of logistics services, increasing customer satisfaction, and ensuring compliance with international standards. Finally, human resources and training highlight the need to strengthen workforce competencies by identifying training needs and cultivating qualified personnel for the sector.

The classification of logistics literature according to these thematic topics is presented in

Table 3. Categorization of logistically screened studies according to thematic issues.

References	Sustainability and Green Logistics	Technological Innovations and Digitalization	Supply Chain Integration	Cost Optimization	Quality Management and Standards	Human Resources and Training
[2,13,14,53]	✓	✓	✓	✓
[4,86]	✓	✓	✓
[8,62]	✓	✓		✓
[9]	✓	✓	✓	✓	✓
[10,12,74,103]		✓
[16,56,89,97,108]	✓	✓
[42,44,45,51,54,57]		✓		✓
[76]		✓	✓			✓
[75,109]		✓	✓
[80]		✓	✓	✓
[81,110,111]		✓	✓	✓	✓
[85]		✓		✓	✓	✓
[98]		✓		✓	✓
[106]	✓	✓				✓
[112]				✓
[113]						✓
[114]		✓				✓
[115]	✓	✓			✓
[116]		✓	✓	✓	✓	✓

Note: The ✓ mark in the table denotes that the subject specified in the column is covered in the study listed in that row.

.

To provide a unifying perspective, Figure 6 presents a conceptual framework that illustrates the interaction between deep learning methods, autonomous air cargo vehicles, and logistics planning.

3.3.2. Application Areas and Success Examples

A growing body of research has demonstrated the effectiveness of deep learning algorithms in surpassing conventional methods across various domains, including route optimization, cargo loading, air traffic control, and security. In the context of route optimization, deep learning has emerged as a particularly promising solution for AACVs, enhancing delivery efficiency by reducing transit times and improving fuel consumption. A comprehensive review reported that deep learning-based route optimization systems achieved, on average, a 15% improvement over traditional methods [83]. Recent research has also underscored the importance of cooperative task allocation in multi-UAV systems. For instance, a multi-strategy clustering ant colony optimization (KMACO) algorithm was proposed, which significantly enhanced convergence speed and reduced operational costs in complex urban rescue environments [117]. Such approaches highlight the potential of integrating swarm intelligence methods with deep learning-based logistics planning in AACVs. Moving beyond single-vehicle optimization, multi-agent coordination has emerged as a critical research frontier. A comprehensive review on synergistic UAV motion emphasized collective path planning, formation control, and trajectory tracking as essential components for scalable and resilient UAV collaboration [118]. Collectively, these insights demonstrate how swarm intelligence principles can complement deep learning-based logistics planning in AACVs. The potential of deep learning also extends to cargo loading, where algorithmic implementations can enhance transport capacity and reduce delivery times by ensuring more efficient cargo arrangements. A comprehensive literature review reported that a deep learning-based cargo loading system can accommodate, on average, 20% more cargo than conventional methods [65]. Moreover, the potential of deep learning extends to air traffic control, where the implementation of algorithms can improve the effectiveness of air traffic management, thereby reducing delays and enhancing safety. Existing literature indicates that deep learning-based approaches to air traffic control achieve, on average, approximately 10% higher performance compared to conventional methods [119]. Furthermore, deep learning can be applied to enhance the security and monitoring of AACVs. In this context, algorithms can support timely intervention by detecting potential security threats. A literature review reported that deep learning-based security systems detect, on average, 25% more threats compared to traditional methods [120].

3.3.3. Challenges Encountered and Solution Suggestions

A range of challenges must be addressed in applying deep learning to the logistics planning of AACVs. These include technical, regulatory, social, and economic factors. On the technical side, key challenges involve data requirements, computational demands, and explainability. Training deep learning models requires substantial amounts of labeled data, which are both costly and time-consuming to acquire. In addition, the advanced hardware needed for model development is expensive. The opaque nature of deep learning models also raises security and regulatory concerns due to their limited explainability. Promising avenues for addressing these issues include the use of synthetic data generation and augmentation techniques in simulation environments that closely replicate real-world conditions. Moreover, the field of explainable AI has made significant progress, with numerous studies exploring methods to enhance model transparency. Studies have proposed potential solutions such as automating data collection and labeling processes and employing augmentation techniques. Among the regulatory barriers, two key issues stand out: establishing common international rules and standards, and creating legal frameworks for the integration of autonomous vehicles into existing air traffic systems. Addressing these challenges requires stronger international cooperation, standardization efforts, and the development of comprehensive legal regulations. Public confidence in autonomous aerial vehicles is also a critical factor for achieving social acceptance and security. Therefore, raising public awareness of the reliability and safety of these vehicles is essential. Enhancing public confidence is also essential for safeguarding communication and data systems against cyber threats. Potential solutions include raising public awareness of autonomous aerial vehicles, developing training programs, and implementing advanced security measures.

Economic challenges are primarily linked to high investment costs, profitability, and efficiency. The development and deployment of AACVs require substantial financial resources. To address these challenges, it is necessary to explore investment and financing opportunities and to develop appropriate business models. Furthermore, integrating deep learning into the logistics planning of autonomous air cargo vehicles can help identify and resolve these issues, thereby ensuring the system’s operational effectiveness and safety.

4. Discussion

The synthesis of findings from the reviewed literature highlights both the opportunities and challenges of applying deep learning (DL) to logistics planning for AACVs. Prior studies demonstrate potential benefits in areas such as route optimization, demand forecasting, and inventory management. However, the overall evidence base remains heterogeneous, with substantial variation in both outcomes and methodological rigor.

A primary area of concern is energy efficiency in DRL-based routing approaches. While several studies report substantial improvements in fuel and energy consumption compared to conventional heuristics, others find only marginal or inconsistent gains, often influenced by network scale and environmental conditions. These inconsistencies underscore the need for further research into the performance of DRL algorithms under real-world uncertainties, including weather variability, fluctuating energy costs, and heterogeneous fleet compositions.

A second issue concerns data limitations. Although deep learning methods have demonstrated strong performance in demand forecasting and cargo loading optimization, many studies rely on small-scale or synthetic datasets. In a related field, a comprehensive survey on UAV-based data collection in wireless sensor networks highlighted challenges such as trajectory planning, energy consumption, and scalability [121]. These findings underscore parallel challenges of data availability and efficiency that are highly relevant to AACV logistics planning. Such reliance raises concerns about the generalizability of results. Moreover, the lack of open-access benchmark datasets limits reproducibility and hinders meaningful cross-study comparisons. Consequently, developing and sharing AACV-specific datasets that capture diverse operating conditions and long-term performance should be considered a critical research priority.

Third, technical, regulatory, and social aspects warrant more comprehensive examination. Technically, CNNs and DRL algorithms have achieved high accuracy in perception and planning tasks, yet their lack of transparency complicates safety certification processes. The limited exploration of explainable AI in this context highlights a significant research gap. From a regulatory perspective, although the role of agencies such as the FAA and EASA is frequently acknowledged, concrete strategies for integrating AACVs into existing air traffic management systems remain underdeveloped. Addressing these challenges will require cross-disciplinary collaboration among AI specialists, aerospace engineers, and policymakers.

Social acceptance also remains an underexplored dimension. Only a limited number of studies have examined how factors such as trust, privacy, and perceived safety shape public attitudes toward AACVs. Systematic empirical research—through surveys, experiments, or case studies—could yield valuable insights into user perceptions and help identify strategies to strengthen societal trust in autonomous logistics systems.

Finally, although the literature acknowledges the substantial economic barriers to AACV deployment, concrete business models remain scarce. Research that systematically evaluates cost–benefit trade-offs across different logistics contexts—such as e-commerce, humanitarian operations, and medical supply chains—would be particularly valuable. Comparative case studies and techno-economic simulations could further help identify circumstances in which AACVs provide clear advantages over traditional transport modes.

In conclusion, while DL-driven AACV applications show considerable promise, they remain largely confined to experimental settings. Future research should focus on:

• Assessing the robustness of DRL algorithms under dynamic and uncertain conditions.
• Creating and disseminating benchmark datasets for AACV logistics.
• Incorporating explainable AI approaches to support regulatory approval.
• Investigating public acceptance through systematic empirical studies.
• Developing sustainable business models and conducting comparative cost analyses.

By moving beyond descriptive summaries toward critical evaluations and actionable research directions, scholars can provide stronger guidance to industry practitioners and policymakers in realizing the full potential of AACVs in logistics planning.