Unlocking the Ocean 6G: A Review of Path-Planning Techniques for Maritime Data Harvesting Assisted by Autonomous Marine Vehicles

Abstract

Seamless integration of both terrestrial and non-terrestrial networks is crucial to providing full-dimensional wireless and ubiquitous coverage, particularly catering to those engaged in marine activities. Compared to terrestrial networks, wireless communications in the marine domain are still not satisfactory for ubiquitous connectivity. Featuring agile maneuverability and strong adaptive capability, autonomous marine vehicles (AMVs) play a pivotal role in enhancing communication coverage by relaying or collecting data. However, path planning for maritime data harvesting is one of the most critical issues to enhance transmission efficiency while ensuring safe sailing for AMVs; yet it has rarely been discussed under this context. This paper provides a comprehensive and holistic overview of path-planning techniques custom-tailored for the purpose of maritime data collection. Specifically, we commence with a general portrayal of fundamental models, including system architectures, problem formulations, objective functions, and associated constraints. Subsequently, we summarize the various algorithms, methodologies, platforms, tools, coding environments, and their practical implementations for addressing these models. Furthermore, we delve into the burgeoning applications of path planning in the realm of maritime data harvesting and illuminate potential avenues for upcoming research endeavors. We believe that future research may focus on developing techniques to adapt more intricate and uncertain scenarios, such as sensor failures, inaccurate state estimations, complete modeling of communication channels, ocean dynamics, and application of heterogeneous systems.

1. Introduction

1.1. Background

Facilitated by advanced mobile broadband technology, recent developments in massive machine-type communications and reliable low-latency communications have propelled the progression towards what is commonly referred to as the fifth generation (5G) of wireless systems, officially designated as International Mobile Telecommunications 2020 [1]. However, the continuous deployment of 5G cellular systems is progressively revealing the inherent limitations of this technology, in contrast to its initial vision as a catalyst for Internet of Everything applications. These drawbacks associated with 5G are catalyzing global efforts aimed at defining the forthcoming 6G wireless system, capable of seamlessly integrating an extensive array of applications, ranging from autonomous systems to extended reality [2]. To meet the 6G prerequisites for comprehensive coverage and pervasive connectivity, it is imperative to devise a large-scale network that connects terrestrial networks and non-terrestrial networks, including four distinct tiers: space, aerial, terrestrial, and oceanic networks [3].

As an extension of terrestrial networks, oceanic networks establish connections among marine objects equipped with sensing, communication, and computing capabilities [4]. They are deemed as promising tools to enhance the efficiency of marine activities for humans, giving rise to the concept of Smart Oceans [5]. Within oceanic networks, the Underwater Wireless Sensor Network (UWSN) stands out as an intelligent network with self-learning and computational abilities. It is part of the networks through which “things” in the earth’s water bodies (oceans/sea, rivers, lakes, and streams) are digitally connected. It comprises a large number of sensor nodes, including anchored nodes, mobile nodes, and surface sinks, each designed for environmental sensing, data collection, and transmission. However, when considering the unique characteristics of the oceanic environment, Smart Oceans encounter several significant challenges. The underwater sensor nodes situated on the seabed are typically reliant on batteries, and this reliance has a substantial impact on their ability to sustain data collection operations. The energy consumption associated with transmitting data underwater through acoustic waves exceeds the computational overheads, and recharging these batteries is impractical due to the constraints imposed by seawater environments. Hence, there is a compelling need to devise methods to extend the operational life of UWSNs. Furthermore, as evidenced by research outlined in [6], underwater acoustic transmission is limited by prolonged propagation delays and a high probability of packet loss, resulting in diminished communication efficiency. Consequently, there remains unresolved challenges of achieving dependable maritime communication as we venture towards 6G technology in ocean realm.

The deployment of connected autonomous marine vehicles (AMVs) is one of the primary drivers for the Smart Ocean due to their strong flexibility and mobility [7,8], including autonomous underwater vehicles (AUVs) and unmanned surface vehicles (USVs) [9]. On the one hand, AUVs offer a significant advantage in underwater data collection for UWSNs due to their ability to serve as underwater mobile agents. This mobility allows for dynamic data collection in response to changing environmental conditions, enabling more comprehensive and timely data acquisition. AUVs also address the energy efficiency challenges faced by UWSNs, as they can incorporate energy-harvesting mechanisms or rechargeable batteries to extend their operational lifetimes. Moreover, the utilization of USVs for achieving fast maritime wireless communications is being recognized as a promising contributor to future maritime wireless communication systems. This approach presents a viable solution for establishing regenerative maritime non-terrestrial networks. USVs offer the capability to deliver fast wireless communication services to maritime devices without the need for traditional infrastructure coverage or specialized high-gain satellite antennas. In comparison to terrestrial and satellite-based systems, USV-assisted maritime wireless communication have several merits, including cost-effectiveness, robust adaptability to varying environments, support for data relaying and collection, as well as the integration of mobile edge computing capabilities.

Remarkably, to deploy AMVs in UWSN, the most challenging issue is to plan the cruising path for each AMV, such that a data harvesting task can be accomplished effectively. Feasible path design holds the potential to enhance transmission performance while ensuring the safety of navigation. Typically, path planning can be conceptualized as an optimization problem with various real-world constraints, including quality of service (QoS), obstacle avoidance, and geographical limitations. Notably, energy efficiency is also a central concern of AMV path planning, necessitating the consideration of factors like surrounding ocean currents or wind patterns to reduce energy consumption. Nevertheless, practical challenges arise due to the unpredictability of meteorological conditions and the inherent uncertainty in weather forecasts, rendering conventional models, such as A* [10] or fast marching [11] ineffective when confronted with the dynamic maritime environment. Furthermore, the intricate communication channels between AMVs and sensor nodes pose substantial hurdles in the realm of path planning. Our previous research [12] has underscored the profound impact of the distance between USVs and sensors on the signal-to-noise ratio (SNR), a pivotal determinant of data transmission efficacy throughout missions. This mechanism is important in achieving efficient data collection, yet it has rarely been considered in path planning. Moreover, the choice of communication protocols can also have a profound impact. For instance, the communication range should be incorporated into the algorithm when planning routes. The vehicle may need to adjust its path to stay within the communication range if the range is insufficient for data transmission. However, these factors are rarely considered by existing studies. Consequently, path planning for maritime wireless communication remains an open and demanding issue, warranting continued research and attention.

1.2. Survey of Surveys

In the domain of path planning for AMVs, there has been a large number of review articles. In accordance with the constraints associated with the USV path-planning problem, the work in [13] summarizes the problem models and the corresponding solutions under various modality constraints. In accordance with the extent of environmental awareness, authors in [14,15] systematically examine the distinguishing features of problem models for global/local path planning and collision avoidance, alongside representative algorithms. In addressing the issue of collision avoidance for USVs, the work in [16,17] provides a comprehensive overview of the regulations established by the International Maritime Organization (IMO) pertaining to USVs. Furthermore, it systematically discusses collision-avoidance navigation technologies, spanning from academic to industrial perspectives. In consideration of the varying types of AUV missions, work in [18] conducts distinct discussions on the problem models and corresponding algorithmic facets concerning coordinated path planning, cooperative path planning, and formation control. Similarly, work in [19] extensively discusses the issues related to task allocation and path planning for multi-AUVs in underwater search for multiple targets.

However, it is clear that despite researchers discussing path planning from different perspectives, such as the availability of environmental knowledge, the number of AMVs, and algorithm types, previous review articles have predominantly portrayed path planning as a point-to-point planning model. This concept has been inherited from traditional robotics and has been studied in the academic community for nearly 60 years, with its formal definition found in [20]. However, the applications of modern marine robots are no longer limited to simple point-to-point movement tasks. For instance, surveying tasks require coverage of an area, sampling/collecting tasks involve accessing multiple locations, and search tasks need to consider the probability of the object’s appearance (see Figure 1). This indicates that different types of applications pose unique requirements and constraints for path planning that traditional models did not consider, and this is what we define as mission-oriented path planning (MOPP). The wireless data harvesting discussed in this paper also poses its unique requirements for the path-planning model. For instance, AMVs need to simultaneously perform data collection tasks during motion, involving multiple points of access. Moreover, AMVs need to consider underwater communication mechanisms and energy consumption. These are the factors not accounted for in traditional path-planning models. It is worth highlighting the crucial role that path planning plays within maritime wireless communication systems, where it exerts a profound influence on both communication quality and task efficiency. Nevertheless, despite its evident significance, there is no comprehensive summary or review reported on this unique group of path planning.

Motivated by the foregoing discussion, this paper presents the pioneering work on reviewing the application of path-planning techniques in AMV-assisted wireless communication. The logical structure and research questions we investigated are shown in Figure 2. The outline of the salient contributions can be summarized as follows:

• This is the first review article on a holistic vision of the path-planning techniques for AMV-assisted oceanic data harvesting missions.
• An extensive search of existing literature has been undertaken, enabling us to present a general portrayal of the current state of research within the academic community.
• Our review encompasses detailed insights into system architectures, problem formulations, objective functions, and the pertinent constraints associated with data harvesting in the marine domain.
• A thorough investigation and consolidation of various algorithms, methodologies, platforms, tools, coding environments, and their practical implementations are meticulously presented, providing a comprehensive knowledge resource for readers.
• We introduce and explore the emerging challenges of path planning techniques in the context of oceanic data harvesting. Additionally, we engage in a forward-looking discussion concerning potential trends for future research.

The remainder of this article is organized as follows. The fundamental models and literature search are presented in Section 2. Section 3 conducts a comprehensive review of the path-planning methods and algorithms. The challenges and open issues are presented in Section 4. Section 5 concludes this research.

2. Fundamental Models and Literature Search

In the content of this section, we will present some fundamental models for data collection along with their characteristics. Concurrently, we conducted a literature search, providing a macroscopic overview and summary of existing research.

2.1. System Model

2.1.1. Distributed Network

The distributed network comprises individual sensor nodes that are positioned across the water, whether it be in water bodies or seabed, as depicted in Figure 3. To facilitate the acquisition of data from these sensor nodes, the AMVs need to establish connections through acoustic waves with all the nodes while remaining within their communication range. Consequently, this approach has the potential to decrease energy consumption in UWSNs, as a substantial portion of operational tasks can be offloaded onto the AMVs. Nonetheless, the deployment of distributed networks presents certain challenges. On the one hand, the relatively slow movement of AMVs leads to extended travel times, resulting in increased data collection delays and a reduction in network throughput. On the other hand, the limited power capacity of AMVs places constraints on the overall efficiency of the data collection process. The necessity for AMVs to access all sensor nodes makes it challenging to accomplish their mission without supplementary recharging methods.

2.1.2. Collection with Fixed Path

The second approach involves the AMV following a predetermined trajectory at a consistent velocity, during which it gathers data from the sensor nodes, see Figure 4. The AMV’s task includes completing this route within a user-defined timeframe, after which it returns to a central control station to transmit the accumulated data. In contradistinction to the strategy involving a distributed network, the fixed-path collection method exhibits greater simplicity and efficiency. This characterization is due to the confinement of the AMV’s locomotion to a predefined path, executed within the confines of prescribed power parameters. Nevertheless, the effectiveness of these approaches hinges significantly on the selection of the AMV’s path. A fixed AMV path greatly limits the flexibility of the mission, as it may not be adjustable to dynamic or adaptive environmental conditions. This is particularly important in scenarios where the underwater terrain is subject to variability, necessitating real-time adjustments to the data collection trajectory for optimized coverage and efficient resource utilization.

2.1.3. Clustering or Layering Network

The final approach divides the network nodes into distinct clusters, thereby empowering the AMV to exclusively gather data from cluster heads, see Figure 5. This innovative scheme yields a substantial reduction in the travel time of the AMV, contributing to enhanced operational efficiency and data collection speed. However, the effectiveness of this method is intricately tied to the clustering or layering algorithms. Cutting-edge algorithms capable of efficiently grouping nodes based on diverse criteria, such as spatial position, communication patterns, or resource utilization, can foster seamless coordination among cluster members and facilitate data exchange. Conversely, if the chosen algorithms are not meticulously designed or optimized, network performance may suffer. Suboptimal clustering may lead to imbalanced data distribution, increased communication overhead, or inefficient travel routes.

2.2. Planning Model

In oceanic data collection, it is commonly observed that, owing to the a priori knowledge of sensor node or cluster head locations, the prevalent path planning models involves the utilization of integer programming models. Among these, the Traveling Salesman Problem (TSP) model stands out as the most extensively employed. However, unlike typical combinatorial optimization problems, the process of data transmission involves highly intricate mechanisms. These complexities include diverse factors, notably communication range, communication delay, noise perturbations, AMV and sensor energy consumption, and value of information (VoI). Consequently, it becomes evident that models for oceanic data collection are much more complex than TSP in nature. In this section, we compile various optimization models mentioned in existing literature, introducing their established optimization objectives and constraints. We aim for these contents to provide readers with a general portrayal of some representative knowledge.

2.2.1. Objectives

• Travel length [21,22,23,24]: The consideration of travel length holds paramount significance in path planning for oceanic data collection. By strategically planning routes that minimize travel distance, various operational advantages are attainable. Firstly, reduced travel distances result in lower energy consumption for vehicles or sensor nodes, thereby prolonging their operational lifespan. Moreover, shorter travel distances enable faster data collection and response times, crucial for real-time monitoring and decision-making in applications such as environmental monitoring or disaster response.
• Energy consumption of nodes [25,26,27]: Energy consumption of the network consists of three parts, the transmission energy $E_t$, the receiving energy $E_r$, and the idle energy $E_{idle}$. $E_t$ refers to the energy expended when nodes send data packets to each other. The amount of transmission energy used is influenced by factors such as the frequency at which data are transmitted and the distance between the devices. $E_r$ refers to the energy used by nodes when they receive data packets from other nodes. The energy consumption for receiving depends on factors like the data rate, modulation schemes, and processing complexity. $E_{idle}$ represents the energy consumed by nodes when they are not actively transmitting or receiving data packets. Even when nodes are not actively sending or receiving, they still use a certain amount of energy to maintain their operational state, listen for incoming messages, and stay connected to the network. This idle energy consumption is often referred to as “quiescent” or “standby” energy. For a more detailed model explanation, readers may refer to [28,29]. Moreover, some scholars have expanded the energy consumption to network lifetime [30,31]. This refers to the duration until the first node becomes non-operational. Consequently, the node exhibiting the most substantial energy consumption emerges as the pivotal factor influencing this lifespan.
• Energy consumption utility [32]: The energy consumption utility parameter, denoted as $\beta$, encapsulates the balance between energy consumption ($E$) and the cumulative network throughput ($I$). The quantification of $E$ denotes the energy consumption of the sensor network, while the network throughput hinges on the magnitude of data packet transmission during the operational sequence. Fundamentally, $\beta$ is calculated as the ratio of network throughput $I$ to energy consumption $E$. By encapsulating this trade-off, $\beta$ provides a comprehensive measure of the network’s efficiency in converting energy inputs into valuable data outputs. Rather than solely focusing on minimizing energy usage, optimizing the energy consumption utility leads to a greater proportion of energy being allocated to data transmission, ultimately enhancing overall efficiency.
• Data collection time/Collection delay [26,33,34]: Collection delay is one of the most important optimization terms. In contrast to electromagnetic waves, the conveyance of underwater signals predominantly relies upon acoustic waves. However, the velocity of sound waves in an underwater environment is constrained to 1500 m/s. In order to ensure the quality of online data, the AMV needs to complete data collection tasks within a shorter timeframe. Evidently, the most straightforward approach involves increasing the AMV’s velocity. Nevertheless, this method is hindered by the limitations imposed by water flow resistance and hardware capacities, and achieving such improvement could be challenging. Therefore, the most effective way to reduce data collection time is by optimizing the path length. Generally, data collection time comprises two components. The first component is the travel time of the AMV, while the second component involves the time slot allocation for time-division multiple access (TDMA), which constitutes the total residence time of the AMV at the cluster head nodes.
• Energy consumption of AMV [35,36,37]: The energy consumed by the AMV is made up of three main components: hotel energy, propulsion energy, and maneuvering energy. The first one refers to the energy consumed by the AMV for all its non-movement-related activities or systems. These activities could include maintaining the necessary onboard systems, sensors, communication devices, and other equipment that are essential for the AMV’s operation but are not directly related to its movement along the path. The propulsion energy is required for the AMV to move along the designated path segment. It involves motors, propellers, or any other mechanisms that generate thrust and propel the AMV forward through the water. Furthermore, the work in [35] has demonstrated that the maneuvering of the AMV within the turning line segment is considerably influenced by the intricate ocean environment. The process of maneuvering encompasses alterations in direction, depth, or orientation, which are necessitated by obstacle avoidance, bearing currents, and other variables affecting the trajectory of the AMV. This dynamic can potentially result in a significant rise in both travel time and energy consumption. Therefore, the significance of maneuvering energy cannot be understated either.
• Value of information (VoI) [38,39,40]: The VoI is used to reflect the importance of data and its sensitivity to time. Many studies drive an AMV to select the nearest sensor (or data collector) based on geographic proximity for its next visit. However, this geographical closeness does not inherently capture the VoI related to physical parameters registered by sensors. These parameters might significantly differ concerning event significance and timeliness [41]. Consequently, an optimal path determined solely by geographic closeness might yield limited value, making a negligible contribution to the overall accumulation of valuable data. Some applications exhibit time-sensitive characteristics, implying that VoI peaks when sensor data are captured, gradually diminishing as time elapses until the data reach the base station. Delayed data transmission leads to reduced VoI at the base station, hindering the effective analysis and response to monitoring necessities.

2.2.2. Constraints

The constraints in the marine data collection path-planning problem arise from a combination of factors, with the task model’s requirements playing a significant role. To begin with, the problem must adhere to constraints related to the starting and ending locations of the data collection mission [32,42]. Moreover, there are constraints on how frequently each node can be visited during the path-planning process [26,35]. Additionally, a constraint governs the total number of nodes that can be visited along the path, with certain literature suggesting that nodes with lower data value might be excluded from the visitation plan, further optimizing the data collection process [43].

The physical characteristics of sensors include constraints that must be accounted for in the path-planning problem. Sensor communication range determines the distance over which sensors can effectively transmit and receive data [32]. This range influences the layout of the path and the arrangement of nodes to ensure adequate coverage. Also, there are constraints on both the maximum energy that can be stored in each sensor node and the energy required to transmit or receive a data packet [41]. These energy-related limitations impact the lifetime and operational capabilities of the sensors. Moreover, signal-to-interference-plus-noise ratios (SINRs) and outage probabilities, as highlighted in the work [34], play a critical role in maintaining reliable communication within the sensor network.

AMVs’ physical system poses distinct constraints to the path-planning problem. AMVs are subject to constraints in energy supply [44], which directly affects their operational duration and the distance they can cover during the data collection mission. Communication distance is a vital constraint that dictates the range over which AMVs can effectively communicate with other nodes or the central control station [45]. Also, the constraint on maximum diving depth and the number of permissible turning points was introduced in [34,35], respectively. The constraint of energy consumption during data packet transmission or reception is also considered [41]. Finally, behavioral constraints specific to AMVs [39] add complexity, requiring careful consideration of how AUVs interact with and approach the nodes during the data collection process.

2.3. Literature Search

We performed an extensive search of the existing literature utilizing the Scopus (https://www.scopus.com/, accessed on 5 August 2023) academic search tool. The search was executed using the prescribed topic framework and keywords outlined in Equation (1). The symbol ““ was employed to ensure precise term matching. The Scopus platform was instrumental in retrieving relevant papers on this topic across diverse fields of study, encompassing ocean engineering, computer science, automation and control, and electrical engineering. This literature search was carried out in August 2023. By employing the search structure and keywords enumerated in Equation (1), a total of 85 articles pertaining to this subject were identified and subsequently evaluated. Importantly, it should be emphasized that no filtering or exclusion of papers relevant to this section was undertaken.

TITLE-ABS-KEY (“USV” OR “ASV” OR “AUV” OR “unmanned surface vehicle” OR “autonomous underwater vehicle” OR “autonomous surface vehicle”) AND TITLE-ABS-KEY (“data collection” OR “data harvesting” OR “data gathering” OR “data acquisition”) AND TITLE-ABS-KEY (“path planning” OR “trajectory planning” OR “route planning”)

(1)

Figure 6 depicts the categorization of 85 research documents along with a visual representation of keyword cloud. As indicated in Figure 6, these documents encompass a diverse spectrum of 15 distinct categories, underscoring the extensive purview of knowledge in oceanic data collection. It encompasses fields such as engineering, computer science, mathematics, geology, and environmental science, among others. This body of work reflects a comprehensive multidisciplinary engagement. Notably, the Engineering and Computer Science categories constitute the majority, accounting for 35.5% and 33.1%, respectively. Following is the mathematics category, representing 7.8%. This observation underscores the pivotal role of knowledge derived from the domains of engineering and computer science in shaping the field of oceanic data collection.

Figure 7 illustrates the distribution of these literature sources among journals and conferences, as well as the variation in the number of publications over the years. From the graph, it can be observed that there is a general upward trend in the quantity of literature on this topic. Compared to the publications before 2020, the number of papers in the last three years has notably increased. We attribute this phenomenon to the rapid advancements in wireless communication and marine robotics, which have stimulated a growth in the number of papers related to ocean data collection. We anticipate that this trend will continue to expand in the future. Furthermore, the pie chart reveals that a significant portion of these publications are featured in IEEE journals (28.6%) and conferences (16.7%), closely followed by journals (15.5%) pertaining to wireless communication. Additionally, some other literature is published in journals focusing on marine engineering, robotics, and sensor domains. This highlights the predominant advantage of IEEE publications in the realm of wireless data transmission research, followed by certain journals specializing in ocean engineering and robotics.

Besides considering the quantity of published articles, another significant measure for assessing the quality and influence of these publications is through citation analysis.

Table 1. Identification of highly cited articles.

Rank	References	Pub. Year	No. of Citations	Journal Title
1	[46]	2010	160	International Journal of Robotics Research
2	[47]	2011	158	Journal of Field Robotics
3	[48]	2012	156	IEEE Journal on Selected Areas in Communications
4	[49]	2013	115	Journal of Field Robotics
5	[41]	2018	100	IEEE Systems Journal
6	[32]	2020	76	IEEE Internet of Things Journal
7	[50]	2010	58	2010 IEEE International Conference on Robotics and Automation (ICRA)
8	[24]	2019	42	Journal of Network and Computer Applications
9	[51]	2020	34	IEEE Internet of Things Journal
10	[35]	2018	34	IEEE Sensors Journal

showcases the top 10 most extensively referenced research works in this specific domain. Among these, the most cited piece is a study from 2010, which employs predictive ocean models to devise path planning for AUVs, facilitating underwater data collection missions. This research appeared in the International Journal of Robotics Research, and its citation frequency on Scopus has reached 160 times. Moreover, we also noted that a substantial portion of the highly cited papers originates from the research group headed by Professor Ryan N. Smith and Professor Arvind Pereira at the University of Southern California. This underscores the far-reaching impact of their pioneering contributions to the field of marine data collection during its early stages.

Figure 8 shows the national distribution of the research institutions to which these references belong. From the figure, it can be observed that research teams from China constitute the majority (37.8%), followed by the United States at 25.2%. In third place is Canada, accounting for 5.4%. This indicates that China and the United States are leading in research and investment in maritime data collection compared to other countries worldwide. However, it is worth noting that a larger number of published papers does not necessarily indicate stronger research capabilities. The above is just the author’s interpretation of the graphical data.

3. Review of Path-Planning Techniques for Data Harvesting

The path-planning algorithms employed in ocean data harvesting diverge significantly from conventional problems. As depicted in Figure 9, the distribution of the path-planning methodologies is presented. In the process of data collection, AMVs are required to access multiple sensor nodes while concurrently optimizing data transmission efficiency. As a result, conventional point-to-point path-planning methodologies, typified by heuristic paradigms such as A* [52,53], D* [54,55], and theta* [56,57,58], as well as sampling-based techniques including Rapidly-exploring Random Trees (RRT*) [59,60,61] and Probabilistic Roadmaps (PRMs) [62,63], show a relative scarcity of application within this field. Contrastingly, meta-heuristic techniques, such as genetic algorithms (GA) [64,65], ant colony optimization (ACO) [66,67], particle swarm optimization (PSO) [68,69], simulated annealing (SA) [70,71], alongside AI-driven approaches exemplified by Q-learning and deep reinforcement learning, exhibit a remarkable ability to accommodate the intricate nature of the problem formulation. Notably, these methodologies not only present a capacity to optimize diverse sets of objective functions but also show an adeptness at navigating the complex constraints. Therefore, they find extensive use in ocean data collection. Additionally, some researchers have also explored the application of online path planning and coverage path-planning techniques for data collection, yielding promising results. In this section, we will thoroughly review these novel approaches, discuss their application scenarios, and evaluate their strengths and limitations.

A summary of the existing research is presented in

Table 2. Summary of existing research.

References	Platforms	Method Category	Method	Comments	Environment	Implementation
[72]	AUV	AI-based method	Deep reinforcement learning	Integrating both acoustic and optical technologies	Python 3.8	Simulation
[73]	Multiple AUVs	AI-based method	Deep reinforcement learning	Data collection with multiple complex tasks simultaneously	Python	Simulation
[74]	USV	AI-based method	Deep reinforcement learning	Stochastic arrived data and the random emergence and movement of dynamic obstacles	Python	Simulation
[75]	USV	AI-based method	Deep reinforcement learning	Arrived data and dynamic obstacles are stochastic	Python	Simulation
[76]	AUV	AI-based method	Q-learning	Improve energy efficiency of sensor nodes	Not mentioning	Simulation
[36]	Multiple AUVs	AI-based method	Q-learning	Balance the energy consumption of a network and extend the network’s lifetime	MATLAB	Simulation
[12]	Multiple USVs	AI-based method	Reinforcement learning	In fully unknown environments, and integrating communication channels	MATLAB	Simulation
[77]	AUV	AI-based method	Reinforcement learning	Dynamic underwater conditions comprising obstacles and small drifts	MATLAB	Simulation
[24]	AUV	Coverage planning	Lawn mower path	AUV is responsible for nominating the Cluster Head	Not mentioning	Simulation
[78]	AUV	Coverage planning	On-the-way fair data freshness	Improved the overall data freshness at the cost of data collection delay	Not mentioning	Simulation
[35]	AUV	Coverage planning	Lawn mower pattern path-planning	Support the long-range AUVs and the uniform energy consumption of sensor nodes	MATLAB	Simulation
[79]	AUV	Coverage planning	Lawn mower pattern path-planning	Maximize the duration of the cooperative works	MATLAB	Simulation
[47]	AUV	Coverage planning	Tranquill ocean path planner	Maximize the information value while minimizing deviation due to ocean currents	/	Experiment
[80]	AUV	Exact approach	Iterative local search	Maximize the completeness of the data collected	Not mentioning	Simulation
[81]	Multiple AUVs	Exact approach	Dynamic programming	Optimize safety, energy consumption, and cooperation	Not mentioning	Simulation
[48]	AUV	Exact approach	Concorde	Robust to channel variations and interference	C++	Simulation
[82]	AUV	Exact approach	Concorde	Communicating with multiple nodes at once	C++	Simulation
[83]	AUV	Heuristic method	Greedy algorithm	Consider the energy consumption of an AUV and value of information (VoI)	Not mentioning	Simulation
[84]	Multiple AUVs	Heuristic method	A* Energy	Reduce the energy consumption of multi-AUV data acquisition systems	MATLAB	Simulation
[85]	AUV	Heuristic method	A*	With predictions of ocean currents	Not mentioning	Simulation
[49]	AUV	Heuristic method	Minimum expected risk planner	Consider uncertainty in the ocean current predictions	\	Experiment
[86]	AUV	Heuristic method	A*	With predictions of ocean currents	Not mentioning	Simulation
[87]	AUV-USV	Meta-heuristic	Shuffled frog-leaping algorithm	Optimize the interaction schedule between USV and multiple AUVs	Not mentioning	Simulation
[22]	AUV	Meta-heuristic	Grey wolf optimizer (GWO)	Validated on the real-world dataset (Canadian NEPTUNE [88])	MATLAB	Simulation
[21]	Multiple USVs	Meta-heuristic	Improved partheno-genetic algorithm	Path length and the workload balance between USVs are considered	MATLAB	Simulation
[89]	Multiple AUVs	Meta-heuristic	Backtracking search optimization	High coverage ratio and low delay	Castalia	Simulation
[90]	Multiple USVs	Meta-heuristic	Immune algorithm	Include obstacle avoidance	MATLAB	Simulation
[91]	AUV	Meta-heuristic	Firefly algorithm	Consider the communication range of sensor nodes and the kinematics of AUVs	Not mentioning	Simulation
[92]	AUV	Meta-heuristic	Genetic algorithm	Reduce energy consumption with mobility consideration	Not mentioning	Simulation
[93]	AUV	Meta-heuristic	Ant colony optimization	Consider the angle control in AUV operation	Not mentioning	Simulation
[32]	AUV	Meta-heuristic	Ant colony optimization	Introduce the energy-consumption utility	MATLAB	Simulation
[94]	AUV	Meta-heuristic	Genetic algorithm	Overcome the kinematic nonholonomic constraints	MATLAB	Simulation
[95]	AUV	Meta-heuristic	Genetic algorithm	Consider non-holonomic constraints	Not mentioning	Simulation
[96]	AUV	Meta-heuristic	Genetic algorithm	Capture the time-sensitive nature of collected information	Not mentioning	Simulation
[97]	AUV	Meta-heuristic	Genetic algorithm	Good stability and convergence	Not mentioning	Simulation
[98]	AUV	Online-planning method	Grouping-based dynamic trajectory planning	Feasible for sensors with no position information	Not mentioning	Simulation
[99]	AUV	Online-planning method	Dynamic trajectory planning	Reduce energy consumption	Not mentioning	Simulation
[41]	AUV	Online-planning method	Dynamic value-based path planning	Balance energy consumption and prolong the underwater network lifetime	MATLAB	Simulation
[100]	Multiple AUVs	Online-planning method	Dynamic path planning	Dynamically choose the next node with recent network status information	Python 2.7	Simulation
[26]	AUV	Other	Joint optimized data collection algorithm	Consider residual energy of nodes and VoI	MATLAB	Simulation
[101]	AUV	Other	Markov decision processes	Delay-aware and energy-efficient	MATLAB	Simulation
[102]	UAV	Other	Fermat point algorithm	A vehicle routing problem (VRP) with pickup	Not mentioning	Simulation
[103]	AUV	Other	Spatial geometric-based	Reduce data collection delays and prolong network lifetime	MATLAB	Simulation
[104]	AUV	Other	Feedback planning	Consider ocean currents and localization uncertainty	Not mentioning	Simulation
[105]	AUV	Other	Artificial potential field (APF)	Obstacle avoidance with energy-saving	Not mentioning	Experiment
[51]	Multiple AUVs	Other	Artificial potential field (APF)	Improve the flexibility and controllability of the AUV flock-based networks	MATLAB	Simulation
[106]	AUV	Other	Extended Euler circuit	Minimize the average data reporting delay through resurfacing	Not mentioning	Simulation
[50]	AUV	Other	Centroid-tracking, waypoint selection algorithm	Able to track a dynamically evolving ocean feature	\	Experiment
[46]	Multiple AUVs	Other	Centroid-tracking, waypoint selection algorithm	Extend [50] by incorporation of multiple vehicles	\	Experiment
[107]	AUV	Other	Sequential quadratic programming	Takes motion constraints into consideration	Not mentioning	Simulation
[108]	AUV	Sampling-based method	Probabilistic roadmap	Inspect multiple goals while maintaining communication with the USV	Not mentioning	Simulation
[109]	AUV	Sampling-based method	Probabilistic roadmap	Consider the time duration limits	Not mentioning	Simulation

.

3.1. Meta-Heuristic Methods

Meta-heuristic methods draw inspiration from natural and social phenomena, often imitating the behavior of biological process or simulating collective behaviors observed in nature. These methods, which include GA, ACO, PSO, SA, and many others, operate by iterative refining and evolving potential solutions over generations or iterations. By integrating randomness and probabilistic components, they navigate the solution space in a manner that balances the exploration and exploitation. Due to their ability to handle complex, multi-dimensional, and non-convex optimization problems, meta-heuristic methods find extensive application in diverse domains such as engineering [110], economics [111], and, as mentioned earlier, in path planning for intricate tasks like ocean data collection.

Due to the inherent capacity to manage intricate problems, there has been a growing trend towards the utilization of meta-heuristic algorithms in oceanic data collection. In support of the trajectory planning procedure, a genetic algorithm-driven approach was formulated in the work [96] to determine the optimal resurfacing schedule. The optimization criteria employed include the value of information (VoI), which captures the time-critical nature of gathered data. Addressing the non-holonomic restrictions of the AUVs, the study presented in [95] introduced a 3D path-planning model based on Dubins curves. This model, solved using a genetic algorithm, proposes a continuous path configuration that eliminates abrupt direction changes at sensor nodes. Subsequently, building upon the concepts in [95], the authors of [94] expanded this research by incorporating considerations for energy efficiency. This was achieved by introducing continuous Bezier curves in the Z-axis, building upon the 2D Dubins curves framework. In order to achieve enhanced energy efficiency while accounting for mobility factors, the study conducted in [92] introduces a dynamic clustering algorithm independent of location and a genetic-driven methodology. In comparison to conventional approaches, the solution results in a network lifetime improvement of twofold, a nearly 30% reduction in total energy consumption, and a 15% increase in delivery ratio under conditions of low densities. In the context of energy management within nodes and angle control, the research in [93] presents an energy-conscious clustering protocol and a novel ACO algorithm that integrates with a Markov Reward Process (R-ACO). This integration is employed to optimize both distance and angle considerations in path planning. Addressing the dual goals of balancing energy consumption and network throughput, the authors of [32] divided the optimization challenge into four distinct segments. To find the optimal solution, they harnessed the power of ant colony optimization techniques. Furthermore, the communication range has been taken into account by research in [91]. Their contribution includes the proposal of an AUV-assisted data collection approach centered on Fermat’s spiral (FS-PPS). This method employs an enhanced firefly algorithm to ascertain the sequence in which nodes are traversed. In the context of addressing challenges like imbalanced energy consumption, prolonged delays, and partial coverage, the study outlined in [89] employs the backtracking search optimization (BSO) technique. This is utilized to attain an optimal selection of coverage-aware target nodes and to devise trajectories for multiple AUVs. Introducing a heterogeneous USV-AUV system to the data collection, the authors of [87] introduced a modified shuffled frog-leaping algorithm (SFLA). This alteration is applied to optimize the coordination schedule between USVs and multiple AUVs.

3.2. AI-Based Methods

AI-based methods, such as reinforcement learning (RL) [112,113,114] and self-organizing maps (SOMs) [115,116,117], have demonstrated remarkable strengths that elevate their effectiveness. On the one hand, RL algorithms can learn optimal paths by iteratively interacting with an environment, receiving feedback in the form of rewards, see Figure 10. This capability is particularly advantageous in complex and dynamic environments, where traditional algorithms might struggle to adapt. RL’s ability to handle uncertainty and adapt in real time makes it a strong contender for tasks like robotic navigation, where the optimal path can change due to unforeseen obstacles or changing conditions. On the other hand, SOMs have proven their worth in path planning through their ability to analyze and simplify complex spatial data. SOMs can reduce high-dimensional input, such as maps or sensor data, into a lower-dimensional representation while preserving the topological relationships of the data. This makes them invaluable for tasks like route optimization and exploration, where understanding the underlying structure of the environment is crucial. In summary, AI-based approaches can provide more efficient, adaptable, and contextually aware navigation solutions across many scenarios and applications.

RL’s capacity to manage uncertainty and adjust in real time has established its popularity in the realm of ocean data collection path planning. In the work of [77], researchers introduced a reinforcement learning technique for AUVs within a hierarchical underwater wireless sensor network (HUWSN). The proposed algorithm dynamically calculates optimal paths toward targets, employing a local search strategy that leverages image acquisition and segmentation. Additionally, to achieve low latency and minimal power usage, authors in [56] put forth an AUV path-planning approach rooted in Q-learning, leading to enhancements in energy efficiency and network longevity. To further optimize sensor node energy consumption, the work in [76] formulated a dynamic routing algorithm based on Q-learning. This algorithm optimizes node routing selection, wherein a novel method based on potential-game theory and optimal rigid graph concepts is introduced to achieve balance between energy consumption and network resilience.

In contrast to conventional RL techniques, deep reinforcement learning (DRL) demonstrates a strong capability to effectively manage intricate, high-dimensional state spaces by utilizing neural networks. In the context of marine mooring systems, researchers in [75] have introduced a DRL approach aimed at minimizing weighted data loss and energy consumption during data collection operations. Building upon the foundation laid by the study in [75], a subsequent contribution by authors of [74] introduced a target-oriented double deep Q-learning network (D2QN)-based algorithm. This algorithm is specifically tailored for collision avoidance and trajectory planning, particularly for USVs. Notably, the proposed algorithm equips the USVs to effectively navigate through stochastic incoming data and dynamically shifting obstacles. Addressing the intricacies of multi-modal transmission and trajectory planning in complex underwater settings, researchers of [73] have developed a distributed trajectory planning algorithm for multiple AUVs using deep reinforcement learning. Collaborative decision-making among the AUVs considers various factors, including transmission conditions, ocean currents, and submerged obstacles. The primary objective of this approach is to maximize the collection rate while ensuring energy efficiency. Moreover, in the work [72], a synthesis of acoustic and optical technologies was proposed to optimize AUV motion trajectories via a deep reinforcement learning. This methodology is used to select optimal communication strategies, while an innovative angle steering algorithm is also introduced to enhance AUV navigation under diverse communication scenarios, leading to reduced energy consumption.

Lately, the self-organizing map has gained significant attention among researchers due to its capacity to reveal the underlying topological arrangement of environments. Given the heterogeneity of task types, the work in [118] introduced an enhanced version of the self-organizing map. It used a fitness function to assign multi-type tasks to specific USVs. Additionally, for specific tasks that require sequential execution, like sea sweeping tasks, the method of tasks treatment list (TTL) is presented. Incorporating energy consumption and path-smoothing techniques, a hierarchical two-layer framework was proposed by the authors of [119]. This framework addressed cooperative path planning for multi-USV systems. In this framework, the self-organizing map served as the first layer to prevent collisions and efficiently access multiple targets in an optimal sequence. Addressing task urgency and uncertain ocean environments, a double-layered bio-inspired self-organizing map (DLBSOM) algorithm was introduced by the study mentioned in [120]. This algorithm facilitates collaborative search within a 3D environment, where multiple AUVs balance factors such as task completion time, distance, energy consumption, system stability, and security to achieve optimal task planning. To enhance cooperative operational capabilities in formation control and cluster search tasks, the authors of [121] proposed the hybrid bio-inspired self-organizing map (HBSOM) algorithm. This algorithm improves the speed, stability, and accuracy of multi-AUVs. Though SOMs are powerful unsupervised learning techniques used for clustering and dimensionality reduction, they also come with certain drawbacks and challenges. First, handling high-dimensional problems with SOMs can be difficult due to the “curse of dimensionality”. The computational complexity of SOM training can be an issue when dealing with large problem scales. Moreover, convergence of SOM training is not always guaranteed, and the stopping criteria for training can be difficult to determine. Stopping training too early might result in an underdeveloped map, while training for too long could lead to overfitting.

While the past decade has witnessed promising outcomes from AI-based approaches, it is imperative to acknowledge that these approaches continue to come with challenges that demand resolution: (1) RL algorithms often require a significant amount of data (samples) to learn effective policies. In oceanic environments, data collection can be time-consuming and costly, making it challenging to gather enough data to train RL agents effectively. (2) Ensuring that policies generalize effectively is quite a challenging work. AMVs need to generalize their learning from one situation to another. However, oceanic environments can vary widely, and what works well in one area might not work in another due to differences in underwater topography, currents, and other factors. (3) Oceanic environments are complex and dynamic, with various factors such as currents, weather conditions, and underwater obstacles affecting the path-planning process. Designing AI-based methods that can handle these complexities and make informed decisions could be challenging as well. (4) Transferring knowledge from simulations to real-world oceanic environments can be challenging due to the domain gap between the two. Ensuring that AMVs can effectively learn from simulations and apply that knowledge in the real world is a significant challenge.

3.3. Coverage Planning Techniques

Coverage path planning refers to a strategic approach used in various domains, including robotics and autonomous systems, to efficiently navigate and explore an area while ensuring that every point within the region is visited or observed. The primary goal of coverage path planning is to devise a path or trajectory for a vehicle or agent that maximizes the coverage of a specified area, minimizing redundancy and ensuring systematic exploration. Particularly, coverage path planning is a critical aspect of marine robotics, encompassing a wide range of applications like bathymetry mapping [122,123], underwater surveys [124,125], search and rescue missions [126,127], and mine-countermeasures [128,129]. By intelligently navigating through designated areas, AMVs can efficiently collect data, perform critical tasks, and contribute to a safer and more informed maritime environment.

In the field of oceanic data collection, the research community has extensively investigated coverage path planning. To enhance the freshness of data in wireless sensor networks, the work in [78] introduced an innovative approach called the on-the-way fair data freshness (OFDF) path, utilizing lawn mower patterns (refer to Figure 11). In comparison to traditional lawnmowers and shortest path traversal algorithms, this scheme improved overall data freshness. Building upon the foundation of [78], work in [24] extended the model to incorporate factors, such as residual energy and communication protocols. In this context, the AUV not only gathers data but also undertakes the responsibility of designating the Cluster Head (CH) and establishing a wakeup-sleep schedule for the clusters. Further contributing to this, work in [79] also devised a lawn mower pattern to facilitate prolonged cooperation with sensor nodes. Their efforts culminated in an AUV-based data-gathering protocol that optimizes cooperative work duration. Drawing inspiration from the work in [79], a subsequent study [35] enhanced the feasibility of this concept in practical scenarios by introducing the idea of cluster reconstruction by AUVs to ensure uniform energy consumption. Another method presented in [47] introduced algorithms designed to formulate persistent monitoring missions for underwater vehicles. The coverage planning methodology not only maximizes information value along the path but also minimizes deviations from the predetermined route.

However, it can be observed from the literature review that coverage path planning is not the mainstream approach for oceanic data collection. This is because the distribution pattern of sensor nodes is different from the approximate cellular decomposition pattern described in [24,78]. In reality, the distribution of sensor nodes is more dispersed and disorderly, rendering it mostly unsuitable for conducting data collection using the back-and-forth path (BFP) method. Additionally, there are some issues with coverage path planning in data collection tasks. Foremost among these is the challenge that BFP poses to the energy consumption of AMVs. Due to the influence of ocean currents, AMVs consume much more energy during turning maneuvers than during straight-line travel, as demonstrated in [35]. However, the BFP path includes a substantial amount of turning motions [130], which hinders the ability of AMVs to perform long-distance missions. Secondary to the foregoing, it is pertinent to acknowledge the substantial challenges confronting AMVs when executing turns in underwater environments. Given the complexity of the ocean environment, AMVs struggle to execute precise turning maneuvers [131], often causing deviations from the original path and thus affecting the efficiency of data collection.

3.4. Heuristic and Direct Search

Heuristic and direct search approaches, such as A*, dynamic programming (DP), and greedy search, have drawn substantial attention in classical robot path planning. The appeal of their simplicity and efficiency has captivated numerous researchers, leading to their widespread exploration.

Due to its ease of implementation, A*-related algorithms have attracted interest among researchers. In the context of enhancing oceanic observation using autonomous underwater vehicles (AUVs), work in [85,86] introduced an A* planner that utilizes a locally generated controllability map derived from ocean current predictions. This approach computes a path between predetermined waypoints that maximizes the probability of successful execution. Expanding upon the foundation laid by [85], subsequent research in [49] refined the concept by incorporating considerations of uncertainty associated with ocean currents. They devised a minimum expected risk planner designed to prevent collision risks through sea trials. Furthermore, to enhance the energy efficiency of data acquisition systems employing multiple AUVs, a study documented in [84] introduced the AE (A*-Energy) algorithm. This algorithm was developed for the purpose of path planning for multiple AUVs. Its implementation led to a reduction in total distance traveled and turning angles executed by the AUV system, consequently resulting in improved energy efficiency.

Several traditional optimization algorithms, such as dynamic programming, greedy algorithms, and local search techniques, have been applied to data collection. In addressing challenges associated with AUV yaw change and sensor node movement arising from time-varying ocean currents, a local path-planning algorithm based on greedy algorithmic approach (depicted in Figure 12) was formulated by researchers of [83]. This algorithm effectively resolves issues related to sensor nodes’ movement while simultaneously optimizing the VoI–energy ratio. Drawing on the interval programming concepts in [132], a multiple-objective path planning strategy for data acquisition tasks was introduced by the authors of [81]. This approach optimizes three distinct objective functions: AUV safety, energy consumption, and vehicle coordination. It is important to note that, however, the communication channel model was omitted from consideration in this study. Moreover, some researchers have directed their efforts towards addressing data collection challenges through the utilization of commercial software. To maximize gathered information while minimizing travel time, a novel form of the TSP was proposed and termed the Communication-Constrained Data Collection Problem (CC-DCP) by authors in [82]. This complex problem was solved using the Concorde software (https://www.math.uwaterloo.ca/tsp/concorde/index.html). Building upon the work in [82], researchers of [48] extended this line of research by incorporating data from AUV deployments and incorporating more advanced acoustic communication models. In addressing scenarios where event-generated data dynamically occur, a scheme known as the on-demand path-planning scheme (OD-PPS) was developed by researchers of [91]. This scheme employs local search techniques to maximize the completeness of collected data. It is worth noting that in this approach, the modeling does not account for external oceanic environmental factors.

3.5. Online Path-Planning Techniques

The ever-changing nature of the oceanic environment underscores the need for online path-planning algorithms. Recently, numerous researchers have directed their efforts towards formulating solutions for dynamic path planning. Notably, a study outlined in [100] introduced a dynamic path-planning framework designed to facilitate informed decisions regarding the sequence in which AUVs should visit network nodes. These decisions are informed by real-time network status information. Furthermore, authors of [41] presented a dynamic value-based path-planning strategy. This strategy empowers AUVs to dynamically select data collectors to visit, with the objective of maximizing the value of information (VoI). Addressing the issue of low location accuracy associated with sensor nodes, the work in [99] proposed a grouping-based dynamic trajectory planning (GDTP) approach. This method detects sensor nodes’ presence and orientations, subsequently grouping them based on a proposed common communication area model. Building on this foundation, authors of [98] extended the GDTP [99] concept to encompass tracking sensors mounted on turtles. In this adaptation, the AUV’s cruising direction is dynamically determined to achieve the maximum expected payoff, considering both data collection and energy consumption despite inaccurate detection. In summary, when contrasted with TSP-variant models, dynamic methods offer significant enhancements in mission feasibility through their adaptive response to diverse scenarios. Nevertheless, the intricacies in dynamic planning pose challenges for researchers, including how to address issues such as Doppler frequency, multipath effects, latency, and the rate of data transmissions.

3.6. Others

Apart from the algorithms discussed in the preceding sections, the research community has introduced an array of alternative methods. In addition to some techniques adapted from other domains, such as artificial potential field (APF) and sequential quadratic programming, the majority of these methodologies have been devised by these scholars. In this section, we will undertake a comprehensive review of these methods.

To facilitate obstacle avoidance for AUVs in data acquisition, a path-planning technique based on fluid mechanics is proposed by the work in [105], where the ocean current characteristics were simplified as a stream function to design the artificial potential field. Furthermore, the APF was extended to the application of AUV flocks by the authors of [51]. They defined the concept of the AUV flock and the united control model for the AUV flock based on the artificial potential field theory. The path-planning method performed more efficiently than the normal-distributed path-planning approach.

In order to enhance the performance of AUVs during resurfacing in data collection, a series of studies were conducted based on the Euler circuit [106,133,134]. An Eulerian circuit is a path in a graph that traverses each edge exactly once and returns to the starting vertex, forming a closed loop. In [134], the authors first established the trajectory planning for AUVs as an Euler cycle problem. Then in [106,133], the authors minimized the average data and event reporting delay. Experiments in both the synthetic and real traces are conducted.

To perform data collection in high-valued regions, the work in [46] proposed a centroid-tracking waypoint-selection algorithm. It determines paths for AUVs to track evolving features of interest in the ocean by considering the output of predictive ocean models. The algorithm uses model predictions to assist in solving the motion planning problem of steering an AUV to high-valued locations. Furthermore, authors of [50] extended the application to tracking a freshwater plume. Also, field trials have been conducted using two gliders at the coast of Newport Beach, Los Angeles, CA.

Considering the state uncertainty due to the ocean currents, the work in [104] presented an energy-aware feedback planning method for an LRAUV. It uses ocean dynamics from a predictive ocean model to understand the water flow pattern and introduces a goal-constrained belief space to make the feedback plan synthesis computationally tractable. To maximize the information collected, the work in [135] proposed a model predictive control method that incorporates the posterior Cramér–Rao Lower Bound. Similarly, the work in [136] introduced a generalized optimal control for the information collection in mine countermeasures. To minimize the collection delay and facilitate information sharing between AMVs, a state prediction-based data collection (SPDC) algorithm is devised by authors of [103], where path planning is achieved by a heuristic strategy based on the updated access area. To solve the trajectory planning problem for the UAV-USV system (see Figure 13) in data collection, the work in [102] devised a Fermat-point theory in maritime Internet of Things systems. Delaunay triangles based on the deployment of unmanned surface vehicles were constructed and the Fermat point of each Delaunay triangle is calculated as a hovering point of UAVs to receive data from USVs.

In general, we have compared existing algorithms mentioned in this section, and their advantages and disadvantages are summarized in

Table 3. Comparison of existing methods.

Category	Methods	Highlights	Drawbacks
Meta-heuristic methods	GA, PSO, ACO, SA, etc.	Applied to complex, large-scale, non-convex problem Good at exploring global optima	No guaranteed optimality Sensitive to hyperparameters Slow convergence
AI-based methods	RL	Applied to dynamic environment Interacting with the environment Allow for end-to-end learning	Tedious training process Generalization Knowledge transferring
	SOM	Considering the topological structure of the solution Dimensionality reduction Without labeled data for learning	Hard to guarantee the convergence Slow in large-scale problems Sensitive to hyperparameters
Coverage planning	-	Full coverage for clustering network	Dependent on the distribution of the sensor nodes High energy cost Not suitable for low maneuverability
Heuristic and direct search	DP, greedy search, A*-variants	Global optimality Simple and easy to implement Quick solution for simple problem	High computational complexity Not suitable for complex problems
Online path planning	-	Real-time decision Handle uncertainty and partial information Sensor integration	Limited global optimization Sensitive to sensor accuracy High computational complexity Local minima

.

4. Discussion

Based on the preceding observations, in this section, we will explore the challenges and opportunities associated with the path-planning problem for maritime data collection assisted by AMVs. The following points are highlighted:

• Vessel dynamics and unknown ocean currents: When planning paths for AMVs in maritime data collection, it is essential to consider the vessel dynamics and the influence of unpredictable ocean currents. On the one hand, our previous works [137] have shown that the trajectory design must ensure that the AMVs move in a feasible manner, as infeasible paths would require inaccessible forces/torques, such as propulsive force and steering torque, which are limited. For instance, a sharp turning is infeasible for a USV, as steering maneuvers can lead to significant sideslip, deviating from the original path. Furthermore, navigating the vehicle along a predetermined path encounters complexity arising from the ever-changing nature of ocean currents and their inherent unpredictability. The vehicle’s trajectory is susceptible to divergence caused by motion-induced disturbances, while underwater environments further compound matters with sensor-induced inaccuracies in state estimation. Successfully addressing this challenge would result in more precise and effective data collection even in the face of varying environmental conditions.
• Path planning under more complex communication channels: Prior research has engaged in simplifications when constructing models for underwater communication, and in certain cases, some studies have even omitted these communication models [90]. However, communication within the actual marine environment presents greater intricacies [138,139,140], including factors such as Doppler frequency, multipath effects, latency, and data transmission rates, all of which are influenced by the environment. Developing sophisticated path-planning algorithms that optimize communication paths while factoring in variables like signal strength and interference patterns is crucial. Potential future avenues involve enhancing the control model, particularly in scenarios considering the instability of underwater communication and significant delays inherent in acoustic-based communication.
• Considering onboard sensor failure and uncertainty in state estimation: Previous research commonly operated under the presumption of having complete knowledge of the environment or flawless sensing capabilities. However, these assumptions are often impractical in the vast majority of scenarios, particularly in underwater settings [129,141]. Even when advanced techniques like terrain-relative navigation (TRN) or simultaneous localization and mapping (SLAM) are employed to enhance localization, achieving perfect knowledge or sensing remains a challenging proposition. However, due to the interference posed by unknown ocean currents, the inaccuracy in onboard sensor data and state observations can lead to errors in path planning. Additionally, satellite-based position accuracy might not suffice, particularly in unforeseen circumstances demanding high precision. Some propose the adoption of re-planning methods, e.g., [142], while others suggest the integration of UAVs for aiding in localization, e.g., [6,143]. Addressing these challenges holds significant implications for real-world maritime data collection missions.
• Navigating AMVs with uncertain sensor node positions: Previous studies have assumed accurate knowledge of sensor node locations. However, in dynamic underwater environments, anchored nodes can shift from their initial positions due to water currents [144,145]. This movement within a defined range can impede data collection when an AMV follows a predetermined strategy to reach a specific location. Therefore, the AMV may face stochastic position data from the sensor nodes, which causes a delay in reaching the target. Consequently, designing the AMV’s path to accommodate sensor node position uncertainty emerges as a pivotal challenge.
• Path planning for long-duration data harvesting with energy supply: Due to limited energy resources, AMVs are faced with limitations when undertaking tasks over extended distances. Despite the existence of energy-efficient, path-planning methods proposed by many researchers, the energy capacity of AMVs remains insufficient for tasks that cover vast regions. Consequently, introducing energy supply stations during AMV missions has emerged as a viable solution to this challenge. Some scholars have initiated preliminary exploration in this realm. The work in [146] has developed a dynamic routing strategy employing USVs as energy supply stations for UAVs. Similarly, authors of [147] employ USVs to provide energy to AUVs for search tasks. However, the aforementioned work oversimplifies the problem. In reality, the recharging mechanisms for UAVs and AUVs are intricate, involving factors such as the docking of AUVs [148] during motion and the landing of UAVs, see Figure 14. In summary, considering the issue of energy supply station placement in path planning is crucial for enhancing AMV operational efficiency. Nevertheless, the current state of research in this area remains nascent.
• Path planning for heterogenous systems: Leveraging AUV-USV-UAV heterogeneous unmanned systems can significantly enhance data collection efficiency. Specifically, the incorporation of UAVs offers advantages, such as cost-effectiveness, high flexibility, wide coverage, and line-of-sight (LoS) communication links. AUVs excel in navigating intricate underwater terrains, performing bathymetry mapping, conducting environmental monitoring, and executing sonar-based data collection. However, both UAVs and AUVs are constrained by limited battery capacities. USVs are adept at surface operations and serve as mobile workstations, supplying power and communication networks and serving as landing/docking platforms for UAVs and AUVs. This collaborative multi-robot system holds immense potential for mitigating risks and enhancing efficiency in demanding and perilous maritime tasks. Presently, some scholars have embarked on explorations in heterogeneous unmanned systems, yet their focus has been on foundational theoretical research, including collaborative control [149,150], landing control [151,152,153], and task allocation [154,155]. However, research into the application of heterogeneous unmanned systems in marine engineering remains largely unexplored.
• Field experiments: In Table 2, it is evident that, in preceding research articles, nearly all investigations relied upon numerical simulation methodologies for validation. It is imperative to acknowledge the substantial distinction between experimental and numerical simulation approaches, recognizing that the substantiation of method feasibility and practical applicability necessitates empirical scrutiny. Nonetheless, the execution of authentic maritime or subaquatic experiments is encumbered by formidable challenges. On the one hand, the procurement of mechanical apparatus and sensors is typically accompanied by a high financial outlay. Moreover, maritime experiments entail a significant investment of both manpower and time, with a customary sea expedition enduring from one week to a month, thereby incurring considerable temporal expenses for a singular experimental iteration. On the other hand, maritime experiments are susceptible to the influences of intricate sea conditions, thereby exerting pronounced impacts on the trajectories of AUVs or USVs, consequently yielding diminished data precision and suboptimal experimental outcomes. Consequently, the formulation of efficacious maritime experimental frameworks for the validation of path-planning algorithms stands out as a formidable challenge for the research community.

5. Conclusions

Our exhaustive exploration of the existing literature has allowed us to offer a comprehensive overview of the current research landscape in the field, and the research questions with short answers are shown in Figure 15. We have delved into various aspects, encompassing system architectures, problem models, objective functions, and the constraints associated with data harvesting. Furthermore, we have diligently examined and condensed information on the diverse array of algorithms, methodologies, platforms, coding environments, and implementation strategies utilized in tackling the path-planning challenge. In addition to shedding light on these essential elements, we have introduced emerging applications that harness path planning for oceanic data harvesting, thereby contributing to the advancement of marine services. Lastly, we have contemplated the potential avenues for future research in marine networks, thereby enriching the discourse surrounding this dynamic and evolving domain. We hope that this article will facilitate researchers and developers being able to understand the perspectives and challenges of developing appropriate path-planning methods for AMV-assisted data harvesting.