Path-planning approaches are used in multiple fields, as they enable autonomous robots to navigate efficiently while avoiding obstacles and fulfilling mission objectives [1]. It is, therefore, the process through which autonomous robots determine an optimal trajectory that meets specific mission objectives, typically evaluated by metrics such as path distance, number of turns, and running time [2]. Depending on the level of environmental knowledge and real-time adaptability, path-planning methods can generally be divided into global and local approaches [3].

These path-planning capabilities are highly relevant for Autonomous Underwater Vehicles (AUVs) in deep-sea exploration, environmental monitoring, seabed mapping, and underwater inspection [4]. Due to the unpredictable and dynamic nature of marine environments characterized by ocean currents, communication delays, and complex obstacle distributions, path planning and trajectory optimization remain highly challenging tasks [5,6]. The goal is, therefore, to produce algorithms that can compute a collision-free and energy-efficient trajectory to allow the AUV to reach a target point safely while minimizing travel time, energy consumption, and deviation from the desired path [6].

Despite the substantial body of prior work, there remains a need for comparative analysis of controlled benchmarks in which algorithms are evaluated in the same controlled environment under varying obstacle densities and environmental disturbances. Furthermore, while many studies report path quality metrics such as length or smoothness, fewer works explicitly evaluate success rate, variability, and feasibility preservation as obstacle density increases. Such is the main goal of this work. The examined scenario is designed to enable simplified visualization of trajectories in a two-dimensional environment (which can be extended to higher dimensions). In addition, robustness to current-induced disturbances is investigated by considering simulated ocean-current conditions, including a constant-direction current and a continuously time-varying current.

The remainder of this work is organized as follows. Section 2 reviews representative AUV path-planning families. Section 3 describes the simulation environment and benchmarking methodology, including the evaluated approaches, obstacle-generation procedure, and performance metrics. Section 4 presents and discusses the results across obstacle densities and examines the impact of steady and time-varying current forcing on the algorithm performance. Section 5 concludes the paper by summarizing the main findings and discussing the key trade-offs observed.

Path planning for AUVs sits at the intersection of motion planning, estimation, and control in operating conditions where uncertainty and vehicle and environmental constraints (e.g., bounded actuation, limited sensing, and ocean currents) are the norm [6]. Recent surveys emphasize that no single planner consistently dominates across “predictable” vs. “unpredictable” underwater environments [7,8]. As a result, practical solutions tend to be hybrid, pairing a global planner (for connectivity and long-horizon efficiency) with a local reactive layer or a short-horizon planning layer (for safety and rapid adaptation), often complemented by an optimization layer that accounts for dynamics, constraints, and mission costs [9].

Global path planning assumes prior knowledge of the environment and focuses on computing, before execution, an optimal or near-optimal path from the start to the goal with respect to a chosen cost model. Representative global methods include sampling-based techniques such as the Probabilistic Roadmap (PRM) [10,11] and population-based metaheuristics. The latter involve evolutionary methods such as Genetic Algorithms (GAs) [12,13], including their continuous-encoding variants (CGAs) [14]. Swarm-based methods such as Particle Swarm Optimization (PSO) [15,16] and the Gray Wolf Optimizer (GWO) [17,18], and metaphor-less optimization algorithms such as Jaya [19,20], all of which are used in this work. These methods search for feasible trajectories by optimizing cost functions related to path length, smoothness, and safety, often providing solutions for complex or high-dimensional environments.

In contrast, local path planning focuses on real-time navigation, where the robot dynamically reacts to changing environments and unforeseen obstacles. Potential field approaches such as the Artificial Potential Field (APF) method [21,22] provide reactive obstacle avoidance by combining attractive and repulsive terms, while Model Predictive Control (MPC) integrated with Control Barrier Functions (CBFs) [23,24] has shown good performance by continuously predicting future states and enforcing safety constraints during motion execution. As a result, global and local approaches provide complementary capabilities: global algorithms emphasize route optimization, while local strategies enable adaptive obstacle avoidance and safe maneuvering in dynamic or partially unknown environments.

In recent years, numerous studies have focused on developing path-planning strategies adjusted specifically to AUV operations. Classical geometric and graph-based planners have been used due to their deterministic guarantees, although their scalability becomes limited in large or continuous underwater domains [25]. Sampling-based motion planners have also been adapted for AUV navigation, offering improved performance in high-dimensional spaces and uncertain environments [26,27]. In parallel, bio-inspired metaheuristic algorithms, such as PSO, GA, GWO, and Ant Colony Optimization (ACO), have gained notoriety for their flexibility and ability to escape local minima when navigating around complex obstacle fields or dynamic ocean currents [28,29].

More recently, control-theoretic approaches (such as integrating MPC with CBFs) have shown potential for safety-critical AUV navigation by ensuring constraint satisfaction and real-time collision avoidance in dynamic marine settings [23,30]. Together, these advances demonstrate the diversity and progression of underwater path planning, including sampling-based roadmaps, potential field methods, population-based metaheuristics, and constraint-based predictive control, each contributing complementary strengths. Accordingly, the remainder of this section focuses on these four families, with emphasis on navigation in clutter and under current induced disturbances.

In underwater settings, discretization and the cost model should reflect current-induced advection. Otherwise, nominally short routes can be far from time-efficient or energy-efficient, motivating current-adaptive cost formulations or frequent replanning [31,32]. For global connectivity in cluttered environments, sampling-based roadmaps such as PRM are often used to identify feasible corridors, which can then be refined by an optimizer to improve path quality [7,10].

Potential-field methods remain attractive for AUV path planning because they are simple, local, and computationally efficient, which makes them suitable for reactive obstacle avoidance. The canonical APF formulation combines an attractive term toward the goal with repulsive terms around obstacles [33], but it is susceptible to local minima and oscillations in cluttered geometries [34]. Consequently, APF variants for underwater navigation often add heuristics (e.g., intermediate targets, adaptive gains, or smoothing) or are coupled with higher-level guidance to mitigate trapping [35]. More recent work commonly integrates APF within metaheuristic or multi-objective optimization frameworks to improve global path quality and reduce failure modes associated with local minima [36,37].

Metaheuristic optimization has become widely used in AUV path planning because it can handle nonconvex, multi-criteria objectives (e.g., path length, smoothness, clearance, energy, and time) with relatively weak modeling requirements. GAs [14] and multi-objective evolutionary methods such as NSGA-II [38] are commonly used when a path is encoded as a sequence of intermediate control points and evaluated by a fitness function that aggregates costs and penalties. PSO [15] is widely adopted for continuous control-point parameterizations, while alternatives such as GWO [17] and metaphor-less schemes like Jaya [19] have been explored to reduce tuning effort. A recurring observation in surveys and comparative studies is that these methods can yield high-quality solutions in moderately cluttered environments [7,9], but feasibility and repeatability tend to deteriorate in narrow passages unless diversity preservation, effective constraint handling, or roadmap-based encodings are incorporated [32]. This sensitivity to environment structure motivates standardized benchmarking and reporting beyond mean path length, including success rate, computational time, and run-to-run variability, to enable fair comparisons across planning frameworks [39,40].

In line with these observations, control-point-based metaheuristics are sensitive to constraint handling and may stagnate in highly constrained geometries [9]. Roadmap-based planners explicitly encode connectivity and tend to be more robust when free space is fragmented [10], whereas MPC/CBF formulations prioritize constraint satisfaction over optimality [41,42].

Hybrid approaches have become a common design choice in underwater path planning: potential-field formulations provide fast, smooth, reactive guidance, while a higher-level optimizer such as PSO, GA, or ACO searches over control points, intermediate goals, or tuning parameters to improve global path quality and mitigate local-minima failure modes. Representative examples include the combination of APF and ACO for global planning with an added velocity-obstacle (VO)-based collision-avoidance layer for underwater vehicles [36], the explicit coupling of APF and PSO in a hybrid planner [43], and the use of PSO to optimize potential-field–driven navigation behavior in cluttered environments [44].

The same hybridization principle extends naturally to multi-AUV planning, where potential-field-like constructs can support coordination while an outer optimization layer addresses route quality and task-level objectives. Recent examples include dynamic velocity potential fields for cooperative hunting tasks [45] and improved multi-objective PSO for cooperative 3D planning in complex terrain and vortex-dominated currents [46]. This cooperative setting typically drives formulations toward explicitly multi-objective tradeoffs (e.g., efficiency, collision risk, cohesion) and, in many cases, decentralized or distributed implementations where each vehicle optimizes locally under shared constraints.

Ocean currents are a major driver of AUV-specific planning complexity because advection and spatiotemporal variability can dominate the effective motion and make purely geometric paths suboptimal (or even infeasible) [31]. Early work on evolutionary planning in variable oceans [12] and current–aware planning for gliders and AUVs showed that achievable improvements in energy consumption and transit time reductions depend strongly on the ratio between current speed and vehicle speed, motivating closer integration between planning algorithms and ocean models. Comparative studies of optimization-based planners under currents further emphasize that performance depends strongly on both the algorithm and the path encoding, and that robustness under time-varying currents commonly requires either replanning or cost formulations that account for currents [32].

A recurring conclusion in the literature (see, e.g., [31,32]) is that ocean-current effects should not be treated as a secondary consideration. Practical planners typically use a global model that accounts for currents, update plans frequently as state and current estimates evolve, or apply local feedback that compensates for drift without compromising obstacle avoidance. In parallel, control-based methods have gained visibility in AUV navigation because they offer a systematic way to enforce dynamic and safety constraints in real time. MPC optimizes a finite-horizon trajectory under explicit constraints [41], while CBFs provide formal safety conditions that can be imposed through quadratic programs or integrated into MPC [42]. Recent AUV-focused work combines these ideas in robust MPC formulations with fast-marching (FMM)-based travel time objectives and adaptive CBFs [23].

Compared with purely geometric planners, MPC and CBF approaches impose input and safety constraints via real-time optimization [42] and can be interpreted as optimization-based safety filters that minimally adjust a reference control objective while maintaining safety [47]. This makes them well-suited to respect actuator limits, vehicle footprint, and safety margins while responding to disturbances. However, in densely cluttered environments, they may become conservative or even reach deadlock, and they often require accurate models and real-time optimization [48]. For these reasons, surveys most often describe MPC and CBF as a local safety or tracking layer under a global planner (e.g., roadmaps or metaheuristics), rather than as a self-contained global planner [9].

Surveys also emphasize that high-performing AUV path-planning approaches are increasingly implemented as layered architectures that combine a global planner that explicitly encodes free-space connectivity (e.g., graph or sampling-based methods), an optimization step for path quality or multi-objective tradeoffs (e.g., PSO, GA, ACO, or NSGA-II), and a control layer that accounts for both dynamics and safety, often implemented with MPC and CBFs [7,9]. Key remaining gaps include standardized, reproducible benchmarks that incorporate realistic sensing and current dynamics, where reusable underwater simulation frameworks can play an important role [49,50].

Evaluation practice is also improved by reporting metrics beyond mean path length, such as success rate and performance variability across runs [51]. Evidence of robustness when moving from simulation to open-water trials remains limited, motivating more systematic validation through approaches such as hardware-in-the-loop testing [52].

A recent work by Yin and Xiang [53] proposes an adaptive differential evolution-based multi-source information fusion framework for real-time cooperative navigation of multiple unmanned surface vehicles in dynamic maritime environments. Their approach integrates heterogeneous sources of information, including reciprocal velocity obstacles and goal-alignment cues, with fusion weights optimized via adaptive differential evolution. Unlike traditional optimization-based planners that rely on static assumptions or reinforcement learning methods requiring extensive training, this method achieves real-time decision-making through a training-free mechanism with adaptive parameter tuning. Such methodology can be a suitable next step for this work to integrate multi-agent coordination.

It is, however, important to note that Yin and Xiang [53] primarily target cooperative navigation systems with explicit inter-agent communication and interaction modeling. In contrast, the present work focuses on benchmarking path planning strategies under controlled environmental disturbances and emphasizes comparative evaluation across different planning paradigms.

To ensure reproducibility and enable a controlled comparison, the computing platform and simulation environment are first described. The evaluated approaches are then summarized, followed by the experimental setup, performance metrics, and common parameter settings used throughout the study.

This section reports and interprets the simulation results obtained under the benchmark described in Section 3. It starts with an algorithmic benchmark and then further examines the results with statistical analysis.

This work presented a comparative analysis of representative path-planning approaches for autonomous underwater vehicles, including metaheuristic optimization methods, sampling-based planning, potential-field navigation, and predictive control with safety constraints. The evaluation was conducted in a controlled two-dimensional simulated environment with systematically increasing obstacle density, allowing a consistent assessment of both path quality and feasibility preservation. In addition, the effect of temporally correlated ocean-current disturbances was examined through Gauss–Markov models, with a particular focus on the sensitivity of PSO-based trajectory generation.

The results highlight that algorithm performance is strongly dependent on environmental complexity. In sparse and moderately cluttered scenarios, several methods achieved comparable path quality, with metaheuristic approaches such as PSO, CGA, GWO, and JAYA producing short trajectories with relatively low variability. However, as obstacle density increased, clear differences emerged in robustness. In particular, PRM–GA showed the most consistent behavior, maintaining an ideal success rate across all tested scenarios while exhibiting controlled growth in path length.

The introduction of current disturbances through Gauss–Markov models led to a consistent degradation in PSO performance, with MOD and STR current configurations reducing efficiency and success rate. These results indicate that, within the present formulation, current-induced perturbations increase the difficulty of maintaining stable and feasible trajectories. The statistical analysis further confirmed significant differences among the evaluated algorithms, with CGA achieving the best mean rank in terms of path-length-based performance. However, this ranking does not fully capture practical robustness, as methods such as PRM–GA provided more reliable feasibility preservation under increasing environmental constraints.

Overall, the findings indicate that PRM–GA is the most reliable method for highly cluttered environments due to its explicit encoding of free-space connectivity, while standard optimization methods like PSO can reach an effective balance between efficiency and reliability in moderately complex scenarios. Control-based methods such as MPC–CBF provide high-quality trajectories when feasible, but their success rate decreases in densely constrained environments, whereas APF and CGA are more sensitive to geometric complexity and exhibit reduced robustness under high clutter.

Despite these contributions, the study presents several limitations. The benchmark is based on a simplified two-dimensional workspace with circular obstacle representations, which, although useful for controlled evaluation, does not fully capture the geometric and physical complexity of real underwater environments. Furthermore, the framework adopts a kinematic abstraction and does not explicitly model vehicle dynamics, sensing uncertainty, or partial observability. As a result, the generated trajectories should be interpreted as geometric planning solutions rather than directly executable paths. In addition, the analysis relies on fixed parameter settings and specific path encodings, which may influence the observed performance.

Future work will focus on extending the framework to more realistic scenarios, including three-dimensional environments, dynamic and uncertain obstacle representations, and the integration of vehicle dynamics and trajectory-tracking control. The incorporation of sensing uncertainty and partial observability is also an important direction to improve realism. Finally, the exploration of hybrid architectures, adaptive parameter tuning, and current-aware cost formulations may provide further improvements in robustness and efficiency under complex environmental conditions.