A Review of Research on Autonomous Collision Avoidance Performance Testing and an Evaluation of Intelligent Vessels

Abstract

As intelligent vessel technology moves from the proof-of-concept stage to engineering applications, the performance testing and evaluation of autonomous collision avoidance algorithms have become core issues for safeguarding maritime traffic safety. The International Maritime Organization (IMO)’s Maritime Safety Committee (MSC), at its 109th session, agreed to a revised road map for the development of the Maritime Autonomous Surface Ships (MASS) Code; the field has experienced the development stages of single-vessel collision avoidance validation based on COLREGs, multimodal algorithm collaborative testing, and the current construction of a progressive validation system for the integration of a mix of virtual reality and actual reality. In recent years, relevant studies have achieved research achievements, especially in the compatibility of COLREGs and in accurate collision avoidance in complex situations, and other algorithm tests and evaluations have made great breakthroughs. However, a systematic literature review is still lacking. In this paper, we systematically review the research progress of autonomous collision avoidance performance testing and the evaluation of intelligent vessels; summarize the advantages and disadvantages of virtual testing, model testing, and full-scale vessel testing; and analyze the applicability and limitations of mainstream algorithms such as the velocity obstacle algorithm, the artificial potential field algorithm, and reinforcement learning. It focuses on the key technologies such as diverse scene generation, local scene slicing, and the construction of an evaluation index system. Finally, this paper summarizes the challenges faced by autonomous collision avoidance performance testing and the assessment of intelligent vessels and proposes potential technical solutions and future development directions in terms of virtual–real fusion testing, dynamic evaluation index optimization, and multimodal algorithm co-validation, aiming to provide a reference for the further development of this field.

1. Introduction

Driven by technologies such as artificial intelligence, intelligent vessels, as a key component in the future digital and intelligent transformation, represent a new engine for promoting high-quality development in the shipping industry. They also present unprecedented challenges for the sector. The development and application of intelligent vessels will profoundly change traditional shipping models, enhance shipping efficiency, reduce operational costs, and strengthen safety capabilities. As a typical example of new productivity in the water transport industry, intelligent vessels will have a far-reaching impact on the global shipping landscape [1]. Ref. [2] provides a comprehensive review of intelligent ships. They are ships that automatically perceive and obtain information and data from the ships themselves, the marine environment, logistics, and the port by making use of sensors, communication, the Internet of Things, the Internet, and other technical means, and they achieve intelligent operation in terms of ship navigation, management, maintenance, and cargo transportation based on computer technology, automatic control technology, and big data processing by analyzing these technologies so that ships can become safer, more environmentally friendly, economical, and efficient.

In terms of key performance for intelligent vessels, autonomous collision avoidance is an essential indicator of their core capabilities. Ref. [3] proposes a diverse scenario generation method involving multiple topographies and accompanying ship behaviors and a navigation state pressure grading method for local scene segmentation and establishes an evaluation index system with 13 indicators. Ref. [4] systematically reviews autonomous collision avoidance from three dimensions: research objects (evolving from traditional ships to MASS), technical methods (focusing on the collision risk, the ship domain, and intelligent decision-making), and algorithms (developing from mathematical models to intelligent algorithms and machine learning). It identifies future trends such as hybrid encounters between traditional and intelligent ships, as well as multi-ship collision avoidance in narrow waterways. Ref. [5] systematically reviews ship autonomous collision avoidance strategies, categorizing current decision-making algorithms into three types: alteration of the course alone, alteration of the speed alone, and alteration of both the course and the speed.

In the field of autonomous collision avoidance for intelligent vessels, existing review articles can be categorized into three aspects:

• Algorithmic-focused reviews, which predominantly synthesize the technical details of collision avoidance algorithms but rarely establish connections between algorithmic characteristics and testing/evaluation systems.
• Testing-centric studies, which elaborate on virtual simulations, model tests, or full-scale trials yet lack an in-depth analysis of how algorithmic performance varies across different testing frameworks.
• Standardization-oriented surveys, which discuss the regulatory progress (such as the IMO MASS Code) without integrating algorithmic adaptability and testing system design into the standardization discourse.

In contrast, this review makes three significant contributions:

• It addresses the lack of integrated comparison between algorithms and testing systems. Previous studies either isolate algorithms from testing methods or treat them as independent entities, failing to clarify how to select testing strategies based on algorithmic features or how testing results can guide algorithm optimization.
• It tackles the insufficient emphasis on dynamic evaluation frameworks. Most reviews focus on static indices but overlook the need for dynamic index adjustments in mixed-traffic environments, such as adaptability to human-like maneuvering of conventional ships.
• It fills the absence of synthesis on standardization bottlenecks. While some studies mention international standards, they rarely analyze contradictions between current standards and algorithmic capabilities or propose feasible solutions to bridge these gaps.

For researchers, this review first provides a decision-making matrix for algorithm-testing matching and establishes a framework integrating autonomous collision avoidance behavior prediction, situational awareness, or digital twin strategies for intelligent vessels. For instance, algorithms with high computational complexity are better validated through model tests combined with digital twins, whereas fast-response algorithms can be efficiently verified via simplified virtual simulations. Second, for practitioners, it proposes a progressive testing strategy aligned with vessel intelligence levels, ranging from virtual simulations for basic collision avoidance functions to full-scale trials in high-density traffic areas, with clear thresholds for transitioning between stages. Third, for standardization bodies, it highlights the necessity of establishing algorithm-classification-based standards, such as formulating differentiated testing requirements for rule-based versus AI-driven collision avoidance systems and incorporating “human–machine interaction safety” into evaluation indices to address challenges in mixed traffic environments.

The structure of the paper is as follows: the first section introduces the research background, current status, and core issues; the second section reviews autonomous collision avoidance algorithms, and analyzes their technical evolution; the third section discusses current testing methodologies and evaluation index systems; the fourth section focuses on the multi-level test and evaluation architecture for autonomous collision avoidance; and the fifth section provides a summary and explores the potential for standardization and industrialization of autonomous collision avoidance technology for intelligent vessels.

2. Overview of Autonomous Collision Avoidance Algorithms for Intelligent Vessels

In the field of autonomous collision avoidance algorithms for intelligent vessels, frameworks integrating situational awareness strategies and behavior prediction are crucial for enhancing maritime navigation safety. These frameworks analyze the vessel’s environmental situation, integrate the prediction of autonomous collision avoidance behaviors of intelligent vessels, and make decisions through diverse approaches, each with unique logics and application scenarios.

2.1. Situation Awareness Methods

Situational Awareness (SA) is a core framework for MASS to make safe and compliant collision avoidance decisions in dynamic marine environments. Ref. [6] focuses on enhancing the situational awareness of stand-on ships by estimating the intention of give-way ships, aiming to construct a Stand-on Ship as Second Line of Defense to reduce collision risks. However this method is simplified action modeling, it Assumes give-way ships only alter course to avoid collisions, ignoring speed changes and speed loss during steering, which may lead to inaccurate reachable velocity estimation.

In [7], a local route-planning algorithm that takes collision-avoidance actions in compliance with COLREGs Rules using a fuzzy inference system based on near-collision (FIS-NC), ship domain and velocity obstacle is proposed. The algorithm achieves COLREGs-compliant autonomous collision avoidance by systematically integrating three stages: perception, comprehension, and prediction. In the perception stage, the algorithm utilizes ARPA and AIS to collect key navigation data between MASS and the target ship (TS), such as Distance to the Closest Point of Approach (DCPA), Time to the Closest Point of Approach (TCPA), Variance of Compass Degree (VCD), and relative distance (Dr), these parameters form the basis for subsequent risk assessment. In the comprehension stage (Level 2 SA), the algorithm processes the perceived data to understand collision risks and encounter scenarios. A “Fuzzy Inference System based on Near-Collision Data” (FIS-NC) is adopted to infer a Collision Risk Index (CRI) ranging from 0.00 to 1.00 from DCPA, TCPA, VCD, and Dr. In the prediction stage (Level 3 SA), the algorithm employs a Ship Domain-extended Velocity Obstacle model (SDVO) to generate a safe velocity space, preventing collisions and maintaining a minimum safe distance. For give-way ships, SDVO triggers a starboard turn in head-on, crossing, or overtaking scenarios; for stand-on ships, actions are taken only when CRI equal or greater than 0.33, in compliance with COLREGs Rule 17. Figure 1 shows an illustration of the CPA. The closest point of approach (CPA) is the point where the MASS is closest to the TS at any time. Let the coordinate, course, and velocity of the MASS be $(x_0,y_0)$, ${\varphi }_0$, and $V_0$, and those of the TS be $(x_t,y_t)$, ${\varphi }_t$, and $V_t$. Then, the time until the CPA and the distance between the MASS and TS at the CPA are calculated using an automatic radar plotting aid (ARPA) and an automatic identification system (AIS) equipped in all cargo vessels [7], as follows:

$$D_r=\sqrt{{(x_t-x_0)}^2+{(y_t-y_0)}^2}$$

(1)

$$V_r=V_0\times \sqrt{1+({\displaystyle \frac{V_t}{V_0}}{)}^2-2\times {\displaystyle \frac{V_t}{V_0}}\times \cos ({\varphi }_0-{\varphi }_t)}$$

(2)

$${\varphi }_r={\cos }^{-1}\times ({\displaystyle \frac{V_0-V_t\times \cos ({\varphi }_0-{\varphi }_t)}{V_r}})$$

(3)

$$T_{CPA}=D_r\times \cos ({\varphi }_r-{\alpha }_t-\pi )/V_r$$

(4)

$$D_{CPA}=D_r\times \sin ({\varphi }_r-{\alpha }_t-\pi )$$

(5)

where $D_r$ is the relative distance between the MASS and the TS, $V_r$ is the relative velocity, ${\varphi }_r$ is the relative course, ${\alpha }_t$ is the azimuth of the TS, and ${\alpha }_r$ is the relative bearing. $T_{CPA}$ is the time to the CPA, and $D_{CPA}$ is the distance between the MASS and the TS at that time.

In [8], researcher analyzes AIS data to characterize the maritime traffic features along the coast of Portugal and provided statistical analyses within the Traffic Separation Scheme (TSS). The paper develops an algorithm for assessing risk profiles and the importance of port-related shipping routes. By estimating the future distances between vessels based on their previous position, heading, and speed, and comparing these distances with defined collision diameters, the algorithm calculates collision risks from the evaluation of the number of collision candidates.

In [9], a Closest Point of Approach (CPA) calculation method based on AIS data is proposed. The method constructs a vessel position prediction model by integrating Speed Over Ground (SOG), Course Over Ground (COG), Change Of Speed (COS), and Rate Of Turn (ROT). It employs the hill climbing algorithm and golden section search algorithm to handle asynchronous AIS data, enabling more accurate CPA calculations. This approach can effectively identify abnormal vessel behaviors and provide early warnings 30 s earlier than traditional methods.

In [10], researcher proposes a double GRU-RNN method combining AIS big data and deep learning, providing a new paradigm for intelligent vessel collision-avoidance decision-making. Using the 2015 annual AIS data of Tianjin Port, which includes 22,349 ships and approximately 1.5 billion vessel trajectory points, the method employs the Douglas Peucker (DP) algorithm for trajectory point compression, risk identification, and normalization processing. The GRU units, through their update gate and reset gate mechanisms, effectively mitigate the gradient disappearance problem in long-sequence training, enabling advance prediction of collision risks and generation of collision-avoidance paths to enhance the autonomous navigation capability of intelligent vessels in complex water areas.

In [11], to address the intelligent collision avoidance problem in multi-vessel encounter scenarios, researchers propose a method based on the Spatio-Temporal Edge and Node Attention Graph Convolutional Network (ST-ENAGCN). This approach integrates AIS big data and graph theory for modeling, identifies 25 types of two-vessel encounter patterns through an 8-azimuth map, constructs multi-vessel encounter graph structures via cross-matching, and utilizes GCN and LSTM to extract spatio-temporal features, enabling collision avoidance decision learning for multi-vessel scenarios.

In [12], aiming at the problems of unstable GPS signals in narrow waterways and the limitations of traditional GPS-dependent navigation methods, researchers propose a method to achieve reliable vessel navigation and collision avoidance by integrating perception sensors such as radar data and video data. This approach fuses the detection results of vision and LiDAR, optimizes trajectories through Model Predictive Control (MPC), and considers vessel path tracking, boundary obstacle avoidance, and input constraints (rudder angle limitations), providing a reliable solution for autonomous vessel navigation in narrow water areas.

2.2. Incorporating Behavior Prediction Methods

Behavior prediction has emerged as a key technology for enhancing the collision avoidance capability of MASS by leveraging historical trajectory data and real-time environmental information. By predicting the movement intentions of target ships and embedding such predictions into decision-making frameworks, modern algorithms can achieve more proactive navigation behaviors that better comply with COLREGs [1]. This section systematically reviews the critical advancements in this field, focusing on aspects such as the evolutionary process of autonomous collision avoidance algorithms, evidence-based critical comparison of algorithms, the practical implementation of COLREGs, and the capability and methods for handling multi-ship scenarios.

2.2.1. Evolution Process of Autonomous Collision Avoidance Algorithms

The incorporating behavior prediction methods takes the COLREGs as the core framework, and combines geometric modeling, expert systems and fuzzy logic and other technologies to build a collision avoidance decision-making logic that conforms to the norms of nautical practice [2]. Its development history can be traced back to the 1960s. Early research was dominated by geometric models and rule-based methods, and then expert systems and fuzzy reasoning techniques were gradually integrated to form an interpretable and standardized collision avoidance strategy system. Figure 2 shows the evolution of research related to regulation-driven approaches. Early incorporating of behavior prediction methods was performed to analyze the ship’s motion trajectory through geometric relationships and generate evasive maneuvers in conjunction with COLREGs provisions. Ship Domain is the core concept for collision risk assessment, and its development is closely related to COLREGs. In 1971, Fujii and Tanaka [13] firstly proposed an ellipsoidal ship domain centered on the ship based on the traffic study of the Japanese waterways, which defines the safe space range that the crew needs to maintain, and became the cornerstone of subsequent studies [14,15,16]. In 1975, Goodwin further discretized the ship domain into three dynamic zones, give way zone, straight ahead zone, and pursuit zone, which correspond to the collision avoidance responsibilities under different encounter scenarios [17]. The radius of a ship’s domain is affected by a variety of factors, including the type of water, traffic density, vessel size and speed, and its dynamic adjustment characteristics make it a key basis for collision risk assessment. The concept is still an important theoretical basis for the design of nautical safety and autonomous collision avoidance algorithms [18,19,20,21]. With the development of computer technology, expert systems and fuzzy logic have been introduced into collision avoidance decision-making, which enhances the adaptability and flexibility of regulation-driven.

2.2.2. Autonomous Collision Avoidance Algorithms

The current mainstream behavior prediction-enhanced collision avoidance algorithms include the Dynamic Window Algorithm (DWA), Velocity Obstacle algorithm (VO), Artificial Potential Field algorithm (APF), and Model Predictive Control (MPC). They exhibit distinct characteristics in terms of uncertainty handling capability and rule compliance.

Dynamic Window Approach

The Dynamic Window Approach is a landmark algorithm in the field of robot navigation and was proposed by Fox in 1997 [22]. Figure 3 shows the main principles of the dynamic windowing algorithm. The core mechanism of the thesis lies in real-time optimization of path planning through dynamic velocity windows. The algorithm selects the optimal solution from the set of feasible velocities that satisfy the dynamics constraints based on the current robot motion state, and realizes the dynamic adjustment of the trajectory through the multi-objective evaluation function, which covers obstacle avoidance, path efficiency, and goal convergence. Its efficient computational characteristics and dynamic environment adaptability make it the preferred solution for real-time obstacle avoidance and local path optimization.

In terms of theoretical validation and extension of the algorithm. Aiming at the convergence problem of the traditional DWA, Ref. [23] presents an improved method based on the optimal control discretization and Lyapunov function, proved the convergence of the algorithm, and experimentally verified its effectiveness in complex environments.

Furthermore, Ref. [24] proposes two Unmanned Ground Vehicle (UGV) obstacle avoidance algorithms, DW4DO and DW4DOT, to cope with dynamic obstacles. DW4DO is based on the dynamic windowing algorithm, which improves safety by double occupancy grids, considers historical speeds to stabilize the paths and keeps the computation time stable. DW4DOT is further extended as a global planner, which reduces the number of path evaluations by a tree construction method. These algorithms perform better in the real environment. These algorithms outperform traditional methods in real environments.

The application of DWA in intelligent vessels breaks through the boundaries of traditional robotics field, algorithm needs to solve the dual challenges of vessel-specific kinematic constraints (large inertia, low steering rate) and marine environment uncertainties (ocean current interference, high density of dynamic obstacles). Ref. [25] proposes a hybrid method combining the improved A-algorithm and the optimized Dynamic Window Approach (DWA). The improved DWA combines path smoothing coefficients and a local target selection strategy to successfully avoid dynamic obstacles, which improves the safety and stability of Unmanned Surface Vehicles (USV) local planning.

In order to mitigate the local optimal problem that local path planning may encounter in complex environments, Ref. [26] adopts the Dynamic Window Algorithm (DWA) and formulates a new cost function that takes into account the maneuvering characteristics, dynamic constraints, and environmental information of USVs, which effectively solves the challenges faced by USVs in path planning and collision avoidance in complex environments.

Based on the combination of ontology-based Knowledge Mapping (KM) and Dynamic Window Approach (DWA), Ref. [27] proposes a fusion method called KM-DWA to enhance the navigation and path planning of Maritime Autonomous Surface Ship (MASS) at sea. By comparing the KM-DWA with the basic DWA in single and multi-ship encounter scenarios, the algorithm is verified to be superior in terms of compliance with COLREGs and collision avoidance rates. Figure 4 shows the extension of the traditional DWA to address the motion characteristics and collision avoidance needs of maritime autonomous surface ships (MASS). The core mechanism of the KM-DWA is to adapt the MASS dynamics through the assumption of circular motion, and utilize the predictive trajectory sampling and safety buffer to achieve the collision avoidance decision in accordance with the COLREGs.

The limitations of the current DWA in autonomous collision avoidance for intelligent vessels are mainly reflected in the following aspects. First, the dynamics modeling is insufficient, the vessels’ underactuation feature and complex hydrodynamic effects lead to the accumulation of errors in the traditional DWA kinematic model, which requires the introduction of six-degree-of-freedom equations of motion for trajectory prediction correction. Secondly, environment perception dependence, existing research is mostly based on idealized raster maps, while the actual marine environment needs to integrate multi-source sensor (radar, AIS, vision) data to construct dynamic obstacle field, which puts forward higher requirements on real-time data processing capability. Thirdly, the evaluation function limitation, the incorporation of COLREGs has not been perfected, and the current algorithms do not fully consider the constraints of COLREGs on path decision-making, such as starboard, overtaking. More attention should be paid to the multi-scale planning architecture in future research to construct a hierarchical framework of global A*algorithm, local DWA, regulation-driven, and encode the international maritime regulations as hard constraint terms of the evaluation function. Meanwhile, digital twin validation should be conducted based on a high-fidelity ship motion simulation platform, aiming to quantify the sensitivity of DWA parameters to wind, wave, and current disturbances and establish an adaptive parameter adjustment model. Additionally, efforts shall be made to extend swarm intelligence and integrate multi-agent reinforcement learning, so as to realize distributed DWA strategy optimization for multi-MASS cooperative collision avoidance and avoid the computational bottleneck of centralized planning.

Velocity Obstacle Algorithm

Velocity obstacle algorithm is a motion planning method for dynamic obstacle avoidance. In [28], the VO algorithm avoids collision by calculating the velocity space of an intelligent body, which is applicable to the dynamic obstacle avoidance problem in multi-self-subject systems. Compared with other obstacle avoidance strategies, the velocity obstacle method has the advantages of high computational efficiency and fast decision-making for obstacle avoidance, so it has gained wide attention and application in unmanned aerial vehicles, unmanned surface vehicles, autonomous vehicles and industrial automation. With the continuous development of technology, the velocity obstacle method is also under constant optimization and improvement to adapt to the increasingly complex and dynamically changing environmental requirements.

Figure 5 represents the basic principle of the velocity barrier algorithm, the red conical region represents the velocity barrier region, which in the velocity barrier algorithm is calculated based on the position of the obstacle, its velocity, and the relative position between the intelligent body and the obstacle. If the velocity vector of the intelligent agent falls into this region, it will collide with the obstacle at some point in the future. Assuming that the intelligent agent is moving at a certain velocity, if the vector corresponding to that velocity is within the red conical region, the trajectories of the intelligent agent and the obstacle will intersect over time, resulting in a collision.

In [29], the researchers propose the Reciprocal Velocity Obstacle (RVO) theory. The core innovation lies in having each agent assume 50% of the avoidance responsibility while anticipating symmetric behavior from other agents. This is achieved through a geometric transformation that translates the VO to the midpoint between the agents’ velocities, thereby enforcing coordinated velocity adjustments. The RVO framework effectively resolves the oscillation problem and collision risks inherent in traditional VO approaches while maintaining computational efficiency. Furthermore, it provides a scalable solution for real-time multi-agent navigation that significantly improves trajectory optimality in dense environments compared to conventional methods.

In order to realize collision-free and oscillation-free navigation of multi-mobile robots or virtual agents, Ref. [30] proposes a Hybrid Reciprocal Velocity Obstacle (HRVO) method for the realization of collision-free and oscillation-free navigation of multi-mobile robots or virtual agents. HRVO, based on RVO, expands the obstacle boundaries on the “wrong” side of the passage according to the current velocity position. When a robot selects the “wrong” side, it must completely avoid the obstacle; when it chooses the “right” side, reciprocity is maintained, thereby avoiding the “reciprocal dance” oscillation. This method predicts the trajectories of other robots by having each robot assume half the responsibility for obstacle avoidance, integrating the current position and velocity, and prevents the “reciprocal dance” oscillation by expanding the velocity barrier on the “wrong” side of the passage.

In [31], researchers propose a collision avoidance decision-making system for inland vessels based on the velocity obstacle (VO) algorithm in response to the problem that inland vessel collisions are frequent and the traditional collision avoidance algorithms are not applicable to the inland river environment. The method adopts the grid method to model static obstacles such as shoals and reefs, and constructs the collision avoidance area based on the traditional VO algorithm; in terms of dynamic obstacles, the algorithm replaces the traditional circular domain with an elliptical ship domain, and calculates the common tangent of the two ellipses through analytical geometry to construct the relative collision cone (RCC) and the absolute collision cones (ACC) and improve the collision avoidance accuracy. Figure 6 illustrates the principle of Relative Collision Cone (RCC) and Absolute Collision Cone (ACC) in the Velocity Obstacle (VO) algorithm. When the relative velocity of the vessel falls within the RCC, the two vessels will collide. By analyzing the relative motion, the collision risk area can be identified in advance in the velocity space. In ACC, the collision region formed by the target vessel to the own vessel is the conversion of RCC to ACC. ACC converts the collision risk of relative motion into the collision avoidance decision region in absolute velocity space, which directly guides the speed adjustment of its own vessel.

In [32], in order to resolve the problem of optimizing vessel navigation strategies in dynamic sea ice environments, the author proposes a decision support framework based on the Flexible Velocity Obstacle (FVO) algorithm. The method classifies the safe distance between the ship and the sea ice into three levels, and prioritizes guiding the ship through the edge of the sea ice instead of the central area. The algorithm incorporates five factors, DCPA, TCPA, Distance of sea ice from the ship (Dr), the variance of compass degree (VCD), and area of sea ice (AREA)to establish a multifactorial risk assessment model, and selects the optimal speed and heading according to the prioritization of the risk level within the speed search space to avoid high-risk areas.

In [33], for the collision avoidance problem of MASS in dynamic obstacle environments, the researcher proposes an intelligent decision-making method that integrates deep Q-network (DQN) and velocity obstacle (VO) algorithms. The algorithm calculates the relative collision cone (CC) based on the safety domain of the target vessel, which is converted to the VO cone in absolute velocity space by Minkowski summing to avoid the speed of the ship from falling into the danger zone, and then the algorithm designs the six reward functions including the speed barrier reward to balance the safety of collision avoidance with the regulation compliance. The algorithm is responsible for long-term strategy learning through DQN, and VO provides short-term geometric collision avoidance constraints [34,35,36,37,38,39,40,41]. The combination of the two methods improve the decision robustness in complex scenarios.

Artificial Potential Field Algorithm

The Artificial Potential Field (APF) algorithm is proposed by Khatib in 1986 [42]. The core idea of the algorithm is to model the motion environment of the robot as a virtual potential field, and guide the motion direction through the superposition of the gravitational field and the repulsive field. The gravitational field drives the subject to move towards the target point, while the repulsive field makes it avoid obstacles, and then the direction of the combined force is solved by the gradient descent method, which has the advantages of simple computation and real-time performance.

Ref. [42] use $p_a$ and $p_g$ to represent the coordinates of the agent and goal. The attractive potential field function $U_{att}(p_a)$ and repulsion field function $U_{rep}(p_a)$ can be written as follows:

$$U_{att}(p_a)={\displaystyle \frac{1}{2}}\xi \rho {(p_a,p_g)}^2$$

(6)

$$U_{rep}(p_a)=\left\{\begin{array}{c}{{\displaystyle \frac{1}{2}}\eta ({\displaystyle \frac{1}{D}}-{\displaystyle \frac{1}{D_{rep}}})}^2,D\le D_{rep}\\ 0,D>D_{rep}\end{array}\right.$$

(7)

where $\xi$ and $\eta$ represent the coefficient of attraction and repulsion, and $\rho (p_a,p_g)$ represent the Euclidean distance between the agent and the goal. The Euclidean distance between the agent and the obstacle is represented by D, and $D_{rep}$ represents the distance threshold of the repulsion potential field. The total potential field $U(p_a)$ and the resultant force F can be written as follows:

$$U(P_a)=U_{rep}(P_a)+U_{att}(P_a)$$

(8)

$$F=-\nabla U(p_a)$$

(9)

Figure 7 illustrates the robot path planning process using the artificial potential field algorithm. Figure 7a is the repulsive potential field, whose potential value increases with decreasing distance from the obstacle, keeping the robot away from the obstacle; Figure 7b is the attraction potential field, whose potential value increases with increasing distance from the target point, guiding the robot toward the target; Figure 7c is the total potential field, where the repulsive potential field and the attractive potential field are superimposed to form a potential field distribution of “high potential near the obstacle and low potential near the target point”; Figure 7d is the robot motion path, and the red curve is the trajectory of the robot from the starting point to the target point, and it descends along the negative gradient of the total potential field, avoiding the obstacle and converging to the target; Figure 7e is the gradient vector field, where the blue arrow indicates the direction of the gradient of the potential field, which leads the robot away from the obstacle. The red circle represents the initial motion position, and the green circle represents the position to be reached. The blue arrow indicates the direction of the gradient of the potential field, and the robot moves in the opposite direction (negative gradient) of the arrow. Through the force field simulation of the potential field, the path planning is transformed into a gradient descent problem in the potential field to realize the unification of obstacle avoidance and goal orientation.

The artificial potential field (APF) method provides simple and efficient path planning algorithms for engineering applications and academic research. However, in artificial potential field construction, after the combined gravitational and repulsive forces are zero or pointing to a local minimal superposition, the low-potential region of the non-target point may become “trap”, leading to path jamming, and the path-planning trajectory may be terminated at the local optimum.

In [43], the researcher focuses on the local minimum problem existing in artificial potential field algorithm (APF) and proposes the concept of virtual obstacle and the discrete modeling method. When the robot is trapped in a local minimum, a virtual obstacle is set at the trap point, which generates a repulsive force through the additional potential field to make it escape. This method can effectively help the robot avoid local minima in various environments containing concave and deep channel-shaped obstacles, and successfully reach the target point.

Aiming at the path planning and formation control problem of multiple unmanned aerial vehicles (multi-UAVs) in three-dimensional constrained space, Ref. [44] proposes an improved artificial potential field algorithm (IAPF), which solves the local minimum and oscillations of the traditional APF by introducing a rotating potential field, designs the formation controller by using the leader–follower model, and proves the stability of the system by combining with the Lyapunov function. The method can effectively realize collision-free path planning and maintain the desired formation shape in three-dimensional static, complex obstacle and random obstacle environments.

In [45], the researcher addresses the real-time path planning challenge for autonomous ships operating in complex dynamic maritime environments by proposing a modified artificial potential field (APF) algorithm. The modified method incorporates a partitioned virtual repulsive field function and virtual force constraints aligned with COLREGs, enabling effective collision avoidance against both dynamic and static obstacles. A key contribution is the mitigation of local minimum, particularly in emergency encounter scenarios. The algorithm demonstrates millisecond-level computational efficiency in multi-vessels encounters and dynamically changing environments while ensuring compliance with COLREGS Rules 13–17. Additionally, it robustly handles uncoordinated maneuvers by target vessels, offering a high-performance solution for autonomous vessel collision avoidance systems.

In [46], the researcher proposes a Discrete Artificial Potential Field (DAPF) method, which divides the space into cells labeled as free, obstacle, start, and goal cells, with the obstacle cell potential value set to infinity, the goal cell potential value to zero, the free cell potential value incrementing with the distance from the goal, and with the static obstacles set to polygons, and the dynamic target vessel represented by a hexagonal field. The method is combined with a Path Optimization Algorithm (POA) to generate collision-free paths that comply with the COLREGs rule. The algorithm is verified by three real nautical scenarios, and the algorithm can complete the planning within 0.45–0.7 s, and the paths satisfy the COLREGs Rules 8, 13, 14, and 15, which provides an efficient solution for the automatic collision avoidance system for vessels.

In [47], in terms of the safety and efficiency of ship navigation in narrow inland waterways, the researcher proposes a real-time vessel path planning model based on the improved APF algorithm for guiding inland vessels to avoid collisions autonomously. The method establishes the attraction potential field, obstacle repulsion potential field, velocity repulsion potential field and boundary repulsion potential field, and triggers the obstacle repulsion field by superimposing the elliptic ship domain of the own vessel and the target vessel to determine whether it intrudes into the safety region; meanwhile, the method introduces the attraction steering point for guiding the ship’s navigation in the curved area to reduce the local minimum problem, and successfully plans a safe and smooth autonomous collision avoidance path [48,49,50,51,52,53,54,55,56,57].

Model Predictive Control (MPC)

Model Predictive Control (MPC) is a control strategy that obtains the current control action by solving an open-loop optimal control problem with a finite time horizon online at each sampling moment [58]. The theoretical basis of MPC can be traced back to the theory of optimal control in the 1960s, in [59], researcher provides a theoretical basis for the online computation of MPC by measuring the state of the current control process and then quickly computing the first part of the open-loop control function, using this part for a short period of time, and then measuring the state again and re-computing the control function afterward, an idea that provides a theoretical basis for the online computation of MPC. Ref. [60] presents the first application of MPC to process control by proposing the IDCOM (Identification and Command) methodology, which employs a finite-time impulse response (linear) model, a quadratic cost function, and input and output constraints. Ref. [61] uses the value function as Lyapunov function for the first time, which laid the foundation for the stability analysis of MPC. MPC studies for nonlinear and uncertain systems are gradually increasing, and researchers have proposed methods to introduce robust control Lyapunov functions in the terminal constraints and terminal costs to ensure the stability of the system in the presence of uncertainty [62,63].

Model predictive control (MPC) is a dynamic model-based multivariate optimal control strategy, which is mainly used to achieve optimal control of the system through rolling optimization and feedback correction [64]. Since MPC has the ability of multivariate cooperative control, which can simultaneously optimize the vessel’s trajectory, speed and collision avoidance, and at the same time can explicitly deal with complex constraints such as COLREGs regulations, vessel maneuverability restrictions, etc., and can be adapted to dynamic environments such as multi-vessel encounters, narrow waterway navigation and other scenarios, it is receiving more and more attention and research by researchers in the field of autonomous collision avoidance of intelligent vessels.

In [65], researcher focuses on the problem of vessel collision avoidance under multiple uncertainties. The paper designs a vessel collision avoidance framework that combines robust motion control of own vessel and probabilistic prediction of the behavior of target vessels. In this study, a motion control method based on tube-based MPC is designed to achieve robust trajectory tracking, while considering the uncertainties of vessel motion and external disturbances. A high-precision probabilistic trajectory prediction method based on Gaussian Process Regression (GPR) and incremental theory is put forward to describe the uncertain behavior of target vessels. Meanwhile, this method incorporates the Artificial Potential Field (APF) approach to handle collision avoidance constraints within the tube-based MPC, effectively reducing computational complexity. Simulations validate the effectiveness of the proposed vessel collision avoidance framework.

In [66], researcher proposes a Model Predictive Control (MPC) method that utilizes intention data of the target vessel and quaternion ship domain model to achieve collision avoidance while considering COLREGs. This approach is based on the theory of MPC, aiming to minimize the risk of collision. It integrates a three-degree-of-freedom vessel dynamics model to optimize the heading offset of the own vessel, ensuring that the avoidance maneuvers are both effective and compliant with physical constraints. By leveraging the intention data of the target vessel, the accuracy of trajectory prediction is enhanced. The MPC framework ensures the optimality and safety of the control actions.

To address the challenge of path following and collision avoidance control for underactuated vessels, Ref. [67] presents an integrated approach based on model-free adaptive predictive control (MFAPC). The approach fuses the line-of-sight (LOS) algorithm and a modified velocity obstacle (VO) algorithm, with collision risk assessment (DCPA/TCPA) determining the need for collision avoidance. The improved VO algorithm enhances collision avoidance smoothness and compliance by expanding the boundary of the VO area and incorporating COLREGs. Meanwhile, the MFAPC controller synthesizes the heading angle and yaw rate as the output to strengthen input-output correspondence, enabling model-free precise heading control.

With its capabilities in multivariable control and constraint handling, Model Predictive Control (MPC) has become a core algorithm for autonomous collision avoidance of intelligent vessels. Future research should focus on intelligent-enhanced MPC, which leverages deep learning to predict environmental disturbances and improve model adaptability, while using deep reinforcement learning (DRL) to automatically adjust weight parameters for adapting to multi-variable scenarios [68]. Meanwhile, simulations of autonomous collision avoidance for intelligent vessels need to overcome bottlenecks such as computational efficiency, multi-ship coordination, and rule adaptability. Integrating AI technologies with edge computing will promote the development of autonomous navigation toward safer and more efficient operations.

Based on the above systematic review of autonomous collision avoidance algorithms for intelligent ships, the Dynamic Window Algorithm (DWA), Velocity Obstacle (VO) algorithm, Artificial Potential Field (APF) algorithm, and Model Predictive Control (MPC) algorithm exhibit differences in terms of basic principles, collision avoidance success rates, and computational costs. Detailed comparative information of the above algorithms is presented in

Table 1. Critical Comparison of Autonomous Collision Avoidance Algorithms.

Algorithm	DCPA/TCPA Performance	Computational Cost	Success Rates in Maritime Traffic
Dynamic Window Approach (DWA)	Medium (DWA has trajectory prediction errors affecting DCPA/TCPA accuracy, though improvements with 6-DOF equations can enhance it)	High (Samples velocity commands and constructs paths, but new non-discrete path representation aims to reduce this)	Medium (Prone to convergence and local optimality issues in complex scenarios, despite improved versions)
Velocity Obstacle Algorithm (VO)	Medium (Can calculate DCPA/TCPA from COLREGS-based geometric constraints, yet may not cover all COLREGs scenarios accurately)	High (Complexity spikes in multi-vessel and dynamic-obstacle–rich scenarios)	Low (In real-world, complex marine conditions, prediction inaccuracies can cause jitter or failure in collision-avoidance paths)
Artificial Potential Field Algorithm (APF)	Low (Prone to local minima, which can lead to sub-optimal paths and affect DCPA/TCPA)	Low (Fast-executing with low computational complexity)	Low (Local minima in certain obstacles and multi-vessel coordination issues reduce success rate)
Model Predictive Control (MPC)	High (Explicitly handles COLREGs and other constraints, optimizing for effective DCPA/TCPA control)	Medium (Solves an open-loop optimal control problem online, but efficiency can be optimized compared to some algorithms)	High (Shows good performance in simulations, though practical challenges remain)

. The DWA can optimize paths in real time, but insufficient dynamic modeling affects the accuracy of DCPA/TCPA calculations, resulting in high computational costs. It achieves a moderate success rate in complex scenarios and is often used for local path planning in dynamic environments. The VO algorithm calculates collision avoidance in the velocity space but insufficiently covers complex COLREGs scenarios. Its computational cost increases significantly in multi-ship situations, and it has a low success rate in complex marine environments, making it suitable for dynamic obstacle avoidance among multiple autonomous agents. The APF algorithm constructs a virtual potential field to guide motion but is prone to falling into local minima. With low computational costs, it shows low success rates in complex and multi-ship coordination scenarios, and is commonly applied to path planning for indoor mobile robots and similar systems. The MPC algorithm has strong capabilities in handling complex constraints, with high accuracy in DCPA/TCPA calculations and moderate computational costs. It achieves a high success rate in simulated scenarios, making it suitable for autonomous collision avoidance of ships in complex dynamic marine environments.

2.2.3. The Practical Implementation of COLREGs

Autonomous collision avoidance behavior prediction methods aim to predict a ship’s future behavior by analyzing historical trajectories, surrounding environmental factors, and other relevant data.

With the increasing deployment of MASS, they must be equipped with robust autonomous prediction capabilities, not only to identify collision risks in advance but also to ensure strict compliance with the International Regulations for COLREGs. Traditional path planning frameworks, which mostly rely on reactive or rule-driven mechanisms, have limitations in dynamic environments. Recent research has shifted toward autonomous prediction models, which enable early anticipation and proactive avoidance of collision risks by integrating situational awareness, real navigation data, and regulatory logic.

COLREGs stipulate that vessels shall at all times use all available means appropriate to the prevailing circumstances and conditions to determine the risk of collision. If a collision risk exists, evasive actions shall be taken as early as possible to ensure passing at a safe distance. Figure 8 fully covers the core collision avoidance rules of COLREGs, ranging from “Look-out (Rule 5)” to “Risk of collision (Rule 7)”, then to “Identification of encounter relations (Rules 13–15)”, “Assignment of action roles (Rules 16–17)”, and “Action to avoid collision (Rule 8)”, providing a clear “rule mapping path” for MASS algorithm development.

Ref. [69] systematically defines requirements for optimal local route planning of autonomous ships by bridging navigation practices and COLREGs. The requirements enable timely evaluation of algorithm effectiveness and guide improvements in existing methods. Figure 9 systematically illustrates the correspondence between ship navigation practices and key provisions of Part B of COLREGs. The left side presents navigation practices summarized based on a survey of 29 Officers of the Watch (OOW), covering seven core dimensions: ship anti-collision responsibility, environmental interference, ship maneuverability, target ship movement, ship domain, expert experience (e.g., good seamanship), and collision avoidance actions. It also distinguishes between scenarios of open waters and restricted waters, as well as two-ship and multi-ship encounters. The right side corresponds to the core rules of COLREGs Part B (Rules 5–10, 13–18), including look-out (Rule 5), collision risk assessment (Rule 7), navigation in narrow channels (Rule 9), encounter situations (overtaking, head-on, crossing, Rules 13–15), and requirements for collision avoidance actions (Rules 8, 16–17), etc. This relationship diagram can serve as a fundamental framework for subsequent verification of the compliance of autonomous collision avoidance algorithms such as APF and VO, helping to determine whether the algorithms meet the core requirements of navigation practices and regulations.

Ref. [70] proposes a Fuzzy Inference System based on Near-Collision data (FIS-NC), which embeds COLREGs-compliant decision logic into a fuzzy inference framework. This system achieves autonomous risk prediction through four key predictive variables: Distance to the Closest Point of Approach (DCPA), Time to the Closest Point of Approach (TCPA), Variance of Compass Bearing (VCD), and relative distance (Dr). These variables are processed by an Adaptive Neuro-Fuzzy Inference System (ANFIS), trained on 1972 real AIS near-collision datasets from Korean waters, to output a Collision Risk Index (CRI) ranging from 0.00 to 1.00. The CRI is mapped to four predictive levels: “Attention,” “Threat,” “Danger,” and “Collision.”

In accordance with COLREGs Rules 5, 7, 8, and 13–17, the system autonomously sets trigger thresholds: CRI greater than or equal to 0.01 for give-way vessels and CRI greater than or equal to 0.33 for stand-on vessels, enabling proactive autonomous collision avoidance. Comparisons with existing studies demonstrate that FIS-NC overcomes the limitation of previous research [71,72,73,74,75,76,77,78], which failed to incorporate all vital variables for COLREGs-compliant collision avoidance. Furthermore, the system can infer the CRI at optimal positions and timings, allowing MASS to take early collision avoidance actions in any scenario [79,80,81,82]. Consequently, it provides decision-makers with more time to implement appropriate collision prevention measures.

2.2.4. Multi-Ship Scenario

In dense traffic environments, behavior prediction is particularly crucial for handling nonlinear interactions. In [7], this study focuses on the collision avoidance problem of MASS. Addressing the deficiency that existing research fails to fully comply with Articles 5, 7, 8, and 13–17 of the COLREGs, a local route-planning algorithm, SDVO + FIS–NC, is proposed. It is based on FIS–NC (Near-Collision Fuzzy Inference System), SD (Ship Domain), and VO (Velocity Obstacle). In multi-ship encounter scenarios, SDVO + FIS–NC processes multi–ship encounters through CRI (Collision Risk Index)-based ranking, dynamically adjusts the collision-avoidance sequence and predicts the behavior of target ships. Meanwhile, FIS-NC leverages AIS data to learn historical interaction patterns and predict cooperative behaviors, reducing the collective collision risk by 30% in 5–ship scenarios.

Figure 10 presents the simulation results of multi-ship collision avoidance scenarios, illustrating the dynamic process of MASS from its initial state to gradual collision avoidance and approaching the target waypoint in a multi-target ship environment, sequentially from Figure 10a–d. This set of figures dynamically validates the multi-ship collision avoidance logic of the SDVO + FIS-NC algorithm from a spatial dimension: it follows hierarchical responses to COLREGs rules (adopting different turning strategies for head-on, crossing, and overtaking scenarios); ensures safety distance by extending Velocity Obstacle (VO) with SD (avoiding near-collisions caused by excessively small safety zones in conventional VO); and quantifies risks via FIS-NC (CRI precisely controls collision avoidance timing to reduce misoperations). Ultimately, it enables the MASS to achieve safe collision avoidance and efficient route recovery in complex multi-ship scenarios.

3. Overview of Autonomous Collision Avoidance Performance Testing and Evaluation System for Intelligent Vessels

One of the core functions of intelligent navigation systems is their collision avoidance decision-making and planning capabilities, which directly impact navigation safety and traffic efficiency, particularly for cargo vessels. In recent years, the rapid development of intelligent ships in various countries and the MASS code proposed by IMO have propelled autonomous collision avoidance technology from theoretical research to engineering applications. However, the safety and reliability of autonomous collision avoidance systems depend on a rigorous testing and evaluation system. At its 101st session of the Maritime Safety Committee (MSC), Ref. [83] aims to enhance the safety and efficiency of technical trials for MASS by specifying requirements for their sea trials. Nevertheless, compared to land transportation, the research level on testing and evaluation of autonomous collision avoidance for intelligent vessels still has a significant gap. Therefore, to ensure the safety and reliability of vessel autonomous collision avoidance algorithms, comprehensive and in-depth testing and evaluation research is urgently needed. This paper systematically reviews the progress in testing and evaluation of autonomous collision avoidance for intelligent vessels, filling the gap in this field and providing research and practical guidance for improving navigation safety and autonomy.

3.1. Construction of Autonomous Collision Avoidance Testing Scenarios

In autonomous collision avoidance testing activities for intelligent vessels, according to the IMO MASS code and the testing procedures for intelligent vessels in various countries, the process can be divided into the following stages: testing scenario construction, virtual simulation testing, model testing, full-scale vessel testing, and evaluation and optimization [84].

According to the current research progress, the methods for constructing autonomous collision-avoidance test scenarios for intelligent vessels can be classified into following categories: Real Data-Driven Testing Scenario Extraction, Mechanism Modeling-Based Testing Scenario Reconstruction and Machine Learning-Based Testing Scenario Derivation [85].

1. Real Data-Driven Testing Scenario Extraction.

Starting from historical vessel navigation data, a testing scenario extraction method based on real data is proposed to extract testing scenarios from various sensor data such as AIS and radar.

In [86], a testing method for autonomous collision avoidance systems based on AIS data is proposed. The method extracts 56,952 valid encounter scenarios from real-world AIS data, and randomly generates 2900 testing scenarios based on the probability distributions of parameters such as relative distance and speed. Collision risk is quantified and scenarios are classified using DCPA and TCPA. By integrating navigation data, this approach enhances testing coverage and scenario authenticity, providing a systematic verification process for intelligent vessel testing. While the method demonstrates effective performance in two-vessel scenario testing, multi-vessel scenario testing reveals limitations: 37 out of 1000 three-ship encounter scenarios failed. The primary failure cases occur in crossing situations, where the algorithm’s delayed course-changing maneuvers lead to collisions when target vessels are positioned on the port side with higher speed. These results highlight the need for improved decision-making algorithms in complex multi-ship interactions.

To address the insufficient applicability of autonomous vessel collision avoidance algorithms in complex traffic scenarios such as multi-vessel encounters, Ref. [87] proposes a method for generating vessel encounter scenarios and analyzing their complexity based on AIS data. This method abstracts each encounter scenario as a dynamic topological graph, where nodes represent vessels, edges represent encounter relationships, and edge weights are calculated by integrating distance risk, DCPA risk, and TCPA risk. Scenario complexity is quantified using Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS and the entropy weight method (EWM), considering both fundamental scenario attributes and dynamic evolution characteristics of vessel interactions. Experimental results demonstrate that the method effectively generates realistic scenarios covering dynamic interactions of multiple vessels. In tests of three typical scenarios (head-on, crossing, and overtaking), the algorithm achieved a pass rate of 97.3%, providing a systematic solution for validating autonomous vessel collision avoidance algorithms in complex maritime traffic environments.

2. Mechanism Modeling-Based Testing Scenario Reconstruction.

This approach leverages COLREGs and vessel kinematic models to generate standardized testing scenarios through parametric modeling. By understanding, abstracting, and reconstructing the underlying principles of vessel encounter scenarios, this method effectively addresses the data dependency limitations inherent in real-data-driven scenario extraction methods [88].

In [89], a method for automatically generating collision avoidance testing scenarios for intelligent vessels based on Sobol sequence sampling and clustering analysis is proposed. The method defines navigation parameters such as vessel speed, heading, and meteorological conditions, uses Sobol sequences to generate uniformly distributed potential scenarios, filters hazardous scenarios using COLREGs, constructs a risk vector including distance risks (DCPA/TCPA), velocity obstacle area and meteorological conditions, identifies scenario clusters via k-means clustering, and finally selects the riskiest scenario within each cluster as the testing case. This approach can reduce 10,000 sample scenarios to 181 hazardous scenarios, with a post-clustering scenario reduction of over 60%, significantly improving test efficiency. Notably, the method does not rely on real AIS data, enabling the generation of interaction scenarios involving non-AIS vessels and covering edge cases that traditional methods cannot capture.

In [90], three typical scenarios are constructed to validate the autonomous collision avoidance performance of intelligent vessels: open-water crossing encounters, multi-vessel avoidance, and narrow-channel head-on encounters. The testing scenarios incorporated COLREGs compliance checking mechanism while considering vessel kinematic constraints, dividing hazardous areas into a safety-distance prohibited zone and a rule-constrained zone. A Motion-Planning, Guidance, and Control (MPGC) architecture is developed to simulate challenges such as sudden course changes in dynamic obstacles and sensor noise. The testing scenarios demonstrated the algorithm’s robustness and real-time response capability in complex dynamic environments.

3. Machine Learning-Based Testing Scenario Derivation.

In recent years, significant progress has been made in applying deep reinforcement learning (DRL) to testing scenarios for autonomous collision avoidance in intelligent vessels [91,92,93].

In [94], researcher tests an autonomous collision avoidance algorithm for intelligent vessels based on the Imazu problem and proposed a vessel automatic collision avoidance algorithm based on deep reinforcement learning (DRL) and long short-term memory network (LSTM). By expanding the Obstacle Zone by Target (OZT) and introducing bow crossing range detection, this method achieves compliant collision avoidance with the safe passing distance increased from 0.3 nautical miles to 0.5 nautical miles in continuous action spaces. The algorithm has passed 22 standard Imazu testing scenarios and demonstrated robustness in complex scenarios with five target vessels, achieving a minimum passing distance of 0.753 nautical miles. This study provides a novel approach for autonomous navigation in dynamic multi-vessel environments [95].

Focusing on the problem of generating high-risk test scenarios for vessel autonomous collision avoidance algorithms, Ref. [96] proposes an adaptive generation method based on reinforcement learning (RL). This approach constructs an RL testing framework based on the Markov Decision Process (MDP) and Heterogeneous Multi-Agent System (HMS), where the interfering vessel acts as an RL agent to dynamically adjust strategies using the Independence D3QN algorithm and generate diverse high-risk scenarios. The testing scenarios cover typical situations such as head-on, crossing, overtaking situation, and complex multi-vessel interactions. Compared with traditional data-driven and mechanism-based modeling methods, this method achieves “targeted pressure” through dynamic interaction and reward mechanisms, more accurately exposing algorithmic flaws. It provides a scalable and high-fidelity scenario generation solution for virtual testing of intelligent vessels.

3.2. Virtual Simulation Testing

Virtual simulation testing for autonomous collision avoidance of intelligent vessels is the most direct and commonly used testing method in the research, development, and application of intelligent vessels. Most testing work for autonomous collision avoidance algorithms begins with virtual simulation testing. In early research, virtual simulation testing methods mainly generated test scenarios based on COLREGs. Among these, the Imazu problem has been widely used to testing vessel collision avoidance performance. It includes 22 different vessel encounter scenarios, as shown in Figure 11. In the Figure, the number in the upper-left corner of each box indicates the scenario number of the Imazu problem, and the short lines beside triangles (own vessel) or circles (target vessels) represent the velocity vectors of each ship. Ref. [97] proposes that the mathematical encoding of COLREGs provides a basic rule library for compliance verification modules in virtual simulations. The vessel relative motion diagram model, as a geometric analysis tool for dynamic interaction scenarios in virtual simulations, is used to generate initial testing scenarios for multi-vessel encounters. The “Early, Large, Wide, Clear” collision avoidance strategy is embedded in the evaluation criteria for collision avoidance strategies in the simulation system to compare the performance differences between intelligent algorithms and traditional strategies.

In [98], DNV GL presents a complete simulation testing system for intelligent vessel autonomous collision avoidance. The system includes simulation of environmental conditions, geographic information, and interaction with other maritime traffic. Through the simulation of digital twin and virtual world, the system generates testing scenarios and evaluates their results. The testing system accurately reproduces the vessel’s dynamic characteristics and control logic through digital twin, and simulates complex marine meteorology, waterway topography, and dynamic traffic in combination with a virtual operating environment. In terms of scenario coverage, the system generates typical COLREG scenarios based on historical AIS data, efficiently explores boundary conditions through Bayesian optimization algorithms, and injects uncertainties such as sensor noise and communication delays through robustness testing to verify the COLREG compliance and safety of autonomous navigation systems (ANS). This provides a complete testing system prototype for the reliable deployment of autonomous vessels.

Ref. [99] proposes an autonomous collision-avoidance virtual simulation testing method for intelligent vessels based on collision grids and deep reinforcement learning. The method constructs a dynamic collision grid that divides the vessel’s observation field into 15 m×15 m grid cells, real-timely mapping static obstacles, moving obstacles, and their future trajectory predictions, and embeds COLREGs weights to form an interpretable environmental risk map. Researchers use the Proximal Policy Optimization (PPO) algorithm to train the agent, with a 208-dimensional grid vector as input and discrete speed and steering commands as output. The simulation testing integrates a first-order Nomoto ship dynamics model to validate collision-avoidance performance. The model achieves a collision-avoidance success rate of 94.69% in dense scenarios and adapts to COLREGs. By compressing observation data and explicitly encoding rules, this method improves the real-time performance and decision interpretability of virtual testing, providing an efficient framework for validating autonomous navigation algorithms in complex waterways.

To address the shortage of testing scenarios for autonomous collision avoidance of inland vessels, Ref. [100] proposes a parametric modeling method for testing scenarios of autonomous collision avoidance for inland waterway vessels. This method constructs vessel navigation data acquisition system through the fusion of AIS and radar data. Taking the Three Gorges-Gezhouba waterway as an example, it analyzes and obtains key scenario elements such as vessel traffic flow density, trajectory, speed, and minimum encounter distance. Based on the MMG hydrodynamic model, vessel motion control model is established to achieve the automatic generation of multi-vessel encounter scenarios. However, the testing scenarios are concentrated in specific waterways and do not cover other typical inland waterway scenarios such as canals, bridge areas, and traffic separation areas, so the universality of the model in different inland waterway environments needs further verification.

3.3. Physical Model Testing

The physical model testing of autonomous collision avoidance for intelligent vessels involves using physical vessel models in real water environments to verify the practical performance of autonomous collision avoidance algorithms, control systems, and vessel motion capabilities. Through the operation of physical entities under real or near-real conditions, it evaluates the safety, reliability, and adaptability of collision avoidance strategies [3].

Research [101] presents model-scale testing of the autonomous navigation system for the ARAGON unmanned surface vehicle (USV) in real marine environments. The testing configuration includes an 8 m-long USV platform, three target vessels, and multiple types of sensors. The USV platform is illustrated in Figure 12. Experiments are designed with scenarios such as port-crossing, starboard-crossing, and head-on approaching in the Asan-ho sea area (western Korea) and Suyoung Bay (southern Korea). Complex scenarios are created, including low radar update rates, environmental noise, and parameter sensitivity. These are validated through asynchronous updates, deep learning-based detection, and dynamic threshold filtering. The collision avoidance strategies complied with COLREGs. The experiments verified the system’s reliability in complex traffic conditions, providing model experimental support for the practical application of unmanned vehicles.

Research [102] established a 1:24 scale physical model of the vehicle. The parameters, such as model dimensions, displacement, and maximum speed, are designed based on the scaling of real vessels to ensure that their dynamic characteristics are consistent with those of real vessels. This is to guarantee the reliability and safety of autonomous collision avoidance strategies. The scenario design includes standard collision avoidance maneuvers (CA) and last moment maneuvers (LMM), simulating encounters between two or more ships to test the system’s ability to generate safe trajectories. This method verifies the feasibility of the system architecture and algorithms at the hardware-software integration level through physical model testing, reducing the risks and costs of real-vessel testing.

3.4. Full-Scale Vessel Testing

Full-scale testing of autonomous collision avoidance for intelligent vessels involves deploying vessels with autonomous navigation and collision avoidance capabilities in real marine environments to verify the comprehensive performance of their perception, decision-making, and control systems.

In [103], the world’s first fully electric autonomous container ship Yara Birkeland entered operation, undergoing multiple collision avoidance tests, and achieved fully autonomous navigation in March 2023, verifying its collision avoidance capability in complex scenarios [104,105,106,107].

In [108], the Japanese container ship Suzaku completed a 790 km round-trip autonomous voyage in Tokyo Bay. Suzaku uses an on-board navigation control system to analyze the dynamics of surrounding vessels in real time, generate optimal collision avoidance paths, and is remotely monitored by a shore-based support center. It successfully avoided 400–500 vessels with 107 collision avoidance maneuvers, marking a major breakthrough in autonomous collision avoidance of intelligent vessels in congested waters.

In [16], Korean shipbuilding enterprises and research institutions have conducted a series of sea trials focusing on autonomous collision avoidance technologies. In July 2023, a 15,000 TEU container ship equipped with SAS (Self-autonomous Navigation System) and SVESSEL (Samsung’s intelligent ship solution) completed a 6-day test covering 1500 km. It accurately identified over 9000 obstacles (including vessels and buoys) within a 50 km radius, provided safe detour recommendations during 90 encounters, and its planned routes matched those designed by experienced navigators with an accuracy exceeding 90%. Systems such as SAS and HiNAS 2.0 have achieved autonomous decision-making in compliance with COLREGs, including executing starboard turns in head-on situations and taking active evasive actions in crossing scenarios. The test on the “Segyero” further verified the system’s capability to handle multiple concurrent collision risks.

3.5. Virtual–Real Fusion Testing

Virtual-reality fusion testing is a hybrid verification method combining virtual simulation and real-vessel testing [109]. This method has become a critical component of intelligent vessel technology research and development, serving as a key means to bridge the gap between theoretical autonomous collision avoidance algorithms and their engineering applications for intelligent vessels [110].

To address the challenges of scene distortion in virtual simulation, high costs and risks in real-vessel testing for autonomous collision avoidance algorithm verification, Research [111] proposes a virtual reality fusion testing method for autonomous collision avoidance of vessels in open water. The researcher develops a virtual simulation platform and a real-vessel testing platform. The virtual testing platform, based on the Transas simulator, integrates modules for virtual vessel motion modeling, AIS data simulation, and radar image generation. It can batch-generate standardized testing scenarios covering encounter situations defined by COLREGs. For real-vessel testing, a 46.3 m fishery engineering vessel is used as the carrier, equipped with sensors including an AIS simulator, radar, electric compass, and CCTV to collect real-time navigation data. The specific details of the vessel are introduced in Figure 13.This verifies the real-time performance and compliance of collision avoidance algorithms in actual water environments, addressing the “scene distortion” issue in virtual simulation.

The virtual-reality fusion approach balances the low cost and high safety of virtual simulation with the authenticity of full-scale vessel testing. In [112] virtual-reality fusion method leverages digital twin technology to construct repeatable and accelerable scenarios, avoiding the high costs and safety risks of real vessel testing, while enhancing scenario authenticity through real historical data. However, current virtual-reality fusion testing primarily operates in open waters under calm sea conditions, without validating the impact of harsh conditions such as strong winds, heavy waves, and strong currents on collision avoidance algorithms. This limitation prevents a comprehensive assessment of system robustness in complex marine environments. Future research should conduct virtual-reality fusion tests under terrible sea states, simulating the effects of strong winds, waves, and ocean currents on ship motion to verify algorithm adaptability in extreme conditions. Additionally, multi-vessel dynamic interaction scenarios should be introduced to test the system’s collision avoidance decision-making capabilities in high-density traffic flows.

3.6. Evaluation Metrics for Autonomous Collision Avoidance of Intelligent Vessel

Autonomous collision avoidance (CA) for intelligent vessels refers to the technology by which vessels avoid dynamic and static obstacles through environmental perception, navigation situation cognition, and autonomous decision-making. In [89], the evaluation of this technology aims to quantify the safety, compliance, and efficiency of algorithms in complex scenarios. The core indicators include the degree of compliance with the COLREGs, navigation efficiency, and scenario adaptability.

In the field of autonomous collision avoidance algorithm evaluation, research [113] systematically constructs an evaluation system encompassing spatial efficiency, temporal efficiency, protocol compliance, and safety. Specifically for the protocol compliance evaluation against COLREGs, the regulations are categorized into 11 classes, each equipped with independent scoring criteria. These criteria include quantitative indicators such as timeliness of actions, saliency of maneuvers, and rule matching degree, reflecting the core role of COLREGs in maritime collision avoidance decision-making.

It is noteworthy that as a traditional maritime regulatory system, parts of COLREG contain subjective expressions. Specifically, ambiguous requirements such as “take early and substantial action” and the overlapping definition of rights and responsibilities under different encounter situations are difficult to directly transform into quantifiable standards executable by machines. The open-ended nature of regulations interpretation poses fundamental obstacles of non-unified standards and unclear boundaries for algorithm evaluation based on COLREGs.

To address this contradiction, Ref. [113] adopts Woerner’s score-penalty model to convert COLREGs rules into mathematical functions, constructing scoring functions through the relative bearing and contact angle at CPA. The researcher of [114] adopts Nakamura’s anxiety estimation method, which divides collision risks into Danger Area, Caution Area, and Safety Area, calculating penalty scores based on the dwell time of vessels in each area to quantify the timeliness and effectiveness of collision avoidance actions [98]. It should be noted that existing evaluation parameters (CPA safety threshold, course change magnitude) are fixed values, unable to dynamically adapt to differences in vessel types (e.g., tankers vs. small crafts) or environments (ports vs. open seas). For example, the CPA threshold needs to be reduced in port scenarios, but current models lack such adaptive adjustment.

Research has [115] proposed a systematic framework for evaluating automatic collision avoidance algorithms for vessels, aiming to construct a standardized test suite covering all possible scenarios specified in COLREGs. The study classifies ship encounter scenarios based on COLREGs, designs 13 one-on-one and 55 one-on-two basic scenarios by integrating geometric analysis of collision courses, and extends to complex conditions like emergency avoidance and multi-target conflicts through dynamic parameters such as TCPA and speed ratios. Mathematical analytical indices including DCPA, TCPA, and rate of relative bearing are proposed to quantify risks and verify algorithm compliance. This framework provides a systematic solution for safety assessment of autonomous vessels.

4. Discussion

4.1. Autonomous Collision Avoidance Algorithms for Intelligent Vessels

The autonomous collision avoidance algorithm for intelligent vessels is a core technology to ensure the safe navigation of intelligent vessels. This paper systematically combs the algorithm principles, application scenarios, and limitations of autonomous collision avoidance algorithms for intelligent vessels. Based on the review and combing, a qualitative evaluation is conducted for each algorithm, and the results are shown in

Table 2. Autonomous Collision Avoidance Algorithms for Intelligent Vessels.

Category	Algorithms	Principle	Scenario	Adaptability	Efficiency
Regulation-driven	Geometric Model	geometric relationship analysis of COLREGs provisions and ship domain	single target vessel in open waters	Low	Strong
	Expert System and Fuzzy Logic	combining expert experience and fuzzy rule base	Simple encounter scenarios	Low	Medium
Physical model-driven	Dynamic Window Approach (DWA)	Optimizes feasible velocity window based on current motion state to achieve dynamic obstacle avoidance	Local dynamic obstacle avoidance	Strong	Strong
	Velocity Obstacle (VO)	Calculates velocity obstacle areas and selects collision-free velocities	dense port waterways	Strong	Medium
	Artificial Potential Field (APF)	Drives ships toward targets and away from obstacles using virtual potential fields	Simple dynamic collision avoidance	Low	Strong
	Model Predictive Control (MPC)	Implements rolling optimization of multivariable control, combining COLREGs constraints and environmental prediction	multi-vessel collaboration	Strong	Low
Data-driven	Maritime Data-Driven Approach	Generates collision avoidance strategies based on collision risk analysis of AIS, Radar and video data	Risk warning in open waters	Low	Strong
	Artificial Neural Network	Captures vessel behavior patterns using time-series data models and generates end-to-end collision avoidance decisions	Dynamic multi-ship collaboration	Strong	Low
	Deep Reinforcement Learning (DRL)	Trains agents through reward functions (including COLREGs constraints) to achieve collision avoidance decisions in high-dimensional spaces	High-risk complex scenarios	Strong	Low
Hybrid methods	Knowledge Graph + DWA	Fuses knowledge graph (encoded with COLREGs) and DWA to enhance situational awareness and regulation compliance	Multi-ship encounter and dynamic obstacle scenarios	Strong	Medium
	DQN + VO	Combines DRL and VO algorithm to prioritize high-risk target vessels	Highly dynamic multi-ship environments	Strong	Medium
	PSO + APF	Optimizes global paths via particle swarm optimization and achieves local obstacle avoidance with APF	Path planning in complex static environments	Strong	Strong
	Multi-Agent Deep Reinforcement Learning	Enables multi-ship collaborative collision avoidance by designing joint reward mechanisms compliant with COLREGs	Multi-ship collaboration in narrow waters	Strong	Medium

.

Regulation-driven methods take COLREGs as the core, combining geometric models and fuzzy logic, emphasizing compliance and interpretability. However, they face challenges such as regulation ambiguity and insufficient adaptability to complex scenarios. Physical model-driven methods achieve precise control through dynamic modeling. Typical algorithms include the DWA, VO, APF, and MPC, featuring real-time performance and dynamic adaptability. Nevertheless, they need to address the issues of rule integration and complex environment modeling. Data-driven methods rely on machine learning and DRL, using AIS, Radar and video data and multi-modal perception to realize end-to-end decision-making, with strong generalization capabilities. However, they are constrained by data quality and model interpretability. Hybrid methods integrate the advantages of regulation-driven, model-driven, and data-driven approaches, adopting a hierarchical architecture of global planning and local collision avoidance as well as multi-algorithm collaboration, significantly enhancing robustness in complex scenarios. However, they need to overcome bottlenecks in real-time performance and multi-agent collaboration [3].

Based on the systematic review in the second part of the paper, methods centered on situational awareness (SA) integrate regulation-driven algorithms to ensure compliance with COLREGs and physical model-driven algorithms (e.g., DWA and VO, leveraging dynamic modeling). These methods utilize environmental perception to enable real-time decision-making. While they can adapt to dynamic scenarios, they face challenges in rule integration. Methods incorporating behavior prediction combine data-driven algorithms and hybrid methods. By incorporating behavior prediction through data analysis and collaborative strategies, they can enhance collision avoidance performance in complex multi-vessel scenarios.Future research will focus on the deep coupling of rules and AI, multi-vessel collaborative collision avoidance, digital twin verification, and lightweight deployment, promoting the leap of intelligent vessels from “compliance” to “intelligence” and achieving safe and efficient autonomous navigation [80].

4.2. Autonomous Collision Avoidance Performance Testing and Evaluation System for Intelligent Vessels

For the safety and reliability verification of autonomous collision avoidance algorithms for intelligent ships, this paper systematically reviews a multi-level technical framework covering test scenario generation, virtual simulation, model experiments, full-scale ship testing, and virtual–real fusion in the testing and evaluation stages. Qualitative evaluations are presented in

Table 3. Testing Scenario Generation Methods for Autonomous Collision Avoidance of Intelligent Vessels.

Scenario Generation Method	Principle	Scenario Authenticity	Regulation Coverage	Scenario Generation Capability	Environmental Coupling	Application Stage
Real Data-Driven Testing Scenario Extraction	Based on real navigation data	High	Limited	Weak	Low	Initial verification
Mechanism Modeling-Based Testing Scenario Reconstruction	COLREGs provisions, ship kinematic models	Moderate	Medium	Medium	Low	Regulation compliance verification
Machine Learning-Based Testing Scenario Derivation	DRL algorithms virtual simulation environment	Variable	Partial	Strong	Medium	High-risk scenario stress testing

and

Table 4. Testing Experimental Methods for Autonomous Collision Avoidance of Intelligent Vessels.

Test Experimental Method	Test Environment Fidelity	Scenario Coverage Capability	Cost and Risk	Regulation Compliance Verification	Application Stage
Virtual Simulation	Medium	Strong	Low	High	Initial algorithm verification
Physical Model Testing	Medium	Limited	Medium	Medium	Mid-term physical performance verification
Full-scale vessel testing	High	Weak	High	High	Final engineering verification
Virtual–Real Fusion testing	High	Strong	Medium	High	Full cycle, transitional verification from algorithm optimization to ship deployment

, respectively.

In Table 4, the scenario generation methods include real data-driven, mechanism modeling-based, optimization search-based and machine learning-based. Although the current testing scenario construction technologies have achieved realistic scenario reproduction, repeatable verification, and diverse generation in the field of vessel autonomous collision avoidance, the following problems still exist in the application of collision avoidance algorithm testing.

• Lack of a complete scenario structure system. Existing methods have insufficient coverage of dynamic elements in collision avoidance scenarios, such as sudden vessel failures and non-standard navigation behaviors.
• Combinatorial explosion in multi-vessel encounters. When the testing scenario involves more than three dynamic target vessels, the parameter combinations of heading, speed, and relative position grow exponentially, leading to a sharp decline in the testing coverage of collision avoidance algorithms.
• Disconnection of scenario dynamics evolution. Traditional static scenarios fail to simulate dynamic interactions such as wind-current interference and inter-vessel effects during collision avoidance, making it impossible to verify the algorithm’s adaptability to sudden situational changes.

Aiming at the testing requirements for high-level autonomous collision avoidance, future research should focus on the following breakthrough directions.

• Construction of dynamic collision avoidance scenarios. Integrating real-time AIS trajectories with hydrodynamic models to generate dynamic collision avoidance scenarios involving wind-current coupling and vessel-shore interactions.
• Big data-driven mining of dangerous scenarios. Identifying high-frequency risk patterns from historical collision avoidance cases based on reinforcement learning to directionally generate high-risk test scenarios such as dense fog and bridge areas.
• Reconstruction of uncertain environments. Constructing non-ideal collision avoidance scenarios with sensor errors and communication delays by combining navigation risk quantification models.
• Scenario evolution driving methods. Designing state transition rules conforming to collision avoidance logic to achieve automatic deduction of full-process scenarios from routine encounters to critical dangers.

Virtual simulation relies on digital twin technology to simulate complex marine environments, while scaled model and full-scale vessel tests verify the algorithm’s robustness in physical environments. Through the above review, it is found that current challenges include maritime data noise, incomplete rule coverage, and insufficient virtual–real interaction consistency. In future autonomous collision avoidance testing, it is necessary to emphasize the integration of data-driven and physical models, develop adversarial testing generation technologies, and establish a multi-dimensional evaluation index system covering safety, rule compliance, and navigation efficiency, providing support for the transition of intelligent vessels from theoretical verification to engineering application.

4.3. Testing and Evaluation Framework for Autonomous Collision Avoidance of Intelligent Ships

In the maritime field, research on autonomous collision avoidance algorithms, testing and evaluation for intelligent vessel is still in its preliminary stage [2]. This paper conducts a detailed review of autonomous collision avoidance algorithms for intelligent vessels and comprehensively surveys the research status of autonomous collision avoidance testing and evaluation. A process framework for the testing and evaluation of autonomous collision avoidance of intelligent vessels is established, filling the gap in this research area. In Figure 14, the testing and evaluation framework for autonomous collision avoidance consists of four key components: the Testing Functional Stage, Testing Scenarios Generation, Testing Tasks Decomposition, and the Evaluation Index System. The specific contents will be introduced below.

1. Evolution of Autonomous Collision Avoidance Capabilities.

The phased division of assisted collision avoidance, remote collision avoidance, and autonomous collision avoidance aligns with the gradual development of intelligent vessel technology. In the assisted stage, the foundation of human–machine collaboration is of great significance. For remote collision avoidance, the reliability of data transmission and the coordination mechanism between shore-based centers and vessels are crucial. As we move towards full autonomy, verifying the robustness of the “perception-decision-execution” loop without human intervention is the core. This evolutionary path reflects the necessity of step-by-step verification to ensure that each stage can effectively support the next, avoiding potential risks associated with a direct leap to full autonomy.

2. Significance of Diverse Scenario Generation.

The generation of both basic and extreme-environment scenarios serves as the basis for comprehensive testing [20]. Basic scenarios, such as open waters, narrow channels, and port areas, provide a foundation for verifying the general collision-avoidance logic in regular navigation environments. They are essential for ensuring compliance with COLREGs and normal navigation efficiency.

Extreme scenarios, including low-visibility conditions, adverse sea states, ice-covered waters, and shallow-water areas, test the system’s performance under stress [85]. In low-visibility situations, the fusion of multiple sensors (radar, AIS, video) and the accuracy of perception with limited information are critical. Adverse sea states challenge the adaptability of the vessel motion model and the anti-interference capability of collision-avoidance strategies. Ice-covered and shallow-water areas test the coordination between the vessel’s maneuvering boundaries and collision-avoidance decisions. These extreme scenarios help identify the system’s limitations and potential failure points, which is invaluable for enhancing the overall reliability of the autonomous collision-avoidance system.

3. Decomposition and Validation of Testing Tasks.

Decomposing testing tasks into compliance testing, system performance testing, and human–computer interaction testing covers all aspects of the autonomous collision-avoidance system [109].

Compliance testing, focusing on COLREGs verification and safety-boundary verification, ensures that the system operates within the legal and safety framework. As the cornerstone of maritime collision avoidance, COLREGs must be strictly adhered to by intelligent vessel systems. Any deviation may lead to maritime accidents and legal disputes. Safety-boundary verification defines the vessel’s “safety domain”, and ensuring that the collision-avoidance strategy does not cross these boundaries is crucial for preventing collisions.

System performance testing, including response-time testing, path-optimization testing, and redundancy testing, reflects the system’s efficiency and reliability. Response time directly affects the timeliness of collision-avoidance actions, and excessive response time can render the collision-avoidance strategy ineffective. Path optimization needs to balance collision avoidance and navigation efficiency to avoid unnecessary detours that increase costs. Redundancy testing ensures that the system can still operate in a degraded mode when components fail, which is essential for mission-critical systems like intelligent vessels.

Human–computer interaction testing, involving operator-intervention capability and alert management, addresses the role of humans in the autonomous system [112]. In some cases, human intervention may still be necessary, and the smooth transition of authority between humans and machines, as well as the effectiveness of human intervention, is important. Effective alert management can prevent information overload for operators while ensuring that critical alerts are not missed, maintaining the proper operation of human–machine collaboration.

4. Rationality and Improvement of the Evaluation Index System.

The evaluation index system, covering the stages of collision-avoidance perception, collision-avoidance decision-making, and collision-avoidance execution, provides a comprehensive and quantifiable evaluation method [115].

In collision avoidance perception, the target-recognition capability, dynamic-prediction accuracy, and environmental-awareness coverage directly affect the quality of subsequent decision-making. High-precision target recognition and accurate dynamic prediction can provide more reliable input for decision-making, while sufficient environmental-awareness coverage can reduce the risk of “perception blind spots”.

The evaluation of collision-avoidance decision-making, including COLREGs compliance, path-optimization efficiency, and collaborative collision-avoidance capability, ensures the legality and effectiveness of the decision-making process. Compliance with COLREGs is a basic requirement, and path optimization needs to consider both collision avoidance and navigation efficiency. In multi-vessel encounter situations, the collaborative collision-avoidance capability is particularly important for avoiding mutual interference in collision-avoidance actions.

For collision-avoidance execution, real-time control, execution accuracy, and emergency reliability determine whether the decision can be effectively implemented. Any deviation in execution may lead to the failure of the entire collision-avoidance process, and emergency reliability ensures that the system can still respond effectively in extreme situations.

However, there is still space for improvement in the current research. For example, the current evaluation indices may not fully consider the impact of emerging technologies such as artificial intelligence-based adaptive collision-avoidance algorithms. Future research can further refine the indices to adapt to the development of intelligent vessel technology. In addition, the integration and mutual verification of indices among different stages can be strengthened to form a more coherent evaluation system.

5. Conclusions

This paper systematically reviews the core algorithms, testing and evaluation systems, and key research progress in the field of autonomous collision avoidance technology for intelligent vessels.

Firstly, regarding mainstream autonomous collision avoidance algorithms (e.g., DWA, VO, APF, and MPC), this paper not only elaborates on their basic principles and applicable scenarios but also clarifies the advantages and limitations of each algorithm through a structured comparative analysis. This addresses the deficiency of previous isolated descriptions. Secondly, focusing on the critical issue of COLREGs compliance, this paper delves into the practical embedding methods of COLREGs rules (particularly Rules 13–17) in algorithms. By linking navigation practices with regulatory provisions, it provides a clear framework for verifying algorithm compliance. Furthermore, to overcome the limitation of traditional research that focuses excessively on two-ship encounter scenarios, this paper supplements collaborative collision avoidance strategies for multi-ship scenarios. Taking the SDVO + FIS-NC algorithm as an example, it analyzes in detail the dynamic collision avoidance sequence adjustment and behavior prediction mechanisms based on the Collision Risk Index (CRI), verifying the algorithm’s effectiveness in complex traffic environments.

In terms of structure and innovation, this paper reconstructs the review framework and proposes an integrative classification system centered on situational awareness, behavior prediction, and regulatory compliance. Additionally, it sorts out testing and evaluation methods for autonomous collision avoidance performance and constructs a full-process testing framework covering capability evolution, scenario verification, task decomposition, and index evaluation, providing systematic guidance for translating theoretical technologies into engineering applications. Future research can further focus on optimizing algorithm robustness in extreme environments, refining the modeling of multi-ship dynamic interactions, and advancing virtual–real fusion testing technologies. These efforts will promote the development of autonomous collision avoidance technology for intelligent vessels toward higher standards of safety compliance, intelligence, and efficiency.