Path planning and collision avoidance for autonomous surface vehicles II: a comparative study of algorithms

Abstract

Artificial intelligence is an enabling technology for autonomous surface vehicles, with methods such as evolutionary algorithms, artificial potential fields, fast marching methods, and many others becoming increasingly popular for solving problems such as path planning and collision avoidance. However, there currently is no unified way to evaluate the performance of different algorithms, for example with regard to safety or risk. This paper is a step in that direction and offers a comparative study of current state-of-the art path planning and collision avoidance algorithms for autonomous surface vehicles. Across 45 selected papers, we compare important performance properties of the proposed algorithms related to the vessel and the environment it is operating in. We also analyse how safety is incorporated, and what components constitute the objective function in these algorithms. Finally, we focus on comparing advantages and limitations of the 45 analysed papers. A key finding is the need for a unified platform for evaluating and comparing the performance of algorithms under a large set of possible real-world scenarios.

Introduction

There is growing appeal for autonomous systems in multiple fields, including manufacturing, transportation, routine work, and work in dangerous environments. In the wake of progress in the domain of autonomous cars, much attention is also given to autonomous surface vehicles (ASVs). In an accompanying article in this journal [1], we present a review on theory and methods for path planning and collision avoidance of ASVs. We attempt to unify and clarify relevant terminology and concepts such as autonomy and safety, as well as models for guidance, navigation, and control. Moreover, we propose a classification scheme for distinguishing and comparing algorithms for path planning and collision avoidance.

Here, we extend this scheme to classify state-of-the-art algorithms presented in 45 different peer-reviewed scientific papers. Several kinds of algorithms are covered, including evolutionary algorithms, sampling-based algorithms, cell decomposition methods, directional approaches, and roadmap methods. We have also included some algorithms for unmanned surface vehicles (USVs).

As for any literature review paper, it is impossible to cover everything in the literature within the scope of a single paper. The number of papers studied before arriving at the shortlist of the 45 papers presented here is probably in the ballpark of several hundreds. We have carefully selected papers that we ultimately found useful to include.

Moreover, whereas much of what we present is general across vessel size, other considerations will differ whether the vessel is a small boat or a large ship. In such cases, the reader should note that larger ships are our main focus. Likewise, although some elements of path planning and collision avoidance are common across congested waters and open sea, we are mainly concerned with shorter time frames and congested waters in the papers we study here.

The rest of the paper is organised as follows: Sect. 2 provides a timeline of some of the most influential algorithms for path planning and collision avoidance for ASVs or USVs. Section 3 extracts distinguishing properties of the algorithms from the literature, and analyses and compares papers based on these properties. Section 4 analyses the proposed algorithms based on their properties whilst focusing mainly on two aspects: (1) safety and collision risk assessment (CRA), and (2) choice of objective function. Section 5 extracts the advantages and limitations of the algorithms used in the different papers. Finally, Sect. 6 presents a discussion, whilst some concluding remarks are drawn in Sect. 7.

Timeline of algorithms

The first use of some of the most influential algorithms used for path planning and collision avoidance for ASVs or USVs is shown in Table 1. Notably, these algorithms have also been successfully used at earlier times for guidance of autonomous underwater vehicles (AUVs), unmanned aerial vehicles (UAVs), or autonomous ground vehicles (AGVs). Note that Table 1 is by no means an exhaustive list but highlights some dominating algorithms that have been commonly employed, directly or in some derivative form, or in combination with others.

Table 1 Timeline of the first time use of dominating algorithms for USV/ASV guidance applications

Properties of algorithms

Although some algorithms in the literature clearly separate the tasks of path planning and collision avoidance, others do not, and attempt to solve both problems with overlapping modules [1]. Furthermore, it is generally not easy to compare path planning and collision avoidance algorithms for ASVs due to the variety of constraints and objectives that exist. One example is the use of regulations such as the International Regulations for Preventing Collisions at Sea (COLREGs) [13]: whereas some algorithms successfully generate paths for avoiding obstacles whilst simultaneously obeying several COLREG Rules [e.g., 14,15,16,17,18], others fully or partially ignore these regulations [e.g., 11, 19, 20,21,22]. For adoption in the future, fully autonomous surface vessels must comply with all the rules of COLREGs. We appreciate, however, that algorithms that comply only with a subset of COLREGs are still a step towards this goal and a contribution towards full COLREGs compliance in the future.

The literature analysis in Vagale et al. [1] shows that there are several properties of path planning and collision avoidance algorithms that can be used for classification and analysis of the algorithms:

  • Compliance with COLREGs: partial/full consideration of COLREGs for collision avoidance.

  • Environmental disturbances: taking into account wind, waves, currents, and tides.

  • Planning type: global and/or local planning.

  • Obstacle type: whether a vessel can deal with static and/or dynamic obstacles (including single or multiple encounter situations at the same time).

  • Environment type: discrete or continuous environment.

  • Type of avoidance action: course change or speed change, or a combination of both.

  • Testing of algorithm: simulation or field test.

  • Traffic category: congested waters (areas crowded with static/dynamic obstacles, including harbour areas, lead to low own vessel speed), open waters (minimal number of static and dynamic obstacles, lead to high own vessel speed), riverines (manoeuvring is limited, current is present), and coastal areas (mostly static obstacles, such as land, islands, and shallow water, lead to varying speed).

  • Predictability of environment: known or unknown environment.

  • Planning time: real-time (online) or offline.

  • Control horizon: infinite or receding horizon control.

  • Number of encountered obstacles: single or multiple target vessel encounter situations.

  • Vessel dynamics and kinematics: maximum ship turning rate, maximum vessel speed, other vessel’s motion constraints, torque of the vessel, etc.

  • Subject of research: type of the researched vessel or system [ASV, USV, and decision support system (DSS)].

  • Safe zones: safety margin, virtual safety zone, ship domain, ship arena, or circle-of-rejection, around the own vessel or static/dynamic obstacles.

Note that Tsou and Hsueh [5] define ship domain as “the sea around a ship that the navigator would like to keep free of other ships and fixed objects.” This criterion has been widely used in ships’ collision avoidance, marine traffic simulation, calculation of encounter rates, vehicle tracking system (VTS) design, and so forth. It differs from ship arena, which is a bigger area around the vessel used to determine the time of taking collision avoidance actions [23]. Similarly, a safety zone can be assumed around each obstacle instead of the own ship, called the circle-of-rejection (COR) [24].

Based on the aforementioned properties, eight properties have been chosen for a comparative study of 45 papers containing algorithms for path planning and collision avoidance of ASVs (see Table 2). The choice of these eight properties is based on the most common available, and relevant, information in algorithm descriptions. Some other properties were neglected due to many papers excluding the very same information regarding such properties. The proposed comparison is an attempt to analyse these state-of-the art algorithms and benchmark them using the proposed criteria. Table 3 compares the ship- and environment-related properties across the chosen papers. The algorithms in the comparison of Table 3 are grouped in three groups, separated by white space, based on the “planning type” property. Each row of the table includes the paper reference (‘Ref.’), the type of path planning, and/or collision avoidance algorithm(s) employed, followed by an analysis of how the 8 properties in Table 2 are taken into account.

Although the focus of this study is on methods for ASVs, papers related to USVs are also considered. The databases used for finding journal and conference papers were IEEE Xplore Digital Library and ScienceDirect. Additionally, the NTNU library was consulted using the search tool Oria, as well as suggestions from the reference organisation tool Mendeley. The keywords used for search were “ASV,” “USV,” “autonomous ships,” “path planning,” “collision avoidance,” and “guidance.” The papers included in the comparison are from the years 2010–2020, and the language was limited to English. The distribution of the analysed papers over the years is represented in Fig. 1. The number of papers with respect to each of the eight selected properties is represented graphically in Fig. 2.

Fig. 1
figure1

Annual distribution of the papers

Fig. 2
figure2

Distribution of properties among the papers

Table 2 Selection of algorithm properties
Table 3 Comparison of situation/environment and ship-related properties of different algorithms in 45 selected papers

We discuss each of the eight properties P1–P8 in turn, before making some general observations, mainly with reference to Table 3 and Fig. 2.

P1. Planning type: The analysis of the selected papers shows that 13 (29%) of the examined algorithms perform global planning and, hence, are mainly concerned with path planning; 17 (38%) algorithms perform local planning and collision avoidance; and 15 (33%) algorithms perform both global and local planning. We also found that in most of the cases, local planning is performed in real time, whereas global planning is often performed offline, prior to departure. In the hybrid cases, when both local and global planning is used, the algorithm is generally a combination of both real-time and offline planning and covers both path planning and collision avoidance. Hence, with this close correlation between local/real-time planning and global/offline planning, a separate property of the algorithm being real time or offline is not considered necessary in Table 3.

P2. COLREGs: The comparison table shows that compliance with COLREGs is included only in the path planning approaches that consider local path planning and collision avoidance (algorithms with property GL and L). Most often, algorithms take into consideration only up to four of the main encounter situations, described in the three COLREG Rules 13–15 [e.g., 40, 54]. These rules are usually implemented as constraints in algorithms and indicate which collision avoidance scenario should be used in the current situation. Eriksen et al. [41], on the other hand, have implemented a cost function penalising gentle turns and small speed changes for obeying COLREG Rule 8, which states that “action taken to avoid collision should be positive, obvious and made in good time.” Hence, the ASV’s behaviour should be obvious and makes sense to human captains. Szlapczynski [39] has proposed an extended method that additionally focuses on COLREG Rule 19, planning the path in restricted visibility conditions. Johansen et al. [53] additionally have also implemented several other COLREG Rules, namely 8, 16, 17, and 18. These rules have been implemented as components of the cost function or as penalty functions. Some papers emphasise that, according to good seamanship practice, course change is preferred over speed change in collision avoidance scenarios [38, 49].

P3. Traffic category: Concerning the traffic categories considered in the papers, one part of the papers focuses on the “open waters” category (13 papers, or 29%), considering an area free from static obstacles such as land and islands. The same amount of papers are dealing with “congested waters” category (13 papers, or 29%) where the traffic most often is busy, such as harbour areas, where both multiple dynamic obstacles and static obstacles are present. However, most of the papers are considering the “coastal area” type of environment/traffic (19 papers, or 42%), where the environment is mostly cluttered with several static obstacles, but there is almost no presence of dynamic obstacles.

P4. Obstacle type: The analysed papers consider different types of obstacles. In the simplest cases, 16 (36%) of papers use algorithms that avoid only static obstacles, including land, islands, and underwater objects. Most of these algorithms are global path planning approaches. For dynamic obstacles, 29 (64%) of the papers consider moving target vessels, underwater vehicles, and icebergs, with 7 (16%) considering single dynamic obstacle situations and 22 (49%) considering more complicated situations involving avoidance of multiple dynamic obstacles. The high number of papers that focus on avoiding dynamic obstacles might be explained by the increased need for real-time collision avoidance solutions. The dynamic obstacle avoidance problem is more complicated, since knowledge of the target object movement is required, and therefore, the consideration of a time parameter must be included. In cases when there is no communication between the own vessel and the target vessels, the examined algorithms perform avoidance action by predicting future positions of target vessels. This can be done by assuming that the own vessel can observe and estimate the dynamics of the target object (velocity and course) and its size; inferring compliance with COLREGs; or by obtaining information from third parties, e.g., from Automatic Identification System (AIS) data.

P5. Testing type: Most of the papers, 38 (84%) in total, test the proposed algorithms only by means of simulations in a simulated test environment built for this reason, for example using simulation software and high-level programming languages such as Matlab. The testing environment varies depending on the papers’ objective, and may include the geographic area, traffic data, obstacles, and other parameters related to ship dynamics of both own and target vessels. Some papers perform tests in several scenarios for representing the flexibility of the algorithm adapting to different situations. Sometimes, the performance of an algorithm is compared with some other under the same environment. A common practice is to use real map data for simulations [e.g., 33, 35,36,37,38]. The remaining 7 (16%) papers are verified in both field tests and simulations. In these cases, small vessels, equipped with GNC systems, e.g., Springer USV [31, 24] and ARCIMS USV [48], are used. An outstanding project with thorough testing is represented in Varas et al. [48] where tests have been performed both on desktop simulations, on a bridge simulator, and on sea trials using a USV. In this paper, testing is performed using Monte Carlo simulations to detect weaknesses of the proposed method and using historical collision incident data for more realistic scenarios.

P6. Environmental disturbances: Table 3 shows that when it comes to environmental disturbances, more than half (27, or 60%) of the papers do not take any environmental disturbances into consideration. Several papers [e.g., 25, 26, 27] are focusing only on the effect of current on the vessel (7, or 16%), some consider both current and wind (4, or 9%) [47, 49, 53, 54], and only two papers consider both current, wind, and waves [29, 35]. None of the papers consider waves as the only environmental disturbance affecting the ship’s movement; however, waves are included in two papers together with wind and current. Out of all of the environmental disturbances, current is the most often included one (13 papers), followed by wind (6 papers) and waves (2 papers).

P7. Vessel dynamics: Vessel dynamics have been considered in most of the cases (32 papers, or 71%). Some of the ship’s parameters included in the papers are dynamics of the vessel, a manoeuvring model, a kinetic model, turning ability, maximum steering angle or speed, and other vessel motion constraints or limitations. The remaining 13 papers (29%) do not consider vessel dynamics.

P8. Safety domain: To enhance safety, a safety zone (domain) is required for ensuring the respect of the closest area around the own vessel, target vessels, or obstacles. Across the applied algorithms, safety zones take a variety of shapes, including circle, ellipse, rectangle, shipshape, and inverted cone. An own ship domain has been implemented using various parameters in 13 (29%) papers. A safety domain around target vessels or a safety zone around obstacles has been implemented in the same number of papers (29%). Finally, 5 (11%) of the algorithms have implemented both an own ship domain and a domain, whereas 14 (31%) algorithms do not include a safety domain.

Hybrid approaches: The study shows that most of the algorithms are using a hybrid approach for path planning and collision avoidance that combines two or more methods to improve the performance and cover different sides of real-life situations. For example, Niu et al. [25] combine Voronoi diagram with visibility graph and Dijkstra’s search, creating a hybrid Voronoi-visibility algorithm; Wu et al. [20] combine artificial potential field method with ACO algorithm for global planning and uses a multi-layer obstacle-avoidance framework for local planning; Xie et al. [22] combine Dijkstra’s algorithm with APF method; and Candeloro et al. [37] merge Voronoi diagram with Fermat’s spiral (FS) to ensure curvature-continuous paths. In most cases, the purpose of the hybrid approach is to be able to solve both local and global path planning.

Single- vs. multiple-vessel control: Most papers are focusing on single-vessel path planning methods, whereas a few authors are considering path planning of a formation or a fleet of more than one vessel [e.g., 16, 56,57,58]. Notably, for formation path planning in a static environment, conflicting collision avoidance situations between formation members also need to be considered, turning the environment into a dynamic one.

Safety and objective functions

A crucial aspect of ASVs is the ability to navigate safely in open waters, coastal areas, and congested waters like harbours. To achieve safe manoeuvring, multiple components should be considered, such as COLREGs, situational awareness (consideration of both dynamic and static obstacles), dynamic properties and limitations of the vessel, environmental disturbances, and safety domain [59]. One way of ensuring the safety of the own vessel considering the dynamic target vessels in the vicinity is to include some safety aspects when searching for collision-free paths, thus evaluating risk of collision. Hence, safety of the own and target vessels should be incorporated, or at least considered, when generating paths based on optimisation of an objective function.

In Fig. 3, we highlight what we have identified as being the four most often used safety components across the examined literature, namely (1) safety conventions, (2) collision risk assessment (CRA), (3) safety domain, and (4) environmental disturbances.

Fig. 3
figure3

Safety components

In the following subsections, however, we limit our study to analysing the employment of (1) collision risk assessment (CRA) and (2) objective function in the algorithms proposed in the selection of literature.

Collision risk assessment

CRA is one of the key factors that aids in evaluating the safety of the path to be taken. It is an assessment tool that may include several safety criteria based on the current and predicted situation, own or target vessels’ parameters, and their mutual relationship.

An often used risk evaluation criterion for CRA in the literature is the closest point of approach (CPA), which can be measured both in time and distance, as illustrated in Fig. 4.

Fig. 4
figure4

The concept of time and distance of CPA

The CPA is the position at which two dynamically moving vessels will have the shortest distance between them at a specific time. The time to the closest point of approach (TCPA) is the time when this position is reached. The distance of the closest point of approach (DCPA) is the distance between both CPA points on the trajectory of each vessel.

Both TCPA and DCPA are proposed for the maritime field by Kearon [60], and they are used mainly for collision risk assessment and navigation safety enhancement. The TCPA and DCPA parameters, however, have a drawback. As noted by the authors in Nguyen et al. [61], both parameters do not adequately represent the danger of a collision when moving into head-on situations and overtaking situations.

CRA parameters are not limited only to TCPA and DCPA, although these are the most commonly used ones. Other papers also consider parameters such as the distance of the last-minute avoidance, distance to the target vessel, ratio of speed, relative bearing, safe passage circle, and distance of adopt avoidance action [15, 47, 62].

Objective function

There are many possible criteria for path evaluation using an objective function. Some of the most often used criteria which we have identified are:

  • Path length: length of the obtained path (either before or after smoothing of the path).

  • Voyage time: time required to reach the target position when traversing the obtained path.

  • Smoothness: connection of waypoints in an optimal way taking into consideration limited curvature or turning radius of the ship. This property partly reflects whether the path is feasible from the ship’s perspective. Reduced number of sharp turns or a path smoothing module are some examples of a smoothness component.

  • Tractability: the practicality of the path, especially in dynamic environments when some waypoints have to be relocated during the journey [63].

  • Energy consumption: a criterion that might be influenced by several other factors, including path length, vessel’s speed, or the effect of sea currents on the vessel, in terms of economy.

  • Path precision: how precisely does the designed path pass through waypoints [63].

The comparison of (1) CRA components and (2) objective function criteria included in papers is presented in Table 4. Here, CRA analysis includes only the most often used criteria, namely TCPA and DCPA. The analysis of the objective function considers only the four most often implemented components: length, time, smoothness, and energy efficiency. For all columns, the presence/absence of the criteria is indicated with ‘+’/’–’, respectively. The analysis is performed for the same 45 papers that were chosen and analysed in Sect. 3 with the same sequence of papers and the division based on “planning type” property. The last row of the table summarises the number of papers that have included each of the criteria.

Table 4 The use of CRA and objective function components in 45 selected papers

Analysis

Table 4 shows that the most often used CRA criterion is DCPA, used in 21 (47%) papers, whereas TCPA was used in 15 (33%) papers. 14 (31%) papers use both TCPA and DCPA, whereas half the papers (23, or 51%) use neither TCPA nor DCPA as a CRA criterion. Most of these 23 papers are dealing with static obstacles; therefore, there is no need for calculating CPA. Instead, authors in Tam and Bucknall [54] use a two-step CRA process by (1) determining the type of encounter, and (2) calculating the dimensions of the safety area. The rest of the 23 papers that do consider dynamic encounters use other ways to ensure safety, and collision-free paths, such as considering COLREGs [16, 24, 40, 53], applying a safety domain around own or target vessels [24, 53, 54], or calculating the probability of collision [52].

Regarding the objective function, path length (27 papers, or 60%) is the component taken into account the most, followed by smoothness (13 papers, or 29%), time (12 papers, or 27%), and lastly energy efficiency (8 papers, or 18%). 10 papers use none of the four objective function components, and no paper uses all four. In most of these cases, the papers are dealing with collision avoidance [9, 15, 47, 51, 52, 53, 55]; therefore, authors do not prioritize optimization of the path’s length, energy efficiency, or other parameters but instead focus on safety of the collision-free path. Other components included in objective functions by some authors are tractability [31]; cost on deviating from the relative nominal trajectory, and on control input [41]; and navigation restoration time and angle during collision avoidance manoeuvre as well as optimal safe avoidance turning angle [18].

Algorithms based on reinforcement learning (RL) [e.g., 49] do not use a standard objective function but rather a reward function. This means that standard objective parameters are not optimised directly. Instead, the reward function helps the agent to learn and improve based on the dynamics of an agent and the practicality and safety of the path. Therefore, even though RL algorithms do not optimise smoothness directly, they might generate a path that is smooth.

Statistics of CRA and objective function components included in the papers are summarised in Figs. 5 and 6, respectively.

Fig. 5
figure5

Usage of the CRA components TCPA and DCPA in 45 selected papers

Fig. 6
figure6

Usage of the objective function components Length, Time, Smoothness, and Energy in 45 selected papers

Advantages and limitations

To further enhance our comparative study of path planning and collision avoidance algorithms for ASVs and USVs, we summarise the advantages and limitations (room for improvement) of the algorithms proposed in the 45 selected papers, as shown in Table 5. The criteria of the analysis include computational complexity, convergence, planned path features (particularly optimality and smoothness), the ability to re-plan, operation in real time, the complexity of the environment, consideration of the local minima trap, and others.

The analysis of the advantages and limitations of the proposed algorithms is based purely on the information provided by the authors of each one of the analysed papers. Therefore, this evaluation is inherently subjective, and in most cases, the authors have not stated the limitations of the algorithms at all even if they exist (noted in the table as ‘N/D’) or they have been extracted from the future work section.

Table 5 Advantages and limitations of the analysed algorithms

The analysis of the algorithms in Table 5 shows that in many papers, authors do not state their limitations in a straightforward manner. In many cases, the limitations of the proposed algorithms have been extracted from the future work section of the paper. This section often gives a better comprehension of the current state of the proposed method and its limitations and parts that have to be improved.

In some cases, the conventional version of an algorithm has been extended and improved to form promising derived algorithms that avoid limitations of the conventional algorithm. For example, a well-known limitation of the conventional APF algorithm is the local minima problem. However, for derived algorithms that are based on the conventional APF, authors often state avoiding local minima trap as their advantage, additionally to other improvements.

To sum up, many of the proposed algorithms are trying to overcome different problems connected with developing an autonomous system that performs well in real-life applications. However, the analysis shows that even when the limitations of the algorithms are not stated clearly by the authors, they still exist. That is, although researchers demonstrate knowledge about which components should be included in an ASV path planning and collision avoidance system, there inevitably will still be difficulties in implementing the system in real life.

Finally, we wish to point out that, according to our knowledge, several other path planning algorithms used for mobile robots, ground vehicles, aerial vehicles, or underwater vessels have not been applied to surface vessels yet, e.g., bug algorithm [64], Voronoi fast marching method [65], symbolic wavefront expansion [66], probabilistic roadmaps [67], and fast marching* (FM*) [68]. Even though these algorithms have been applied for path planning in various other fields, it would be possible to adapt these algorithms also to applications for ASVs. Moreover, interested readers should note that additionally to our own comparison of algorithms, and a comparison of performance of the A* algorithm and derivative algorithms (A*PS, Theta*, and A*GB) used for path planning for autonomous inland vessels is provided by Chen et al. [34].

Discussion

The timeline of algorithms for the latest decade shows an increased interest of researchers for solving path planning and collision avoidance problems for surface vessels by experimenting with, and developing new, methods and algorithms from the AI domain. However, this comparative study shows that there is still no unambiguous model of how “the ultimate” autonomous ship should be designed, which components it should include, and how it should act. The analysed papers offer various solutions to example problems, but these solutions are often limited to perform well under specific and restrictive conditions.

Through the analysis, we have identified a number of limitations in recent solutions for path planning and collision avoidance of ASVs (some of these limitations have also been pointed out in other review papers in the field, as described in our accompanying paper [1]):

  • The variety of algorithms used for solving path planning issue is wide, with researchers continuously exploring different options and trying to find better and more general solutions.

  • Many developed algorithms that appear to be efficient theoretically have not been tested in a real environment or with real traffic data; hence, it is not possible to evaluate their efficiency in handling real-world issues.

  • Some algorithms deal only with static obstacles, excluding dynamic ones.

  • In many cases, the developed algorithms do not take into account external disturbances such as wind, waves, or current, which means that the modelled environment is not complete and the performance of the algorithms under realistic conditions would differ.

  • Some algorithms assume that the velocity of target ships (that need to be avoided) is constant, and/or that target vessels do not follow COLREGs, meaning that the controlled vessel is not observed and is ignored by other vessels, which is not very realistic.

  • Although many researchers agree that safety is the top priority when navigating vessels, not all solutions are considering COLREGs as part of their safe collision avoidance or path planning algorithm.

  • Collision risk assessment is typically based only on one or two factors that do not represent the full comprehension of the safety situation of the own vessel in the environment.

Several of these shortcomings lead to the consideration of non-realistic testing environments for vessels, which, in turn, might cause situations where the behaviour of the vessel at sea will differ from the one in simulation tests.

Regarding the limitations of this comparative study, we wish to highlight the following:

  • It could be argued that the algorithms in the selected papers should be sorted depending on whether they are solving a path planning (on the global level) or a collision avoidance (on the local, reactive level) problem. The reason for not doing this is the difficulty in distinguishing the algorithms based on this division, as some algorithms are used both for solving path planning and collision avoidance issues.

  • Another limitation is that the comparison of the considered properties only gives a partial understanding of the performance of different algorithms in action.

  • Finally, it is difficult to extract sufficient details about the properties of the algorithms because of the incomplete or vague descriptions in some of the papers, thus requiring interpretation by the reader.

Future work should try to address these limitations, and examine in more depth some of the properties in Sect. 3 left out in this study, especially “predictability of environment” and planning with uncertainty.

Conclusions

The extent of this research is large and fills in some gaps in the field by comparing existing path planning and collision avoidance algorithms of ASVs in a manner they have not been compared before.

ASVs clearly have a big potential in future maritime transportation, but their limitations should also be considered and treated with caution. In this study, we extracted a set of defined properties and characteristics that was used for comparison of the proposed algorithms in 45 carefully selected papers. These properties can be used later by other researchers for benchmarking and for comparing their own algorithm to others’. With respect to the analysis of the 45 papers, the main contribution is threefold and consists of: (1) a comparison of the usage of eight important ship- and environment-related properties; (2) an analysis of how safety has been incorporated, and what components constitute the objective function; and (3) an analysis of advantages and limitations of the proposed algorithms. We consider this comparative study a good attempt at comparing the current state-of-the-art and believe that it can serve as the basis for a deeper performance evaluation system of path planning and collision avoidance algorithms for ASVs.

Future research should be dedicated to simulation as well as real-world field tests that evaluate the actual performance of algorithms in various scenarios under different conditions for a more precise comparison of the developed methods. Such testing systems might aid in evaluating the reliability, durability, and the flexibility of the methods, and in designing appropriate algorithms for specific applications and needs. Testing a large number of different scenarios might be performed using Monte Carlo simulation methods.

Another interesting direction of future research is the evaluation of safety and collision risk assessment of the own ship navigating realistic environments. Components that should be considered when evaluating safety and collision risk are obedience to COLREGs, environmental disturbances, static and dynamic obstacles, and safety domain.

Finally, quantitative and objective evaluation of ASV behaviour should be supplemented by qualitative and subjective evaluation by domain experts such as pilots that could observe ASV behaviour in simulated and real environments. This would lead to improved safety evaluation and could help with designing new quantitative performance measures for evaluating safety and risk in ASV operations.

Notes

  1. 1.

    https://www.ntnu.edu/amos.

  2. 2.

    https://www.ntnu.no/blogger/cpslab/.

References

  1. 1.

    Vagale A, Oucheikh R, Bye RT, Osen OL, Fossen TI (2021) Path planning and collision avoidance for autonomous surface vehicles I: a review. J Mar Sci Technol. https://doi.org/10.1007/s00773-020-00787-6

  2. 2.

    Ito M, Zhnng Feifei, Yoshida N (1999) Collision avoidance control of ship with genetic algorithm. IEEE Int Conf Control Appl 1791–1796

  3. 3.

    Hwang C-N, Yang J-M, Chiang C-Y (2001) The design of fuzzy collision-avoidance expert system implemented by H-infinity autopilot. J Mar Sci Technol 9(1):25–37

    Google Scholar 

  4. 4.

    Loe ØAG (2008) Collision avoidance for unmanned surface vehicles. PhD thesis, Norwegian University of Science and Technology

  5. 5.

    Tsou MC, Hsueh CK (2010) The study of ship collision avoidance route planning by ant colony algorithm. J Mar Sci Technol 18(5):746–756

    Google Scholar 

  6. 6.

    Chen C, Chen X-Q, Ma F, Zeng X-J, Wang J (2019) A knowledge-free path planning approach for smart ships based on reinforcement learning. Ocean Eng 189(August)

  7. 7.

    Medyna P, Ma̧ka M (2012) Determination of the shortest path as the basis for examining the most weather favorable routes. Sci J 32(104):29–33 (Maritime University of Szczecin)

    Google Scholar 

  8. 8.

    Candeloro M, Lekkas AM, Sørensen AJ (2017) A Voronoi-diagram-based dynamic path-planning system for underactuated marine vessels. Control Eng Pract 61:41–54

    Article  Google Scholar 

  9. 9.

    Kuwata Y, Wolf MT, Zarzhitsky D, Huntsberger TL (2014) Safe maritime autonomous navigation with COLREGS using velocity obstacles. IEEE J Oceanic Eng 39(1):110–119

    Article  Google Scholar 

  10. 10.

    Wu P, Xie S, Liu H, Li M, Li H, Peng Y, Li X, Luo J (2017) Autonomous obstacle avoidance of an unmanned surface vehicle based on cooperative manoeuvring. Indust Robot 44(1):64–74 1

    Article  Google Scholar 

  11. 11.

    Liu Y, Bucknall R (2015) Path planning algorithm for unmanned surface vehicle formations in a practical maritime environment. Ocean Eng

  12. 12.

    Cheng Y, Zhang W (2018) Concise deep reinforcement learning obstacle avoidance for underactuated unmanned marine vessels. Neurocomputing 272:63–73

    Article  Google Scholar 

  13. 13.

    International Maritime Organization (1972) COLREGs—international regulations for preventing collisions at sea

  14. 14.

    Campbell S, Naeem W, Irwin GW (2012) A review on improving the autonomy of unmanned surface vehicles through intelligent collision avoidance manoeuvres

  15. 15.

    Zhao Y, Li W, Shi P (2016) A real-time collision avoidance learning system for Unmanned Surface Vessels. Neurocomputing

  16. 16.

    Naeem W, Henrique SC, Hu L (2016) A reactive COLREGs-compliant navigation strategy for autonomous maritime navigation. IFAC-Papers Online

  17. 17.

    Beser F, Yildirim T (2018) COLREGS based path planning and bearing only obstacle avoidance for autonomous unmanned surface vehicles. Procedia Comput Sci 131:633–640

    Article  Google Scholar 

  18. 18.

    Tsou MC, Kao SL, Su CM (2010) Decision support from genetic algorithms for ship collision avoidance route planning and alerts. J Navig 63(1):167–182

    Article  Google Scholar 

  19. 19.

    Wang N, Gao Y, Zheng Z, Zhao H, Yin J (2018) A hybrid path-planning scheme for an unmanned surface vehicle. In: 8th International Conference on Information Science and Technology, pp 231–236

  20. 20.

    Wu P, Xie S, Luo J, Qu D, Li Q (2015) The USV path planning based on the combinatorial algorithm. Revista Tecnica de la Facultad de Ingenieria Universidad del Zulia

  21. 21.

    Nguyen M, Zhang S, Wang X (2018) A novel method for risk assessment and simulation of collision avoidance for vessels based on AIS. Algorithms 11(12):204

    MATH  Article  Google Scholar 

  22. 22.

    Xie S, Wu P, Liu H, Yan P, Li X, Luo J, Li Q (2015) A novel method of unmanned surface vehicle autonomous cruise. Indust Robot 43(1):121–130

    Article  Google Scholar 

  23. 23.

    Xue Y, Clelland D, Lee BS, Han D (2011) Automatic simulation of ship navigation. Ocean Eng 38(17–18):2290–2305 12

    Article  Google Scholar 

  24. 24.

    Naeem W, Irwin GW, Yang A (2012) COLREGs-based collision avoidance strategies for unmanned surface vehicles. Mechatronics 22(6):669–678

    Article  Google Scholar 

  25. 25.

    Niu H, Lu Y, Savvaris A, Tsourdos A (2018) An energy-efficient path planning algorithm for unmanned surface vehicles. Ocean Eng

  26. 26.

    Song R, Liu Y, Bucknall R (2019) Smoothed A* algorithm for practical unmanned surface vehicle path planning. Appl Ocean Res 83:9–20 2

    Article  Google Scholar 

  27. 27.

    Bitar G, Breivik M, Lekkas AM (2018) Energy-optimized path planning for autonomous ferries. IFAC-Papers Online 51(29):389–394

    Article  Google Scholar 

  28. 28.

    Xu Q, Meng X, Wang N (2010) Intelligent evaluation system of ship management. Trans Nav Int J Mar Navig Saf Sea Transport 4(4):479–482

    Google Scholar 

  29. 29.

    Xia G, Han Z, Zhao B, Liu C, Wang X (2019) Global path planning for unmanned surface vehicle based on improved quantum ant colony algorithm. Math Prob Eng 1–10(4):2019

    Google Scholar 

  30. 30.

    Chen LJ, Huang LW, Li-Jia C, Li-Wen H (2012) Ship collision avoidance path planning by PSO based on maneuvering equation. Lecture Notes Elect Eng 2:675–682

    Google Scholar 

  31. 31.

    Liu Y, Song R, Bucknall R (2015) A practical path planning and navigation algorithm for an unmanned surface vehicle using the fast marching algorithm. In: MTS/IEEE OCEANS 2015—Genova: discovering sustainable Ocean energy for a new world

  32. 32.

    Song R, Liu Y, Bucknall R (2017) A multi-layered fast marching method for unmanned surface vehicle path planning in a time-variant maritime environment. Ocean Eng

  33. 33.

    Yang JM, Tseng CM, Tseng PS (2015) Path planning on satellite images for unmanned surface vehicles. Int J Naval Archit Ocean Eng

  34. 34.

    Chen L, Negenborn RR, Lodewijks G (2016) Path planning for autonomous inland vessels using A*BG. Int Conf Comput Logist 9855:65–79

    Article  Google Scholar 

  35. 35.

    Kim B, Kim TW (2017) Weather routing for offshore transportation using genetic algorithm. Appl Ocean Res 63:262–275 (2)

    Article  Google Scholar 

  36. 36.

    Wang YH, Cen C (2016) Research on optimal planning method of USV for complex obstacles. In: International Conference on Mechatronics and Automation, IEEE ICMA, pp 2507–2511, Harbin

  37. 37.

    Candeloro M, Lekkas AM, Sørensen AJ, Fossen TI (2013) Continuous curvature path planning using voronoi diagrams and Fermat’s spirals. In: 9th IFAC Conference on Control Applications in Marine Systems, pp 132–137, Osaka

  38. 38.

    Hu L, Naeem W, Rajabally E, Watson G, Mills T, Bhuiyan Z, Raeburn C, Salter I, Pekcan C (2019) A multiobjective optimization approach for COLREGs-compliant path planning of autonomous surface vehicles verified on networked bridge simulators. IEEE Trans Intell Transp Syst 1–12

  39. 39.

    Szlapczynski R (2015) Evolutionary planning of safe ship tracks in restricted visibility. J Navig 68:39–51

    Article  Google Scholar 

  40. 40.

    Chiang H-TL, Tapia L (2018) COLREG-RRT: an RRT-based COLREGS-compliant motion planner for surface vehicle navigation. IEEE Robot Autom Lett 3(3):2024–2031

    Article  Google Scholar 

  41. 41.

    Eriksen B-OH, Bitar G, Breivik M, Lekkas AM (2020) Hybrid collision avoidance for ASVs compliant with COLREGs rules 8 and 13–17. Front Robot AI 7:18 (2)

    Article  Google Scholar 

  42. 42.

    Lyu H, Yin Y (2018) Fast path planning for autonomous ships in restricted waters. Appl Sci (Switzerland) 8(12):6–8

    Google Scholar 

  43. 43.

    Hongdan L, Sheng L, Zhuo Y (2015) Application of adaptive wolf colony search algorithm in ship collision avoidance. Int J Simul Syst Sci Technol 16(2A):1–14

    Google Scholar 

  44. 44.

    Xu H, Wang N, Zhao H, Zheng Z (2018) Deep reinforcement learning-based path planning of underactuated surface vessels. Cyber-Phys Syst, 1–17 (11)

  45. 45.

    Zhao L, Roh M-I (2019) COLREGs-compliant multiship collision avoidance based on deep reinforcement learning. Ocean Eng 191:106436 (11)

    Article  Google Scholar 

  46. 46.

    Lyu H, Yin Y (2019) COLREGS-constrained real-time path planning for autonomous ships using modified artificial potential fields. J Navig 72:588–608

    Article  Google Scholar 

  47. 47.

    Kufoalor DKM, Brekke EF, Johansen TA (2018) Proactive collision avoidance for ASVs using a dynamic reciprocal velocity obstacles method. IEEE Int Conf Intell Robots Syst 2402–2409

  48. 48.

    Varas JM, Hirdaris S, Smith R, Scialla P, Caharija W, Bhuiyan Z, Mills T, Naeem W, Hu L, Renton I, Motson D, Rajabally E (2017) MAXCMAS project: autonomous COLREGs compliant ship navigation. In: 16th Conference on Computer Applications and Information Technology in the Maritime Industries (COMPIT), pp 454–464

  49. 49.

    Shen H, Hashimoto H, Matsuda A, Taniguchi Y, Terada D, Guo C (2019) Automatic collision avoidance of multiple ships based on deep Q-learning. Appl Ocean Res 86:268–288 5

    Article  Google Scholar 

  50. 50.

    Lee YI, Kim SG, Kim YG (2015) Fuzzy relational product for collision avoidance of autonomous ships. Intell Autom Soft Comput

  51. 51.

    Tam C, Bucknall R (2013) Cooperative path planning algorithm for marine surface vessels. Ocean Eng 57:25–33

    Article  Google Scholar 

  52. 52.

    Blaich M, Köhler S, Reuter J, Hahn A (2015) Probabilistic collision avoidance for vessels. IFAC-Papers Online 28(16):69–74

    Article  Google Scholar 

  53. 53.

    Johansen TA, Perez T, Cristofaro A (2016) Ship collision avoidance and COLREGS compliance using simulation-based control behavior selection with predictive hazard assessment. IEEE Trans Intell Transp Syst 17(12):3407–3422

    Article  Google Scholar 

  54. 54.

    Tam C, Bucknall R (2010) Path-planning algorithm for ships in close-range encounters. J Mar Sci Technol 15(4):395–407

    Article  Google Scholar 

  55. 55.

    Perera L, Carvalho J, Soares C (2010) Autonomous guidance and navigation based on the COLREGs rules and regulations of collision avoidance. In: Advanced Ship Design for Pollution Prevention. Taylor & Francis Group

  56. 56.

    Campbell S, Abu-Tair M, Naeem W (2014) An automatic COLREGs-compliant obstacle avoidance system for an unmanned surface vehicle. Proc Inst Mech Eng Part M 228(2):108–121

    Article  Google Scholar 

  57. 57.

    Liu Y, Bucknall R (2016) The angle guidance path planning algorithms for unmanned surface vehicle formations by using the fast marching method. Appl Ocean Res 59:327–344

    Article  Google Scholar 

  58. 58.

    Liu Y, Bucknall R, Zhang X (2017) The fast marching method based intelligent navigation of an unmanned surface vehicle. Ocean Eng 142(April):363–376

    Article  Google Scholar 

  59. 59.

    Lazarowska A (2016) A trajectory base method for ship’s safe path planning. Procedia Comput Sci 96:1022–1031 1

    Article  Google Scholar 

  60. 60.

    Kearon J (1979) Computer programs form collision avoidance and traffic keeping. In: Conference on mathematical aspects of marine traffic, pp 229–242, London

  61. 61.

    Nguyen MD, Nguyen VT, Tamaru H (2012) Automatic collision avoiding support system for ships in congested waters and at open sea. In: International Conference on Control, Automation and Information Sciences (ICCAIS), pp 96–101

  62. 62.

    Xu Q, Zhang C, Wang N (2014) Multiobjective optimization based vessel collision avoidance strategy optimization. Math Prob Eng 2014:9

    Google Scholar 

  63. 63.

    Lekkas A, Fossen TI (2013) Line-of-sight guidance for path following of marine vehicles. In: Oren G (ed) Advanced in marine robotics. Lambert Academic Publishing, pp 1–29

  64. 64.

    Buniyamin N, Ngah WW, Sariff N, Mohamad Z (2011) A simple local path planning algorithm for autonomous mobile robots. Int J Syst Appl Eng Dev 5(2):151–159

    Google Scholar 

  65. 65.

    Garrido S, Moreno L, Blanco D (2008) Exploration of a cluttered environment using Voronoi transform and fast marching. Robot Auton Syst 56(12):1069–1081

    Article  Google Scholar 

  66. 66.

    Soulignac M, Taillibert P, Rueher M (2009) Time-minimal path planning in dynamic current fields. In: IEEE International Conference on Robotics and Automation, pp 2473–2479, Kobe, 5 2009. IEEE

  67. 67.

    Kavraki LE, Švestka P, Latombe J-C, Overmars MH (1996) Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Trans Robot Autom 12(4):566–580

    Article  Google Scholar 

  68. 68.

    Petres C, Pailhas Y, Patron P, Petillot Y, Evans J, Lane D (2007) Path planning for autonomous underwater vehicles. IEEE Trans Rob 23(2):331–341

    Article  Google Scholar 

Download references

Acknowledgements

This work is partly sponsored by the Research Council of Norway through the Centre of Excellence funding scheme, Project Number 223254, AMOSFootnote 1. The work was also supported by the European Research Consortium for Informatics and Mathematics (ERCIM), which provided funding to Rachid Oucheikh for his postdoctoral fellowship in the Cyber-Physical Systems LaboratoryFootnote 2 at NTNU in Ålesund.

Funding

Open Access funding provided by NTNU Norwegian University of Science and Technology (incl St. Olavs Hospital - Trondheim University Hospital).

Author information

Affiliations

Authors

Corresponding author

Correspondence to Anete Vagale.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

This article is published under an open access license. Please check the 'Copyright Information' section either on this page or in the PDF for details of this license and what re-use is permitted. If your intended use exceeds what is permitted by the license or if you are unable to locate the licence and re-use information, please contact the Rights and Permissions team.

About this article

Verify currency and authenticity via CrossMark

Cite this article

Vagale, A., Bye, R.T., Oucheikh, R. et al. Path planning and collision avoidance for autonomous surface vehicles II: a comparative study of algorithms. J Mar Sci Technol (2021). https://doi.org/10.1007/s00773-020-00790-x

Download citation

Keywords

  • Autonomous surface vehicle (ASV)
  • Path planning
  • Collision avoidance
  • Algorithms
  • Safety