723

1 INTRODUCTION

The prospect of merchant ships that sense their

environment, plan their passage and execute

manoeuvres with reduced or no direct human

intervention has reshaped the research agenda of the

maritime sector. Early conceptual and project work,

exemplified by the Maritime Unmanned Navigation

through Intelligence in Networks initiative, established

the architectures and operational concepts through

which an unmanned merchant vessel might be

supervised from a shore control centre [1]. What began

as a small number of feasibility studies has become one

of the most active frontiers in maritime science and

technology, described under the umbrella term

Maritime Autonomous Surface Ships, or MASS [2].

The interest is driven by a convergence of

expectations. The principal argument advanced in

support of autonomy is navigational safety, on the

reasoning that removing the crew removes the human

errors implicated in the majority of maritime accidents

[3]. Economic arguments have been developed in

parallel, with exploratory cost-comparison work

examining the conditions under which unmanned

operation could be commercially advantageous [4].

The technological enablers have matured alongside

these expectations, including ship power and

The Governance Lag in Maritime Autonomous Surface

Ship Research: A Bibliometric Analysis of Growth,

Themes and Gaps, 2010-2025

V. Milin, P. Ivančić, M. Čorić & D. Medić

University of Split, Split, Croatia

ABSTRACT: Research on Maritime Autonomous Surface Ships (MASS) has expanded rapidly, yet the operational

introduction of autonomous vessels is widely held to depend less on technical capability than on risk governance,

regulation and the human element. Whether the research effort reflects that dependency has not been tested. This

study asks whether governance-oriented MASS research is keeping pace with technical research. A reproducible

corpus of 1,874 peer-reviewed journal articles and reviews published between 2010 and 2025 was retrieved from

the open OpenAlex database and screened transparently. Growth was modelled by log-linear regression, the

intellectual base was identified by co-citation, the intellectual structure was mapped by keyword co-occurrence

and Louvain clustering with a strategic diagram, and thematic composition was tracked across three periods.

Comparative growth rates were then estimated for seven lexically defined topics. The field grew at 46.9 per cent

per year between 2013 and 2025, doubling every 1.8 years. Its composition shifted decisively away from platform

and propulsion engineering, which fell from 34.3 to 17.0 per cent of output, towards risk, safety, security and

operations, which rose from 25.7 to 39.0 per cent. Against this, research on regulation, law and liability grew at

only 16.9 per cent per year, less than half the rate of perception research, and its share fell after 2021 even as the

International Maritime Organization advanced its MASS instrument. The field is consolidating around risk and

operations while its regulatory foundations lag, which constitutes a measurable governance lag and the principal

frontier for future work.

http://www.transnav.eu

the International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 20

Number 3

September 2026

DOI: 10.12716/1001.20.03.19

724

propulsion architectures suited to highly automated

operation [5] and the connectivity infrastructure

required to link vessels with shore [6]. Yet the ocean-

going application of these concepts continues to raise

unresolved challenges and threats [7], and an

influential strand of the literature cautions that

autonomy displaces rather than eliminates risk,

transferring it toward software assurance, sensing

reliability, cyber exposure and the changed role of a

remote operator [3].

This tension defines the practical position of the

field. It is widely accepted that the binding constraints

on operational MASS are not exclusively technical. The

regulatory framework remains under construction: the

regulatory scoping exercise conducted by the

International Maritime Organization identified

numerous provisions across the existing conventions

that require reinterpretation or amendment before

higher degrees of autonomy can be certificated, and it

explicitly flagged the treatment of the master, the crew

and the responsible person as unresolved [8]. If

deployment depends on governance as much as on

technology, one would expect the research effort to

reflect that dependency. Whether it does is an

empirical question, and it is the question this study

addresses.

The MASS literature has already attracted several

reviews. Narrative and focused syntheses have

surveyed autonomous-ship applications and business

models [9], the costs and benefits of autonomous

shipping [10], and human and artificial-intelligence

interaction in autonomous ship systems [11].

Quantitative mapping has also been undertaken.

Bibliometric reviews now exist for the risk, safety and

reliability of autonomous ships [12], for MASS risk and

reliability analysis over 2015 to 2022 [13], for intelligent

ship collision avoidance [14], for maritime

cybersecurity [15], and for machine learning in

maritime safety for autonomous shipping [16]. A field-

wide bibliometric analysis of autonomous shipping

technologies has also been published, drawing on

records indexed in Web of Science [17]. The claim that

the MASS literature has not been mapped

quantitatively can therefore no longer be sustained,

and any new contribution must be positioned carefully

against this body of work.

Three limitations of the existing quantitative

literature nevertheless remain. First, the sub-domain

reviews are, by construction, bounded: a bibliometric

study of risk and reliability [12], [13] or of collision

avoidance [14] can describe the internal development

of its own topic with authority, but it cannot compare

that topic with the others, because the comparison

requires a corpus that contains them all. The question

of whether one part of the field is advancing faster than

another is thus systematically outside the reach of the

studies best placed to care about the answer. Second,

the field-wide analyses rely on proprietary databases

whose coverage is subscription-dependent and whose

queries cannot be re-executed by readers without a

licence [17], which limits reproducibility at a time

when it is increasingly expected of bibliometric work

[18]. Third, existing studies map the literature without

synthesising what it has established, so the reader

learns which topics are large without learning what is

known about them.

This study addresses those three limitations. Its

contribution is fourfold. It analyses a field-wide corpus

drawn entirely from an open database, so that the

search, the screening and the analysis can be re-

executed by any reader without a subscription. It

models growth by log-linear regression rather than by

endpoint ratios, which yields a defensible estimate of

the expansion rate and its uncertainty. It introduces a

comparative growth analysis across lexically defined

topics, which permits the central question, whether

governance-oriented research is keeping pace with

technical research, to be answered quantitatively rather

than asserted. Finally, it combines the mapping with a

synthesis of what each theme has established, so that

the structural findings are interpretable to a maritime

readership rather than to bibliometricians alone.

Bibliometric analysis is the appropriate instrument

for this purpose. By applying quantitative methods to

large volumes of publication metadata it characterises

a research field objectively and traces its development

transparently [18], while science-mapping techniques

expose the latent thematic structure of a literature and

how it evolves [19]. The approach is descriptive of the

research effort rather than evaluative of individual

studies, a distinction to which we return when

discussing limitations.

Accordingly, this study addresses five research

questions. RQ1: how has the volume of MASS research

evolved between 2010 and 2025, and at what rate. RQ2:

which works constitute the intellectual base on which

the field draws, and which studies have been most

influential within it. RQ3: what is the intellectual

structure of the field, and how mature is each theme.

RQ4: how has the thematic composition of the field

changed over time, and which topics are emerging or

receding. RQ5: is governance-oriented research

growing as fast as technically oriented research, and if

not, what follows for the research agenda. The

remainder of the paper sets out the data and methods,

reports the results, discusses what the field has

established theme by theme and what the growth

asymmetry implies, and concludes.

2 MATERIALS AND METHODS

2.1 Research design

The study combines three complementary strands of

bibliometric enquiry. Performance analysis quantifies

the volume, growth and citation impact of the

literature. Science mapping reconstructs its intellectual

and conceptual structure through co-citation and co-

word networks. A comparative growth analysis,

developed for this study, tests whether the constituent

topics of the field are expanding at similar rates. All

three strands draw on a single, transparently

constructed corpus, and all figures and indicators are

computed directly from the retrieved metadata.

2.2 Data source

Bibliographic data were obtained from OpenAlex, an

open index of scholarly works, authors, sources and

institutions [20]. Three considerations governed this

choice. First, coverage: OpenAlex indexes a very large

share of the scholarly record, and a recent reference-

725

coverage analysis benchmarking it against Web of

Science and Scopus found its coverage to be broadly

comparable, while noting differences in metadata

completeness that are relevant to affiliation-level

analysis [21]. Second, reproducibility: the application

programming interface allows a query to be specified,

executed and re-executed exactly, and the retrieval

date can be reported so that any divergence in a later

replication is attributable to database growth rather

than to an undocumented change in method. Third,

accessibility: because the database is unrestricted,

readers can reproduce the analysis without an

institutional subscription, which is not the case for the

proprietary sources used in previous field-wide

mapping of this literature [17]. All records were

retrieved on 17 July 2026.

2.3 Search strategy

A Boolean query was applied to the title and abstract

fields, combining the canonical term and its principal

variants: "maritime autonomous surface ship" and its

plural, together with "autonomous ship", "autonomous

ships", "autonomous vessel", "autonomous vessels",

"autonomous shipping", "unmanned ship" and

"unmanned cargo ship". The query was deliberately

restricted to the maritime-transport framing of

autonomy. Terms denoting underwater vehicles or

purely robotic surface platforms were not included as

search terms, because the wider unmanned-vehicle

literature addresses a different engineering problem

and would dilute the corpus. This decision has a cost,

since some work on unmanned surface vehicles is

directly relevant to MASS navigation, and it is revisited

in the limitations. The search period was 2010 to 2025

inclusive.

2.4 Screening and corpus construction

The screening procedure follows the reporting logic of

the PRISMA 2020 framework, adapted to a bibliometric

context in which records are screened by metadata

rather than assessed for eligibility by full-text reading

[22]. The steps are shown in Figure 1. The query

retrieved 4,213 records. Records outside the 2010 to

2025 window (458) and records not in English (184)

were removed, leaving 3,571. The corpus was then

restricted to peer-reviewed journal articles and

reviews, which excluded 1,697 records, of which the

largest group was conference papers (1,056), followed

by book chapters (210), preprints (155) and a residual

of theses, reports, editorials and similar material (276).

The final analytical corpus comprises 1,874 journal

articles and reviews.

Two features of this procedure warrant comment.

First, the exclusion of conference papers is

consequential in a field where proceedings carry a

substantial part of the discourse, and it reduces the

corpus by more than a third. It is nevertheless justified

on grounds of metadata consistency: conference

records in the source database have markedly less

complete keyword, affiliation and citation metadata,

and every analysis reported here depends on those

fields. The excluded conference literature is therefore

reported as context rather than analysed, and the

corpus should be read as the peer-reviewed journal

record of the field rather than as its complete output.

Second, this definition is stricter than that used in some

earlier work; the document-type taxonomy of the

source database now separates conference papers from

journal articles explicitly, and studies that predate this

separation may include proceedings within counts

described as articles.

Figure 1. Flow diagram of the corpus construction and

screening procedure.

2.5 Performance indicators and growth modelling

Standard indicators were computed: annual and

cumulative output, total and mean citations, the

number of contributing sources, authors, institutions

and countries, and the share of internationally co-

authored articles. The concentration of output across

sources was examined following the logic of Bradford's

law of scattering, by identifying the number of sources

required to account for the first third of all articles.

Growth was modelled by ordinary least-squares

regression of the natural logarithm of annual output on

the publication year over 2013 to 2025, the interval

during which the field sustains a non-trivial annual

volume. The fitted slope yields a constant annual

growth rate and an implied doubling time, and the

coefficient of determination indicates how closely the

expansion follows an exponential path. This approach

is preferred to the compound annual growth rate

computed from the first and last years of an interval,

which is widely used in bibliometric work but is

determined entirely by two observations and is

therefore unstable when the base year is small. Both

quantities are reported so that the difference can be

judged.

2.6 Science mapping

The intellectual base of the field, meaning the body of

prior work on which it draws, was identified by

counting how often each reference cited by corpus

articles is cited within the corpus. This local citation

count distinguishes works that are foundational for

MASS research specifically from works that are merely

well cited in general [23]. The most influential studies

within the field were identified separately by their

global citation counts.

726

The conceptual structure was reconstructed from a

keyword co-occurrence network. Keywords supplied

by the source database were normalised by removing

generic discipline labels, merging evident synonyms

and discarding terms with a relevance score below

0.30; the sixty most frequent surviving terms were

retained and linked when they co-occurred in at least

three articles. Communities were detected with the

Louvain modularity-maximisation algorithm [24], and

each cluster was positioned on a strategic diagram

according to its centrality, defined as the mean strength

of its external links and interpreted as the theme's

relevance to the field as a whole, and its density,

defined as the mean strength of its internal links and

interpreted as the theme's internal development [25],

[26]. Each article was assigned to the cluster

contributing most of its keywords, which permits the

thematic composition of the corpus to be tracked across

the periods 2010 to 2017, 2018 to 2021 and 2022 to 2025.

The relative recency of individual terms was

summarised by the mean publication year of the

articles carrying them.

2.7 Lexical topic probes and comparative growth

The cluster structure describes how the vocabulary of

the field organises itself, but it does not isolate policy-

relevant topics that may be distributed across several

clusters. Regulation, for instance, is discussed in

articles whose dominant vocabulary is operational. To

address RQ5, seven topics were therefore defined a

priori from the practical concerns identified in the

introduction: collision avoidance and path planning;

perception and situational awareness; risk and safety

assessment; cyber security; the human element and

training; regulation, law and liability; and techno-

economics and the environment. Each topic was

operationalised as a regular-expression probe over the

title and keywords of every article, and an article

matching a probe was counted under that topic. For

each topic the prevalence, the share within each period,

the mean citation count and the annual growth rate

over 2018 to 2025, estimated by the same log-linear

regression used for the field as a whole, were

computed.

The probes are lexical instruments and their

properties should be understood before the results are

read. They are deliberately inclusive, so an article is

counted under a topic if it engages with it at all, not

only if it is principally about it; the topics consequently

overlap, and their shares do not sum to unity. Their

absolute levels depend on the breadth of the

expressions used and should not be over-interpreted.

Their comparative and temporal behaviour is more

robust, because each probe is applied identically across

years, so that a difference in growth rates between two

topics is not an artefact of how broadly either is drawn.

The probe definitions are reported in full so that they

can be varied and the analysis repeated.

2.8 Reproducibility

The query, the retrieval date, the screening counts, the

normalisation rules and the probe expressions are

reported in this paper. The analysis uses only fields

available through the public interface of the source

database, so the corpus and every indicator reported

here can be regenerated independently. Because the

database is updated continuously, a later replication

will retrieve a slightly larger corpus; the structural

findings, which concern proportions and rates, are not

sensitive to this growth.

3 RESULTS

3.1 Growth and volume of the field

The MASS literature has expanded almost without

interruption over the study period, as shown in Figure

2. Annual output stood at two articles in 2010 and

remained in single figures until 2015. It rose to 29

articles in 2016, paused at 27 in 2017, and then

accelerated sharply, reaching 76 in 2018, 207 in 2021

and 418 in 2025. The log-linear model fitted over 2013

to 2025 gives an annual growth rate of 46.9 per cent

with a coefficient of determination of 0.936, implying

that the volume of MASS research has doubled every

1.8 years. The compound annual growth rate

computed from the endpoints of the same interval is

42.4 per cent, so the two estimates agree closely and the

conclusion does not depend on the choice of estimator.

Over the shorter and more recent interval 2018 to 2025,

growth settles at 25.4 per cent per year with a

coefficient of determination of 0.956, indicating that

expansion continues at a high but moderating rate

from a much larger base.

Figure 2. Annual and cumulative output of MASS journal

articles and reviews, 2010 to 2025, with the fitted exponential

growth model.

Table 1. Summary indicators of the analytical corpus.

Indicator

Value

Journal articles and reviews

1,874

Publication period

2010-2025

Distinct sources

570

Distinct authors

4,206

Total citations received

35,333

Mean citations per article

18.9

Internationally co-authored articles

20.0%

Single-authored articles

15.7%

Annual growth rate, 2013-2025 (log-linear)

46.9% (R² = 0.936)

Implied doubling time, 2013-2025

1.8 years

Annual growth rate, 2018-2025 (log-linear)

25.4% (R² = 0.956)

Sources carrying the first third of output

10 of 570 (1.8%)

The summary indicators of the corpus are reported

in Table 1. The 1,874 articles appeared in 570 sources

and involved 4,206 distinct authors, and they had

accumulated 35,333 citations at the retrieval date, a

mean of 18.9 citations per article. One article in five was

internationally co-authored, and 15.7 per cent were

single-authored, a low share consistent with a

727

laboratory-based engineering field. The acceleration

after 2017 coincides with the period in which the

autonomous-ship concept moved from isolated

demonstration projects onto the agendas of regulatory

and classification bodies, most visibly through the

regulatory scoping exercise initiated at the

International Maritime Organization [8].

3.2 Contributors and outlets

Output is geographically concentrated, as Table 2

shows. China contributes to 528 articles, or 28.2 per

cent of the corpus, followed at a distance by Norway

with 169. Norway's position is disproportionate to the

size of its research system and reflects a sustained

national concentration of effort on autonomous

shipping. South Korea, the United Kingdom and the

United States form a second tier, ahead of a group of

northern European maritime nations. The institutional

profile mirrors this pattern: Dalian Maritime

University and Wuhan University of Technology lead,

followed by the Norwegian University of Science and

Technology, with Aalto University, Delft University of

Technology and Shanghai Maritime University

forming the next group. Specialised maritime

universities and technical institutes therefore anchor

the field.

Table 2. Ten most productive countries and institutions by

article count (n = 1,874).

Rank

Country

Art.

Institution

Art.

China

528

28.2

Dalian Maritime University

158

Norway

169

9.0

Wuhan University of Technology

133

South Korea

143

7.6

Norwegian University of Science

and Technology

109

United

Kingdom

135

7.2

Aalto University

United States

5.2

Delft University of Technology

Poland

3.9

Shanghai Maritime University

Finland

3.5

Liverpool John Moores University

Netherlands

3.1

Korea Maritime and Ocean

University

Japan

2.9

Gdynia Maritime University

Russia

2.5

China Ocean Shipping (China)

Table 3. Ten most productive sources by article count.

Rank

Source

Articles

(%)

Ocean Engineering

197

10.5

Journal of Marine Science and Engineering

160

8.5

TransNav the International Journal on Marine

Navigation and Safety of Sea Transportation

2.9

Applied Sciences

2.1

IFAC-PapersOnLine

2.1

Reliability Engineering & System Safety

2.0

IEEE Access

1.7

Journal of Marine Science and Technology

1.4

Journal of Marine Engineering & Technology

1.4

Safety Science

1.2

Publications are dispersed across 570 sources but

concentrate sharply within them, as Table 3 shows.

Ocean Engineering and the Journal of Marine Science

and Engineering lead by a wide margin, followed by

the journal devoted to marine navigation and safety of

sea transportation, and by Reliability Engineering and

System Safety and Safety Science. Applying the logic of

Bradford's law of scattering, only ten sources, or 1.8 per

cent of all outlets, account for the first third of the

corpus, which indicates a well-defined disciplinary

core surrounded by a long tail of occasional venues.

The presence of safety and reliability journals

alongside the engineering titles shows that risk

questions occupy an established place among the

principal outlets of the field.

3.2.1 The intellectual base

Counting how frequently each cited reference is

cited by articles within the corpus identifies the works

on which the field actually draws, as distinct from

works that are merely well cited in general. The corpus

cites 32,245 distinct references; the most locally cited

are reported in Table 4.

The structure of this base is informative. Its single

most cited work is the study by Wróbel and colleagues

on the potential impact of unmanned vessels on

maritime transportation safety [3], cited by 145 articles,

or nearly eight per cent of the entire corpus. That a

safety study, rather than a control-engineering or

propulsion study, anchors the field is a significant

observation, and it indicates that the field understands

itself as a safety problem. The second most cited work

is Fossen's handbook of marine craft hydrodynamics

and motion control [27], which supplies the modelling

foundation on which guidance and control research is

built and is the only monograph in the base. The

remainder is dominated by risk and safety

contributions, including risk assessment of MASS

operations [28], a framework for identifying factors

influencing navigational risk [29], supervisory risk

control [30], the applicability of established ship-risk

models to MASS [31], system-theoretic safety

assessment [32] and early risk assessment for an

unmanned merchant ship [33], alongside the collision-

avoidance strand represented by a state-of-the-art

review [34], COLREGS-constrained path planning [35],

simulation-based control behaviour for COLREGS

compliance [36] and a review of collision-avoidance

navigation systems [37]. Human-centred and economic

foundations are also present through operator task

analysis for collision avoidance [38], human-system

concurrent task analysis [39], the MUNIN concept [1],

exploratory cost analysis of unmanned operation [4],

an overview of unmanned surface vehicle

developments [40] and an examination of

implementation factors for autonomous merchant

vessels [41].

Table 4. Fifteen most co-cited references constituting the

intellectual base of the field, ranked by local citations, that

is, citations received from articles within the corpus.

Rank

Cited work

Year

Local citations

Wróbel et al. [3]

2017

145

Fossen [27]

2011

117

Chang et al. [28]

2021

101

Fan et al. [29]

2020

Huang et al. [34]

2020

Lyu et al. [35]

2018

Utne et al. [30]

2020

Abilio Ramos et al. [38]

2019

Burmeister et al. [1]

2014

Kretschmann et al. [4]

2017

Johansen et al. [36]

2016

Thieme et al. [31]

2018

Zhang et al. [37]

2021

Wróbel et al. [32]

2018

Rødseth et al. [33]

2015

728

3.3 Most influential studies within the field

The most-cited articles within the corpus, reported in

Table 5, describe a different distribution from the

intellectual base. Influence measured by global

citations attaches most strongly to enabling

technology: the leading article reviews the design and

control of hybrid power and propulsion systems for

smart ships [5], and it is followed by a survey of video

processing from electro-optical sensors for maritime

object detection and tracking [42]. Collision avoidance

and path planning are represented by a state-of-the-art

review [34], COLREGS-constrained real-time path

planning [35], a reinforcement-learning approach to

path planning that dispenses with prior knowledge

[43], a review of collision-avoidance navigation

systems for MASS [37] and a generalized velocity

obstacle algorithm [44]. Safety and risk are represented

by the assessment of the safety impact of unmanned

vessels [3] and by risk assessment of MASS operations

[28], while the digital and connectivity dimension

appears through work on the digital transformation of

the maritime transport sector [45], a review of sensors

and artificial-intelligence techniques for situational

awareness [46] and a survey of the Internet of Ships [6].

The contrast between Tables 4 and 5 is itself a finding:

the works the field builds upon are predominantly

about safety, whereas the works that attract the widest

citation are predominantly about technology.

Table 5. Twelve most-cited articles within the corpus,

ranked by global citations at the retrieval date.

Rank

Study

Year

Source

Citations

Geertsma et al. [5]

2017

Applied Energy

581

Prasad et al. [42]

2017

IEEE Trans. Intell. Transp.

Syst.

468

Huang et al. [34]

2020

Safety Science

419

Lyu et al. [35]

2018

Journal of Navigation

356

Wróbel et al. [3]

2017

Reliab. Eng. Syst. Saf.

328

Chang et al. [28]

2021

Reliab. Eng. Syst. Saf.

299

Tijan et al. [45]

2021

Technol. Forecast. Soc.

Change

274

Thombre et al.

[46]

2022

IEEE Trans. Intell. Transp.

Syst.

269

Aslam et al. [6]

2020

IEEE Internet Things J.

253

Chen et al. [43]

2019

Ocean Engineering

233

Zhang et al. [37]

2021

Ocean Engineering

232

Huang et al. [44]

2019

Ocean Engineering

225

3.4 Intellectual structure and thematic maturity

The keyword co-occurrence network, shown in Figure

3, comprises 60 terms and 417 links and resolves into

five communities with a modularity of 0.299. The first

gathers the engineering platform, combining marine

engineering, propulsion, hull and the vessel itself as an

object of design. The second centres on collision and

obstacle avoidance together with the learning and

data-driven methods increasingly applied to it. The

third, by far the largest with 24 terms, assembles risk

assessment, cyber security, port operations, the

maritime industry, automation and the crew. The

fourth covers perception and sensing, including

computer vision, deep learning, remote sensing,

positioning and radar. The fifth concerns guidance and

control, including path planning, control theory,

optimisation, model-predictive control and fuzzy

logic.

Figure 3. Keyword co-occurrence network of the MASS

literature; node size reflects term frequency, colour denotes

thematic cluster.

Figure 4. Strategic diagram of MASS research themes; the

horizontal axis represents centrality and the vertical axis

density, with bubble size proportional to keyword volume.

Positioning the clusters on the strategic diagram,

Figure 4, distinguishes their maturity, and the

summary coordinates are given in Table 6. The

platform cluster and the collision-avoidance cluster

combine high centrality with high density and occupy

the motor-theme region: they are central to the field

and internally well developed. The guidance and

control cluster is the densest of all, with a density of 82

against a median of 73, while its centrality falls exactly

on the median, placing it at the boundary of the motor

region; this is the signature of a well-consolidated and

methodologically self-sufficient specialty that is highly

developed internally without being the most

outwardly connected. The perception cluster sits low

on both dimensions. The most consequential position

is that of the risk, safety, security and operations

cluster. It is the largest theme in the field by every

volume measure, with 24 terms, a keyword volume of

1,445 and 614 assigned articles, yet it records the lowest

centrality (40) and the lowest density (39) of any

cluster. A theme can only be simultaneously the largest

and the least cohesive if it is internally heterogeneous:

it is not one research programme but many, loosely

joined, spanning risk analysis, cyber security, port

729

operations, regulation and the human element without

a shared method.

Table 6. Thematic clusters with their strategic coordinates

and assigned articles.

Theme

Terms

Vol.

Art.

Cen.

Den.

Position

Platform and propulsion

691

298

140

Motor

Collision avoidance and

learning

629

194

118

Motor

Guidance and control

881

324

Motor

(boundary)

Perception and sensing

332

159

Emerging /

peripheral

Risk, safety, security and

operations

1445

614

Emerging /

peripheral

3.5 Thematic evolution, 2010 to 2025

Tracking the composition of the corpus across three

periods reveals a decisive reordering of the field,

shown in Figure 5(a). In the formative period 2010 to

2017, when only 70 articles carried assignable

keywords, the platform cluster was the single largest

theme at 34.3 per cent of output, consistent with a field

then preoccupied with whether an unmanned vessel

could be built at all. By 2018 to 2021 its share had fallen

to 20.5 per cent, and by 2022 to 2025 to 17.0 per cent, a

halving in relative terms across the period. The risk,

safety, security and operations cluster moved in the

opposite direction, rising from 25.7 per cent to 39.7 per

cent and settling at 39.0 per cent, and it is now

comfortably the largest theme. Collision avoidance

grew from a negligible 2.9 per cent to 13.4 per cent.

Guidance and control remained conspicuously stable

at close to one fifth of output throughout, moving only

from 21.4 to 20.2 per cent, and perception fluctuated

between 8.2 and 15.7 per cent without a clear trend.

The field has therefore not simply grown; it has

changed its subject. The question that organises MASS

research has shifted from whether an autonomous

vessel can be engineered to whether its operation can

be shown to be safe and managed. This is a maturation

signature, and it is consistent with the finding that a

safety study anchors the intellectual base.

Figure 5. (a) Thematic composition of the corpus across three

periods, by share of articles assigned to each cluster. (b)

Annual growth rate of each lexically defined topic over 2018

to 2025, compared with the growth of the field as a whole

over the identical interval.

3.6 Emerging and receding terms

Summarising each term by the mean publication year

of the articles carrying it identifies which parts of the

vocabulary are arriving and which are receding, as

shown in Figure 6. The most recent terms are liability

(mean year 2023.14), environmental resource

management (2023.12), software deployment (2023.06),

training (2023.03), architecture (2022.92), the

autonomous surface ship itself as a named object

(2022.80), reinforcement learning (2022.80) and

machine learning (2022.73). The terms with the earliest

mean years are propulsion (2020.91), hull (2021.15),

control engineering (2021.19), operations management

(2021.28), shore (2021.29) and the automatic

identification system (2021.32).

The pattern is coherent with the thematic evolution.

The vocabulary of the physical platform, propulsion

and hull, is receding, while the vocabulary of software,

learning, training and legal responsibility is arriving.

The appearance of liability as the single most recent

term of any consequence is notable, and it indicates

that questions of legal responsibility have begun to

enter the technical literature rather than remaining the

preserve of legal scholarship. It should be read against

the volume of that work, however, which the next

subsection quantifies.

Figure 6. Emerging and receding terms, measured by the

mean publication year of the articles carrying each term;

marker size is proportional to the number of articles.

3.7 Comparative growth of technical and governance

research

The seven lexical probes permit a direct test of whether

the constituent topics of the field are expanding at

comparable rates. Their prevalence, citation impact

and growth are reported in Table 7 and their growth is

compared with the field in Figure 5(b). Risk and safety

assessment is the most prevalent topic, engaged by 22.3

per cent of articles, followed by cyber security at 21.5

per cent, regulation, law and liability at 19.8 per cent

and collision avoidance and path planning at 19.4 per

cent. Risk and safety assessment also attracts the

highest mean citation count at 30.2, ahead of cyber

security at 28.9 and collision avoidance at 25.7, while

the human element attracts the lowest at 19.4.

The growth rates are the central result of this study.

Over 2018 to 2025 the field as a whole grew at 25.4 per

cent per year. Perception and situational awareness

grew fastest at 36.9 per cent, followed by collision

avoidance and path planning at 33.9 per cent and the

human element and training at 30.9 per cent, all of

them faster than the field. Cyber security at 25.9 per

cent, risk and safety assessment at 25.5 per cent and

techno-economics and the environment at 24.4 per cent

tracked the field closely. Research on regulation, law

and liability grew at 16.9 per cent per year, two thirds

730

of the rate of the field as a whole and less than half the

rate of perception research. It is the only topic that grew

materially more slowly than the literature that

surrounds it.

The temporal profile sharpens the finding. The

share of articles engaging regulation rose from 15.1 per

cent in 2010 to 2017 to 23.8 per cent in 2018 to 2021, the

period spanning the regulatory scoping exercise, and

then fell back to 18.4 per cent in 2022 to 2025.

Regulatory research thus responded strongly to the

opening of the international regulatory agenda and

then lost ground, in relative terms, precisely during the

years in which that agenda advanced toward a formal

instrument. No other topic exhibits this rise-and-

retreat profile.

Table 7. Prevalence, citation impact and growth of the seven

lexically defined topics. Shares do not sum to unity because

the probes are inclusive and overlapping.

Topic

Art.

(%)

Mean

cit.

2010-

17 (%)

2018-

21 (%)

2022-

25 (%)

Growth

(%/yr)

Perception and

situational

awareness

211

11.3

20.0

8.1

9.7

12.2

36.9

Collision avoidance

and path planning

363

19.4

25.7

9.3

17.9

20.7

33.9

Human element

and training

322

17.2

19.4

14.0

17.5

17.3

30.9

Cyber security

403

21.5

28.9

15.1

22.0

21.7

25.9

Risk and safety

assessment

417

22.3

30.2

20.9

23.1

22.0

25.5

Techno-economics

and environment

347

18.5

18.1

23.3

18.8

18.1

24.4

Regulation, law

and liability

371

19.8

21.9

15.1

23.8

18.4

16.9

4 DISCUSSION

The results describe a field that has grown at a rate few

areas of maritime science have matched, that has

decisively reoriented itself from building autonomous

vessels to governing their operation, and whose largest

theme is also its least cohesive. They also isolate a

specific asymmetry: the one topic growing materially

more slowly than the field is the one concerned with

regulation, law and liability. This section first

synthesises what each theme has established, then

interprets the asymmetry, positions the study against

prior quantitative work, and sets out a research

agenda.

4.1 What the field has established

4.1.1 Guidance, control and collision avoidance

This is the most mature part of the field, and the

strategic diagram registers it as such: guidance and

control is the densest cluster in the network, and

collision avoidance sits squarely among the motor

themes. Its maturity rests on an unusually stable

foundation. The hydrodynamic and motion-control

models codified by Fossen [27] provide a common

vessel representation, which allows methods

developed by different groups to be compared on

comparable terms, and the international collision

regulations supply an agreed behavioural specification

against which a manoeuvre can be judged correct or

incorrect. The literature has consequently converged

on a well-posed problem: plan and execute a

COLREGS-compliant manoeuvre in real time among

dynamic targets.

Within that problem, distinct method families have

matured. Potential-field formulations offer

deterministic real-time planning with collision-

regulation constraints embedded in the virtual forces

[35]. Simulation-based control behaviour selection

evaluates candidate manoeuvres against predicted

trajectories and compliance criteria [36]. Velocity

obstacle formulations generalise the geometry of

encounter to produce feasible avoidance sets [44]. More

recently, reinforcement learning has been applied to

path planning in a manner that dispenses with hand-

coded knowledge [43], which is consistent with the

arrival of reinforcement learning and machine learning

among the most recent terms in the field's vocabulary.

Comprehensive reviews now exist both of collision-

avoidance methods generally [34] and of collision-

avoidance navigation systems for MASS specifically

[37], and a dedicated bibliometric review of intelligent

collision avoidance confirms the topic's internal

consolidation [14]. The open problems are

correspondingly well defined: the validation of learned

policies, behaviour in mixed traffic where most vessels

remain conventionally crewed, and the interpretation

of collision regulations written for human judgement

in terms a machine can execute and a court can review.

4.1.2 Perception and situational awareness

Perception is the fastest-growing topic in the field at

36.9 per cent per year, yet its cluster occupies a low

position on both axes of the strategic diagram, which

marks it as expanding but not yet consolidated. The

literature has established the basic sensing repertoire,

with electro-optical object detection and tracking

surveyed early and influentially [42] and the broader

combination of sensors and artificial-intelligence

techniques for situational awareness reviewed

subsequently [46]. What it has not established is a

shared basis for evaluation. Perception research lacks

the two assets that made guidance and control mature:

a canonical model and an agreed correctness criterion.

There is no maritime equivalent of the collision

regulations for the question of what a vessel must be

able to see, in what conditions, at what range and with

what confidence, and no common benchmark dataset

comparable to those that disciplined progress in other

perception domains. This is why a topic can grow

faster than any other and still register as immature, and

it is a concrete gap rather than a diffuse one.

4.1.3 Risk, safety and reliability

This is the field's centre of gravity, and its

intellectual anchor. The founding contribution

established that the safety case for unmanned vessels is

not self-evident: removing the crew removes the

human errors attributed to those on board, but it also

removes the human recovery actions that arrest

developing accidents, so the net effect on safety is an

empirical question rather than a premise [3]. That

framing has structured the subsequent literature. Work

has since asked whether established ship-risk models

transfer to autonomous operation, and found that they

do so only partially [31]; developed system-theoretic

safety assessment suited to software-intensive systems

731

whose failures are not component failures [32];

formalised risk assessment of MASS operations [28];

identified the factors influencing navigational risk [29];

and advanced supervisory risk control, in which the

vessel or its shore supervisor adapts behaviour as risk

evolves [30]. The origins of the strand lie in early risk

assessment for an unmanned merchant ship [33].

The paradox that this theme is the largest yet the

least cohesive is explained by what it contains. The

cluster joins risk analysis to cyber security, port

operations, the maritime industry and the crew, which

are not variants of one problem but distinct problems

sharing a vocabulary of safety. Its low density thus

records a genuine feature of the literature:

methodological pluralism without convergence. Two

dedicated bibliometric reviews of this sub-domain

reach compatible conclusions about its internal

fragmentation [12], [13]. The consequence is practical.

A regulator asking for the accepted method by which

the safety of a given degree of autonomy is

demonstrated would find several candidate methods

and no settled answer.

4.1.4 Cyber security

Cyber security is engaged by more than a fifth of the

corpus and attracts high citation, which reflects the

recognition that connectivity is constitutive of

autonomy rather than incidental to it: a shore-

supervised vessel is by definition a networked vessel,

and the architectures that make this possible have been

surveyed comprehensively [6]. The literature has

produced dedicated cyber-risk assessment methods for

ship systems [47], and a recent bibliometric review

maps the sub-domain in detail [15]. Its growth at 25.9

per cent per year, however, merely tracks the field.

Given that cyber vulnerability is the failure mode with

the clearest potential to be simultaneously

catastrophic, deliberate and difficult to attribute, its

failure to outgrow the field is itself worth noting.

4.1.5 The human element

The human element grows at 30.9 per cent per year,

faster than the field, and training is among the most

recently arriving terms in the vocabulary. The

literature has moved beyond the assumption that

autonomy removes the human, and now analyses the

tasks that remain. Detailed task analyses have

decomposed the operator's role in collision avoidance

and identified the associated human failure events [38],

and concurrent task analysis has been developed for

MASS operation and shore control [39]. A systematic

review of human and artificial-intelligence interaction

consolidates the strand [11], while implementation-

oriented work has examined the organisational factors

conditioning adoption [41]. Yet this topic attracts the

lowest mean citation of the seven, at 19.4 against 30.2

for risk assessment, which indicates that the field

acknowledges the human element without building on

it as heavily as it builds on technical results. Given that

the regulatory scoping exercise identified the roles of

the master and the responsible person as unresolved

[8], the human element is precisely where the technical

and legal agendas must eventually meet.

4.1.6 Regulation, liability and the economic case

The economic case was made early: exploratory cost

analysis examined the conditions under which

unmanned operation might yield a benefit [4], and

subsequent review consolidated the evidence on costs

and benefits [10]. The broader digital transformation of

the sector has been analysed in one of the most cited

works in the corpus [45], and business-model

implications have been surveyed [9]. The regulatory

literature, by contrast, remains comparatively thin

relative to the field. The operational, regulatory and

quality challenges of autonomous shipping were

identified early [2], and the challenges of ocean-going

operation catalogued [7], but the field has not

developed regulatory scholarship at a rate

commensurate with its technical output. The arrival of

liability as the most recent term of consequence in the

vocabulary indicates that the question is now being

asked; its aggregate growth rate indicates that it is not

yet being asked at scale.

4.2 The governance lag and why it matters

The comparative growth analysis quantifies what has

hitherto been asserted. Research on regulation, law and

liability grew at 16.9 per cent per year over 2018 to 2025

against 25.4 per cent for the field, and its share of

output fell from 23.8 to 18.4 per cent after 2021. The

interpretation is not that regulatory research is

stagnant, since it grows in absolute terms and its

constituent vocabulary is arriving rather than receding.

The interpretation is relative: regulatory research is

being outpaced by the technical research whose

deployment depends on it, and it lost ground during

precisely the period in which the international

regulatory agenda was most active. We term this the

governance lag.

The rise-and-retreat profile suggests a plausible

mechanism. The regulatory scoping exercise opened a

well-defined question and drew a burst of scholarly

attention between 2018 and 2021 [8]. Once the process

moved into the drafting of an instrument within an

institutional forum, the intellectual problem may have

appeared to be delegated and hence closed, while the

technical problems remained open and continued to

absorb effort. If that reading is correct, the retreat is an

artefact of how research attention responds to

institutional processes, and it is likely to reverse only

when a draft instrument creates concrete compliance

questions that require scholarly answers.

The lag matters for three reasons. First, it is

asymmetric in consequence: a shortfall in collision-

avoidance research delays a capability, whereas a

shortfall in regulatory research delays the certification

of every capability, because approval is a gate through

which all of them pass. Second, it compounds with the

fragmentation of the risk theme. The regulator's

question, namely what method demonstrates the

safety of a given degree of autonomy, is answered by

the risk literature only plurally, and the theme's low

density measured here is the structural expression of

that plurality. Technical capability may therefore

arrive without an agreed means of demonstrating that

it is safe, which is a governance failure rather than an

engineering one. Third, the human element sits at the

junction: the unresolved status of the master and the

732

responsible person is simultaneously a legal question,

a task-design question and a training question, and it is

being addressed by a literature that grows quickly but

is cited least.

It should be stressed that the finding concerns the

research effort, not the regulatory process itself.

Regulatory work proceeds substantially within

institutional forums and produces instruments rather

than journal articles, and that output is not captured by

a corpus of peer-reviewed literature. The claim

advanced here is narrower and defensible: within the

scholarly literature on MASS, governance-oriented

enquiry is growing more slowly than technical

enquiry. Since the scholarly literature is where

methods of demonstration are developed, contested

and refined before regulators can adopt them, that

shortfall is consequential even though it is not the

whole picture.

4.3 Relation to previous bibliometric studies

The findings reported here are consistent with, and

extend, the existing quantitative literature. The sub-

domain reviews of risk, safety and reliability describe

an internally diverse research area [12], [13], which

corresponds to the low density recorded for the risk

cluster in this analysis; the difference is that a field-

wide corpus permits that diversity to be compared

with the cohesion of the guidance and control cluster,

and thereby to be interpreted rather than merely

described. The bibliometric review of intelligent

collision avoidance reports a consolidating topic [14],

which matches its motor-theme position here. Reviews

of maritime cybersecurity [15] and of machine learning

in maritime safety [16] identify rapidly developing

technical areas, consistent with the growth rates

estimated here for cyber security and for perception.

The field-wide analysis of autonomous shipping

technologies based on Web of Science likewise reports

sharp growth concentrated in the late 2010s [17], in

agreement with the inflection identified here.

The contribution of the present study is therefore

not the discovery that MASS research is growing,

which is established, but three additions. It compares

the growth of the field's constituent topics on a

common corpus and identifies a governance lag that no

single-topic study could detect. It reconstructs the

intellectual base by local citation and shows that a

safety study, not a technical one, anchors the field. And

it does so from an open database with a fully reported

procedure, so that the analysis can be re-executed

without a subscription.

4.4 A research agenda

The structure identified here supports a specific

agenda rather than a general call for more research.

Table 8 sets out five priorities, each derived from a

result reported above and stated so that progress can

be recognised. The priorities are ordered by the

strength of the evidence supporting them.

Table 8. Research agenda derived from the structural

findings.

Priority

Evidence from this

study

What would constitute

progress

Converge on

methods for

demonstrating

safety

The risk theme is the

largest yet least

cohesive cluster

(centrality 40, density

39), indicating plural

methods without

convergence.

Comparative evaluation of

candidate safety-

demonstration methods

against a common MASS

reference case, leading to

an agreed basis for

approval of a stated

degree of autonomy.

Define a

perception

performance

specification

Perception grows

fastest of all topics

(36.9 per cent per year)

yet remains

structurally immature,

lacking an agreed

correctness criterion.

A performance

specification and shared

benchmark stating what a

vessel must detect, at what

range, in what conditions

and with what confidence,

comparable in function to

the collision regulations

for manoeuvring.

Scale regulatory

and liability

scholarship

Regulation, law and

liability grows at 16.9

per cent per year

against 25.4 per cent

for the field, and its

share fell from 23.8 to

18.4 per cent after 2021.

Sustained scholarship on

the allocation of

responsibility, the legal

status of the remote

operator and the

evidentiary basis for

compliance, developed

alongside rather than after

the drafting of

instruments.

Model the tasks

and competences

of the remote

supervisor

The human element

grows quickly (30.9 per

cent per year) but

attracts the lowest

citation impact (19.4),

while training is

among the most

recently arriving terms.

Validated task and

competence models for

shore-based supervision,

and the training and

certification frameworks

that follow from them.

Validate learned

control policies

in mixed traffic

Reinforcement learning

and machine learning

are among the most

recently arriving terms,

applied to a motor

theme whose

validation practice

predates them.

Methods for verifying and

assuring learned collision-

avoidance policies in

traffic shared with

conventionally crewed

vessels, sufficient to

support certification.

4.5 Implications for maritime research communities and

education

The distribution of effort has a practical reading for

maritime faculties, particularly smaller ones. The

mature motor themes are dominated by large, well-

resourced groups with established modelling and

simulation infrastructure, and the barriers to a

marginal contribution there are high. The areas

identified here as lagging or fragmented, namely

regulatory scholarship, the demonstration of safety,

the human element and training, are more accessible,

because they require access to operational expertise,

seafarers, simulators and institutional knowledge

rather than to large-scale computation. These are

exactly the assets that maritime universities hold. The

finding that the human element is the topic with the

lowest citation impact yet among the fastest growing

suggests an area where careful work can establish

position. The same reasoning applies to training: its

arrival among the most recent terms indicates a

literature still forming, and curricula for the remote

operation and supervision of vessels are a natural

733

contribution for institutions whose core activity is

seafarer education.

4.6 Limitations

Several limitations qualify these findings. The corpus

is drawn from a single database, and although its

coverage is broadly comparable to the proprietary

alternatives [21], absolute counts would differ under

another source; the analysis therefore emphasises

proportions, rates and structure rather than exact

totals. The restriction to peer-reviewed journal articles

and reviews excludes 1,056 conference papers, which

is material in a field where proceedings carry an

important part of the discourse, and the excluded

literature may be distributed unevenly across themes,

plausibly favouring control and perception; the growth

comparison could be affected if that distribution has

changed over time. The exclusion of unmanned surface

vehicle terminology from the query keeps the corpus

focused on maritime transport but omits

methodologically relevant robotics work. The lexical

probes are inclusive and overlapping, so their absolute

levels are sensitive to how each expression is drawn,

and only their comparative and temporal behaviour

should be relied upon; a probe based on titles and

keywords will also under-count articles that address a

topic substantively in the body of the text without

signalling it in the metadata, which may particularly

affect regulation, since a technical article may discuss

compliance without announcing it. The keyword

analysis depends on database-supplied terms, whose

quality is uneven and was mitigated but not eliminated

by normalisation. Citation-based indicators favour

older work, so recent contributions are necessarily

under-represented in Tables 4 and 5. Cluster

assignment excluded 285 articles that carried none of

the sixty retained terms. Finally, bibliometric mapping

characterises a literature's structure and cannot assess

the merit of individual studies, which remains the task

of focused critical review.

5 CONCLUSIONS

This study mapped the research landscape of Maritime

Autonomous Surface Ships through a reproducible

bibliometric analysis of 1,874 peer-reviewed journal

articles and reviews published between 2010 and 2025

and drawn from an open database. The field grew at

46.9 per cent per year between 2013 and 2025, doubling

every 1.8 years, and continued to expand at 25.4 per

cent per year from 2018. It is geographically

concentrated, with China contributing more than a

quarter of the corpus and Norway a disproportionate

share relative to the size of its research system, and it is

served by a compact disciplinary core in which ten of

570 sources carry the first third of all output.

Three structural findings follow. First, the field has

changed its subject. Platform and propulsion research

fell from 34.3 to 17.0 per cent of output while risk,

safety, security and operations rose from 25.7 to 39.0

per cent, and the vocabulary of the physical vessel is

receding while that of software, learning, training and

liability is arriving. The organising question has moved

from whether an autonomous vessel can be built to

whether its operation can be shown to be safe. Second,

the field understands itself as a safety problem: the

work on which it most draws is a study of the safety

impact of unmanned vessels, not a technical

contribution, even though the works attracting the

widest citation are predominantly technical. Third, the

largest theme is the least cohesive. Risk, safety, security

and operations records the lowest centrality and

density of any cluster despite being the largest by every

volume measure, which identifies methodological

pluralism without convergence and means that the

question of how the safety of a given degree of

autonomy is to be demonstrated has several candidate

answers and no settled one.

The central conclusion concerns the balance of

effort. Research on regulation, law and liability grew at

16.9 per cent per year against 25.4 per cent for the field,

and its share of output fell from 23.8 to 18.4 per cent

after 2021, even as the international regulatory agenda

advanced. Every technically oriented topic either

outgrew the field or tracked it. This governance lag is

measurable rather than merely asserted, and it matters

disproportionately, because regulatory approval is the

gate through which every technical capability must

pass. On present trends the field is likely to arrive at

demonstrable technical capability before it arrives at an

agreed means of demonstrating that such capability is

safe and of allocating responsibility when it fails.

The agenda that follows is specific. It requires

convergence on methods for demonstrating the safety

of a stated degree of autonomy, a benchmark and

performance specification for maritime perception,

regulatory and liability scholarship at a scale

commensurate with the technical literature, task and

competence models for the remote supervisor, and

validation methods for learned control policies in

mixed traffic. For smaller maritime research

communities, these priorities are also opportunities,

since they draw on operational, human and

institutional expertise rather than on large-scale

computational infrastructure. Closing the governance

lag is not a supplement to the technical programme of

MASS research. On the evidence presented here, it is

the condition of its arrival.

REFERENCES

[1] Burmeister, H. C., Bruhn, W., Rødseth, Ø. J., & Porathe, T.

(2014). Autonomous Unmanned Merchant Vessel and its

Contribution towards the e-Navigation Implementation:

The MUNIN Perspective. International Journal of e-

Navigation and Maritime Economy, 1, 1-13.

https://doi.org/10.1016/j.enavi.2014.12.002

[2] Komianos, A. (2018). The Autonomous Shipping Era.

Operational, Regulatory, and Quality Challenges.

TransNav, the International Journal on Marine

Navigation and Safety of Sea Transportation, 12(2), 335-

348. https://doi.org/10.12716/1001.12.02.15

[3] Wróbel, K., Montewka, J., & Kujala, P. (2017). Towards the

assessment of potential impact of unmanned vessels on

maritime transportation safety. Reliability Engineering &

System Safety, 165, 155-169.

https://doi.org/10.1016/j.ress.2017.03.029

[4] Kretschmann, L., Burmeister, H. C., & Jahn, C. (2017).

Analyzing the economic benefit of unmanned

autonomous ships: An exploratory cost-comparison

between an autonomous and a conventional bulk carrier.

Research in Transportation Business & Management, 25,

76-86. https://doi.org/10.1016/j.rtbm.2017.06.002

734

[5] Geertsma, R., Negenborn, R., Visser, K., & Hopman, J.

(2017). Design and control of hybrid power and

propulsion systems for smart ships: A review of

developments. Applied Energy, 194, 30-54.

https://doi.org/10.1016/j.apenergy.2017.02.060

[6] Aslam, S., Michaelides, M. P., & Herodotou, H. (2020).

Internet of Ships: A Survey on Architectures, Emerging

Applications, and Challenges. IEEE Internet of Things

Journal, 7(10), 9714-9727.

https://doi.org/10.1109/jiot.2020.2993411

[7] Felski, A., & Zwolak, K. (2020). The Ocean-Going

Autonomous Ship—Challenges and Threats. Journal of

Marine Science and Engineering, 8(1), 41.

https://doi.org/10.3390/jmse8010041

[8] International Maritime Organization. (2021). Outcome of

the regulatory scoping exercise for the use of Maritime

Autonomous Surface Ships (MASS) (MSC.1/Circ.1638).

London: IMO.

[9] Munim, Z. H. (2019). Autonomous ships: a review,

innovative applications and future maritime business

models. Supply Chain Forum: An International Journal,

20(4), 266-279.

https://doi.org/10.1080/16258312.2019.1631714

[10] Ziajka-Poznańska, E., & Montewka, J. (2021). Costs and

Benefits of Autonomous Shipping—A Literature Review.

Applied Sciences, 11(10), 4553.

https://doi.org/10.3390/app11104553

[11] Veitch, E., & Andreas Alsos, O. (2022). A systematic

review of human-AI interaction in autonomous ship

systems. Safety Science, 152, 105778.

https://doi.org/10.1016/j.ssci.2022.105778

[12] Chaal, M., Ren, X., BahooToroody, A., Basnet, S., Bolbot,

V., Banda, O. A. V., & Gelder, P. V. (2023). Research on

risk, safety, and reliability of autonomous ships: A

bibliometric review. Safety Science, 167, 106256.

https://doi.org/10.1016/j.ssci.2023.106256

[13] Li, Z., Zhang, D., Han, B., & Wan, C. (2023). Risk and

reliability analysis for maritime autonomous surface ship:

A bibliometric review of literature from 2015 to 2022.

Accident Analysis & Prevention, 187, 107090.

https://doi.org/10.1016/j.aap.2023.107090

[14] Zhu, Q., Xi, Y., Weng, J., Han, B., Hu, S., & Ge, Y. E.

(2024). Intelligent ship collision avoidance in maritime

field: A bibliometric and systematic review. Expert

Systems with Applications, 252, 124148.

https://doi.org/10.1016/j.eswa.2024.124148

[15] Peng, P., Xie, X., Claramunt, C., Lu, F., Gong, F., & Yan,

R. (2025). Bibliometric analysis of maritime cybersecurity:

Research status, focus, and perspectives. Transportation

Research Part E: Logistics and Transportation Review,

195, 103971. https://doi.org/10.1016/j.tre.2025.103971

[16] Xue, J., Yang, P., Li, Q., Song, Y., Gelder, P. H. A. J. M. V.,

Papadimitriou, E., & Hu, H. (2025). Machine Learning in

Maritime Safety for Autonomous Shipping: A

Bibliometric Review and Future Trends. Journal of

Marine Science and Engineering, 13(4), 746.

https://doi.org/10.3390/jmse13040746

[17] Özkan, E. D., & Sevgili, C. (2025). Future horizons: a

bibliometric analysis of autonomous shipping

technologies. Ships and Offshore Structures, 20(9), 1407-

1421. https://doi.org/10.1080/17445302.2025.2454306

[18] Donthu, N., Kumar, S., Mukherjee, D., Pandey, N., &

Lim, W. M. (2021). How to conduct a bibliometric

analysis: An overview and guidelines. Journal of Business

Research, 133, 285-296.

https://doi.org/10.1016/j.jbusres.2021.04.070

[19] Aria, M., & Cuccurullo, C. (2017). bibliometrix : An R-

tool for comprehensive science mapping analysis. Journal

of Informetrics, 11(4), 959-975.

https://doi.org/10.1016/j.joi.2017.08.007

[20] Priem, J., Piwowar, H., & Orr, R. (2022). OpenAlex: A

fully-open index of scholarly works, authors, venues,

institutions, and concepts. arXiv.

https://doi.org/10.48550/arXiv.2205.01833

[21] Culbert, J. H., Hobert, A., Jahn, N., Haupka, N., Schmidt,

M., Donner, P., & Mayr, P. (2025). Reference coverage

analysis of OpenAlex compared to Web of Science and

Scopus. Scientometrics, 130(4), 2475-2492.

https://doi.org/10.1007/s11192-025-05293-3

[22] Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I.,

Hoffmann, T. C., Mulrow, C. D., Shamseer, L., Tetzlaff, J.

M., et al. (2021). The PRISMA 2020 statement: an updated

guideline for reporting systematic reviews. BMJ, n71.

https://doi.org/10.1136/bmj.n71

[23] van Eck, N. J., & Waltman, L. (2009). Software survey:

VOSviewer, a computer program for bibliometric

mapping. Scientometrics, 84(2), 523-538.

https://doi.org/10.1007/s11192-009-0146-3

[24] Blondel, V. D., Guillaume, J. L., Lambiotte, R., &

Lefebvre, E. (2008). Fast unfolding of communities in

large networks. Journal of Statistical Mechanics: Theory

and Experiment, 2008(10), P10008.

https://doi.org/10.1088/1742-5468/2008/10/p10008

[25] Callon, M., Courtial, J. P., & Laville, F. (1991). Co-word

analysis as a tool for describing the network of

interactions between basic and technological research:

The case of polymer chemsitry. Scientometrics, 22(1), 155-

205. https://doi.org/10.1007/bf02019280

[26] Cobo, M., López-Herrera, A., Herrera-Viedma, E., &

Herrera, F. (2011). Science mapping software tools:

Review, analysis, and cooperative study among tools.

Journal of the American Society for Information Science

and Technology, 62(7), 1382-1402.

https://doi.org/10.1002/asi.21525

[27] Fossen, T. I. (2011). Handbook of Marine Craft

Hydrodynamics and Motion Control. Chichester: John

Wiley & Sons. https://doi.org/10.1002/9781119994138

[28] Chang, C. H., Kontovas, C., Yu, Q., & Yang, Z. (2021).

Risk assessment of the operations of maritime

autonomous surface ships. Reliability Engineering &

System Safety, 207, 107324.

https://doi.org/10.1016/j.ress.2020.107324

[29] Fan, C., Wróbel, K., Montewka, J., Gil, M., Wan, C., &

Zhang, D. (2020). A framework to identify factors

influencing navigational risk for Maritime Autonomous

Surface Ships. Ocean Engineering, 202, 107188.

https://doi.org/10.1016/j.oceaneng.2020.107188

[30] Utne, I. B., Rokseth, B., Sørensen, A. J., & Vinnem, J. E.

(2020). Towards supervisory risk control of autonomous

ships. Reliability Engineering & System Safety, 196,

106757. https://doi.org/10.1016/j.ress.2019.106757

[31] Thieme, C. A., Utne, I. B., & Haugen, S. (2018). Assessing

ship risk model applicability to Marine Autonomous

Surface Ships. Ocean Engineering, 165, 140-154.

https://doi.org/10.1016/j.oceaneng.2018.07.040

[32] Wróbel, K., Montewka, J., & Kujala, P. (2018). Towards

the development of a system-theoretic model for safety

assessment of autonomous merchant vessels. Reliability

Engineering & System Safety, 178, 209-224.

https://doi.org/10.1016/j.ress.2018.05.019

[33] Rødseth, Ø., & Burmeister, H. C. (2015). Risk Assessment

for an Unmanned Merchant Ship. TransNav, the

International Journal on Marine Navigation and Safety of

Sea Transportation, 9(3), 357-364.

https://doi.org/10.12716/1001.09.03.08

[34] Huang, Y., Chen, L., Chen, P., Negenborn, R. R., & van

Gelder, P. (2020). Ship collision avoidance methods: State-

of-the-art. Safety Science, 121, 451-473.

https://doi.org/10.1016/j.ssci.2019.09.018

[35] Lyu, H., & Yin, Y. (2018). COLREGS-Constrained Real-

time Path Planning for Autonomous Ships Using

Modified Artificial Potential Fields. Journal of

Navigation, 72(3), 588-608.

https://doi.org/10.1017/s0373463318000796

[36] Johansen, T. A., Perez, T., & Cristofaro, A. (2016). Ship

Collision Avoidance and COLREGS Compliance Using

Simulation-Based Control Behavior Selection With

Predictive Hazard Assessment. IEEE Transactions on

735

Intelligent Transportation Systems, 17(12), 3407-3422.

https://doi.org/10.1109/tits.2016.2551780

[37] Zhang, X., Wang, C., Jiang, L., An, L., & Yang, R. (2021).

Collision-avoidance navigation systems for Maritime

Autonomous Surface Ships: A state of the art survey.

Ocean Engineering, 235, 109380.

https://doi.org/10.1016/j.oceaneng.2021.109380

[38] Abilio Ramos, M., Utne, I. B., & Mosleh, A. (2019).

Collision avoidance on maritime autonomous surface

ships: Operators’ tasks and human failure events. Safety

Science, 116, 33-44.

https://doi.org/10.1016/j.ssci.2019.02.038

[39] Ramos, M., Thieme, C. A., Utne, I. B., & Mosleh, A.

(2020). Human-system concurrent task analysis for

maritime autonomous surface ship operation and safety.

Reliability Engineering & System Safety, 195, 106697.

https://doi.org/10.1016/j.ress.2019.106697

[40] Liu, Z., Zhang, Y., Yu, X., & Yuan, C. (2016). Unmanned

surface vehicles: An overview of developments and

challenges. Annual Reviews in Control, 41, 71-93.

https://doi.org/10.1016/j.arcontrol.2016.04.018

[41] Hogg, T., & Ghosh, S. (2016). Autonomous merchant

vessels: examination of factors that impact the effective

implementation of unmanned ships. Australian Journal

of Maritime & Ocean Affairs, 8(3), 206-222.

https://doi.org/10.1080/18366503.2016.1229244

[42] Prasad, D. K., Rajan, D., Rachmawati, L., Rajabally, E., &

Quek, C. (2017). Video Processing From Electro-Optical

Sensors for Object Detection and Tracking in a Maritime

Environment: A Survey. IEEE Transactions on Intelligent

Transportation Systems, 18(8), 1993-2016.

https://doi.org/10.1109/tits.2016.2634580

[43] 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

Engineering, 189, 106299.

https://doi.org/10.1016/j.oceaneng.2019.106299

[44] Huang, Y., Chen, L., & van Gelder, P. (2019). Generalized

velocity obstacle algorithm for preventing ship collisions

at sea. Ocean Engineering, 173, 142-156.

https://doi.org/10.1016/j.oceaneng.2018.12.053

[45] Tijan, E., Jović, M., Aksentijević, S., & Pucihar, A. (2021).

Digital transformation in the maritime transport sector.

Technological Forecasting and Social Change, 170,

120879. https://doi.org/10.1016/j.techfore.2021.120879

[46] Thombre, S., Zhao, Z., Ramm-Schmidt, H., Vallet Garcia,

J. M., Malkamaki, T., Nikolskiy, S., Hammarberg, T.,

Nuortie, H., et al. (2022). Sensors and AI Techniques for

Situational Awareness in Autonomous Ships: A Review.

IEEE Transactions on Intelligent Transportation Systems,

23(1), 64-83. https://doi.org/10.1109/tits.2020.3023957

[47] Bolbot, V., Theotokatos, G., Boulougouris, E., & Vassalos,

D. (2020). A novel cyber-risk assessment method for ship

systems. Safety Science, 131, 104908.

https://doi.org/10.1016/j.ssci.2020.104908