611

1 INTRODUCTION

Autonomous shipping is no doubt a future of the

maritime industry [21]. According to Rolls-Royce, a

fully automatic ship will go to its first voyage by 2035

[21] which means the need for related research now.

However, autonomous shipping poses a great

challenge for information technology and research in

the field of onboard control systems. In contrast with

autonomous cars, autonomous ships will operation in

much larger environments travelling by much longer

distances. The following types of ships with regard to

autonomy are defined in research: ordinary ships,

where all decisions are made by the crew and all

operations are performed by the crew, smart ships,

where the crew still makes a decision, but the

operations are performed automatically and

autonomous ships where both decisions and

operations are automatic [12]. Smart ships are not

guaranteed to have the crew on board. Such ships can

be controlled from a remote control centre [6, 37].

However, it seems doubtful that smart ships with

fully remote control can be implemented because

such ships would require stable network connection

with the control centre along the entire route of the

voyage, which may be difficult to secure in the case of

trans-ocean voyages [37]. Therefore, smart ships are

expected to have some degree of autonomy to make

them able to operate amid the absence of connection

with the control centre. It is expected that an

autonomous ship can also be controlled remotely,

therefore there can be stages of fully autonomous

operation and remotely controlled stages within a

single voyage [35]. Remote control stages may be

required to navigate the ship in the most dangerous

or difficult circumstances, e.g. at nearshore areas and

Towards Autonomous Shipping: Benefits and

hallenges in the Field of Information Technology and

elecommunication

I. Yanchin & O. Petrov

State Marine Technical University

, Saint-Petersburg, Russia

ABSTRACT: This paper is dedicated to an overview of components of an onboard control system of an

autonomous ship. This system controls and operates the ship. Therefore, this system needs to be able to analyze

the ship’s state, predict its future development and analyze the consequences of its own decisions. The paper

focuses on software aspects of the onboard control system, not the hardware. The paper provides an overview

of technologies that can be used to implement the components of such a system responsible for planning new

routes, handling the ship during the voyage, ensuring its seaworthiness and safety during the voyage,

monitoring an autonomous ship from an onshore control centre, ensuring the robustness of the onboard control

system, and collective operations of multiple autonomous ships. The paper describes benefits the maritime

industry would gain from deploying some of the technologies developed for autonomous ships on ordinary,

human-controlled ships. The paper also describes some challenges, especially in the field of automatic decision

and reasoning, arising from the emergence of autonomous and smart ships. The main contribution of the paper

is that it summarizes existing research in different areas of autonomous ship technology.

http://

www.transnav.eu

the

International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 14

Number 3

September 2020

DOI:

10.12716/1001.14.03.12

612

at ports [37]. In this paper we focus on fully-

autonomous ship operations, thus regarding

autonomous ships and smart ships that operate in the

autonomous regime.

This paper is dedicated to the challenges in the

field of information technology and

telecommunication with regard to smart and

autonomous ships. It is expected that the success in

these fields will significantly affect the maritime

industry as being the most important milestone in the

journey towards autonomous shipping [35].

Within this paper, we assume that an autonomous

ship is controlled and monitored by a single onboard

control system. This system runs on the onboard

computers that are connected into the onboard

computational network. The paper describes the

design neither of the onboard control system nor the

computational network, but refers to them when

describing their components responsible for tackling

the challenges arising from the emergence of

autonomous and smart ships.

This paper is organized as follows. Section 2

provides an overview of techniques that can be used

to plan safe and optimal routes for autonomous ships.

Section 3 is dedicated to automatic ship handling,

ensuring its seaworthiness and safety while following

the route, keeping the ship on the route, and collective

operations of multiple autonomous ships. Section 4

describes the remote monitoring of a ship. Section 5 is

dedicated to the robustness of the onboard control

system that controls the ship. Section 6 concludes the

paper.

2 AUTONOMOUS NAVIGATION

Route planning is an important feature of the ship’s

onboard control system. The crew uses this feature to

plot efficient routes taking into account weather

conditions, legal regulations, fuel efficiency and other

requirements [17]. If the system fails to plot an

effective route or the resulting route is not suitable for

some reason, the crew still can intervene and

introduce corrections. In case of an autonomous ship,

there is no crew to correct the inefficient route.

Planning an efficient route for an autonomous ship is

a more complicated task than in the case of an

ordinary ship [6]. The onboard control system needs

to be able to plan as efficient routes as possible

without any human intervention.

Within this paper, we define an efficient route as a

route that is the shortest route possible in terms of

travel time that does not put the ship in danger in any

way. This definition can be extended to assume

additional requirements. For example, the route

planning system can take fuel consumption into

account, in this case, the system would try to achieve

a compromise between travel speed (and thus travel

time) and fuel consumption possibly by using slow

steaming.

In case of an autonomous ship, there can be two

approaches for planning a route. In the first case, the

route is planned by the onboard control system of the

ship, in the second case, the route is planned by a

shore control centre and then transmitted to the ship

[36, 41]. In the latter case, the route could be evaluated

by an expert before being transmitted to the ship [41].

In both cases, the input data for the route planning

procedure is the destination point and the additional

constrains for the route (e.g. fuel consumption, arrival

time and so on).

Weather conditions affect the ship’s safety and its

ability to follow the route. Taking the weather into

account not only helps to avoid negative effects on the

ship [35,41] but also reduces the economical costs of a

voyage by reducing the travel time and fuel

consumption [16, 43]. Therefore, the ability to take

weather conditions into account is essential for the

route planning system [17, 23, 38].

One of the most common algorithms used for

pathfinding is A*. This algorithm is able to find the

shortest path between two vertices of a graph or grid.

It is possible to tune the algorithm to take into account

requirements specific for a particular task, e.g.

weather conditions [17], ice [45] or in order to reduce

the computation time [28]. However, A* is known to

be highly dependant on the heuristic function used to

determine which vertices to investigate and for its

high memory consumption in case of the large search

spaces (grid with many vertices) [33]. One can reduce

the complexity by reducing the number of points the

grid has, however, this reduces the freedom of the

algorithm to choose movement direction thus

reducing the optimality of the route.

Planning a route for a ship is a complex task that

requires attention to many factors affecting routes

safety and optimality. The navigators plan routes

using their experience and background [35]. Building

this experience into a formal route planning algorithm

like A* or Dijkstra’s is a difficult task. We argue that

the maritime industry could benefit from employing

algorithms and techniques that belong to the fields of

Artificial Intelligence, Machine Learning and Deep

Learning. We can divide these promising techniques

into two groups: those that use a kind of agent that is

trained from its own or external experience and those

that are heuristic, fuzzy, or competitive approaches to

find the solution for the task in question [30].

Two of the most well-known techniques of the first

family are Reinforcement Learning and its variant,

Deep Reinforcement Learning. Reinforcement

Learning uses a table π, called policy, that helps

understand why the agent took particular action

being in a particular state, thus making the decision

process interpretable. This table is designed during

the preparation stage of the training process and must

contain all of the possible agent states and all of the

possible actions the agent can take. Deep

Reinforcement Learning does not require any

predefined states and table π, instead, it uses a deep

neural network that takes state representation as

input. Deep Reinforcement Learning has

demonstrated good results in training an agent to

play Go [39], chess [40] or video games [44], making it

possible to beat a champion human player. However,

in practice, even an expert often fails to design a good

reward function for the agent, thus making it possible

to maximize the reward without actually completing

the task [22]. Moreover, currently, there is not enough

research in the field of Reinforcement Learning about

solving problems other than games [22]. Another

613

difficult here is that the training process assumes that

the agent needs to observe all possible states and take

all possible actions for any of them to test what

happens, thus making the training process

computationally difficult.

We argue that there is still not enough research to

employ Reinforcement Learning to planning a route

for a ship. Another obstacle on this path is that the

agent needs an environment that it interacts with

while solving the task [22]. As we have already

pointed out, the agent needs to observe all possible

situations, including those that lead to maritime

accidents, moreover, it needs to observe these

situation several times with slightly different

properties. Therefore, using natural environments as

training environments and real ships as agents to

train the route planning system has unacceptably

high costs, not to mention the related dangers.

Another approach is to simulate the natural

environment. In this case, all of the processes that

take place in Nature and affect the ship are modelled

and this modelled environment is used to train the

agent. However, this approach requires the model to

represent the real processes as precisely as possible so

that the agent can use its knowledge gained from the

simulation in the real-life tasks. This poses a difficult

task.

We, therefore, argue that we should not use agents

that are trained from their own experience for

planning ship routes due to high costs of real

experiments and extremely difficult development of

simulators.

We, therefore, think that the route planning

algorithms should be based on approaches that can

adapt to a changing environment and can handle

complex success criteria but do not require

preliminary training. Genetic algorithms are one of

such approaches [20]. The fitting function can be

expressed as a weighted sum of values describing

different route’s properties thus making the algorithm

to generate a good route through maximizing these

value. Genetic algorithms can be effectively

parallelized using the island model, taking advantage

of distributed onboard computation systems [7, 18].

Since in every moment a genetic algorithm operates

on a set of possible solutions, it is possible to get a set

of Paretooptimal routes as a result and choose the

final one using another examination technique [25].

Another advantage of genetic algorithms is that it is

possible to restrict the total computation time allotted

for planning the route and still get an acceptable sub-

optimal solution [18].

As we have already said, navigators use their

experience to judge on the route’s quality. Experience

is a very subjective and personal concept so it could

not be formalized to be used by an automatic route

planning systems. We propose to use a set of

characteristic coefficients that can be used to measure

route’s properties thus making it possible to compare

routes and select the best one. The fitting function of a

genetic algorithm can be represented as a weighted

sum of this coefficient values. These coefficients

express how close the route is to the ideal route.

Therefore, professional navigators could be

interviewed about what is the optimal and safe route

and what is the ideal route, and then these qualities

can be expressed through formalized coefficients.

Here we propose several basic coefficients that can

be used as a starting point. This set can be

extended with additional coefficients designed for a

particular ship. Moreover, different routes for the

same ship could be planned while taking into account

different coefficients that were selected according to

the cargo, technical conditions or other factors.

The safety coefficient can be defined as

( )

Sr P= −

(1)

where

( )

P is the probability of a negative encounter

(e.g. collision, rough weather, severe pitching) to

occur while the ship follows the route r. The shorter

route is a better route, therefore the distance

coefficient can be defined as

( )

(2)

where

( )

,sd

is the length of the line segment

between start and destination points and l

r is the

length of the route being examined. The closer the

route to the straight line, the greater the value of D is.

Since it is not possible to find a shorter way between

two points than the straight line, this coefficient can

be used to measure the optimality of the route with

regard to its length. Another coefficient related to

optimality is the time coefficient

( )

,sd

(3)

where

(

)

,sd

is the time needed for the ship to travel

along the straight line

( )

,sd

with the maximum

technic-ally possible speed and tr is the time needed

to travel following the route r. This coefficient helps

ensure that the route is not only short in terms of

distance but that it will not take long to follow it.

According to [?], the number of manoeuvres the ship

needs to perform during the voyage also affects the

route’s optimality and the most important manoeuvre

is the change of direction. The change of direction

usually happens at a waypoint that joins two edges of

the route each of which defines new direction. There-

fore, we can define the simplicity coefficient as

( )

where p

r is the number of waypoints in the route.

Every route has at least two points: start and

destination, and the simplicity coefficient encourages

the route to have as little points as possible.

Route planning algorithm needs to be robust and

able to handle unexpected situations. It also needs to

be predictable and its output needs to be

interpretable. That is why designing such an

algorithm requires extensive testing and verification

[35]. Formal algorithms like A* or Dijkstra’s are

proven to be able to find the best solution but they

614

pose strong requirements for the environment and the

input data [11]. Non-formal algorithms like genetic

algorithms may pose less strict requirements but there

can be no formal proof of correctness. Therefore,

testing and verification of such algorithms require

simulation of the common and extreme situations in

order to validate their behaviour. There are several

techniques that can be used to verify and validate the

correctness of models and algorithms while using

simulation modelling, the Balchi’s scheme is one of

them [1, 2].

3 AUTONOMOUS SHIP HANDLING

After the ship has acquired the route, it needs to

follow it. This means the onboard system needs to be

able to handle the ship autonomously. This is

expected to be the most difficult component of the

onboard control system since it needs to take into

account different aspects of ship handling, including,

but not limited to, keeping the ship on the route and

ensuring that the ships seaworthiness is satisfactory

[35].

3.1 Keeping the ship on the route

Keeping the ship on the route is the first responsibility

of the onboard control system when underway. The

system analyzes the ship’s motion and checks that the

actual route does not deviate from the prescribed one

more than allowed. In the case of discrepancy, the

control system orders the propulsion system of the

ship to perform actions to resolve this discrepancy

[24].

Since it is expected that the prescribed route allows

some sort of freedom (i.e. it does not prescribe the

exact waypoints), another important aspect is

planning the exact manoeuvres prescribed by the

route [37]. For example, if a route prescribes the ship

to go under a bridge, the onboard control system

must issue orders to make the ship go under the

bridge’s arc by a safe distance from the columns and

ensure that the arc is high enough. Although bridges,

canal locks, and other infrastructure facilities can be

detected during the planning stage, the actual route

affects the moment of time when the ship reaches the

facility and ship’s exact position relative to it. That is

why the ship needs to be able to plan the manoeuvres

itself before performing them thus requiring the

onboard control system to analyze the current

situation and issue orders with regard both to the

manoeuvre and the environment. Therefore, this

component of the control system needs to be

universal enough to be able to deal with different

manoeuvres and surrounding situations.

It is expected that the onboard control system can

be based on AI approaches [35] thus being able to

handle situations that differ from those observed

during the design stage.

The onboard control system can benefit from tree-

based situation classification approach [31]. In case of

such an approach, there is a set of classifiers

organized as a tree. Each classifier takes the

representation of the situation that includes, but is not

limited to, images from the cameras that capture

images or video of the surroundings, radars that

observe the surroundings, gyroscopes and

accelerometers that measure the ship’s state and other

sensors. The root classifier that corresponds to the

root vertex of the tree performs the initial

classification of the situation, whether it corresponds

to an open area, a bridge, locks, and so on. This initial

classifier makes the top-level decision detecting which

type the situation belongs to and does not detect its

internal properties. Once the generic type of the

situation is determined, the situation representation is

passed to the next level of the tree. Classifiers of this

level are specialized to observe situations of a specific

type and are able to detect different subtypes of them.

Figure 1 illustrates the tree-based classification

approach.

Figure 1. Sample tree-based situation classifier. Each level

refines the output of the previous one until the final

decision is made.

Therefore, each level of the classification tree

refines the decision made by the previous level thus

enabling the onboard control system to capture all

important properties of the situation [31]. The

refinement process finishes when the tree traversal

process reaches a leaf that contains navigation

instructions suitable for the detected situation [31].

The use of a classification tree helps to keep each

classifier specialized on a particular task thus making

it easier to design it, train it, maintain its operation,

interpret the output or insert new tree branches (thus

making the tree aware of new situations) leaving

other branches intact. In addition, the classifiers can

be independently replaced or updated according to

the requirements.

The classifiers could be based on Deep Neural

Networks or use simpler machine learning models

suitable for the task being solved by the classifier.

Convolution Neural Networks have demonstrated

good results when processing image and video data

or other spatiallydistributed values [15]. If the inputs

for the network contain both image and non-image

data, it is possible to train a two-branch neural

network (figure 2), where one branch handles image

data and another one non-image data and then they

join their outputs for further processing. Twobranch

neural networks have demonstrated good results

when processing mixed data of different nature [46].

615

Figure 2. An example of neural network with two branches.

3.2 Ensuring the route correctness

As we have already said, the route’s quality is

expressed through a set of characteristic coefficients

that describe the route’s properties. These coefficients

can also be used to ensure that the actual route does

not deviate from the prescribed one in a way that may

lead, in the long run, to an accident.

Consider the prescribed route and the actual route

that is being plotted during the voyage. In order to

check that the actual route is not going to cause an

accident we can to compute its characteristic

coefficients during the voyage taking into account the

actual locations of the waypoints and their properties,

and then compare these coefficients to those

computed for the prescribed route. This lets to

determine which properties of the actual route are

worse than expected. Another approach is to plot the

sum of the characteristic coefficients and thus decide

whether the trend is positive or stable (the route

quality either increases or stays the same) or negative

(the route quality decreases) [30]. Figure 3 shows an

example of a route quality plot.

Figure 3: Route quality decreases over time, this indicates

negative trend.

According to the plot 3, the route quality decreases

with time, which means that if the ship does not re-

plan the route, an accident may happen [30].

3.3 Ship collective intelligence

Autonomous and smart ships are expected to be

equipped with an onboard control system that makes

all of the decision related to ship handling. As inputs,

this system takes data from the sensors installed

onboard. In some situations, however, an

autonomous ship needs to take into account the

actions of the surrounding ships in order to maintain

safety and comply with COLREGs [13, 35]. Consider

two possible cases: in the first case there are only

ships that operate fully automatically, and the second

case when at least one of the ships is operated by

humans. In both cases, the ships need to pass safely

through a particular area.

In the first case, when there are only smart ships,

these ships can communicate in order to infer a

decision [36]. The main difficulty here is to make all of

the ships to come to the same decision, i.e. to make all

of the ships agree with it. All ships are considered

equal, i.e. there are no primary ships that make

decisions and secondary ships that obey. In this case,

they can use a distributed consensus algorithms like

PAXOS [26] or Raft [34] that are widely used in

clustered and distributed computing. These

algorithms help a set of distributed computing nodes

get the same result for a particular task, thus coming

to a consensus. These algorithms can be used by

autonomous ships to come to a consensus about their

actions in the case of an encounter. Both PAXOS and

Raft guarantee that all distributed nodes end up with

the same state of the finite automaton that forms the

core of the algorithms. This means that all of the ships

in question are able to agree on how to behave in

order to safely pass through the area.

In the second case, the autonomous ships and

ordinary ship also need to communicate in some way

in order to agree on their actions [36]. However, in

this case, the ordinary ship is operated by humans

that need to be able to communicate with machines

operating other ships. This issue can be resolved by

installing onto ordinary ships a special

communication unit that is able to communicate with

smart ships (standardization required). The crew uses

this unit to explain their plans on the safe passage

through the area in question so that the autonomous

ships are aware of the human’s decisions and thus can

infer their own. The smart ships can still use PAXOS

and Raft to come up with a decision taking into

account human’s opinion. Another approach is to

make this onboard unit a part of the consensus

framework. The ordinary ship also takes part in the

decision process and is considered as a member of a

distributed consensus network. However, in this case,

the crew of the ordinary ship needs to obey the

collective decision.

3.4 Ensuring the ship’s seaworthiness

One of the most important tasks the crew needs to

care about is the ship’s seaworthiness. Here we

concentrate on seakeeping performance criteria, i.e.

the ability of a ship to effectively and safely deal with

the environmental conditions, i.e. pitch and roll,

buoyancy and stability. In the case of an ordinary

ship, the crew is responsible for maintaining these

criteria, but in the case of an autonomous or smart

ship, this needs to be done automatically. Since there

is no crew on board, no one can perform actions to

prevent an accident that is about to happen. The

onboard control system should preliminarily detect

such situations and take actions not just to prevent

them, but to avoid the related risks [30]. There is a

known controllability issue: under some

circumstances, the ship may not be under a full

control [35], therefore the onboard control system

616

needs to take into account that the issued orders may

be performed only partially, or have no effect.

Another difficulty is the uncertainty in the ship’s state

data: the data that is thought to be actual and correct

may be out of date or incomplete [42].

Waves affect the seakeeping characteristics

inducing rolling and pitching, which, in turn, affect

the ship’s stability. Waves are highly stochastic, non-

stationary process and therefore predicting its

parameters is a difficult task [8]. Although wave

parameters can be drawn from a wave spectra [19],

each spectrum provides approximate results and thus

needs to be tuned for a particular area [4]. The tuning

requires long time series of wave parameters, which

could not be collected during the voyage. Moreover, a

ship, in general, does not have the machinery

required to measure wave parameters, e.g. period or

wavelength. The possible solution is to install

accelerometers and gyroscopes onboard and directly

measure rolling and pitching parameters. The data

received from these sensors form a time series that

can be analyzed with Deep Neural Networks,

possibly predicting future parameters [10,48],

anomaly detection techniques or other approaches to

identify patterns and make decisions accordingly.

Hull consistency affects the ship’s buoyancy and

stability. In case of a breach, the hull’s compartment

gets flooded thus changing the list angle and reducing

ship’s stability. Additional water volume inside the

hull also reduces the ship’s buoyancy. Therefore it is

important that the onboard control system is able to

monitor the hull’s integrity and compartment

flooding. This can be achieved using Hull Monitoring

System [14] and flooding detection sensors installed

in all compartments. Hull Monitoring System consists

of several sensors located around the hull measuring

tensions and stress of the hull caused by the sea state

[14, 47]. The collected data is used to analyze the

effects of the sea state and the ship’s movements. The

data can be used to predict possible breaches caused

by extreme wave effects, e.g. slamming and detect

existing breaches. Flooding detectors provide the data

whether a compartment is flooded and to what extent.

The onboard control system uses the data from the

Hull Monitoring System and flooding detectors to

ensure the hull’s integrity. In case of flooding the

control system uses stability curves to select the

appropriate actions (e.g. to flood the opposite

compartment).

Since Machine Learning is known to be able to

handle data with missing values [10] (e.g. because of

temporary failure of a sensor) and adapt to changes in

data patterns happened since training [27], it is

expected that Machine Learning and Deep Learning

will form a general framework for the analysis of the

ship’s state needed to ensure it’s seaworthiness. To be

able to analyze its state the ship needs to measure its

static and dynamic motion parameters, like roll

angles, roll acceleration, and amplitude. Therefore, it

should be equipped with sensors, accelerometers and

detectors and Hull Monitoring System that measure

the ship’s state. These sensors form the digital model

of the ship. With this model available it is possible not

only to analyze the instant ship’s state but also to

predict the future development of the current

situation [32].

4 AUTONOMOUS SHIP REMOTE MONITORING

In order to ensure seaworthiness, the smart and

autonomous ships will be equipped with a great

number of sensors that measure their dynamics. This

data can be transmitted to the onshore control centre

thus making it possible to remotely analyze the ship’s

behaviour [12, 13, 37]. Because of the distributed

nature of the onboard ship machinery, a ship can be

thought of as an IoT-enabled multiagent system

where each machinery unit is an independent agent

[35]. The amount of data generated by these IoT

agents for a single ship is expected to make up a large

dataset [35]. If a shipowner has a fleet of smart ships,

the amounts of data increase by orders of magnitude.

However, in the latter case, the shipowner can benefit

from the large datasets collected from the entire fleet

by getting the overall statistics computed using the

data from all of the ships, which increases prediction

accuracy. This, in turn, requires the ship owners to

have a Data Science or Data Engineering departments

to successfully handle, analyze and gain insights from

this data using Big Data and Machine Learning

techniques. It is expected that the entire marine

industry would benefit from these datasets and

related software [3].

Collecting datasets with the data related to the

ship’s operations enables ship owners to employ

maintenance prediction techniques [6, 37]. These

techniques let the ship owners analyze the behaviour

of the onboard machinery and predict whether it is

going to go out of order. This lets the owners

optimally plan the maintenance operations [35, 37].

Actually, such a monitoring system can be installed

on an ordinary ship and the monitoring data can

available both to the shipowner and the crew.

5 ONBOARD CONTROL SYSTEM ROBUSTNESS

The main source of trouble onboard is the crew [12].

Since there is no crew on board, its effect on the ship’s

safety is reduced, making technical issues the main

concern. In the case of an autonomous ship, the

onboard control system becomes a single point of

failure [13]. If the onboard control system goes out of

order, the ship is not able to operate. Since the ship

has no crew on board, there is no way to repair the

failed systems. Therefore, the onboard control system

should be able to detect its own failures and either

entirely recover from them or rebalance the workload

in such a way that the components that are still alive

can keep the system operating.

To ensure robustness, the development team needs

to use distributed approaches, where a computer

system consists of several network nodes connected to

accomplish a single task. If a single node goes out of

order, another one takes its workload. One possible

solution for that is to adopt the multi-agent approach,

also known as actor approach. There are different

actor implementations available for different

programming languages and environments [5, 9, 29].

In the case of the actor approach, the software

system is represented as a set of independent

components (called actors). The actors communicate

by sending messages, not through direct interaction

617

by function calls like in traditional software systems.

Since the actors communicate through message

passing, different actors can be located on different

network nodes [5] enabling effective development of

distributed applications. Here, the sender actor does

not need to know where (on which network node) the

receiver is located and the underlying actor

framework hides all the complexity of the network

communication [5,9] making it easier for the

developer to concentrate on the domain, not

infrastructure.

The distributed nature of applications built using

the actor approach makes them more error-resistant

than traditional centralized applications. Once a

network node goes down, the actor framework

detects this failure and deploys the actors that have

been on the failed node on the nodes that are still

alive and routes the messages to the new instances [5].

Each actor is responsible for persisting its internal

state and loading it once it is deployed, thus making

the application survive through node failures without

data loses.

Since the ability to operate regardless of the

failures is essential for the onboard control systems of

autonomous ships, we argue that they can benefit

from adopting the actor approach making them more

error-resistant and letting to gracefully recover from

failures. It is important to distribute the network

nodes across the entire ship in order to ensure that

physical damage or flooding of a single compartment

does not lead to the destruction of the entire network.

Nevertheless, the actor model is not able to

overcome hardware failures if the particular

hardware exists as a single instance. Therefore,

sensors that are used to measure the ship’s state need

to be duplicated in order to increase their robustness.

However, it is still possible that the onboard

control system or the onboard machinery goes out of

order. It is expected that in this case, a ship should

activate a kind of fail-to-safe mode, that lets it safely

operate using the machinery it still has and wait for

the assistance from the onshore services [37].

6 CONCLUSION

There is still much to do before an ocean-going

autonomous ship goes to its first voyage. This means

the need for research on topics related to ship

architecture, legal regulations, port operations,

information technology, and others. In this paper, we

have discussed the information technology aspects of

the autonomous and smart ships. It is expected that

such ships will have an onboard control system

responsible for planning route, handling the ship

during the voyage and monitoring its state.

We argue that the route planning component

should not be implemented using techniques that

require training because this needs a great number of

real-life experiments including those that can lead to

an accident; otherwise, it requires a complicated

simulation environment that precisely models the

ocean or the agent knowledge would be wrong.

Ship handling during the voyage is expected to be

based on intelligent techniques that analyze the ship’s

state and the environment around it making decisions

on what actions to perform in order to keep the ship

on the route and maintain its seaworthiness.

The need for the onboard control system to

monitor the ship’s state and the environment means

that the autonomous and smart ships will be

equipped with a great number of sensors that capture

the properties of the ship’s movements and the

surrounding environment. This enables for remote

monitoring of the ship’s state and actions thus making

it possible to make decisions based on collected

datasets. This also enables predictive maintenance

thus reducing the downtime due to spare part

delivery and reducing costs. The maritime industry

will also benefit from the large datasets collected

during autonomous ship operations.

In order to ensure the robustness of the onboard

control system, it can be developed using the actor

approach thus making it inherently parallel and

distributed.

Most of the technology needed to support the

onboard control system already exists but needs to be

adapted to the requirements of the maritime industry

to make the future of ocean transportation closer.

REFERENCES

[1] Osman Balci. Validation, verification, and testing

techniques throughout the life cycle of a simulation

study. Annals of Operations Research, 53(1):121– 173,

Dec 1994.

[2] Osman Balci. Verification, validation, and accreditation.

In Proceedings of the 30th Conference on Winter

Simulation, WSC ’98, pages 41–4, Los Alamitos, CA,

USA, 1998. IEEE Computer Society Press.

[3] C.L. Benson, P.D. Sumanth, and A. P. Colling. A

quantitative analysis of possible futures of autonomous

transport. In INEC 2018 Conference, 2018.

[4] Lars Bergdahl. Comparison of measured shallowwater

wave spectra with theoretical spectra. Proceedings of the

8th European Wave and Tidal Energy Conference,

Uppsala, Sweden, 2009, 2009.

[5] Phil Bernstein, Sergey Bykov, Alan Geller, Gabriel Kliot,

and Jorgen Thelin. Orleans: Distributed Virtual Actors

for Programmability and Scalability. March 2014.

[6] Mogens Blanke, Michael Henriques, and Jakob Bang. A

pre-analysis on autonomous ships. Technical University

of Denmark, 2016.

[7] E. Cantu´-Paz. A survey of parallel genetic

algorithms. CALCULATEURS PARALLELES,

RESEAUX ET SYSTEMS REPARTIS, (10), 1998.

[8] A. Castro, R. Carballo, G. Iglesias, and J.R. Rabun˜al.

Performance of artificial neural networks in nearshore

wave power prediction. Applied Soft Computing, 23:194

– 201, 2014.

[9] Dominik Charousset, Thomas C. Schmidt, Raphael

Hiesgen, and Matthias Wa¨hlisch. Native Actors– A

Scalable Software Platform for Distributed,

Heterogeneous Environments. In Proc. of the 4rd ACM

SIGPLAN Conference on Systems, Programming, and

Applications (SPLASH ’13), Workshop AGERE!, pages

87–96, New York, NY, USA, Oct. 2013. ACM.

[10] Zhengping Che, Sanjay Purushotham, Kyunghyun Cho,

David Sontag, and Yan Liu. Recurrent neural networks

for multivariate time series with missing values.

Scientific Reports, 8(1):6085, 2018.

618

[11] Thomas H. Cormen, Charles E. Leiserson, Ronald L.

Rivest, and Clifford Stein. Introduction to Algorithms.

MIT Press and McGraw-Hill, 2009.

[12] John F. Cross and Gordon Meadow. Autonomous ships

101. Journal of Ocean Technology, Vol 12:23– 27, 09

2017.

[13] V. I. Dmitriev and V. V. Karetnikov. Methods of

ensuring the safety of navigation when implement

unmanned technology (in russian). Vestnik

Gosudarstvennogo universiteta morskogo i rechnogo

flota imeni admirala S.O. Makarova 9.6 (2017), 2017.

[14] Global Maritime Engineering. Hull Stress Monitoring

System Electric Sensors’ Solution. Measure stresses,

torsions, caused by Waves, Cargo Operations and

Motions.

[15] Ian Goodfellow, Yoshua Bengio, and Aaron Courville.

Deep Learning. MIT Press, 2016.

http://www.deeplearningbook.org.

[16] M. Grifoll, F. X. Mart´ınez de Ose´s, and M. Castells.

Potential economic benefits of using a weather

shiprouting system at short sea shipping. WMU Journal

of Maritime Affairs, 17(2):195–211, Jun 2018.

[17] Manel Grifoll, Llu´ıs Martorell, Marcel·la Castells, and

Francesc Xavier Mart´ınez de Ose´s. Ship weather

routing using pathfinding algorithms: the case of

Barcelona – Palma de Mallorca. In XIII Congreso de

Ingenieria del Transporte, pages 1–11. Universidad de

Oviedo, 2018.

[18] A. Hassani and J. Treijs. An overview of standard and

parallel genetic algorithms. IDT workshop on interesting

results in computer science and engineering, 2009.

[19] Klaus Hasselmann, T. Barnett, E. Bouws,H.

Carlson, D. Cartwright, K Enke, J Ewing, H

Gienapp, D. Hasselmann, P. Kruseman, A Meerburg,

Peter Muller, Dirk Olbers, K Richter, W. Sell, and H.

Walden. Measurements of wind-wave growth and swell

decay during the joint north sea wave project (jonswap).

Deut. Hydrogr. Z., 8:1–95, 01 1973.

[20] John H. Holland. Adaptation in Natural and Artificial

Systems. University of Michigan Press, Ann Arbor, MI,

1975. second edition, 1992.

[21] Rolls-Royce Ship Intelligence. Autonomous ships. the

next step. Rolls-Royce Marine, 2016.

[22] Alex Irpan. Deep reinforcement learning doesn’t work

yet. https://www.alexirpan.com/2018/02/14/ rl-

hard.html, 2018.

[23] Hinnenthal Jo¨rn and Harries Stefan. A systematic

study on posing and solving the problem of pareto

optimal ship routing. In 3rd International Conference on

Computer and IT Applications in the Maritime

Industries, COMPIT’04, pages 27–36, 2004.

[24] V. V. Karetnikov, I. V. Pashchenko, and A. I. Sokolov.

Prospects of introducing unmanned navigation on

inland waterways of the Russian Federation (in russian).

Vestnik Gosudarstvennogo universiteta morskogo i

rechnogo flota imeni admirala S.O. Makarova 9.3 (2017),

2017.

[25] Abdullah Konak, David W. Coit, and Alice E. Smith.

Multi-objective optimization using genetic algorithms: A

tutorial. Reliability Engineering & System Safety,

91(9):992 – 1007, 2006. Special Issue Genetic Algorithms

and Reliability.

[26] Leslie Lamport. Time, clocks, and the ordering of

events in a distributed system. Commun. ACM,

21(7):558–565, jul 1978.

[27] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton.

Deep learning. Nature, 521:436 EP, May 2015.

[28] R. Leigh, S. J. Louis, and C. Miles. Using a genetic

algorithm to explore a*-like pathfinding algorithms. In

2007 IEEE Symposium on Computational Intelligence

and Games, pages 72–79, April 2007.

[29] Hugh McKee and Oliver White. Akka A to Z. An

Architect’s Guide To Designing, Building, And Running

Reactive Systems. Lightbend, Inc., 2018.

[30] Yu. I. Nechaev. Catastrophe theory: the modern

approach to decision-making (in Russian). ArtExpress,

Saint-Petersburg, 2011.

[31] Yuri I. Nechaev and Oleg N. Petrov. Intellectual

technology for control of dynamic unsinkability on the

unmanned vessels (in Russian). Marine intellectual

technologies, 3(1), 2020.

[32] Nikitas Nikitakos. Fourth industrial revolution in

maritime sector. University of the Aegean, 2019.

[33] Masoud Nosrati, Ronak Karimi, and Hojat Allah

Hasanvand. Investigation of the * (star) search

algorithms: Characteristics, methods and approaches.

World Applied Programming, 2(4):251–256, April 2012.

[34] Diego Ongaro and John Ousterhout. In search of an

understandable consensus algorithm. In 2014 USENIX

Annual Technical Conference (USENIX ATC 14), pages

305–319, Philadelphia, PA, jun 2014. USENIX

Association.

[35] Lokukaluge Perera. Deep learning towards autonomous

ship navigation and possible colregs failures. Journal of

Offshore Mechanics and Arctic Engineering, 05 2019.

[36] Alexandr Pinskiy. E-navigation and autonomous ship

handling (in russian). Transport of the Russia

Federation. A journal about science, practice, economics,

2016.

[37] MUNIN Project. Research in maritime autonomous

systems project results and technology potentials, 02

2016.

[38] Wei Shao, Peilin Zhou, and Sew Kait Thong.

Development of a novel forward dynamic programming

method for weather routing. Journal of Marine Science

and Technology, 17(2):239–251, Jun 2012.

[39] David Silver, Aja Huang, Chris J. Maddison, Arthur

Guez, Laurent Sifre, George van den Driessche, Julian

Schrittwieser, Ioannis Antonoglou, Veda

Panneershelvam, Marc Lanctot, Sander Dieleman,

Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya

Sutskever, Timothy Lillicrap, Madeleine Leach, Koray

Kavukcuoglu, Thore Graepel, and Demis Hassabis.

Mastering the game of Go with deep neural networks

and tree search. Nature, 529(7587):484–489, Jan 2016.

[40] David Silver, Thomas Hubert, Julian Schrittwieser,

Ioannis Antonoglou, Matthew Lai, Arthur Guez,

Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore

Graepel, Timothy P. Lillicrap, Karen Simonyan, and

Demis Hassabis. Mastering chess and shogi by self-play

with a general reinforcement learning algorithm. CoRR,

abs/1712.01815, 2017.

[41] Martin Hjorth Simonsen, Erik Larsson, Wengang Mao,

and Jonas W. Ringsberg. State-of-the-art within ship

weather routing. In ASME 2015 34th International

Conference on Ocean, Offshore and Arctic Engineering,

page 11, 2015.

[42] Alexander N. Suslov, Olga V. Odegova, and Sun

Tanshi. Approaches for the stability automation

monitoring creation (in Russian). Marine intellectual

technologies, 3(1), 2020.

[43] K. Takashima, B. Mezaoui, and R. Shoji. On the Fuel

Saving Operation for Coastal Merchant Ships using

Weather Routing. TransNav, the International Journal

on Marine Navigation and Safety of Sea Transportation,

3:401–406, 2009.

[44] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki,

Michae¨l Mathieu, Andrew Dudzik, Junyoung Chung,

David H. Choi, Richard Powell, Timo Ewalds, Petko

Georgiev, Junhyuk Oh, Dan Horgan, Manuel Kroiss, Ivo

Danihelka, Aja Huang, Laurent Sifre, Trevor Cai, John P.

Agapiou, Max Jaderberg, Alexander S. Vezhnevets,

Re´mi Leblond, Tobias Pohlen, Valentin Dalibard, David

Budden, Yury Sulsky, James Molloy, Tom L. Paine,

Caglar Gulcehre, Ziyu Wang, Tobias Pfaff, Yuhuai Wu,

Roman Ring, Dani Yogatama, Dario Wu¨nsch, Katrina

McKinney, Oliver Smith, Tom Schaul, Timothy Lillicrap,

Koray Kavukcuoglu, Demis Hassabis, Chris Apps, and

David Silver. Grandmaster level in StarCraft II using

619

multi-agent reinforcement learning. Nature,

575(7782):350–354, Nov 2019.

[45] Anna Ja. Voitkunskaia and Petr N. Zvyagin. A routing

of a vessel in drifting ice (in russian). In Marine

Intellectual Technologies. Proceedings of the Saint-

Petersburg State Marine Technical University., volume 2

(40), 1, pages 166–172, 2018.

[46] Liwei Wang, Yin Li, and Svetlana Lazebnik. Learning

two-branch neural networks for image-text matching

tasks. CoRR, abs/1704.03470, 2017.

[47] Wa¨rtsila¨. Stress monitoring system, hull stress

surveillance system. In Wa¨rtsila¨ Encyclopedia of

Marine Technology, 2020.

[48] Haimin Yang, Zhisong Pan, and Qing Tao. Robust and

adaptive online time series prediction with long short-

term memory. Computational Intelligence and

Neuroscience, 2017:1–9, 12 2017.