Thesis Huarong Zheng

Delft University of Technology
Coordination of waterborne AGVs
Zheng, Huarong
DOI
10.4233/uuid:f6aefbb0-1b95-44e9-a4dc-8e6c02d94f37
Publication date
2016
Document Version
Final published version
Citation (APA)
Zheng, H. (2016). Coordination of waterborne AGVs. https://doi.org/10.4233/uuid:f6aefbb0-1b95-44e9-
a4dc-8e6c02d94f37
Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.
Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.
This work is downloaded from Delft University of Technology.

For technical reasons the number of authors shown on this cover page is limited to a maximum of 10.
Coordination of Waterborne AGVs
Huarong Zheng
.
Coordination of Waterborne AGVs
Proefschrift
ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus prof. ir. K.C.A.M. Luyben,
voorzitter van het College voor Promoties,
in het openbaar te verdedigen op
maandag 12 december 2016 om 10:00 uur
door
Huarong ZHENG
Master of Science in Intelligent Transportation Engineering,

Wuhan University of Technology, Wuhan, P.R. China
geboren te Jingshan, P.R. China.
This dissertation has been approved by the promotors:
Promotor: Prof. dr. ir. G. Lodewijks
Copromotor: Dr. R.R. Negenborn
Composition of the doctoral committee:
Rector Magnificus chairman

Prof. dr. ir. G. Lodewijks Delft University of Technology, promotor
Dr. R.R. Negenborn Delft University of Technology, copromotor
Independent members:
Prof. dr. ir. E. Theunissen Netherlands Defence Academy
Prof. dr. R. Dekker Erasmus University Rotterdam
Prof. dr. ir. B. De Schutter Delft University of Technology
Dr. C. Ocampo Martinez Technical University of Catalonia
Prof. dr. ir. L.A. Tavasszy Delft University of Technology
The research described in this dissertation is fully supported by China Scholarship Coun-
cil under Grant 201206950021, and partially by the VENI project “Intelligent multi-agent
control for flexible coordination of transport hubs” (project 11210) of the Dutch Technol-
ogy Foundation STW, a subdivision of the Netherlands Organization for Scientific Research
(NWO).
TRAIL Thesis Series T2016/25, the Netherlands TRAIL Research School

P.O. Box 5017
2600 GA Delft, The Netherlands
Email:[email protected]
Published and distributed by: Huarong Zheng

E-mail: zhr [email protected]
ISBN 978-90-5584-218-6
Keywords: Waterborne AGVs, model predictive control, cooperative distributed control,

robust control, closed-loop scheduling and control, Inter Terminal Transport.
Copyright c 2016 by Huarong Zheng
All rights reserved. No part of the material protected by this copyright notice may be re-
produced or utilized in any form or by any means, electronic or mechanical, including pho-
tocopying, recording or by any information storage and retrieval system, without written
permission of the author.
Printed in the Netherlands

Preface
It has been almost four years since I landed at Schiphol Airport for the first time! I was
told by the customs officer that the PhD study in the Netherlands would be boring when he
asked about the reason why I came here. I was scared. However, there was no way back!
Therefore, I headed on and started this uncertain journey. All kinds of incredible things,
expected and unexpected, happened along the journey. Now, standing at the very end of my
PhD and looking back, I can clearly see how different I am, which gives me the greatest
sense of achievement, even greater than that from the research work presented in this thesis.
Four years ago, I did not know much about control. I spent quite some time following
a master course, Control Theory, from the control department, but still failed the first exam.
It was the first time in my study career to retake an exam. Nevertheless, I understood the
field of control better and was able to develop some simple control algorithms after all. In
small steps, I had my first simulation run successfully, and later even had some interesting
results to present at internal meetings, international conferences, and even the world control
congress. Acceptances for publication in high impact factor peer-review journals and the
award in a poster session further confirm the research and industrial contributions of my
work. Eventually, I finished this PhD thesis. At this moment, I owe a lot of thanks to those
who have helped me make all these possible.
First of all, I would like to acknowledge the financial support from China Scholarship
Council for my four-year living in the Netherlands, and from the Department of Maritime
Transport & Technology, Faculty of Mechanical, Maritime and Materials, Delft University
of Technology for all the research relevant expenses.
Secondly, I would like to thank sincerely my PhD supervisors, Prof. Gabriël Lodewijks
and Dr. Rudy R. Negenborn. Prof. Lodewijks has always been critical about my research,
which always makes me feel nervous before the meeting with him. Although he does not
have many comments, those raised by him are always sharp and can pinpoint the deficiency
of my algorithms. I benefit a lot from his rich academic and industrial experiences and in-
sights into engineering problems. The group lunch times when he shared his life experiences
all over the world are also the most enjoyable.
Dr. Negenborn is the one that I could not thank too much. As my daily supervisor, he
has always been very patient, encouraging, and helpful in training me into an independent
researcher. Every time we had a progress meeting, he would ask tens of questions, at the
same time guiding me to the right research directions. Every time I sent him a draft paper,
he would read and correct it so carefully. I am always the one to feel proud when a group
of PhDs are comparing supervisors in terms of the quality and the waiting time of feedback
on papers. His very organized working style has also impressed me deeply. I gained and
learned so much by working with him, not only the knowledge necessary for my research,
v
vi
but also the way of effective working and supervision. Definitely, I will benefit from these
influences in my future academic career. Special thanks go to him for translating the thesis
summary into Dutch.
Moreover, it has been a delight for me to work with all the colleagues in the Section of
Transport Engineering and Logistics. I enjoyed all the lunch breaks together with them in
Aula. I will not miss the food in Aula, but will, for sure, miss all the happy times we spent
there.
In addition, I would like to thank all my Chinese friends in Delft who have encour-
aged and supported me all the way. Especially, I would like to mention Xiaoyan Wei and
Xiangrong Wang, two lovely Chinese girls from whom I gained so much fun and positive
energy. I will miss those times that we hang out together, shopping, movie or discover-
ing nice Chinese food in Rotterdam and the Hague. It is their encouragements to exercise
more that make me stay healthy physically during the PhD. My three housemates, Jinhu
Wang, Changgong Zhang, and Dong Liu, who have excluded me from the list to take out
the garbage, are the cutest housemates one could have.
Last but not the least, I owe my parents numerous thanks for their unconditional love,
and my cute nephew for the laughter over the weekend video calls. I should also thank my
boyfriend who has always been patient and tolerant when I poured my junk emotions to
him.
I appreciate and treasure this precious experience in my life. It is the end. It is the start.
Facing the future, I feel uncertain about everything again. However, I will head on again
with more confidence this time.
Huarong Zheng,
Delft, August 2016.
Contents
Preface v
1 Introduction 1
1.1 Waterborne AGVs for Inter Terminal Transport . . . . . . . . . . . . . . . 1
1.2 Scope and assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Literature review and background 9

2.1 Relevant intelligent vehicles . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Automated guided vehicles . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 Intelligent marine surface vehicles . . . . . . . . . . . . . . . . . . 11
2.1.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Motion control literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 General control techniques . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 Model predictive control . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Scheduling approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Inter terminal transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Dynamic models of waterborne AGVs 27

3.1 Modeling of waterborne AGVs for ITT . . . . . . . . . . . . . . . . . . . . 27
3.1.1 Waterborne AGV model . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.2 Network model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Successively linearized models . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Predictive path following with arrival time awareness 39

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Modeling in path coordinate systems . . . . . . . . . . . . . . . . . . . . . 40
4.3 Two-level double integrator dynamics . . . . . . . . . . . . . . . . . . . . 42
4.4 Predictive switching logic . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Receding horizon control . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.6 Simulation results and discussion . . . . . . . . . . . . . . . . . . . . . . . 51
4.6.1 Simulation experiment 1: Feasible ITT task . . . . . . . . . . . . . 54
4.6.2 Simulation experiment 2: Infeasible ITT task . . . . . . . . . . . . 58
vii
viii Contents
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5 Cooperative distributed waterborne AGVs 65

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 A centralized formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Cooperative distributed waterborne AGVs . . . . . . . . . . . . . . . . . . 67
5.3.1 Derivation of ADMM for waterborne AGVs . . . . . . . . . . . . . 67
5.3.2 Convergence analysis . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3.3 Stopping criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Fast ADMM for cooperative distributed waterborne AGVs . . . . . . . . . 70
5.5.1 Safely carrying out ITT tasks . . . . . . . . . . . . . . . . . . . . . 73
5.5.2 Convergence behavior in distributed computations . . . . . . . . . 75
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6 Cost-effective robust distributed control of waterborne AGVs 81

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Cost-effective robust centralized formulation . . . . . . . . . . . . . . . . . 82
6.2.1 Parameterized uncertainty bounds . . . . . . . . . . . . . . . . . . 82
6.2.2 Tube-based MPC for successively linearized models with paramenter-
ized bounded uncertainties . . . . . . . . . . . . . . . . . . . . . . 84
6.2.3 Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3 Relaxed cost-effective RDMPC with ADMM . . . . . . . . . . . . . . . . 88
6.3.1 Derivation of ADMM for the relaxed RDMPC problem . . . . . . . 88
6.3.2 Convergence analysis and stopping criteria . . . . . . . . . . . . . 90
6.4 Efficient exact cost-effective RDMPC . . . . . . . . . . . . . . . . . . . . 92
6.4.1 B&B with SOS1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4.2 Integrated B&B with SOS1 and ADMM . . . . . . . . . . . . . . . 94
6.5.1 Safely fulfilling ITT tasks . . . . . . . . . . . . . . . . . . . . . . 97
6.5.2 Convergence of distributed decision making . . . . . . . . . . . . . 98
6.5.3 Cost-effective robust tubes . . . . . . . . . . . . . . . . . . . . . . 101
6.5.4 Efficient B&B search . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7 Closed-loop scheduling and control for autonomous Inter Terminal Transport 109
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3 Energy efficient scheduling of ITT using waterborne AGVs . . . . . . . . . 112
7.3.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.3.2 Mixed integer programming problem . . . . . . . . . . . . . . . . 113
7.3.3 Transformations into linearity . . . . . . . . . . . . . . . . . . . . 115
7.4 Real-time closed-loop scheduling and control . . . . . . . . . . . . . . . . 117
7.4.1 Modeling interactions and real-time speed assignment . . . . . . . 117
7.4.2 Closing the real-time loop . . . . . . . . . . . . . . . . . . . . . . 120
7.5 Experiments and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Contents ix
7.5.1 From the waterborne AGV perspective . . . . . . . . . . . . . . . 122

7.5.2 From the ITT request perspective . . . . . . . . . . . . . . . . . . 125
7.5.3 From the berth perspective . . . . . . . . . . . . . . . . . . . . . . 128
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8 Conclusions and future research 131

8.1 Conclusions and contributions . . . . . . . . . . . . . . . . . . . . . . . . 131
8.1.1 Answering the research questions . . . . . . . . . . . . . . . . . . 131
8.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.2 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.2.1 Directions for coordinating waterborne AGVs . . . . . . . . . . . . 135
8.2.2 Additional directions for future research . . . . . . . . . . . . . . . 137
Bibliography 139
Glossary 151
TRAIL Thesis Series publications 157
Samenvatting 159
Summary 163
Curriculum vitae 167

Chapter 1
Introduction
Efficient and sustainable container handling is critical for large ports to improve compet-
itiveness in the increasingly globalized economy. Terminal level operations in the port of
Rotterdam have benefited significantly from innovative technologies such as Automated
Guided Vehicles (AGVs) [135] and fully automated terminals. This dissertation proposes
a new type of container transporter, the waterborne Autonomous Guided Vessels (water-
borne AGVs) for smarter port level logistics. The main focus of this thesis is on developing
control and scheduling strategies for coordinated waterborne AGVs to move containers au-
tonomously between terminals, the so-called Inter Terminal Transport (ITT) [129].
In this chapter, the background and the motivation for the research on coordination
of waterborne AGVs are first presented in Section 1.1. Section 1.2 formally defines the
research scope and assumptions, followed by the research questions that will be addressed
in this thesis in Section 1.3. This chapter is concluded in Section 1.4 with an overview of
the contents of this thesis.
1.1 Waterborne AGVs for Inter Terminal Transport

In 2013, the Port of Rotterdam opened a new port area, Maasvlakte II, with approximately
2,000 hectares reclaimed from the sea [102]. Maasvlakte II together with Maasvlakte I
forms a global container hub complex, as shown in Figure 1.1. It is the vision of the Port
Authority to affirm its leading position in the field of efficiency and sustainability by 2030
[101]. Especially for the new port area, the Port Authority leaves space for innovative
technologies aiming at developing it into the smartest and most sustainable port in the world.
The port of Rotterdam has a long tradition of applying advanced technologies. The port of
Rotterdam was the first port in the world adopting AGVs at the ECT Delta terminal in 1993.
The first fully automated terminals, APMT, opened in 2015. A 40% increase in productivity
is foreseen by APMT due to automation [3].
The port area of the Maasvlakte in Figure 1.1 has several features that are representative
for general large ports. Firstly, it is expected that the throughput of containers will increase
to more than 30 million Twenty-foot Equivalent Unit (TEU) per year by 2035 [101]. They
need to be handled more efficiently in order to increase the competitiveness of the port.
Movement of containers happens both inside terminals, likely handled by land-side AGVs,
1
2 1 Introduction
Figure 1.1: Overview of terminals at Maaasvlakte I and II; adapted from [32].
and between terminals via various modalities (e.g., road, rail, sea), i.e., ITT. At present,
ITT is realized mainly by road vehicles, e.g., multi-trailer systems. Secondly, the reclaimed
land from the sea in the new port area is limited, which poses challenges on expanding
the existing physical transportation infrastructure to accommodate increasing traffic flow by
land. Thirdly, for complex geographical layouts like the Maasvlakte, the distances between
some terminals are much longer by land than by water. This is illustrated in Figure 1.1 by
the example involving Terminals 8 (ECT Delta Barge Feeder terminal), 20 (Rhenus), and 2
(Euromax terminal). Distances among the three terminals by water indicated by the green
lines are much shorter than by land indicated by the red lines. Last but not the least, the
Maasvlakte can be seen approximately as a confined water area which has relatively unso-
phisticated traffic so far and has reliable advanced ICT systems supporting the development
of intelligent infrastructures.
A new type of container transporter over water, waterborne AGVs, is proposed for ITT
in this thesis. Literally related with conventional AGVs, waterborne AGVs share similari-
ties but are different from conventional AGVs. Both waterborne AGVs and AGVs are un-
manned software controlled vehicle systems for logistics distributions; and both of them are
favorable for relatively simple environments with repeating transportation patterns. How-
ever, waterborne AGVs differ with AGVs on several aspects: a) waterborne AGVs are for
transportation over water and AGVs for transportation over land; b) waterborne AGVs can-
not be navigated by following markers, wires, or magnets etc. in the floor, which is the case
for most of the existing AGVs. According to [87] on the definitions of levels of autonomy,
waterborne AGVs have a higher level of autonomy than AGVs in that waterborne AGVs not
only have built-in functionality but also have goal-directed reaction and behavior. There-
fore, the word “Autonomous” has been used for waterborne AGVs while “Automated” for
conventional AGVs; c) it is not possible to specify an area especially for waterborne AGVs
1.2 Scope and assumptions 3
without external traffic as has been done for AGVs that are applied in manufacturing in-
dustries, warehouses, and container terminals [135]; d) temporal requirements for water-
borne AGVs are more stringent than AGVs since for ITT, the most important criterion is
“non-performance” which happens when the completion time of ITT tasks is later than the
permitted latest arrival time [21].
In general, the potential benefits of developing waterborne AGVs for transport in port
areas are summarized as follows:
• Waterborne AGVs could be almost labor cost free since no mariners are necessarily
on board;
• Waterborne AGVs could offer another transport mode to handle the expected large
throughput instead of exploiting the limited land in port areas for road traffic;
• Waterborne AGVs, comparable to land-side AGVs, could be optimally operated 24/7

with reliable performance and improve port efficiency exploiting automation;
• For terminals with longer distances by land than by water, waterborne AGVs could
save energy compared to road vehicles; and
• Waterborne AGVs are in line with the development of smart ports and are deemed as
very relevant to the ITT practice in the port of Rotterdam [24].
To develop a transport system using waterborne AGVs involves multi-discipline tech-

nological and methodological advancement. This dissertation focuses on the coordination
of waterborne AGVs for ITT as to be further clarified next.
1.2 Scope and assumptions

We consider an autonomous waterborne ITT system: a fleet of waterborne AGVs that han-
dles a set of ITT requests to transport autonomously specified amounts of containers be-
tween specified origins and destinations with temporal constraints in an energy efficient
way.
Typically, a transportation decision-making system is hierarchically partitioned into
three levels: long-term strategic, mid-term tactical, and short-term operational control deci-
sions. Analogous levels of an autonomous ITT system using waterborne AGVs are shown
in Figure 1.2. Within such an ITT system, a port authority runs a fleet of waterborne AGVs
shuttling between terminals internally in the port area to transport containers. Strategic deci-
sions regarding locations of berths for waterborne AGVs, fleet size, and composition issues,
etc. are long term decisions in the order of years. Tactical and operational levels determine,
for each waterborne AGV, the chronological events that occur at the hours time scale and the
speeds, accelerations, or amount of power to input that occurs at the seconds time scale in
order to assure those events are executed as scheduled, respectively. The tactical scheduling
and operational control level problems as circled by the red dashed line in Figure 1.2 are of
interest in this dissertation for coordinating waterborne AGVs.
Assumptions are made that:
4 1 Introduction
Strategic level
Long-term: berth layout, fleet size and composition, etc.
ITT network/fleet
Tactical level
Mid-term: adjust fleet size, routing, berth allocation, etc.
Schedules
Operational level
Short-term: speed selection, tracking, load/unload, etc.
Forces/moments
Communication
Sensor
Waterborne
AGVs Processing
Physical system level & environment
Figure 1.2: Different levels of an ITT system using waterborne AGVs; adapted from [16].
• The waterborne AGV fleet size and composition have been decided by the strategic
level in a way that there is a sufficient number of waterborne AGVs available for ITT
requests;
• The ITT network has also been designed at the strategic level. The network includes:
berths that can accommodate waterborne AGVs by providing charging, maintenance,
parking, etc., and routes as shortest paths connecting berths;
• Each terminal has one waterborne AGV berth with available load/unload equipment
so that the berth allocation problem at the tactical level and the load/unload problem
at the operational level vanish;
• Each waterborne AGV is equipped with sensors, communication devices, and pro-
cessing units to measure its own system states, communicate with other waterborne
AGVs within a certain range, and perform certain computations; and
• The number of containers that need to be transported from each ITT request is smaller
than the capacity of waterborne AGVs, and split of ITT requests is not allowed.
From a control perspective, waterborne AGVs should be able to comprehend the sur-
roundings and determine what to do autonomously in order to fulfill ITT tasks, i.e., pick-up
and deliver containers at specified terminals at specified times. Challenges for the con-
trol level problems arise from various aspects: a) waterborne AGVs, like other marine
surface vehicles, have limited maneuverability. Therefore, they cannot respond timely to
environmental changes, which could lead to undesirable or even dangerous behavior, e.g.,
collisions; b) multiple conflicting operational control objectives exist including tracking,
energy efficiency, low “non-performance” rate etc. even for one waterborne AGV; c) sys-
tem constraints on inputs and outputs due to limited engine power, mechanical maximum
1.3 Research questions 5
deflections/revolutions or spatial no-sailing zones, etc., need to be satisfied; d) complex

waterborne AGV dynamics that model waterborne AGV behavior could render applicable
control techniques limited; e) when multiple ITT tasks are scheduled for multiple water-
borne AGVs, waterborne AGVs should preferably compute in a distributed and cooperative
way; and f) if the influences of environmental disturbances, e.g., wind, waves, and currents,
cannot be known perfectly beforehand, waterborne AGVs should react in a robust way and
maintain safety, e.g., satisfying physical limitation and collision avoidance constraints, with
possible reasonable performance deterioration.
From a scheduling perspective, the goal is to determine, for each waterborne AGV, a
sequence of terminals to visit, the corresponding arrival times and loading/unloading vol-
umes. The main challenge lies in developing a closed-loop and tightly integrated scheduling
and control scheme for waterborne AGVs since the scheduling and control levels share the
common aim of making economical and environmentally friendly decisions. Although both
levels largely rely on mathematical models and optimization techniques and both aim at
either maximizing profit or minimizing cost, the inherently different time-scale nature pro-
hibits an integrated and computationally tractable solution.
1.3 Research questions

Following the scope and assumptions for the considered research problems with challenges
presented above, this dissertation aims to develop advanced control and scheduling strate-
gies for coordinated waterborne AGVs applied to ITT. This main research goal will be
achieved by addressing the following five Key Research Questions:
1. Which technique is suitable for the control of waterborne AGVs?
2. What performance criteria should be considered in optimizing the process of one
waterborne AGV carrying out one ITT task and how can the optimal performance be
achieved?
3. How can multiple waterborne AGVs be coordinated for multiple ITT tasks with wa-
terborne AGVs making decisions locally while minimizing the overall cost in a coop-
erative and distributed way?
4. How can environmental disturbances due to wind, waves, and current be systemati-
cally handled by cooperative and distributed waterborne AGVs?
5. In what way can the scheduling and control loop for waterborne AGVs be closed in
order to obtain an energy-efficient autonomous ITT system?
For answering these Key Research Questions, a comprehensive literature review will be
carried out, which also further motivates the research discussed in this thesis. System and
control theories, model predictive control (MPC) in particular, will be utilized extensively.
Waterborne AGV dynamic trajectories considering various system constraints will be con-
trolled and optimized satisfying possibly conflicting design objectives using deterministic,
distributed, and robust control tools. Coordination at the tactical level will also be consid-
ered and will be tightly integrated with the control problems to build an autonomous ITT
system using waterborne AGVs.
6 1 Introduction
1.4 Thesis outline

The road map of this dissertation is presented in Figure 1.3, illustrating connections of
chapters and a suggested order in which the chapters can be read. The contents for Chapters
2 – 8 are summarized briefly as follows:
• Chapter 2 reviews relevant literature on control and scheduling techniques that are
applied to improve intelligence of ground vehicles and marine surface vehicles. Ex-
isting approaches and applications regarding MPC, distributed MPC, and robust MPC
techniques are discussed. This chapter will partially answer Key Research Question
1.
• Chapter 3 defines the dynamic models for waterborne AGVs that are used in the latter
chapters of this dissertation. A nonlinear dynamic model of marine surface vehicles
with three degrees of freedom is used to simulate waterborne AGV behaviors. Suc-
cessively linearized dynamic models based on this nonlinear model are used to predict
waterborne AGV trajectories over a future horizon for controller development. This
chapter will partially answer Key Research Questions 2 – 5.
• Chapter 4 proposes a predictive path following with arrival time awareness con-
troller for one energy-efficient waterborne AGV. Control goals will be achieved by a
proposed connected reference path coordinate system, a switching logic for avoiding
overshoots, and a two-level double integrator scheme being aware of the arrival time.
This chapter will partially answer Key Research Question 2.
• Chapter 5 extends the proposed controller for one waterborne AGV in Chapter 4 to
scenarios involving multiple waterborne AGVs. Cooperative distributed waterborne
AGVs will be realized using the alternating direction method of multipliers (ADMM).
A fast ADMM algorithm is further proposed to improve convergence rates. This
chapter will partially answer Key Research Question 3.
• Chapter 6 considers the influences of environmental disturbances and proposes a

novel cost-effective robust and cooperative distributed control approach for multiple
waterborne AGVs. We define system robustness levels, propose and solve a cost-
effective robust distributed MPC problem for waterborne AGVs based on results from
Chapter 5. This chapter will partially answer Key Research Question 4.
• Chapter 7 closes the loop of scheduling and control of waterborne AGVs for an au-
tonomous energy-efficient ITT system. A novel pick-up and delivery scheduling ap-
proach considering safe intervals between berthing time slots of different waterborne
AGVs is also proposed. This chapter will partially answer Key Research Question 5.
• Chapter 8 concludes the research in this thesis and outlines directions for future
research.
1.4 Thesis outline 7
1. Introduction
2. Literature review
3. Waterborne AGV
dynamic models
5. Cooperative 6. Waterborne AGVs

4. Single waterborne
distributed waterborne with environmental
AGV path following
AGVs uncertainties
7. Closed-loop
scheduling and control
8. Conclusions and
future research
Figure 1.3: Road map. Arrows indicate read before relations.

Chapter 2
Literature review and

background
This chapter presents an overview of the literature relevant for the development of water-
borne Autonomous Guided Vessels (waterborne AGVs). Section 2.1 introduces two types of
intelligent vehicles, land-based Automated Guided Vehicles (AGVs) and intelligent marine
surface vehicles, which are closely related to waterborne AGVs. Coordinating technolo-
gies regarding motion control and scheduling that could be applicable to waterborne AGVs
are then subsequently reviewed in Section 2.2 and Section 2.3, respectively. Section 2.4
discusses briefly the research work on Inter Terminal Transport (ITT). Conclusions of this
chapter are presented in Section 2.5.
2.1 Relevant intelligent vehicles

Broadly speaking, intelligent vehicles should possess certain capabilities in the so-called
observe, orient, decide and act loop [10]. Waterborne AGVs belong to the class of intel-
ligent vehicles since they are able to observe the surroundings and determine what to do
autonomously to fulfill ITT tasks. This section introduces the development of two types
of existing intelligent vehicles, land-based AGVs and intelligent marine surface vehicles,
which are closely related to waterborne AGVs.
2.1.1 Automated guided vehicles

Land-based AGVs are driverless vehicles that were first introduced in the 1950s simply
towed by a chain [83]. They came into industrial and commercial use in 1976 by adopt-
ing invisible markers on the floor [23], and saw wide applications in the late 20th century
[135]. They are especially designed to move materials in certain areas, e.g., manufacturing
industries, warehouses, and container terminals with a certain level of intelligence. Spec-
ifications of AGVs differ from one another depending on their application scenarios. The
proposed waterborne AGVs can be seen as an extension of AGVs in container terminals
in the sense that AGVs move containers internally in terminals over land and waterborne
9
10 2 Literature review and background
Figure 2.1: The ECT Delta Terminal Rotterdam (Courtesy of ECT).
Figure 2.2: Terex Gottwald AGVs [128].
AGVs move containers between terminals within a port area over water. Therefore, we
confine our introduction to AGVs to those applied to container transportation in container
terminals.
The first container carrying AGV was adopted at the ECT Delta terminal (see Figure
2.1), in the port of Rotterdam in 1993, almost 40 years later than its first introduction and
almost 20 years later than its commercial use in other industries. Ever since then, AGVs
have been widely used in semi- and fully automated container terminals. Cleaner, safer, and
more efficient AGVs are developed to improve container terminal internal automation and
efficiency. Figure 2.2 shows currently adopted AGVs working in container terminals.
In a transport system using AGVs, four parts are identified as fundamental [135]: 1) the
vehicles; 2) the transportation network; 3) the physical interfaces (pick-up/delivery points)
that link the storage and transport systems; and 4) the coordinating system. A number of es-
sential decisions have to be made for each part when designing an AGV system. In terms of
2.1 Relevant intelligent vehicles 11
vehicles, the problem investigated the most is the minimum fleet size [136] required for the
system to, on the one hand, satisfy transport demands, and on the other hand, be economical
and avoid congestion. Other vehicle design issues include deciding on features related to
capacity, speed, power, costs, and more recent self-lifting or non-lifting [21], fixed path or
free-ranging [143], which all interact with the fleet size problem and other system parts.
The transportation network in container terminals connects pick-up (quay side) and deliv-
ery (stacking area) points [142] and defines the guide paths that AGVs follow. Markers or
wires in the floor are usually necessary for navigating AGVs through the network. Decision
problems in physical interfaces concern location and equipment choices [119]. Quay cranes
and stacking cranes are commonly used for loading and unloading containers to and from
AGVs, respectively.
The last part of an AGV system, the coordinating system, is critical in guaranteeing
system performance, e.g., maximizing throughput, minimizing makespan, delays, and cost
while avoiding conflicts among AGVs. Collisions and deadlocks [135] are the two likely
conflicts for fixed path AGVs. A proper layout of the guide paths could prevent conflicts,
which could degrade system performance though. Two more commonly employed ap-
proaches are the so-called zone-control [44] and advanced scheduling strategies [130]. The
idea of zone-control is similar to using traffic lights at crossroads by dividing the network
into several control zones and allowing one vehicle to occupy a zone at one time. Although
simple to implement, zone-control could also sacrifice system performance. Extensive re-
search has been done on scheduling problems that aims at constructing routes an AGV can
take and schedules that give times when the AGV can traverse corresponding routes, see
[135] and references therein. Operations research based vehicle routing problems (VRPs)
[130] or flow shop problems [142] are widely used to formulate the scheduling problem
either in a static or dynamic way. However, the existing coordination mostly takes place at
the scheduling level (see Figure 1.2), and is only applicable to AGVs with fixed paths and
perfect executions of schedules in a disturbance free environment; dynamics that model the
movement of AGVs are rarely considered by the scheduling problems in the literature. In
[142, 143], hierarchical approaches are proposed for scheduling and trajectory planning of
free-ranging AGVs. One dimensional point-mass dynamics are used to model free-ranging
AGVs.
2.1.2 Intelligent marine surface vehicles

Intelligent marine surface vehicles using automatic co-pilots or even autopilots require less
human supervision or intervention during the voyage. Higher intelligence directly reduces
the workload of mariners, and lower the rate of human errors which are the most important
factor contributing to maritime accidents [42].
In general, intelligence is achieved by sensing the environment, processing the sensed
data, and determining what to do based on given missions and current system states. Intelli-
gent marine surface vehicles are usually equipped with a navigation, guidance, and control
(NGC) software system that is responsible for those capabilities. Figure 2.3 briefly illus-
trates the technological components of such an NGC system. The navigation system deals
with noised or even incorrect measured (partial) system states from various sensors and ob-
tains estimated states for use in other components. Commonly used sensors for intelligent
marine surface vehicles include positioning devices like Global positioning system (GPS),
Environment map Control

Constraints
Missions goals Actuators
Route generation Motion control
Guidance System Control System External
disturbances
Vehicle
Controlled
trajectories
Sensor
Observer
fusion
Sensors
Estimated states
Software Navigation System Hardware
Figure 2.3: Diagram of an NGC system for intelligent vehicles (adapted from [30]).
more accurate Differential GPS, or higher cost inertial navigation systems, vision sensors
using cameras, and maritime detection equipment such as wave or depth sensors [14] using
radars. Radio, Wi-Fi, and Automatic Identification System (AIS) are also seen installed on
maritime vehicles [147] for vehicle-to-vehicle or vehicle-to-control center communication.
The guidance system generates reference routes based on environmental information and
given missions. References are then tracked by the vehicle which is controlled by the control
system. Mostly, there exist possibly conflicting control goals such as tracking accurately,
arriving at the destination on time, and using an as small amount of energy as possible. The
control system also takes care of system dynamical limitations such as maximum maneu-
vering speed, maximum engine deflections or revolutions. The processing in guidance and
control systems relies on theoretical tools of optimization and automatic control as to be
reviewed in the following sections.
Research work on intelligent marine surface vehicles has always been active ever since
the first autopilot for ships was proposed in [79]. Platforms and prototypes are built for
various purposes. Table 2.1 provides an overview of the developed prototypes by differ-
ent organizations worldwide. NGC software associated with prototypes is also developed
for intelligent marine vehicles. The SCOUT team developed a set of open source modules
MOOS-IvP [55] for providing autonomy to general robotic platforms. Including the Cy-
bership II system, shown in Figure 2.4, a comprehensive software library called Marine
Systems Simulator [99] is developed to provide necessary resources for rapid implemen-
tation of mathematical models and controllers for marine systems. Vehicles like SCOUT,
Springer, DelfimX, and Delfia-1 (Figure 2.5) are designed to be operable in both remote
control and fully autonomous control modes. However, high speed military vehicle PRO-
TECTOR still relies significantly upon operator guidance and remote control. Most vehicles
are equipped with one PC for all the processing tasks in the NGC system while Springer
has three PCs on board, each responsible for a module, i.e., navigation, guidance, and con-
trol. Note that, except for Delfia-1, few of the aforementioned intelligent marine vehicle
platforms have been designed in the context of transport and logistics.
2.1 Relevant intelligent vehicles 13
Table 2.1: Prototypes of intelligent surface vehicles.
Organization Vehicle Purpose Year

MIT Sea Grant Program, US ARTEMIS, ACES, Oceanographic data 1993, 1997,
AutoCat, and collection, research 1999, 2004
SCOUT [17]
Instituto Superrior Tècnico, Delfim, Caravelas, Coordination with 2004
Portugal DelfimX [34] underwater vehicles
University of Rostock, Ger- Measuring Dolphin Shallow water sur- 2006
many [72] vey, oceanography
tasks
University of Plymouth, UK Springer [126] Environmental and 2007
geographical survey
Institute of Intelligent Sys- Charlie [13] Mine hunting 2003
tems for Automation, Italy
Eotvos Lorand University, Sillverlit [127] Research 2011
Hungary
Israeli Rafael Advanced De- PROTECTOR [108] Military 2005
fense Systems
United States Navy Spartan Scout [112] Military 2001
Marine Cybernetics Labora- Cybership II [121] Research 2005
tory, Norwegian University
of Science and Technology
Delft University of Technol- Delfia-1 [88] Research 2015
ogy, the Netherlands
Figure 2.4: Cybership II at Norwegian University of Science and Technology [121].

Figure 2.5: Delfia-1 at Delft University of Technology [88].
2.1.3 Summary
Several conclusions can be drawn from the development of land-based AGVs and intelli-
gent marine surface vehicles for waterborne AGVs. Firstly, the design of a transport system
using waterborne AGVs are similar with a system using AGVs. Decision problems on
the four system parts, i.e., the vehicles, the transportation network, the physical interfaces,
and the coordinating system could be comparably identified. Secondly, since all the prob-
lems are essentially coupled and interrelated, simultaneous decisions are preferable for a
transport system. However, hierarchical approaches are generally proposed in the literature
for tractable solutions. Tighter integration of the hierarchical levels could be expected to
achieve economical benefits. Thirdly, few research on intelligent marine surface vehicles
has been targeted for civilian use or transportation. Safer, more sustainable, and efficient
marine transport systems could be expected if more intelligent waterborne transport vehicles
are used.
2.2 Motion control literature

Waterborne AGVs are operated by controllers instead of human beings. When applied to
ITT, it is desirable that the waterborne AGV controller can achieve the following goals:
• A given geometric reference path could be tracked with deviations as small as possi-
ble;
• A given arrival time requirement could be met when a preferable time is feasible
considering system limitations, or a minimal delay with respect to the preferable time
2.2 Motion control literature 15
within a specified time window otherwise;
• The aforementioned two design requirements could be achieved in an energy eco-

nomical way;
• Distributed decision making could be possible even when there exist couplings among
waterborne AGVs; and
• A certain level of robustness could be achieved when uncertainties are involved.
In this section, we review general control techniques that have been applied to marine sur-
face vehicles, and that could be applicable to waterborne AGVs for controllers that meet the
above specifications. Particularly, we review model predictive control (MPC) which will be
used extensively in the later chapters.
2.2.1 General control techniques

The meaning of “control” refers to achieve desired system dynamical behaviors using de-
signed algorithms in this thesis. The system together with the controller is called a con-
trolled system. Since external disturbances, mismatches between the model and the system,
and variations in the system itself ubiquitously exist, control engineers generally prefer feed-
back (closed-loop) control with inherent robustness over open-loop control [4]. Briefly, a
feedback controlled system measures system outputs, compares the outputs with references,
computes corrective inputs probably based on a system model, and applies the inputs to the
system to achieve desired behaviors. However, design and analysis of feedback controllers
can vary significantly depending on control purposes and system characteristics such as
whether the system is deterministic or non-deterministic, with linear or nonlinear dynam-
ics, and with coupled or decoupled subsystems.
Particularly, for motion control of marine surface vehicles, three categories of control
problems with different purposes are recognized [30]:
• Setpoint regulation: In this case, the references to the controlled system are constant,
and the corresponding controller is also called a regulator. Examples are constant
speed regulation [29] , heading control [62], and dynamic positioning [124]. Regula-
tion control is also one of the most widely analyzed and best understood problems in
the field of automatic control.
• Path following: The reference is a geometric path independent of time. The refer-
ence path can be straight lines [28, 93, 150] or curves [19, 139] without temporal
constraints.
• Trajectory tracking: Explicit time parameterized references (e.g., positions, veloci-

ties) must be given. The control goal is then to drive the system to the specified states
at specified time [146].
Different types of control techniques are applied to the above motion control problems
for marine surface vehicles:
• The first recognized and most widely implemented controller until now is proportional-
integral-derivative (PID) [79] which was first proposed for ship steering control. PID
has the advantages of being simple to implement and at low cost [134]. However,
issues such as parameter tuning, overshoots, constraints, and performance guarantees
are recognized in PID design.
• Lyapunov-based control design and analysis are more systematic and sophisticated.
Analytical control laws are usually available with guaranteed Lyapunov stability [4]
for controlled marine surface vehicles [19, 28, 122]. In [19, 28, 122], the analysis
is done based on derived error dynamics. Specifically, the reference path given in
[19] is second-order time differentiable so that second-order system dynamics can be
converted to error dynamics. In [28], the references are straight line segments. A
Line-of-Sight (LOS) guidance method is introduced to provide moving references for
the system to track. A maneuvering problem is proposed in [122] where a geometric
task and a dynamic task are involved. The geometric task guarantees path conver-
gence and the dynamic task tracks an assigned speed along the path. Constraints and
performance regarding cost are not considered in these approaches.
• Sliding mode control, as a nonlinear control technique, has also seen applications [9,
38] to marine surface vehicles which have complex nonlinear dynamics. Sliding mode
control laws are discontinuous and can have “chattering” phenomenon. Moreover,
constraints and cost performance cannot be systematically considered as well.
• Intelligent control methods, e.g., fuzzy logic [91], neural network [15] and genetic
algorithm [84] are model free and are based on heuristics. Usually an intelligent and
a non-intelligent control methods are combined. In [84], the genetic algorithm is used
as an optimization tool in an MPC framework. Challenges with intelligent control are
that heuristics are generally empirically determined, e.g., the fuzzification and de-
fuzzication rules for fuzzy logic control [91], and system properties are difficult to be
analyzed.
• Optimal control [31] differs from other control techniques in that it can formulate
a particular objective function, and thus achieves desired behaviors at an optimal
cost. The intelligent marine surface vehicle prototype Springer is controlled by a
linear-quadratic-Gaussian controller [85]. Simplified linear vehicle dynamics are
used. Broadly speaking, MPC also belongs to optimal control. System constraints
and design indices are explicitly taken into account for marine surface vehicle mo-
tion control problems using MPC in [62, 139, 150]. However, optimal control relies
on solving mathematical optimization problems which can be hard when nonlinear
system dynamics and constraints are present.
An overview of the characteristics of the discussed control techniques are summarized
in Table 2.2.
2.2.2 Model predictive control

MPC is one of the most widely implemented control techniques in process industries [107].
General advantages of MPC include handling system constraints systematically and opti-
mizing system performance quantitatively. In particular, considering the characteristics and
Table 2.2: An overview of different control techniques applied to marine surface vehicles.
Control technique Advantage Disadvantage

PID Simplicity and computation- Parameter tuning, overshoots; con-
ally fast straints and performance not guar-
anteed; non-predictive
Lyapunov-based Analytical control laws and Assumptions on system character-
theoretical properties istics; constraints and performance
not guaranteed; non-predictive
Sliding mode Analytical control laws and Chattering phenomena; constraints
robust to disturbances and performance not guaranteed;
non-predictive
Intelligent methods, Model free, simple and com- Relying on experiences; per-
e.g., fuzzy, neural, putationally fast formance not quantified; non-
genetic algorithm predictive
Optimal control Constraints and perfor- Model-based; computational issues
mance guaranteed; easy with complex models
tuning
the main challenges in motion control problems of waterborne AGVs as discussed in Chap-
ter 1, the following justifications for the suitability of applying MPC to waterborne AGVs
are made:
• Waterborne AGVs are with limited maneuverability and could not respond to envi-
ronmental changes or emergencies timely. MPC makes decisions based on predicted
information over a future horizon. Therefore, it can anticipate and prevent undesirable
and dangerous situations, e.g., likely collisions, at an early stage;
• Waterborne AGVs have physical limitations on input, states, and outputs, e.g., maxi-
mum speed and engine power. Moreover, waterborne AGVs need to be a safe distance
away from each other. MPC can handle these constraints explicitly and systemati-
cally;
• Waterborne AGV systems have multiple inputs and outputs, which can also be han-
dled elegantly by MPC;
• For small magnitude of modeling inaccuracies and external disturbances, MPC has
inherent robustness;
• The desired behaviors of waterborne AGVs are defined considering possibly conflict-
ing safety, economical, and environmental factors. Optimization based MPC makes
optimal and quantified trade-off among conflicting objectives with respect to user
defined criteria.
However, there are also concerns with MPC applications. Firstly, although MPC solves
a finite horizon optimization problem, which relieves computational burden in infinite hori-
zon optimal control [31], optimizations in MPC mostly are necessarily solved online repet-
itively except for explicit MPC [8] which is, however, only applicable to simple low order
Algorithm 2.1 Basic MPC algorithm

1: Measure system states at step k;
2: Predict system trajectories based on the system model;
3: Formulate and solve the optimization problem;
4: Apply the first control input;
5: k ← k + 1 and go to Line 1.
dynamics. This restricts the online optimizations to those that can be solved efficiently. Sec-
ondly, controlled system properties, e.g., recursive feasibility, stability and robustness, are
extensively analyzed and guaranteed by design only for certain classes of systems, in partic-
ular for linear time-invariant systems. For more complex systems not necessarily respecting
assumptions made in theoretical analysis, there are still open issues regarding theoretical
properties.
Next, considering the motion control problems for a singe waterborne AGV, multiple
cooperative waterborne AGVs, and waterborne AGVs with environmental disturbances, we
review relevant research on MPC for individual systems, networked systems, and systems
with uncertainties in the literature.
Individual systems
For individual deterministic systems, MPC solves online optimization problems obtaining
a sequence of optimal control inputs and applies the first control input to the system in
a receding horizon way. The predicted system trajectories over a prediction horizon are
driven by solving the optimization problem to the desired behavior, as shown in Figure 2.6.
Five essential components are recognized for an MPC controller design [110]: 1) System
prediction model; 2) Cost indices; 3) Constraints; 4) Solving optimization problems; and
5) Receding horizon principle. Then a general MPC controller can be designed using these
five components as Algorithm 2.1.
Nominal stability and recursive feasibility properties of Algorithm 2.1 have been well
known since 2000 for specific cases when the system model, cost function, and constraints
satisfy certain conditions, see [36, 76, 77] and references therein. Generally, regulation or
stabilizing problems are considered. Properly defined cost functions and constraints are nec-
essarily present so that the “energy” of the system can be proved to dynamically decrease.
System states are then guaranteed to converge to desired states and the feasibility at a previ-
ous step immediately implies the next step feasibility [110]. Linear time-invariant systems
are extensively studied [77]. Properties of MPC for nonlinear systems in certain conditions
have also been well understood [36]. Various extensions of the basic MPC in Algorithm
2.1 exist including hybrid MPC [58], economic MPC [109], explicit MPC [8], distributed
MPC [71] and robust MPC [78]. Applications of MPC in practice, however, usually do not
satisfy the cost format and assumptions made for theoretical analysis. Moreover, designed
extra constraints that guarantee theoretical properties could probably degrade system per-
formance. Successful applications of MPC are widely seen in process industries [107],
intelligent cars [138, 145], power networks [90], intelligent transport systems [60], auto-
mated container terminals [143], and intelligent buildings [68] mostly without “stablizing”
ingredients.
Past Future
References
Predicted states
Measured states
Computed input sequence
k k +1 k + 2 ... k + Np
Prediction horizon
Figure 2.6: Illustration of MPC.
The first application of MPC to marine vehicles in literature is [137]. An MPC con-
troller based on nonlinear vessel dynamics is designed to track splines representing water-
ways. Computer simulations and experiments on inland vessels show the effectiveness of
the MPC controller. Recently, MPC has been applied to vessel path following [63] and
heading control [62]. In [93], the LOS guidance [28] is integrated in MPC which uses a
linearized model. When the vessel’s heading angle is controlled to converge to the angle
provided from the guidance module, cross-track errors can be proved to converge to zero
realizing path following. However, unrealistic assumptions that velocities and cross-track
errors should be small are made. A similar vessel path following problem is considered in
[98] combining MPC and an LOS guidance law. In both [98] and [93], reference paths are
specified by waypoints. Overshoots during switching of waypoints are observed due to the
use of non-predictive reference information. Few works consider hazardous area avoidance
or timing issues.
Networked systems
In general, system-wide control for networked systems can be approached in four ways
[89, 116]:
• centralized: there is a single controller taking care of the entire networked system;
• decentralized: there are multiple controllers and each controller solves a local sub-
system control problem using only local information, not relying on communication
between subsystems;
• distributed: there are multiple controllers and each controller solves a local subsystem
control problem, using also communicated information from other subsystems;
• hierarchical: controllers are working at different levels possibly with different time
scales.
For many applications, distributed approaches are preferable [71]. Particularly for mul-
tiple waterborne AGVs when coupled by collision avoidance constraints for safety, we ob-
serve six reasons for this preferability: 1) physically distributed by nature; 2) computational
efficiency; 3) limited communication range; 4) modularity for maintenance or expansion of
the system; 5) privacy issues if with different ownerships; and 6) robustness to local failures.
Solutions to distributed collision avoidance include priority [86], potential field [114],
velocity obstacles [54] and optimization [53] based methods. Conflicts disappear if subsys-
tems are prescribed relative priorities by certain rules (e.g., COLREGs) [86], but fixed rules
degrade system flexibility and optimality. Potential field approaches model conflicts as re-
pulsive forces [114] for which it is difficult to consider multiple objectives and constraints.
Velocity obstacles [54] usually assume constant velocities which might not hold in complex
situations. Optimization based approaches are largely embedded in distributed MPC [71]
considering the advantages of MPC as listed in Section 2.2.2.
In the literature, many distributed MPC approaches have been proposed with applica-
tions to various networked systems, e.g., power grids [90], aerial vehicles [53], intermodal
freight transport [60], and traffic networks [64]. Few applications of distributed MPC to
marine surface vehicles are seen in the literature to date. An overview of 35 different
distributed MPC approaches categorized by process, control architecture, and theoretical
properties is provided in [71]. Much of the research realizes distributed control by solving
local problems sequentially using communicated intent trajectories from coupled neighbor-
ing subsystems [22, 47, 50, 53, 59, 65, 111, 132, 133]. Assumptions are then made that
the deviations of communicated intent trajectories from actual trajectories are either small
[53, 111, 132, 133], or compatibility constraints [22], penalty functions [59], bounds [65]
are introduced to make the deviations small. With knowledge of the overall system dynam-
ics, each local controller solves a centralized problem in [50]. The coupling effects from
other subsystems are treated as bounded disturbances in [47]. Subsystems coupled via in-
puts are treated as agents in a cooperative game and distributed control is achieved following
game theory in [70]. However, generally in sequential approaches, when one subsystem is
computing, other subsystems are idle; and the order or priorities of computing agents still
matters.
In terms of control architecture, besides computing in sequential, distributed MPC can
also be achieved in parallel [90]. Parallel distributed MPC treats all agents equally and
usually iterative negotiations between subsystems are required before an overall agreement
is reached [26, 90, 95, 131, 148, 149]. For systems with special structures, the separability of
dual decomposition is exploited to realize parallelism in [33] with accelerated convergence
rate. For more general system couplings, [26] provides two distributed MPC solutions,
i.e., dual decomposition and the alternating direction method of multipliers (ADMM) [12]
which has better convergence properties than dual decomposition. Conventionally, ADMM
is formulated as a 2-block consensus problem of which proof of convergence has been
well established [12]. The 2-block consensus ADMM has been applied to the flocking
problem achieving near-centralized performance [125], communication networks to reduce
congestion [82], and networked road vehicles achieving the inner loop optimality in a two-
loop convex-concave procedure [95]. A large number of iterations are usually required
before convergence to a modest accuracy is obtained. Variants of ADMM, e.g., multi-
block schemes [81] and varying penalty parameters [39], show convergence in numerical
simulations for specific applications though with less rigorous convergence theorems.
Systems with uncertainties
System robustness against uncertainties with performance guarantees and constraint satis-
faction are dealt with in the literature in several ways. Besides systematically considering
system constraints and optimizing performance, MPC, under certain conditions, is inher-
ently robust by solving repetitively online optimization problems with new system outputs
[76]. However, this inherent robustness can only handle sufficiently small uncertainties
[49]. A more reliable approach is to have designed robustness. For bounded uncertainties,
an intuitive option is min-max MPC [47] minimizing a worst-case performance index; con-
servativeness and prohibitive computational time are the concerns. Alternatively, constraint
tightening based MPC [53, 78, 111, 131] has a comparable complexity as that of conven-
tional MPC by solving nominal optimization problems with tightened constraints. In [62]
for ship heading control in wave fields, disturbances are estimated and compensated in a
two-step MPC algorithm, assuming bounded estimation errors.
In practice, however, disturbances such as environmental forces acting on marine ves-
sels often bear stochastic characteristics [30], are not necessarily bounded, and even when
bounded, the bound is typically unknown. For unbounded uncertainties, small constraint
violations are necessary either via soft constraints [144] or stochastic MPC [52, 104]. Two
formulations of stochastic MPC are the expected case [104] and the chance constrained case
with a specified probability of constraint satisfaction [52]. However, similarly with bounded
uncertainties, this probability is still pre-designed. The trade-off between specified uncer-
tainty probabilities and system performance is investigated in an Antarctic krill catch level
control problem [43], showing in results from multiple simulations that increasing constraint
satisfaction probability leads to an exponential decrease of catch levels. Bounds of a subset
of uncertainties with high confidence are first determined with a scenario approach and then
utilized in a robust problem in [74]. The bounds are, however, still fixed in optimizations.
Regarding robustness in distributed MPC, efforts have been made on decomposing the
overall uncertain system based on distributed approaches reviewed in Section 2.2.2 and
solving local robust MPC problems based on robust approaches reviewed before. Bounded
coupling effects from other subsystems are considered in min-max local robust problems
in [47]. By assuming small deviations of communicated intent trajectories from actual
trajectories, [53, 111] solve local constraint tightened nominal problems serially. The intent
and actual trajectory deviations are explicitly penalized in distributed cost functions in [59]
and are explicitly bounded in [65]. External disturbances are then accommodated together
with the coupling penalty function and coupling bounds by robustness constraints in local
problems in [59] and [65], respectively. For linear systems with coupled state constraints,
different tube-based robust distributed MPC (RDMPC) problems have also been proposed,
e.g., single-update scheme [132], parallel-update scheme [131], and hierarchical control
schemes [113]. Most aforementioned approaches solve local robust problems serially except
for [131] which, however, loses cooperativeness. Note that iterative parallel approaches
[33, 148] have only been studied in deterministic cases.
2.2.3 Summary
There are several options of control techniques that could be applicable to motion control
problems in general. MPC turns out to be the most suitable for controlling waterborne
AGVs with special purposes in the context of transport and logistics. Extensive research
on both theoretical analysis and applications of MPC, distributed MPC, robust MPC, and
RDMPC has been done. More advanced techniques still need to be developed for the par-
ticular cases of a single waterborne AGV, networked waterborne AGVs with cooperative
distributed solutions, and waterborne AGVs with environmental disturbances.
2.3 Scheduling approaches

Scheduling in this thesis, as shown in Figure 1.2, refers to determining the reference infor-
mation for the lower level control problem. The reference information contains, for each
deployed vehicle, a sequence of routes to travel and chronological events that occur along
the routes. Approaches that are relevant for scheduling waterborne AGVs are reviewed in
this section.
Essentially, waterborne AGV scheduling for ITT is a pick-up and delivery problem
(PDP) [115] with time windows using capacitated vehicles. PDP is a generalization of a
VRP [130]. Both PDPs and VRPs involve finding a set of optimal routes for a fleet of
vehicles but differ in that PDP deals with transportation between distinct pick-up and de-
livery locations while in VRP, either the pick-up or the delivery location needs to be the
same, i.e., the depot. Within the operational research realm in a logistical context, it is
customary and sufficient to only care for setting schedules on discrete events. Details on
how these events really happen, i.e., the evolution of the lower level system dynamics, are
generally neglected. From a control point of view, however, vehicles concerned in VRPs
or PDPs are actually assumed as dimensionless mass points predominantly with constant
speeds such that any lower level feedback becomes irrelevant in a scheduling problem. We
observe that two variants of VRPs are exceptional. The time-dependent VRPs [27] adopt a
time-dependent speed model which, to some extent, considers lower level information, e.g.,
traffic congestion. But the speed is known a priori rather than being a decision variable that
could be manipulated. The time-dependent VRPs belong to a more generic class of dynamic
VRPs [106] dealing with dynamism such as online requests, dynamic travel times, etc. and
update route responsively. Solutions with acceptable quality and computational efficiency
are largely of concern for dynamic VRPs. Exact solutions are generally only applicable
to small networks [120]. Therefore, considerable research has been done on developing
heuristic methods to solve large network dynamic VRPs efficiently [57]. The second excep-
tion is the pollution-routing problem proposed in [7], which considers factors as load and
speed in producing “environmental-friendly” vehicle routes. The resulting problem is more
difficult to solve but yields lower load and speed dependent energy consumption cost. Still,
the combined route-speed optimization is open-loop and far from being able to consider
lower level complex dynamics.
In the maritime sector, the relation between marine vehicle speed and energy consump-
tion is highlighted even more by both practitioners and researchers. The engine of Maersk
“Triple-E” [69] was designed to sail relatively slowly to reduce 50% of the CO2 emitted
on the Asia and North Europe transport route. Another common practice in the shipping
industry known as “slow-steaming” [73] by cruising at a lower speed than the design speed
to reduce cost has also been widely accepted and implemented [105]. Arrival times are
optimized in [25, 92] to obtain optimal speeds along shipping routes. Results of applying
2.4 Inter terminal transport 23
the method to real shipping routes shows the potential for reducing environmental emis-
sions is substantial. Besides the emphasis on speed, coordination of arrival times of ships
at terminals to avoid unnecessary waiting or conflicts is more critical than for land-based
vehicles. The reasons are twofold. First, ships visit the same terminal more frequently
considering the limited pick-up and delivery locations. This is particularly the case in ITT.
In fact, most PDPs assume distinct pick-up and delivery locations and each vehicle visits
each location exactly once [115], which diminishes the arrival time coordination. Secondly,
loading/unloading of ships could take more time than land-based vehicles, and thus cannot
be neglected. Berthing time clash avoidance is modeled in [96] by constraining, for pick-up
and delivery visits sharing a same berth, the departure time of a visit not to be larger than
the arrival times of a later visit. This is problematic when extra time intervals are imposed
between departure and arrival times which is practically the case if ship dimensions and
safety distances are considered. Another characteristic of maritime logistics is that envi-
ronmental uncertainties are prevalent. These uncertainties include current, waves, wind and
encounters with other moving objects that not only interact with waterborne AGV dynamics
at the operational level but also influence the scheduling level. This calls for a closed-loop
system that makes decisions based on real-time feedback with tightly integrated scheduling
and control levels.
However, scheduling and control, typically as two distinct levels in a transportation
decision-making hierarchy, have been explored independently by researchers in the two
areas [61]. Although both levels largely rely on mathematical models and optimization
techniques and both aim at either maximizing profit or minimizing cost, the inherently dif-
ferent time-scale nature brings technical challenges for an integrated and computationally
tractable solution. On the one hand, discrete decisions involved in scheduling problems
restrict them to nothing but low dimension models solved in low frequency and off-line;
on the other hand, feedback and closed-loop operation in real-time are essential in control
systems to handle disturbances and complex dynamics. Efforts have been made either from
a “Top-down” perspective by considering control elements in a scheduling problem [142]
or from a “Bottom-up” perspective by including scheduling-oriented economic terms in the
cost function of a control problem [2]. In the field of process industry, the economic ben-
efits of integrating scheduling and control have been recently recognized and emphasized
[5]. A so-called “time scale-bridging” model is proposed in [20], but this model counts on
an explicit, low-order representation of the input/output process dynamics which is by all
means hard to derive for general systems. Moreover, operational constraints cannot be in-
corporated. A decent solution to integrated scheduling and control has to date not yet been
proposed.
Summarizing, the scheduling problem of waterborne AGVs can be formulated as a
PDP. Special issues for waterborne AGVs in port areas include emissions and coordinated
berthing times. A closed-loop scheme with more tightly integrated scheduling and control
could be expected to benefit the overall coordination performance of waterborne AGVs.
2.4 Inter terminal transport

Waterborne AGVs are especially proposed and designed for ITT to improve the port level
autonomy and efficiency in logistics. We give a brief introduction to ITT in this section.
ITT refers to the transportation of goods between terminals including dedicated auxil-
iary and value-added logistics service areas (e.g., inventory, packing, cargo bundling, re-
pairing, and cleaning) within a port [41]. ITT can be conducted either by land using trucks,
multi-trailer systems, railway systems, AGVs, and Autonomous lifting vehicles (ALVs) or
by sea using barges. Similar with general transport systems, the goal of an efficient ITT
system is to satisfy customer demands with minimal economical and environmental costs,
such as energy consumption, traffic congestion, and green-house emissions. However, since
ITT forms a complex network and a delay could adversely affect all following operations,
punctuality is deemed as the most important criterion in ITT [21].
At present, ITT, e.g., in the port of Rotterdam, is implemented mainly by means of
multi-trailer systems. Such systems use manned trucks, pulling trains of five trailers. The
performance of using multi-container yard trucks, AGVs, and ALVs for ITT is simulated
and compared [21] with real data from the port of Rotterdam. Automated vehicles (AGVs
and ALVs) are found to be superior to manned vehicles in terms of punctuality. In [129],
barges are also considered and integer programming models are proposed based on a time-
space graph to optimize and analyze ITT performance. Barges in an ITT system are shown
to be beneficial in terms of efficiency for delivering cargo and are even critical in instances
with a large number of containers. Hybrid transport modes for ITT are recommended by
the authors. Aiming at enhancing the management of real-time data and traffic flow of
ITT, [40] proposes a cloud-based information system. The system relies on advanced data-
collection and information exchange technologies and acts as a decision support system to
the port community. A chronological overview of approaches, methods, and contributions
in the area of ITT can be found in [41] and references therein. Research on ITT is there
categorized as simulation, optimization, information system approaches, and case studies.
Most of the work studies ITT scenarios in the port of Rotterdam, Maasvlakte I and II in
particular.
2.5 Conclusions
In this chapter, relevant literature in view of the development of waterborne AGVs is re-
viewed. Two existing intelligent vehicles, land-based AGVs and intelligent marine surface
vehicles, are discussed. The development trend, experiences, and technologies in using
AGVs instead of manned trucks in automated container terminals well motivate the use of
waterborne AGVs instead of manned barges in port areas. Research on intelligent marine
surface vehicles has been mostly targeted for military or research purposes; transportation
oriented design is rare. Safer, more sustainable, and efficient marine transport systems could
be expected if more intelligent waterborne transport vehicles are used.
Research related to the focus of this thesis, coordination of waterborne AGVs, have been
reviewed. There are several options of control techniques that could be applicable to motion
control problems in general. Considering the advantages of MPC and the characteristics
of waterborne AGVs with applications to ITT, MPC turns out to be the most suitable for
the motion control problems of waterborne AGVs. Extensive research on both theoretical
analysis and applications of MPC, distributed MPC, robust MPC, and RDMPC has been
done. Special techniques still need to be developed for the particular cases of a single
waterborne AGV, networked waterborne AGVs with cooperative distributed solutions, and
2.5 Conclusions 25
waterborne AGVs with environmental disturbances. Different scheduling approaches have

also been discussed. The scheduling problem of waterborne AGVs can be formulated as
a PDP. Special scheduling issues for waterborne AGVs in port areas are that emissions
and coordinated berthing times are critical. Based on the literature review in this chapter,
advanced coordinating strategies will be proposed for waterborne AGVs applied in typical
ITT scenarios in Chapters 3 - 7.
Chapter 3
Dynamic models of waterborne

AGVs
As discussed in Chapters 1 and 2, coordination strategies for waterborne AGVs are based
on proper knowledge of the system dynamics. In this chapter, two types of mathematical
models, i.e., simulation and prediction models, that describe dynamic waterborne AGV
behaviors are proposed. Section 3.1 models the scenario of Inter Terminal Transport (ITT)
using waterborne AGVs based on graphs. In Section 3.2, these models are approximated
using a successive linearization approach for predicting system trajectories. Section 3.3
concludes this chapter. The models presented in this chapter will be used in the sequel
throughout the thesis.
The research discussed in this chapter is partially based on [149, 150, 152].
3.1 Modeling of waterborne AGVs for ITT

In an ITT network using waterborne AGVs, a set of transportation tasks are assigned to a
fleet of deployed waterborne AGVs. Each waterborne AGV is required to load/unload a
certain number of containers at an origin terminal, depart at a specified time, travel along a
reference route, arrive at a specified terminal at a specified time, and load/unload a certain
number of containers. Waterborne AGVs are equipped on board with processing, measure-
ment, and communication devices to measure their own system states, communicate with
other waterborne AGVs within a certain range, and process locally. When waterborne AGVs
will not collide over a finite number of predictable steps in any case, they maneuver inde-
pendently to fulfill their assigned ITT tasks. However, when several waterborne AGVs are
involved in a neighborhood, negotiations and alterations of trajectories to avoid collisions
are typically necessary for ensuring the overall safety. In this section, waterborne AGVs for
ITT are modeled using graphs [18] G (k) = (V (k), E (k)) where k is the discrete time step
and relates to continuous time t as t = kTs with Ts as the sampling time. The graph vertices
V (k) representing waterborne AGVs are modeled in Section 3.1.1, and the graph edges
E (k) representing couplings between waterborne AGVs are modeled in Section 3.1.2.
27
28 3 Dynamic models of waterborne AGVs
3.1.1 Waterborne AGV model

This section models the vertices V (k) in graph G (k), i.e., waterborne AGV dynamic mod-
els. Models of marine crafts with different degrees of freedom (DOFs) for different purposes
have been elaborated on in [30]. Generally, a marine surface vehicle experiences motions in
six DOFs, as shown in Figure 3.1a. For tracking problems of surface vehicles, models with
three DOFs in the horizontal plane, as shown in Figure 3.1b, are sufficient to capture the
main system characteristics [30] based on the assumption that the roll and pitch motions are
small. Waterborne AGVs are modeled based on a three DOF maneuvering model in [30].
Consider that a set V (k) = {1, 2, ..., n(k)} of n(k) numbered homogeneous waterborne
AGVs and for each waterborne AGV p ∈ V (k), the dynamics that consider environmental
disturbances due to wind, waves, and current are modeled as:
η̇ R(ψ p (t))νν p (t),

η p (t) =R (3.1)
−1 T T
M RB + M A ) (ττ p (t) + R (ψ p (t))R
ν̇ν p (t) =(M R (ψb (t))b(t)

− C p,RB (t) +C
C p,A (t) ν p (t) + (D
DL + D p,NL (t)) ν p (t)), (3.2)
T
η pT ν pT

where and τ p are system states and inputs, respectively, and
     
xp up τu,p
η p =  y p , ν p =  v p , τ p =  τv,p ,
ψp rp τr,p
where x p , y p and ψ p are coordinates, and heading angle, respectively, in the inertial coor-
dinate {n}. Linear velocities in surge and sway are expressed in the body-fixed coordinate
{b p } of waterborne AGV p as u p and v p , respectively, and the angular velocity of the head-
ing angle is expressed by r p . The control input vector for waterborne AGV p consists of
surge force, sway force, and yaw moment represented by τu,p , τv,p and τr,p , respectively.
Due to physical limitations such as maximum speeds, maximum engine power, etc., con-
straints on states and control inputs are usually imposed as:
ν p min 6 ν p 6 ν p max , τ p min 6 τ p 6 τ p max . (3.3)
In (3.1), R (ψ p ) is a rotation matrix relating motions in {n} and {b p }, defined as:

 
cos(ψ p ) − sin(ψ p ) 0
R (ψ p ) =  sin(ψ p ) cos(ψ p ) 0  .
0 0 1
Environmental disturbances due to wind, waves, and current are modeled as a non-rotational
force b with angle ψb in {n}. The effects of b along three DOFs are mapped by the rotation
T
vector R T (ψb (t)) = cos(ψb (t)) sin(ψb (t)) 0

assuming that the force is acting on
the gravity center. Disturbance effects in {n} are further rotated to {b p } to account for the
force changes with the heading of waterborne AGV p. Note that since all the waterborne
AGVs in a port area experience the same environmental disturbances predicted by the port
authority, the disturbance related parameters are without subscript • p .
3.1 Modeling of waterborne AGVs for ITT 29
xn
winds
On f (roll) xb
(surge)
yn
Ob waves
zn
q (pitch)
(sway)
currents y (yaw) yb
(heave)
zb Rotational and
translational motions
(a) Marine surface vehicle in six DOFs.
Yn
X
b
Y b
u
v
y y
O
r b
Communication
Sensor
Processing
X
O n
n
x
(b) Waterborne AGV in three DOFs.
Figure 3.1: Modeling waterborne AGV dynamics

Rigid-body and added mass matrices are the same for all the homogeneous waterborne
AGVs and are given as:
   
m 0 0 −Xu̇ 0 0
M RB =  0 m mxg  , M A =  0 −Yv̇ −Yṙ  ,
0 mxg Iz 0 −Nv̇ −Nṙ
where subscripts ·RB and ·A stand for rigid body and added force related matrices, respec-
tively; m is the mass of waterborne AGVs1 ; Iz is the moment of inertia in the yaw rotation;
and xg is the distance between the center of gravity of waterborne AGVs to the center of
{b p }. Similarly,
 
0 0 −m xg r p + v p
C p,RB (t) =  0 0 mu p ,
m xg r p + v p −mu p 0
 
0 0 Yv̇ v p + (Nv̇ +Yṙ ) r p /2
C p,A (t) =  0 0 −Xu̇ u p 
−Yv̇ v p − (Nv̇ +Yṙ ) r p /2 Xu̇ u p 0
are rigid-body, and added Coriolis and centripetal matrices of waterborne AGV p, respec-
tively.
Damping forces are separated into two parts: a linear part as
 
−Xu 0 0
DL =  0 −Yv −Yr  ,
0 −Nv −Nr
and a nonlinear part as
−X|u|u |u p | − Xuuu u p 2
 
0 0
D p,NL (t) =  0 −Y|v|v |v p | −Y|r|v |r p | −Y|v|r |v p | −Y|r|r |r p |  .
0 −N|v|v |v p | − N|r|v |r p | −N|v|r |v p | − N|r|r |r p |
Hydrodynamic derivatives follow the notations in [123]. For instance, the hydrodynamic
added mass force X along the x axis due to an acceleration u̇ in the x direction is written as
∂X
X = −Xu̇ u̇, Xu̇ := ,
∂u̇
which implies {MM A }11 = −Xu̇ . Readers are referred to [123] for more details.
Depending on whether environmental disturbances are perfectly known beforehand or
not, models (3.1) – (3.2) are further discussed as follows.
Waterborne AGV models with perfectly known disturbances

In a deterministic case, environmental disturbances are known perfectly beforehand. Partic-
ularly, we consider waterborne AGVs with known constant current which is non-rotational,
1 For the controller design problems in Chapters 4 – 6, the empty waterborne AGV mass m is considered
without any containers on board. In Chapter 7, the waterborne AGV mass is considered as the sum of the empty
waterborne AGV mass m and the mass of all containers on board.
and has fixed speed Vc and angle βc in {n} (V˙c = 0 and β˙ c = 0). A rotation to {b p } is
 
Vc cos βc
ν p,c = R T (ψ p )  Vc sin βc  . (3.4)
0
The influences of current are then expressed by the relative velocity in {b p } between the wa-
T
terborne AGV hull and the fluid as ν p,r (t) = ν p (t) − ν c,p (t) = u p,r (t) v p,r (t) r(t) ,
and satisfy:
C p,RB (t)νν p (t) + (C

M RB ν̇ν p (t) + M A ν̇ν p,r (t) +C C p,A (νν p,r ) + D L + D p,NL (νν p,r )) ν p,r (t) = τ p .
d
Since R satisfies dt {R
R(ψ p )} = r p R(ψ p )SS with
 
0 −1 0
S= 1 0 0 ,
0 0 0
we have ν̇ν p,r = ν̇ν p − r p S T ν p,c . Then, (3.2) can be rewritten as:
M RB + M A )−1 (ττ p (t) − (C

ν̇ν p (t) =(M C p,A (νν p,r ) + D L + D p,NL (νν p,r )) ν p,r (t)
C RB (t)νν p (t) + M A r p (t)SS T ν c,p (t) ,

−C (3.5)
where the added Coriolis and centripetal matrix depends on ν p,r as

 
0 0 Yv̇ v p,r + (Nv̇ +Yṙ ) r p /2
C p,A (vvr ) =  0 0 −Xu̇ u p,r ,
−Yv̇ v p,r − (Nv̇ +Yṙ ) r p /2 Xu̇ u p,r 0
and the nonlinear damping matrix depends on ν p,r as
−X|u|u |u p,r | − Xuuu u p,r 2

 
0 0
D p,NL (ννr ) =  0 −Y|v|v |v p,r | −Y|r|v |r p | −Y|v|r |v p,r | −Y|r|r |r p |  .
0 −N|v|v |v p,r | − N|r|v |r p | −N|v|r |v p,r | − N|r|r |r p |
Waterborne AGV models with not perfectly known disturbances

Generally, waterborne AGVs travel in good weather based on roughly predicted values of
b and ψb . For example, the port authority cooperates with local water management orga-
nizations and installs hydrological and meteorological sensors at different locations in the
port of Rotterdam to provide information on visibility, tides, flow rates, wave heights, wind
speeds and directions [103]. However, predictions based on weather forecast and sensed in-
formation are mostly stochastically uncertain. Therefore, it is assumed that environmental
disturbances b and ψb could in principle be obtained from the port authority’s meteoro-
logical predictions, but with the existence of prediction uncertainties following a normal
distribution on the magnitude of disturbance forces b as, i.e.,
b ∼ N(b̄, Σ), (3.6)

where the mean b̄ is predicted by the port authority and the covariance Σ reflects the pre-
diction accuracy. Similar use and assumptions of weather prediction information have been
made for energy-efficient building systems [94].
Summarizing, waterborne AGVs p ∈ V maneuver independently to fulfill assigned ITT
tasks when far away. The dynamics are defined as (3.1), (3.5) when environmental distur-
bances are perfectly known, and as (3.1) – (3.2), (3.6) when environmental disturbances are
not perfectly known. In both cases, system physical limits (3.3) need to be satisfied.
3.1.2 Network model

This section models the edges E (k) in graph G (k), i.e., the couplings between waterborne
AGVs. When waterborne AGVs are within each other’s communication range Dc , shown
as the black dotted circle in Figure 3.2a, couplings arise as pairwise collision avoidance
constraints to enforce a minimal safety distance Ds (red dashed circle). Note that in real
implementations, Dc and Ds that are related to the ship domain [100] are usually not circular.
However, the determination of the shape and size of a ship domain is out of the scope of this
thesis. The algorithms designed in the latter chapters based on circles with constant radii
are for simplicity and are also applicable to cases where non-circular ship domains with
varying sizes are involved. Based on Figure 3.2a, we first define a subset Vw (k) ⊆ V of
nw (k) working waterborne AGVs. A waterborne AGV p ∈ Vw (k) if and only if its assigned
departure time has been reached and it has not arrived at its assigned destination yet, i.e.,
t > t p , kr p (k) − d p k2 > dtol , (3.7)

T
where k·k2 denotes the two-norm Euclidian distance and r p (k) = x p (k) y p (k) is the
measured position of waterborne AGVs p; t p and d p are the assigned departure time and
destination, respectively; dtol > 0 is a smalltolerance.
The edge set is then defined as E (k) = e p,q (k)|p, q ∈ Vw (k), p < q with
(
1, if d p,q (k) 6 Dc ,
e p,q (k) =
0, otherwise,
where Dc = 2umax Ts Np + l + α with prediction horizon Np ; umax is the maximum surge

speed of waterborne AGVs; l is one waterborne AGV length; α is the margin for uncertain-
ties; and d p,q (k) = r p (k) − r q (k) 2 is the Euclidean distance between waterborne AGV
p and q. Collision avoidance couplings are then imposed to pairwise waterborne AGVs

(p, q) |e p,q (k) = 1, e p,q (k) ∈ E (k) as:
d p,q (k) = r p (k) − r q (k) 2

> Ds . (3.8)
The above logic implies that collision avoidance couplings emerge when there is a possibil-
ity of collision over the next prediction horizon if two waterborne AGVs are sailing at their
maximum speeds. This is a relatively conservative logic but guarantees safety if proper ac-
tions are taken. Coupling constraints (3.8) hinder independent computations of waterborne
AGVs.
It is usually the case that we have groups of waterborne AGVs that are coupled within
a group but are decoupled between groups, as Figure 3.2b shows. In this case, G (k) is
Wind
n
wAGV r
D c
D s Currents
wAGV q
wAGV p
Waves
Container
terminals
(a) Waterborne AGV couplings; communication (black dotted) and safety (red
dashed) ranges.
AG erbo
W
at
V rn
6
#7 e
V# e
AG rborn
Port te
Wa
authority
10
V # ne
AG rbor
AG erb
W
at
e
V orn
t
Wa
#4 e
#9 e
V rn
AG erbo
e Sub-network #3
rborn at
Wate #3 W
V
Waterborne
A G
AGV #5
Sub-network #1
W
e
rn
Wa GV #
bo
er #2
ter 8
Wat
A
at
erb W AGV
bo
AGV orne
rne
#1
Sub-network #4 Sub-network #2
Vessel Sensor Processor Communication Containers
(b) Groups of waterborne AGVs: coupled intra groups and decoupled inter
groups.
Figure 3.2: Waterborne AVG networks for ITT.

disconnected [18]. Connected components that are subgraphs of G (k) can be computed us-
ing algorithms like breadth-first search [45] based on E (k). Let nG (k) subgraphs Gs (k) =
(Vs (k), Es (k)) for s = 1, 2, .., nG (k) be derived, where Vs (k) is the subset of ns (k) water-
borne AGVs and Es (k) defines the set of collision avoidance coupling pairs in Gs (k). Obvi-
ously, Vs (k) ⊆ Vw (k) and 1 6 ns (k) = Vs (k) 6 nw (k) 6 n(k). All the numbered working
nG (k)
waterborne AGVs appear once and only once in one of the subgraphs, i.e., ∪s=1 Vs (k) =
nG (k)
Gn (k)
Vw (k), ∩s=1 Vs (k) = 0/ and ∑ ns (k) = nw (k). At each time step k, subgraphs Gs (k) =
s=1
(Vs (k), Es (k)) are constructed and assumed to be constant over the next prediction horizon.
Cases of Gs (k) include:
1. Subgraphs Gs (k), s = 1, ..., nG (k) are singleton graphs when ns (k) = 1 and Es (k) = 0 ;
2. There are multiple subgraphs Gs (k) with Es (k) 6= 0 and ns (k) > 1; and
3. There is only one subgraph Gs (k) in G (k) when Gs (k) = G (k) and ns (k) = nw (k).
The first case corresponds to the problem for one single waterborne AGV considered
in Chapter 4. The second case corresponds to the problem for multiple waterborne AGVs
with multiple couplings that are considered in Chapters 5 and 7. The third case corresponds
to the problem for multiple waterborne AGVs considered in Chapter 6. All these problems
can be modeled and will be solved based on the above constructed graph structures.
3.2 Successively linearized models

As discussed in Chapter 2, one of the issues concerning model predictive control (MPC)
is that optimization problems need to be solved online repetitively. The complexity of an
MPC controller with certain controller parameters mainly depends on the characteristics
(order, nonlinearities) of the prediction model it uses and imposed system constraints. For
highly nonlinear systems with nonlinear constraints, a dilemma is usually faced by MPC.
On the one hand, stringent demands on system performance generally require a high ac-
curacy prediction model. On the other hand, online prediction and optimization of future
system behavior based on complex prediction models is too time consuming, which is un-
desirable for real-time applications. A compromise between optimality and computational
complexity has to be made for fast and nonlinear waterborne AGV dynamics.
Dynamic models of waterborne AGVs as presented in Section 3.1 with nonlinear dy-
namics and non-convex collision avoidance constraints are not preferable directly for con-
troller design. The reasons are twofold: 1) they are computationally too complex for real-
time control; and 2) distributed and robust control design are difficult if not impossible. A
comparison on performance of using nonlinear and single step linearized models of water-
borne AGVs in MPC has been made in [146]. It was found that computational time is much
longer using nonlinear prediction models than using linearized models. In this section, we
introduce a successive linearization approach and derive approximated dynamic waterborne
AGV models and constraints for MPC controller design. Details on MPC controllers will
be presented in the later chapters.
3.2 Successively linearized models 35
The basic idea is to utilize the whole sequence of control inputs from a previous MPC
step and pre-calculate a shifted system trajectory for linearizations over all prediction steps
at the next MPC step. We generalize the dynamics (3.1) – (3.2) of waterborne AGV p as:
ẋx p (t) = f (xx p (t), u p (t), b(t)), (3.9)
where f : R6 × R3 × R → R6 is a nonlinear smooth function with system states x p =
T T
η p ν Tp and control inputs u p = τ p . For numerical simulations, the continuous time
model (3.9) is discretized with the zero-order-hold assumption as:
Z (k+1)Ts
x p (k + 1) = x p (k) + f (xx p (k), u p (k), b(k))dt. (3.10)
kTs
Then at each time step k, for each waterborne AGV p ∈ V (k), successive linearizations are
implemented as the following three steps:
1. Obtain seed [51] input trajectory u 0p (i|k) whereby, (i|k) denotes the ith prediction step
at time step k, and the superscript •0 denotes seed trajectories. Whenever contextually
clear, prediction step i for control/disturbance inputs is over 0, 1, ..., Np − 1 and for states
over 0, 1, ..., Np . Consider the previous time step k − 1 (k > 1), the calculated optimal
control input sequence is u p (i|k − 1). Conventionally, the first element u p (0|k − 1) is
applied to the system and the rest are disregarded. For linearizations at step k, we make
extensive use of the “tail” which is shifted as:
u 0p (i|k) = u p (i|k − 1) (3.11)
for i = 0, 1, ..., Np − 2 and
u 0p (Np − 1|k) = u p (Np − 1|k − 1). (3.12)
2. Obtain seed state trajectory x 0p (i|k). With an initial state x 0p (k|k) = x p (k) and u 0p (i|k),
apply u 0p (i|k) to (3.10) iteratively or to (3.9) using available ordinary differential equa-
tion solvers (e.g., in MATLAB [75]) which provide higher precision than (3.10). This is
straightforward when disturbances are perfectly known. However, when disturbances are
not perfectly known as modeled in (3.6), the values of b(i|k) remain unknown. There-
fore, we define a seed disturbance input trajectory as b0 (i|k) = b̄(i|k) by utilizing the
predicted mean values from the port authority. One of the justifications for this defi-
nition is that deviations of the real value b(i|k) from b0 (i|k) (or b̄(i|k)) are small for a
high probability, and small deviations satisfy the assumption of Jacobian linearizations
conducted at the next step. Besides, the widely known extended Kalman filter [48] also
conducts linearizations of nonlinear systems about mean values of random variables.
3. Linearize nonlinear dynamics and constraints at seed trajectory x 0p (i|k), u 0p (i|k) in a

deterministic case or x 0p (i|k), u 0p (i|k), b0 (i|k) in an uncertain case. Define small per-

turbations around the seed trajectory as (∆xx p (i|k), ∆uu p (i|k), ∆b(i|k)) (∆b(i|k) = 0 in a
deterministic case) which satisfy:
x p (i|k) = x 0p (i|k) + ∆xx p (i|k), (3.13)
u p (i|k) = u 0p (i|k) + ∆uu p (i|k), (3.14)
0
b(i|k) = b (i|k) + ∆b(i|k). (3.15)
Substituting (3.13) – (3.15) into (3.10), we get
x 0p (i + 1|k) + ∆xx p (i + 1|k) = x 0p (i|k) + ∆xx p (i|k)+

Z (k+1)Ts
f x 0p (i|k) + ∆xx p (i|k), u 0p (i|k) + ∆uu p (i|k), b0 (i|k) + ∆b(i|k) dt.

(3.16)
kTs
The integrator term, by applying Taylor’s theorem and neglecting the higher order terms
than the first order, is approximated as:
Z (k+1)Ts
f x 0p (i|k), u 0p (i|k), b0 (i|k) dt+

kTs
Z (k+1)Ts
A cp (i|k)∆xx p (i|k) + B cp (i|k)∆uu p (i|k) + E cp (i|k)∆b(i|k)dt (3.17)
kTs
where
∂f
A cp (i|k) = ,
∂xx (x 0p (i|k),uu0p (i|k),b0 (i|k))
∂f
B cp (i|k) = ,
∂uu (x 0p (i|k),uu0p (i|k),b0 (i|k))
∂f
E cp (i|k) = ,
∂b (x 0p (i|k),uu0p (i|k),b0 (i|k))
are continuous Jacobian state, input, and disturbance matrices, respectively. Then by
(3.16) and (3.17), we reach the discrete linearized incremental model
∆xx p (i + 1|k) = A dp (i|k)∆xx p (i|k) + B dp (i|k)∆uu p (i|k) + E dp (i|k)∆b(i|k), (3.18)
where A dp (i|k), B dp (i|k), and E dp (i|k) are corresponding discrete Jacobian matrices.
In a similar way, non-convex collision avoidance constraints (3.8) are approximated as:
d 0p,q (i|k) +C
C (i|k)∆rr p (i|k) + D (i|k)∆rr q (i|k) > Ds , (3.19)
where C (i|k) ∈ R 1×2 and D (i|k)∈ R 1×2 are Jacobian matrices of function d with respect
to r p and r q evaluated at r 0p , r 0q , respectively.
Time-varying but linear dynamic models (3.13) – (3.15), (3.18) and convex constraints
(3.19) are then used to approximate original nonlinear dynamics (3.1) – (3.2) and non-
convex collision avoidance constraints (3.8) for later controller design.
3.3 Conclusions
In this chapter, we have presented two types of dynamic models that describe the waterborne
AGV behavior. The first type is for simulation purposes and models waterborne AGVs for
ITT as graphs. Graph vertices represent waterborne AGV nonlinear dynamics and graph
edges represent couplings between waterborne AGVs. These models are able to represent
3.3 Conclusions 37
all the typical ITT scenarios considered in the later chapters. However, waterborne AGV
nonlinear dynamics and non-convex couplings are too complex for real-time applications
and controller design. The second type approximates the first type using successive lin-
earization in the framework of MPC. Linearized models could be beneficial by providing a
trade-off among optimality, computational performance, and the ease of controller design.
The waterborne AGV models used in Chapters 4 – 7 are all based on models presented in
this chapter.
Chapter 4
Predictive path following with

arrival time awareness
Regarding the Key Research Question on performance criteria for one waterborne AGV in
Chapter 1, this chapter considers a single waterborne AGV controlled to carry out an as-
signed Inter Terminal Transport (ITT) task. A predictive path following with arrival time
awareness controller is proposed based on the waterborne AGV dynamic models in Chap-
ter 3. The controller is also able to handle static obstacles and avoid overshoots during
switching of reference path segments.
The research discussed in this chapter has been published in [150].
4.1 Introduction
Currently, container movements are handled by land-based AGVs inside container terminals
and by manned trucks between terminals. Waterborne AGVs are proposed as an alternative
and innovative way for ITT with advantages presented in Chapter 1. A fundamental sce-
nario in which a single waterborne AGV autonomously fulfills one assigned ITT task in a
deterministic case is considered in this chapter. The waterborne AGV departs from an ori-
gin terminal, arrives at a destination terminal at specified times, and moves along a specified
route that has been designed connecting these two terminals over water. The reference route
is determined by waypoints and consists of straight line segments. In navigation and guid-
ance systems for aircraft, waypoints are also often necessarily available to generate control
references that are able to handle overshoot issues [56]. For marine vehicles, the refer-
ence generation approach, Line-of-Sight (LOS) [28], is usually applied in tracking routes
defined by waypoints. However, overshoots during switching of line segments are seen in
LOS based approaches. Besides, since a low rate of “non-performance” which happens
when delays exist, is the most important criterion of ITT, the geometric time-independent
reference route needs to be tracked while keeping the arrival time in mind. Moreover, the
dynamics of a waterborne AGV are typically constrained multi-input multi-output systems,
as seen in Chapter 3. Few control techniques can handle timing, overshoots, system con-
straints, and optimizing system performance quantitatively in a systematic way, as discussed
39
40 4 Predictive path following with arrival time awareness
in Chapter 2.
In this chapter, based on the literature review in Chapter 2 and the models presented in
Chapter 3, a predictive path following with arrival time awareness (PPF-ATA) controller is
proposed based on model predictive control (MPC) that achieves for a single waterborne
AGV that:
1. a given geometric reference path is tracked with deviations as small as possible;
2. a given arrival time requirement is met when a preferable time is feasible considering
system limitations, or a minimal delay with respect to the preferable time within a
specified time window otherwise; and
3. the aforementioned two design requirements are achieved in a speed-dependent en-

ergy economical way.
In particular, connected coordinate systems are established in which system kinematics are
re-modeled. The benefits of doing so are twofold: first, tracking errors can be formulated
more compactly; secondly, the along-track state is utilized in a reference switching logic
so that overshoots are avoided. The switching logic combined with a coordinate transfor-
mation renders a continuous model in one coordinate system still applicable for successive
linearizations. Moreover, a two-level double integrator model for parameterizing reference
paths is proposed to achieve smooth tracking and arrival time awareness. The lower level
is embedded in online MPC optimizations for smooth tracking. The higher level solves an
optimal control problem considering distance-to-go and time-to-go each time step. Simula-
tion results of two industrially relevant ITT case studies in the port of Rotterdam illustrate
the effectiveness of the proposed model and control design for a waterborne AGV.
The remainder of this chapter is organized as follows. In Section 4.2, the connected path
coordinate systems are established in which waterborne AGV kinematics are re-modeled for
path following. The two-level double integrator model for path parameterization is proposed
in Section 4.3. Section 4.4 introduces a switching logic with binary decision variables and
coordinate transformations. The PPF-ATA algorithm is summarized in Section 4.5. Then in
Section 4.6, simulation experiments and results are presented, followed by the concluding
remarks of this chapter in Section 4.7.
4.2 Modeling in path coordinate systems

For controlling the waterborne AGV to move along the reference route of straight line seg-
ments, two system design requirements are first distinguished here: minimizing the cross-
track error which is defined as the distance from the waterborne AGV’s current position to
the reference line, and minimizing the along-track error, defined as the error between the
orthogonal projection point of the waterborne AGV on the reference line and an reference
along path point. Besides the inertial coordinate system {n} and the body-fixed coordinate
system {b p } for modeling waterborne AGVs as in Chapter 3, connected path coordinate
systems are established as shown in Figure 4.1. By re-modeling waterborne AGV kinemat-
ics (3.1) in path coordinate systems, the cross-track and along-track errors are conveniently
formulated as linear functions of system states. Apart from that, the along-track state can
be taken advantage of in a switching logic to avoid overshoots.
4.2 Modeling in path coordinate systems 41
Yn
+1
pj
X
1
lj +
j +1
p
Ypj
Y
y j +1
Y p1 X p j Op (wpt j +1 )
lj j +1
y yj
l1 Op j (wpt j ) Xb
Op (w y1 X p
1 pt1 ) 1
u
Yb v y
Ob
r
Currents
Xn
On x
Figure 4.1: Re-modeling waterborne AGVs in path coordinate systems.
In Figure 4.1, the path coordinate systems {p j } ( j = 1, 2, ...) are based on connected
reference paths; Xp j is along the reference path and Yp j is vertical to the reference path
pointing π/2 counterclockwise; Op j is the origin of the jth path coordinate system located
at the jth waypoint connecting reference path j − 1 and j. Lengths and angles with respect
to Xn of reference path j are denoted as l j and ψ j , respectively. Then kinematics (3.1) 1 are
modeled in path coordinate system {p j } as:
ηp j (t) = R (ψp j (t))νν(t),

η̇ (4.1)
T
where, likewise, η p j = xp j yp j ψp j is the pose expressed in {p j } with ψp j = ψ −
ψ j . R (ψp j ) is a rotation matrix relating motions between coordinate systems {p j } and {b}
and defined as:
cos(ψp j ) − sin(ψp j ) 0
 
R (ψp j (t)) =  sin(ψp j ) cos(ψp j ) 0  .

0 0 1
Kinetics of the waterborne AGV system are still expressed in frame {b} as (3.2). Since
waterborne AGV heading angle ψ is involved in (3.2) to transform current dynamics from
h iT
{n} to {b}, a state vector in R 7 is defined as η Tpj ψ νT .
The new kinematics (4.1) are continuous within one coordinate system. However, the
continuity is lost during the switch of coordinate systems. A transformation including ro-
tation and translation of coordinates is then necessary. More specifically, we consider a
switch from {p j } to {p j+1 }, as shown in Figure 4.1. The angle error between the new ‘x’
axis Xp j+1 and the old ‘x’ axis Xp j is ψ j+1 − ψ j . Since {p j } and {p j+1 } are connected,
1 Since only a single waterborne AGV is considered, the subscript • p indicating waterborne AGV p is dropped
in this chapter. A normal font subscript •p indicates variables in the path coordinate systems.
the new origin Op j+1 has coordinates (l j , 0) relative to the old coordinate system {p j }. A
transformation of coordinates from {p j } to {p j+1 } would then be:
xp j − l j

xp j+1 cos(ψ j+1 − ψ j ) sin(ψ j+1 − ψ j )
= . (4.2)
yp j+1 −sin(ψ j+1 − ψ j ) cos(ψ j+1 − ψ j ) yp j − 0
Meanwhile, a transformation of the heading angle from {p j } to {p j+1 } would be:
ψp j+1 = ψp j + ψ j − ψ j+1 , (4.3)
or simply as:
ψp j+1 = ψ − ψ j+1 . (4.4)
To obtain initial path coordinate states, measured states η need to be transformed from
{n} to {p j }. Similar with (4.2), a transformation of position is:
x − xwpt j

xp j cos(ψ j ) sin(ψ j )
= , (4.5)
yp j −sin(ψ j ) cos(ψ j ) y − ywpt j

where xwpt j , ywpt j is the coordinate of waypoint j, or origin of {p j } in {n}. The transfor-
mation of heading angle is then the same with (4.4). So far, we have re-modeled kinematics
(4.1) with coordinate transformations of (4.2) and (4.4) for transforming coordinates be-
tween {p j } and {p j+1 }, and (4.4) and (4.5) for transforming coordinates between {n} and
{p j }.
4.3 Two-level double integrator dynamics

For an ITT task, the reference path is given as a sequence of time-independent straight-
line segments, i.e., a path following rather than a trajectory tracking problem needs to be
solved. To obtain smoothly tractable along-track references and at the same time satisfying
the timing requirements, the geometric path is parameterized by double-integrator dynamics
modeled at two levels, as shown in Figure 4.2. The lower level is embedded in online MPC
optimizations. The higher level solves an optimal control problem considering distance-to-
go and time-to-go which are fed back from the lower level double-integrator dynamics and
provides the lower level timing aware references over the next receding prediction horizon.
The double integrator dynamics are modeled as:
x s (k + 1) = A s x s (k) + B s us (k), (4.6)

T
where x s = s vs and us = as . The scalar s is introduced as the traveled distance of
the waterborne AGV along the straight-line reference paths with vs as its velocity and as as
its acceleration. State and input matrices are:
2
1 Ts Ts /2
As = , Bs = .
0 1 Ts
Remark : Since s, vs and as are all along the reference paths with one DOF, they are mod-
eled as continuous regardless of the two DOF coordinate systems, e.g., {n} or {p j }. 2
4.3 Two-level double integrator dynamics 43
final time step N and N max

final state [ st 0]T
f
Upper level dynamics

(MIQP)
current time step k

s r (i | k )
current state xs,l ( k )
Lower level dynamics

(MPC online optimization)
Figure 4.2: Two-level double integrator dynamics.
To distinguish notations in the two levels, we use subscript ·s,l denoting lower level
variables and ·s,h indicating higher level variables. In the lower level, given an initial state
x s,l (k), predicted trajectories over a prediction horizon are:
x s,l (i + 1|k) = A s x s,l (i|k) + B s us,l (i|k), (4.7)
for i = 0, 1, ..., Np − 1 with x s,l (0|k) = x s,l (k). Prediction model (4.7) is embedded in online
MPC optimizations that are to be formulated in Section 4.5.
The higher level shares the same double integrator dynamics (4.6) with the lower level.
An MIQP problem is formulated aiming at generating an optimal reference trajectory for
the lower level over the next receding prediction horizon, i.e., sr (i|k), for i = 1, 2, ..., Np .
For the MIQP, we specify main objective as guaranteeing a required arrival at stf at time tf .
Terminal state stf is set as the total length of all the path segments.
Considering limitations of waterborne AGV dynamics, the double integrator dynamics
cannot evolve freely either. A waterborne AGV’s maximum surge speed umax is imposed
as a state constraint for vs,u in MIQP. Due to this speed limit, there could be a feasibility
issue for a specific ITT task: if the scheduled arrival time is too stringent, the waterborne
AGV could not be able to arrive on time even if it sails at its highest speed. In reality, a
time window is often assigned to allow for an acceptable delay ∆t in terms of a preferable
arrival time tr . Finite flexibility is thus set for the arrival time by tf ∈ tf ,tf , where tf = tr
and tf = tr + ∆t and we assume by tf , the arrival can by all means be achieved. In this case,
the problem becomes a constrained optimal control problem with a fixed terminal state
and a minimal arrival time [31]. However, the minimal arrival time should be within the
time window tf ,tf . Next, we show how this can be implemented in MIQP using binary
variables.
In a discrete time setting, we denote Tf (k) as the calculated arrival time step at time step
k, N and Nmax corresponding to continuous time tf and tf , respectively. Therefore, Tf (k), N
and Nmax satisfy Tf (k) ∈ [N, Nmax ]. The cost function regarding the energy and arrival time
is separated into two parts:
Js (k) = Js1 (k) + Js2 (k), (4.8)
where Js1 (k) is written as a summation from the current time step k to time step N − 1, i.e.,
N−1
Js1 (k) = ∑ kus,h (n|k)k2w1 + kxxs,h (n|k)k2w2 (4.9)
n=k
subject to, for n = k, k + 1, ..., N − 1
x s,h (n + 1|k) = A s x s,h (n|k) + B s us,h (n|k), (4.10)

0 ≤ x s,h (n) ≤ umax . (4.11)
Notation k·k2w stands for weighted vector two norms, e.g., kus,h (n|k)k2w1 = us,h (n|k)T w1 us,h (n|k).
Minimization of the two norms of us,h (k) and x s,h (k) aims at optimizing energy efficiency
and smoothness of s dynamics. Symbols w1 and w2 represent the weighting parameter and
matrix for us,h (k) and x s,h (k), respectively.
The second part of Js (k), Js2 (k), is a summation over [N, Nmax ], defined as
Nmax
Js2 (k) = ∑ w3 nbb(n − N + 1|k) + kus,h (n|k)k2w1 + kxxs,h (n|k)k2w 2 , (4.12)
n=N
where b(n − N + 1|k) for n = N, N + 1, ..., Nmax are binary decision variables satisfying
(
1, for Tf (k) = n,
b (n − N + 1|k) =
0, otherwise.
and
Nmax
∑ b (n − N + 1|k) = 1 (4.13)
n=N
to ensure one arrival time step is selected. This selected arrival time is then the minimal
arrival time over [N, Nmax ]. If at time step k, the task is feasible within the preferable arrival
time N, then N will be decided as the terminal time of the MIQP. Before the arrival time,
Tf (k), the same constraints with Js1 (k) are imposed to Js2 (k). Constraints are relaxed after the
selected arrival time since the waterborne AGV has stopped. In addition, terminal constraint
ss,h (Tf (k)) = stf (4.14)
is applied upon Tf (k). We define the above logic as logic constraint C1 which is modeled for
ni = N, N + 1, ..., Nmax as:
 ni
(4.10) and (4.11), for ∑ b (n − N + 1|k) = 0,

C1 = n=N (4.15)
ni
ss,h (n|k) = stf ,
 for ∑ b (n − N + 1|k) = 1.
n=N
A diagram illustrating the timing involved cost of Js2 is shown in Figure 4.3. Therefore,
the final MIQP problem formulated at the higher level is as:
u s,h ∗ (k), b ∗ (k) = argmin Js (k), (4.16)

u s,h ,bb
4.4 Predictive switching logic 45
J s2
w3 N
0 k N N max
n
Figure 4.3: Preferable arrival time and a maximum delay.
subject to, for n = k, k + 1, ..., N − 1,
(4.10) and (4.11),
and for n = N, N + 1, ..., Nmax

(4.13) and (4.15).
Remark : As can be observed from the above derivation, the length of the reference gen-
erated by the higher level is shortened by one each simulation step. The MPC embedded
low level, however, requires an Np -length reference sr (k + i|k), for i = 1, 2, ..., Np each step.
When the current time is still distant with the scheduled arrival time, the generated ref-
erence might remain longer than Np , however, upon arrival, this might cause problems.
Therefore, we introduce Np extra time steps in addition to Nmax , and states x s (n) over
n = Nmax + 1, Nmax + 2, ..., Nmax + Np are then constrained to stay as the terminal state, i.e.,
ss,h (n) = stf . (4.17)
4.4 Predictive switching logic

MPC can take into account future situations so that effective actions can be taken at an early
stage to avoid undesirable system behaviors. This predictive feature of MPC is useful in our
switching reference path segments to avoid overshoots. Based on the path coordinate system
models described in Section 4.2, we formulate a predictive switch logic in this section.
Before proceeding, two relevant definitions are given first.
Definition 4.1 (Position) The position of a waterborne AGV is called in path coordinate
system {p j } at time step k if xp j (k), the along-track state in {p j }, is not larger than the
length of reference path j, i.e.,
xp j (k) ≤ l j . (4.18)
Definition 4.2 (Tracking errors) For the waterborne AGV to track a geometric reference
path j, three kinds of tracking errors are recognized and minimized in online MPC opti-
mizations:
• cross-track error yp j (k) the definition of which has been given in Section 4.2 and
yp j (k) → 0 indicates a convergence to the reference path;
• along-track error s j (k) − s(k) where s j (k) is the total along-track distance the water-
borne AGV has traveled, so its relationship with the along-track state xp j (k) is:
!
j
s j (k) = xp j (k) + ∑ li − l j ; (4.19)
i=1
• and heading angle error ψp j .
Gathering the tracking errors, we define

T
x ep j (k) =

s j (k) − ss,l (k) yp j (k) ψp j . (4.20)
The error vector x ep j , which is coordinate system dependent, is minimized in MPC for track-
ing.
At time step k, considering the waterborne AGV is still in {p j }, the initial states x p j (k)
can be obtained by a transformation of the current measured waterborne AGV states x (k)
from {n} to {p j } according to (4.4) and (4.5). Future system trajectories x p j (i|k) are then
predicted in a linear way as (3.13) – (3.14) and (3.18)
Remark : Note that the successive linearization approach in Chapter 3 is applied to path
dynamics in this chapter. In particular, since the discretization and linearization theories are
not applicable to discontinuous dynamics, all the predicted system trajectories as well as the
seed trajectories for linearizations are defined for path dynamics in {p j }. Therefore, (3.13)
– (3.14) and (3.18) are linearized path dynamics in {p j }. Two modifications are as follows:
• Successive linearizations are implemented for the nonlinear path dynamics which
uses the path coordinate kinematics (4.1) instead of the inertial coordinate (3.1);
• Initial states for path dynamics are not directly measurable, and are transformed from
measured the inertial coordinate states by (4.5).
The successive linearization procedure of path dynamics in this chapter is summarized by

Figure 4.4 based on the three steps in Chapter 3. 2
If the predicted trajectories to be optimized are indeed all within {p j }, e.g., when the
waterborne AGV is far away from a switching waypoint, as Figure 4.5a shows, minimization
of x ep j in online MPC optimizations realizes reference tracking. However, since MPC looks
into the future over a prediction horizon, an initial state close to the switching waypoint
would then result in predicted trajectories dispersed in both {p j } and {p j+1 }. In this case,
minimizations of x ep j will result in overshoots as Figure 4.5b shows. A transformation of
coordinates from {p j } to {p j+1 } is then necessary. Therefore, based on Definition 4.1, the
4.4 Predictive switching logic 47
Step 1: obtain seed input

trajectory
Step 2: obtain seed state

Measure initial
trajectory using path
Transform by states in {n}
dynamics
(4.5)
Step 3: linearize nonlinear

path dynamics at the seed
trajectories
Figure 4.4: Successive linearizations in path coordinates.
following logic is introduced to realize a shift of coordinate system for predicted trajectories
and tracking error x ep j so that overshoots are avoided, as Figure 4.5c illustrates.
Define binary decision variable b p (k) as an Np × 1 vector at time step k with
(
1, for xp j (i|k) ≤ l j ,
b p (i|k) = (4.21)
0, otherwise.
When the waterborne AGV travels to {p j+1 }, i.e., when xp j (i|k) > l j , it is expected to track
reference path j + 1. This logic is expressed as logic constraint C2 as:
(
x ep j (i|k), for b p (i|k) = 1,
C2 = e
(4.22)
x p j→ j+1 (i|k), for b p (i|k) = 0.
where x ep j→ j+1 (i|k) is the predicted tracking error with respect to reference path j + 1 while
the waterborne AGV is still in {p j }, i.e., predicted states x p j (i|k) are still derived in {p j }.
Then according to transformations from {p j } to {p j+1 } as (4.2) and (4.3),
T
x ep j→ j+1 (i|k) = s j+1 (i|k) − ss,l (i|k) yp j+1 (i|k) ψp j+1 (i|k) ,

(4.23)
where !
j+1
s j+1 (k + i|k) = xp j+1 (i|k) + ∑ l j j − l j+1 , (4.24)
j j=1
and
xp j (i|k) − l j

xp j+1 (i|k) cos(ψ j+1 − ψ j ) sin(ψ j+1 − ψ j )
= , (4.25)
yp j+1 (i|k) −sin(ψ j+1 − ψ j ) cos(ψ j+1 − ψ j ) yp j (i|k) − 0
and
ψp j+1 (i|k) = ψ(i|k) − ψ j+1 (i|k). (4.26)
In this way, a solution to the binary variable b p (i|k) will predictively and optimally
determine the waterborne AGV’s predicted position in coordinate system {p j } or {p j+1 }.
Corresponding tracking errors are then minimized in the online MPC optimizations and
overshoots are expected to be avoided as in Figure 4.5c.
+1
pj
X
+1
j +1
lj
Ypj
p
Y
yp j +1
X pj
Op j +1 (wpt j +1 )
lj
yp j
Op j (wpt j )
(a) Predicted trajectories are all within {p j }.
+1
pj
X
+1
lj
j +1
Ypj
p
Y
yp j +1
X pj
Op j +1 (wpt j +1 )
lj
yp j
Op j (wpt j )
(b) Predicted trajectories are dispersed in {p j }

and {p j+1 } and optimized in {p j }.
+1
pj
X
+1
lj
j +1
Ypj
p
Y
yp j +1
X pj
Op j +1 (wpt j +1 )
lj
yp j
Op j (wpt j )
(c) Predicted trajectories are dispersed and optimized in
{p j } and {p j+1 }.
Figure 4.5: Green dot–initial states; red circled dots–states predicted in {p j }; red circled
green dots– States optimized in {p j }; red circled yellow dots – States optimized
in {p j+1 }.
4.5 Receding horizon control 49
4.5 Receding horizon control
This section describes the proposed PPF-ATA algorithm based on MPC. MPC online op-
timizations compute optimal control inputs based on approximated linearized prediction
models. Waterborne AGV behaviors are then updated based on the first element of the op-
timal control input sequence. This process is repeated until the waterborne AGV arrives
at the destination specified by the ITT task. To achieve arrival time awareness and smooth
tracking, double integrator dynamics are introduced for path parameterization and gener-
ating timing-aware references over the prediction horizon by solving MIQPs. Overshoots
are avoided in the proposed MPC framework during switching waypoints by optimizing
switching logic related binary decision variables.
To achieve all the control goals presented in Section 4.1, four cost terms are minimized
in online MPC optimizations:
1. Path tracking errors as defined in Section 4.4 over the prediction horizon;
2. Too large changes in control inputs which could lead to actuator damages;
3. Kinetic energy consumption which is formulated as 12 ν T M ν for a surface waterborne

AGV, where M = M RB + M A is the mass matrix;
4. Differences between the lower level and higher level double integrator dynamics that
might cause delays.
Therefore, for a waterborne AGV in {p j } at time step k, the following MPC optimization
problem is solved:
∆uu∗ (k), u s,l ∗ (k), b p ∗ (k) = argmin J(k), (4.27)

∆uu,uus ,bbp
where
Np −1 2
J(k) = ∑ x ep j (i + 1|k) + k∆uu(i|k)k2w 5 +
i=0 w4

kνν(i + 1|k)k2w6 M / 2 + kss(i + 1|k) − s r (i + 1|k)k2w7 , (4.28)
subject to,
(3.13), (3.14), (3.18), (4.29a)
(4.21), (4.22), (4.29b)
(4.7), (4.29c)
∆xxp j (k|k) = 0, (4.29d)
|uu(i|k)| 6 u max , (4.29e)
x p j ,min 6 x p j (i + 1|k) 6 x p j ,max , (4.29f)

xp j (i + 1|k) 6 obsx p j ,min − ds + Mbobs,1 , (4.29g)

− xp j (i + 1|k) 6 − obsx p j ,max + ds + Mbobs,2 , (4.29h)

yp j (i + 1|k) 6 obsy p j ,min − ds + Mbobs,3 , (4.29i)

− yp j (i + 1|k) 6 − obsy p j ,max + ds + Mbobs,4 , (4.29j)
4
∑ bobs, n 6 3 and bobs, n ∈ {0, 1}. (4.29k)
n=1
where •∗ (k) denote the sequence of optimal variables solved at time step k. In J(k), ref-
erences for the lower level double integrator dynamics, s r (i|k) over the prediction horizon
are calculated by solving an MIQP problem before solving the online MPC optimization
problem. Generally, the time steps involved in the calculated reference s r are longer than
Np . However, only the s r (i + 1|k) is necessarily fed to J(k). Constraints (4.29a) are equal-
ity constraints of the approximated linearized prediction models of nonlinear path coor-
dinate dynamics. Constraints (4.29b) are the logic constraints for formulations of track-
ing errors in different reference path frames, as derived in Section 4.4; initial incremental
state ∆xxp j (0|k) is set to zero as (4.29d) because x p j (0|k) = x 0p j (0|k) and both of them are
equal to x p j (k) which is the current “measured”2 state; system limitations on control in-
puts and states due to system physical limits on maximum actuator forces/moment and
maximum speed, etc, are imposed by (4.29e) and (4.29f), respectively; obstacle avoidance
for static
obstacles which causes delays
to a schedule are formulated as (4.29g) – (4.29k)
where obsxp j ,min , obsyp j ,min and obsxp j ,max , obsyp j ,max are the coordinates in {p j } for
the left-low and right-up corner of a rectangular obstacle, respectively; bobs, n and M are
binary variables and a big value, respectively for an convex obstacle avoidance formula-
tion [117]. Since the avoidance
√ constraints are only applied at discrete time steps, a safety
margin ds = umax Ts /2 2 is implemented to avoid crossings in corners [53]. Note here that
obstacle avoidance constraints are imposed to the center of the waterborne AGV without
considering specific waterborne AGV shapes. However, it is assumed that waterborne AGV
sizes have been taken into account when obstacle areas are defined. Therefore, as long as
the trajectory of waterborne AGV’s center is outside obstacle areas, the waterborne AGV is
safe.
At each time step k, two MIQPs need to be solved: one is the upper level timing-aware
reference generation problem (4.16) – (4.15) and the other is the online MPC optimiza-
2x
p j (k) is not directly measurable but transformed from x (k) by (4.4) and (4.5)
4.6 Simulation results and discussion 51
Algorithm 4.1 Predictive path following with arrival time awareness (PPF-ATA)
1: Initialization at path coordinate system j = 1 at time step k = 0;
2: Solve problem (4.16) – (4.15) to obtain x rs (i) for i = 1, 2, ..., Np ;
3: while x (k) 6= x (tf ) do
4: while xp j (k) ≤ l j do
5: Measure and transform current states x (k) to x p j (k) in {p j };
6: Obtain linearized prediction models as Figure 4.4;
7: Solve optimization problem (4.27)-(4.29) to determine u ∗ (k), u s ∗ (k);
8: Apply the first element u ∗ (k|k) to waterborne AGV dynamics (3.1) and (3.5);
9: k = k + 1;
10: end while
11: j = j + 1;
12: end while
tion problem (4.27)-(4.29). With reasonable problem size for one waterborne AGV, the
two MIQPs can be solved efficiently by standard solvers. Each time a new optimization
problem is formulated given the current new measurements; a sequence of optimal control
inputs u ∗ (k) = ∆uu∗ (k) + u 0 (k) is calculated which drives predicted system outputs close
to set references to achieve design requirements. The first element of this optimal control
sequence, i.e., u ∗ (k|k) is applied to the real system (3.1) and (3.5). Time is then shifted
one step forward and the above procedures are repeated at the new time step to formulate a
receding horizon law. Convergence to the reference path and timing aware of arrival at the
destination in an economical way is thus guaranteed. System constraints are also well con-
sidered in online MPC optimizations. The overall algorithm for the problem of PPF-ATA is
summarized in Algorithm 4.1.
The condition in the outer while loop x (k) 6= x (tf ) means that a waterborne AGV has
not arrived at the final destination yet and
T
x (tf ) = ηtf T vtf T ,

(4.30)
T
where ηtf is the final pose dependent on reference path information and vtf = 0 0 0 .
The PPF-ATA controller based on Algorithm 4.1 designed for waterborne AGVs is shown
in Figure 4.6. Note that the original nonlinear waterborne AGV dynamics in {n} are used
in the closed-loop simulation.
4.6 Simulation results and discussion

In this section, we present simulation results of two typical ITT scenarios to illustrate how
the PPF-ATA controller works and to demonstrate its potential for ITT. For the first sim-
ulation, the controller is given a feasible ITT task, which means the scheduled preferable
arrival time can be achieved by the waterborne AGV. In the second simulation, an infeasible
ITT task is set where the preferable arrival time cannot be met even if the waterborne AGV
sails at the maximum speed all the time without any obstacles. In the latter case, we show
how the waterborne AGV achieves the task with a minimum delay regarding the preferable
arrival time.
ITT task Timing aware

(geometric path, reference calculator
(MIQP)
PPF-ATA
timing)
controller
Trade-off between
tracking and timing
Prediction
dynamics
Linearized models in
{pj} Switching logic
Initial states
Seed Contraints
trajectories obstacle
Transfor- avoidance
mation Optimal
Nonlinear dynamics Physical MPC control input
in {pj} limitations optimization
Shifted Smooth tracking

control inputs
Timing Performance
indices
Fuel economical
Optimal Waterborne AGV system

trajectories (Nonlinear dynamics in {n})
Figure 4.6: PPF-ATA controller for waterborne AGVs.
For both simulations, we set an ITT task from APM terminal to Euromax Terminal in
the port of Rotterdam, as shown in Figure 4.7. Distributing ITT over waterborne AGVs is
of practical interest in this scenario since these two terminals are not connected by land,
and even if they would be connected, the distance by land is much longer than by water.
The reference path consists of several straight-line segments. Simulations are implemented
based on a 1 : 70 small scaled marine surface vehicle model, CSII [121] since all of the
necessary parameters for models in (3.1) and (3.5) have been experimentally identified.
Simulation results based on CSII√are then scaled-up according to Froude scaling law [80],
e.g., 1 : 70 for length (m) and 1 : 70 for time (s), for the real scale quantities. The reference
path information3 for both of the two ITT tasks is then given as Table 4.1.
Reference path details including l j and ψ j can then be calculated from the waypoints
given in Table 4.1. Considering that in a real situation, the waterborne AGV will not stop
with a heading angle decided by Intermediate waypoint 2 and Euromax Terminal, but a
heading angle required by berthing at the terminal, see the red circle in Figure 4.7. Inter-
mediate waypoint 3 is therefore introduced to produce the final reference heading angle.
T
Therefore, final pose in (4.30) is given as ηtf = −334.7 1786.5 3.04 . We show
that this berthing behavior can also be well achieved by our PPF-ATA controller. In addi-
tion, two static obstacles are placed along the path, which causes unexpected delays. One
of them is placed half way of the first line segment, and the other half way of the third line
3 Thepositions in latitude/longitude are obtained from Google Earth and then converted to inertial frame
coordinates with APM Terminal as the origin.
Figure 4.7: APM terminal and Euromax Terminal at Maasvlakte 2 in the port of Rotterdam
from Google Earth [35].
Table 4.1: ITT scenario for a singe waterborne AGV.
Lat./Lon. (xn , yn ) (m)

APM Terminal (51.9578◦ , 4.0417◦ ) (0, 0)
Waypoint 1 (51.9614◦ , 4.0533◦ ) (798.0, 404.7)
Waypoint 2 (51.9655◦ , 4.0538◦ ) (829.1, 852.1)
Waypoint 3 (51.9734◦ , 4.0390◦ ) (-187.5, 1731.5)
Euromax Terminal (51.9739◦ , 4.0368◦ ) (-334.7, 1786.5)
segment.
The experiments also share the same MPC controller settings with a prediction horizon
Np = 20. Weight parameters are given as:

0 0
w1 = 1, w 2 = , w3 = 1000,
0 1
   
1000 0 0 1 0 0
w4 =  0 1000 0  , w 5 = I 3×3 , w 6 =  0 1 0  , w7 = 100. (4.31)
0 0 100 0 0 1
The waterborne AGV is initially positioned at (70, 0) with ψ = π and zero velocity, i.e.,
x 0 = [ 70 0 π 0 0 0 ]T . System sampling time Ts = 1s. System constraints are set
as:
   
0 1.67
 −0.84  6 v 6  0.84  , and |ττmax | = 686000 686000 36015000 T .

−15π/180 15π/180
Algorithms in this chapter are implemented using YALMIP (version 20131002) [66] in
MATLAB 2011b [75]. Optimization problems are solved by Gurobi (version 5.6 academic)
[37]. All the simulations are run on a platform with Intel (R) Core (TM) i5-3470 CPU
@3.20 GHz.
4.6.1 Simulation experiment 1: Feasible ITT task

The total reference path length is 2844.6 m and the preferable arrival time at destination
is 2510 s after departure from the origin. Therefore, if no unexpected events happen, an
average speed of 1.13 m/s should be attained, which is within the maximum speed range of
the waterborne AGV system, and thus is deemed as a feasible ITT task. However, consid-
ering the waterborne AGV cannot sail at this average speed all the time and delays might
still happen due to unexpected events, the higher level MIQP problem is solved based on an
acceptable delay tolerance of 167 s. However, delays with respect to 2510 s are penalized
in the MIQP problem. The first simulation is run based on this task to achieve smooth path
tracking with arrival time awareness in an economical way.
Path tracking performance

Path tracking performance of the PPF-ATA controller is illustrated by smooth convergence
to reference paths when there are off-sets and small deviations when the waterborne AGV is
on track. Besides, overshoots during switching of reference line segments are well avoided,
which also demonstrates the controller’s capability of path following.
Figure 4.8 shows how the waterborne AGV accurately tracks the reference path. Yellow
heptagons represent the controlled waterborne AGV that are plotted according to waterborne
AGV poses at certain time steps. Figure 4.9 further illustrates the path reference tracking
performance by showing cross-track and path heading angle errors which are the second
and third term in the error vector (4.20), respectively. Small cross-track errors around zero
are observed in general with an average value of 1.76 m. Large errors are observed for both
reference path
1600 controlled trajectory
1340
1320
1400 1300
1280
1200 1260
1240
250 300 350
1000
y (m)
880
800 860
840
820
600
800
40 780
400 800 850
20
200 0
−20 0 20 40 60 80
0
−500 0 500 1000
x (m)
Figure 4.8: Tracking performance of task 1.

cross−track error (m)
20
10
0
−10
−20
−30
0 500 1000 1500 2000 2500
heading angle error (rad)
2.5
2
1.5
1
0.5
0
0 500 1000 1500 2000 2500

time (s)
Figure 4.9: Cross-track and heading angle errors.
cross-track and heading at the beginning and also around obstacle areas because there is an
initial offset and obstacle avoidance are implemented as hard constraints to guarantee safety.
Other relatively smaller deviations in Figure 4.9 are due to switches at the three intermediate
waypoints.
The three boxes in Figure 4.8 along the path are zoom-ins of waterborne AGV behaviors
at starting point, switching at intermediate waypoint 2, and around the second obstacle,
respectively. In boxes 1 and 3, it can be observed that the waterborne AGV is able to
Euromax Terminal
1600
1400
1200
1000
y (m)
800
600
400
200
APM Terminal
0
−500 0 500 1000
x (m)
Figure 4.10: Predicted waterborne AGV trajectory over the prediction horizon at one of the
time steps during switching.
converge to the reference path smoothly with an initial offset or after a necessary offset to
avoid obstacles. This is because the lower level double integrator dynamics always “slow
down” to “wait for” the waterborne AGV if the waterborne AGV is in a situation with low
speed, e.g., at the starting point, avoiding an obstacle.
During switching of the reference line segments, as shown in the second box, the con-
trolled waterborne AGV trajectory can also match the reference path well with negligible
deviations and almost no overshoots. This is due to MPC’s predictive feature. In Figure
4.10, predicted waterborne AGV trajectories over the prediction horizon at one time step
close to switching waypoints are plotted. If there are no model mismatches between the real
system model and the model used for prediction, and if there are no external disturbances,
the future system trajectories will be exactly like the one predicted at the current time step,
which means that the real waterborne AGV trajectory will also switch successfully. Al-
though successively linearized prediction models inevitably result in model mismatches,
the successive linearization framework by conducting the linearization of the nonlinear
dynamic system about a shifted optimal trajectory has significantly reduced linearization
errors. Therefore, real waterborne AGV trajectories are also expected to have successful
switches as in Figure 4.10. Box 2 in Figure 4.8 confirms this.
Arrival time awareness
The “slow down” of lower level double integrator dynamics for smooth path tracking causes
delays. However, the delays can be compensated after the waterborne AGV is not so
“lagging-behind” by minimizing the error between lower and higher level double integra-
tor dynamics. The higher level considers current new distance-to-go and time-to-go such
that timing aware references are generated. Figure 4.11 illustrates this. At the starting time
60 along track state and lower level s

lower level s and upper level s
along track errors (m) 50
40
30
20
10
0 500 1000 1500 2000 2500

time (s)
Figure 4.11: Along-track errors.

surge velocity (m/s)
0
0 500 1000 1500 2000 2500
sway velocity (m/s)
−1
0 500 1000 1500 2000 2500
yaw rate (rad/s)
0.2
−0.2
0 500 1000 1500 2000 2500

time (s)
Figure 4.12: Waterborne AGV surge, sway velocities, and yaw rate.
and when the first and second obstacle avoidance happen, both along-track errors and lower-
/higher level tracking errors see some fluctuations, but both of them return to an approximate
zero afterwards. Moreover, the waterborne AGV arrives at the destination at t = 2518 s with
eight second’s delay which is 0.33% of the total time.
Longitudinal force (N)

5
x 10
5
0
−5
0 500 1000 1500 2000 2500

5
x 10
lateral force (N)
5
0
−5
0 500 1000 1500 2000 2500

7
Yaw moment (Nm)
x 10
4
2
0
−2
−4
0 500 1000 1500 2000 2500
time (s)
Figure 4.13: Waterborne AGV surge, sway forces, and yaw moment.
Energy consumption and system constraints
The objectives of good path following performance and arrival time awareness are achieved
in an energy efficient way within system limits. In Figure 4.12, system velocities all
maintain almost constant except for fluctuations at initial, obstacle and reference switch-
ing points. Since all the MPC optimization problems are successfully solved, the velocities
are optimal values in the feasible region defined by system constraints. Comparisons on
energy consumption of the two experiments are presented in Section 4.6.2. System physical
constraints are also well satisfied in our scheme. Actuator inputs are shown in Figure 4.13.
Same as in Figure 4.12, all the parameters are within the system limitations.
4.6.2 Simulation experiment 2: Infeasible ITT task

In this simulation, the waterborne AGV also needs to follow the scaled reference path from
APM Terminal to Euromax Terminal. However, the scheduled preferable arrival time is
set to be only 1673 s after departure. Therefore, even if no unexpected events happen and
the waterborne AGV is right on the path with a heading angle tangent to the path, the
waterborne AGV still needs to sail at an average speed of 1.7 m/s all the time. However, the
waterborne AGV has a maximum surge speed of 1.67 m/s let alone the effects of current, off-
track positions and zero velocities as initial states, and unexpected events such as obstacle
avoidance. Therefore, this ITT task is defined as infeasible. Similarly with Experiment
1, however, we append 167 s as an acceptable maximum arrival time, which results in an
average speed of 1.55 m/s. Simulation results below illustrate how the PPF-ATA controller
also works well in this scenario to achieve smooth path tracking, arrival time awareness,
and energy efficiency.
reference path
1600 controlled trajectory
1400 480
1750 460
440
1200 1700 420
−350−300−250−200 400
1000 380
750 800
y (m)
800 240
220
600 200
180
400 160
350 400 450
200
0
−500 0 500 1000
x (m)
Figure 4.14: Tracking performances.

cross−track error (m)
20
10
0
−10
−20
−30
0 200 400 600 800 1000 1200 1400 1600
heading angle error (rad)
2.5
2
1.5
1
0.5
0
0 200 400 600 800 1000 1200 1400 1600

time (s)
Figure 4.15: Cross-track and heading angle errors.
Path tracking performance
Similar with Experiment 1, accurately tracking of the reference path is observed in this case
as Figure 4.14 shows. Three boxes in this figure along the path are zoom-ins of water-
borne AGV behaviors around the first obstacle, switching at intermediate waypoint 1, and
during the final destination area, respectively. Again, smooth tracking and convergence to
the reference path are achieved including areas around the starting point, obstacle and dur-
ing switches. Figure 4.15 further illustrates the path convergence performance by showing
Euromax Terminal
1600
1400
1200
1000
y (m)
800
600
400
200
APM Terminal
0
−500 0 500 1000
x (m)
Figure 4.16: Predicted waterborne AGV trajectory over the prediction horizon at one of the
time steps during switching.
cross-track and heading angle errors along time. In this experiment, cross-track errors are
with an average value of 2.75 m. Relatively obvious deviations in both sub-figures of Figure
4.15 are due to the initial offset, obstacles, and switches at the three intermediate waypoints.
The second box demonstrates the switching of reference paths has been successful with
almost no overshoots. Compared with the switch box in Figure 4.8, larger deviations are
observed, which can also be observed by comparing the errors caused by switches in Figures
4.9 and 4.15. This is because when the arrival time is set shorter, waterborne AGV needs to
sail at a higher speed, which leads to larger errors. Predicted waterborne AGV trajectories
over the prediction horizon at one time step near switching waypoints for this ITT task are
shown as Figure 4.16.
In the third box, waterborne AGV trajectories can also well follow the last line segment
which has been added for a berthing behavior. Figure 4.17 further shows the heading angle
trajectories which illustrate that the waterborne AGV stops at the destination terminal with
a berthing angle.
Arrival time awareness
Figure 4.18 shows along-track errors in this task. Again, both along-track errors and lower-
/higher level tracking errors see some fluctuations during starting, obstacle, and switching
areas, but both of them return to an approximate 0 afterwards. Compared to Figure 4.11, the
times when fluctuations happen due to obstacles and switches are earlier. This is because in
Experiment 2, the waterborne AGV is sailing at a higher speed.
However, in this experiment, the waterborne AGV is still able to meet the timing re-
quirement and arrives at the destination at t = 1774 s with a 6% delay with respect to the
expected arrival time. Figure 7.10 shows how the arrival times calculated by higher level
3 desired heading angle

vessek heading angle
2.5
heading angle (rad)
1.5
0.5
0 200 400 600 800 1000 1200 1400 1600

time (s)
Figure 4.17: Waterborne AGV heading angles.
60 along track state and lower level s

lower level s and upper level s
50
40
along track errors (m)
30
20
10
0 200 400 600 800 1000 1200 1400 1600

time (s)
Figure 4.18: Along-track errors.
MIQP changes every time an delay event, e.g., obstacles and switches, happens.
Energy consumption and system constraints

The total kinetic energy consumption calculated according to Section 4.5 for task 1 and task
2 is 1.8 × 106 kJ and 2.6 × 106 kJ, respectively. Again, since we are solving repetitive con-
strained optimization problems which are all successfully solved, it is sufficient to conclude
1684
calculated arrival time from upper level (s)
1682
1680
1678
1676
1674
0 200 400 600 800 1000 1200 1400 1600

time (s)
Figure 4.19: Calculated arrival time from the higher level.
18000 task 1
task 2
16000
14000
12000
energy (KJ)
10000
8000
6000
4000
2000
0
0 50 100 150 200 250 300
time (s)
Figure 4.20: Comparison of consumed energy.
that the energy consumption is optimal in a sense that they are the smallest within the system
constraints. Although for the second ITT task, the cumulative time is shorter, it still has a
much larger total energy consumption. Comparisons of the time-wise energy consumption
of them are presented as Figure 4.20. It is clear that the waterborne AGV in the second task
is consuming more energy all the time so that it can fulfill the ITT task on time.
Velocity and actuator force trajectories for this experiment are shown as Figure 4.21 and
4.7 Conclusions 63
surge velocity (m/s)

2
0
0 200 400 600 800 1000 1200 1400 1600
sway velocity (m/s)
−1
0 200 400 600 800 1000 1200 1400 1600
yaw rate (rad/s)
0.2
−0.2
0 200 400 600 800 1000 1200 1400 1600

time (s)
Figure 4.21: Waterborne AGV surge, sway velocities and yaw rate.
Longitudinal force (N)
5
x 10
5
0
−5
0 200 400 600 800 1000 1200 1400 1600

5
x 10
lateral force (N)
5
0
−5
0 200 400 600 800 1000 1200 1400 1600

7
Yaw moment (Nm)
x 10
4
2
0
−2
−4
0 200 400 600 800 1000 1200 1400 1600
time (s)
Figure 4.22: Waterborne AGV surge, sway forces and yaw moment.
Figure 4.22, respectively. Again, all the parameters are within the system limits.
4.7 Conclusions
A predictive path following with arrival time awareness (PPF-ATA) controller for a single
waterborne AGVs carrying out an assigned ITT task has been proposed in this chapter. The
proposed approach answers the second Key Research Question in Chapter 1 by considering
smooth path tracking, arrival time awareness, and energy efficiency as the performance cri-
teria. In our two simulation experiments based on ITT scenarios in the port of Rotterdam,
these conflicting objectives have been achieved in a systematic way by the proposed PPF-
ATA controller. For both experiments, smooth path tracking behaviors are observed with
average cross-track errors 1.76 m and 2.75 m, respectively. The waterborne AGV in the two
simulations has 0.33% and 6% delays with respect to the expected arrival times, respec-
tively. The relatively large delay in simulation 2 is due to the physical speed limits and time
consuming obstacle avoidance. Moreover, overshoots are avoided during switching refer-
ence paths by taking advantage of the predictive feature of MPC and an along-track state
involved switching logic. The method proposed in this chapter provides a comprehensive
solution applicable to problems on path following with timing requirements including but
not limited to waterborne AGVs for ITT.
The PPF-ATA controller proposed for a single waterborne AGV in this chapter will be
extended to multiple waterborne AGVs in Chapter 5 and to waterborne AGVs with uncer-
tainties in Chapter 6.
Chapter 5
Cooperative distributed
waterborne AGVs
This chapter extends the scenario of a single waterborne AGV considered in Chapter 4 to
multiple waterborne AGVs carrying out Inter Terminal Transport (ITT) tasks. The network
of multiple waterborne AGVs is modeled as in Chapter 3. The control goal is that water-
borne AGVs minimize an overall objective in a cooperative distributed way. Simulation
results are presented to illustrate the effectiveness of the proposed algorithms.
The research discussed in this chapter is based on [148, 149].
5.1 Introduction
A single waterborne AGV is controlled by the proposed predictive path following with ar-
rival time awareness (PPF-ATA) controller to fulfill an assigned ITT task in Chapter 4. Con-
trol goals such as smooth path tracking, timing, and energy efficiency are well achieved for
the waterborne AGV when maneuvering independently. However, as discussed in Chap-
ter 2, usually a fleet of waterborne AGVs will be deployed for carrying out multiple ITT
tasks. When waterborne AGVs are in the proximity of each other, couplings arise, as mod-
eled in Chapter 3, which impede independent decision making. System-wide control can
be approached in centralized, decentralized, and distributed ways [89]. The advantages of
distributed control for waterborne AGVs have been discussed in Chapter 2. The main chal-
lenge lies in the satisfaction of collision avoidance couplings in a distributed and preferably
parallel way for equally treated waterborne AGVs.
In this chapter, based on the PPF-ATA controller proposed for one single waterborne
AGV in Chapter 4, a distributed PPF-ATA controller is proposed for cooperative water-
borne AGVs carrying out multiple ITT tasks. ITT using waterborne AGVs is modeled by
time-varying graphs, see Chapter 3 and Figure 3.2. Cooperative distributed computations
based on these graphs are realized in the framework of distributed MPC where parallelism is
achieved following the Alternating Direction Method of Multipliers (ADMM) [12]. A fast
ADMM approach iteratively approximating global information in local problems is pro-
posed to obtain better convergence rates than that of the conventional ADMM. Simulation
65
66 5 Cooperative distributed waterborne AGVs
results for an ITT case study illustrate the effectiveness of the proposed algorithms for dis-
tributed MPC of time varying networks in general and cooperative distributed waterborne
AGVs in particular.
The remainder of this chapter is organized as follows. A centralized formulation for co-
operative waterborne AGVs based on the PPF-ATA controller in Chapter 4 is first presented
in Section 5.2. Distributed solutions based on ADMM and fast ADMM are then proposed in
Section 5.3 and Section 5.4, respectively. In Section 5.5, simulation experiments and results
are discussed, followed by concluding remarks in Section 5.6.
5.2 A centralized formulation

When multiple waterborne AGVs are scheduled for carrying out multiple ITT tasks, be-
sides the control goals of a single waterborne AGV, i.e., smooth path tracking, arrival time
awareness, and energy efficiency, extra requirements are to be satisfied:
• a safety distance between waterborne AGVs;
• minimal overall energy consumption; and
• distributed parallel computations.
A centralized solution that satisfies the first two requirements can be formulated based
on the models introduced in Chapter 3, and the proposed PPF-ATA controller in Chapter 4
as:
nG (k)

min ∑ JGs X Gs (k),U U Gs (k) , (5.1)
s=1
where

U Gs (k) =
JGs X Gs (k),U ∑ J p (xx p (k), ∆uu p (i|k)) (5.2)
p∈Vs (k)
2 2
= ∑ x ep (i|k) w1
+ k∆uu p (i|k)k2w 2 + x νp (i|k) w3
p∈Vs (k)
subject to for s = 1, 2, · · · , nG (k)
∆xx p (i|k) ∈ Cx p (k) ∀p ∈ V (k), (5.3)

∆uu p (i|k) ∈ Cu p (k) ∀p ∈ V (k), (5.4)
(∆rr p (i|k), ∆rr q (i|k)) ∈ Cr p,q (k) ∀e p,q (k) = 1, e p,q (k) ∈ Es (k), (5.5)
where local physical limitations due to (3.3) on perturbation states and control inputs are
represented by convex sets Cx p (k) and Cu p (k) as (5.3) and (5.4), respectively. Similarly,
collision avoidance constraints (3.19) on perturbation position variables are represented by
convex sets Cr p,q (k) as (5.5).
The total cost in (5.1) is a summation of costs over all subgraphs and the subgraph cost
(5.2) is a summation of local costs over all waterborne AGVs in the current subgraph. Local
cost functions are convex minimizing path following errors, control input changes, and
5.3 Cooperative distributed waterborne AGVs 67
kinetic energy consumption. For ease of notation, states and control inputs for p ∈ Vs (k)
are represented compactly by X Gs (k) and U Gs (k), respectively. Centralized problem 5.1 –
5.5 is coupled due to pairwise collision avoidance constraints (5.5). Distributed solutions
are proposed next to decompose this centralized problem to satisfy the third requirement.
5.3 Cooperative distributed waterborne AGVs

Centralized problem (5.1) – (5.5) can be decomposed into nG (k) subgraph problems (5.2) –
(5.5) that are independent of one another. Subgraphs that are not singleton, i.e., ns (k) > 1,
have coupling collision avoidance constraints (5.5) which prohibit a further distributed so-
lution. This section proposes a distributed PPF-ATA control approach based on the iterative
decomposition-coordination procedure of ADMM for the problem over subgraphs.
5.3.1 Derivation of ADMM for waterborne AGVs

For each subgraph Gs (k) = (Vs (k), Es (k)), we assign one of the waterborne AGVs p ∈ Vs (k)
as the coordinator to take care of the couplings. All waterborne AGVs are able to carry
out computations simultaneously and communicate with the coordinator. The coordinator
broadcasts solutions towards which local solutions are regularized by adjusting a Lagrange
multiplier and an augmented quadratic penalty term until consensus is achieved.
For each waterborne AGV p ∈ Vs (k), we introduce a copy of the perturbation position
variables as ∆r̂r p (i|k). Then centralized problem (5.2) – (5.5) is equal to the same problem
subject to an additional constraint
∆r̂r p (i|k) = ∆rr p (i|k). (5.6)
The augmented Lagrangian [12] that relaxes (5.6) is:
Lρ (k) = ∑ J p (xx p (k), ∆uu p (i|k)) + λ p,r (i|k)T (∆rr p (i|k) − ∆r̂r p (i|k))
p∈Vs (k)

+ρ/2 k∆rr p (i|k) − ∆r̂r p (i|k)k22 + IC (k), (5.7)
where λ p,r (i|k) ∈ R 2 is the dual variable with respect to (5.6); ρ is the augmented La-
grangian parameter. The coupling collision avoidance constraint (5.5) has been replaced by
the non-differential indicator function IC (k) for Cr p,q (k) as:
(
0, for (∆r̂r p (i|k), ∆r̂r q (i|k)) ∈ Cr p,q (k), ∀e p,q (k) = 1, e p,q (k) ∈ Es (k),
IC (k) =
∞, otherwise.
j
Given initial values λ p,r (i|k) and ∆r̂r pj (i|k) with j = 0, the ADMM decomposition-
coordination at each iteration j consists of the following three steps:
Step 1: Each waterborne AGV p ∈ Vs (k) solves a local problem with the information ∆r̂r pj (i|k)
sent from the graph coordinator to update the original perturbation position states
∆rr p (i|k):
∆uu pj+1 (i|k), ∆rr pj+1 (i|k) = argmin J p (xx p (k), ∆uu p (i|k)) +

(5.8)
j T j
j 2
λ p,r (i|k) ∆rr p (i|k) − ∆r̂r p (i|k) + ρ/2 ∆rr p (i|k) − ∆r̂r p (i|k) 2
subject to (5.3) – (5.4).

j+1
Step 2: The graph coordinator solves the following problem with the information ∆rr p (i|k)
collected from the waterborne AGVs to update the copied perturbation position states
∆r̂r p (i|k):

j
∆r̂r pj+1 (i|k) = argmin IC (k)+ ∑ (i|k)T ∆rr pj+1 (i|k) − ∆r̂r p (i|k)

λ p,r (5.9)
p∈Vs (k)
2

+ρ/2 ∆rr pj+1 (i|k) − ∆r̂r p (i|k) 2
.
Step 3: Each waterborne AGV p ∈ Vs (k) updates the local dual variables λ p,r (i|k) with new
information ∆r̂r pj+1 (i|k) from the coordinator as:
j+1 j
(i|k) + ρ ∆rr pj+1 (i|k) − ∆r̂r pj+1 (i|k) .

λ p,r (i|k) = λ p,r (5.10)
Step 1 and Step 3 are both carried out in parallel on board of each waterborne AGV
p ∈ Vs (k). The coordinator problem at Step 2 is implemented as Euclidean projections onto
Cr p,q (k) implemented as:
2
j+1
∆r̂r pj+1 (k) := argmin ∑ ∆r̂r p (k) − ∆rr pj+1 (i|k) + λ p,r (k)/ρ (5.11)
2
p∈V (k)
subject to
(∆r̂r p (i|k), ∆r̂r q (i|k)) ∈ Cr p,q (k), ∀e p,q (k) = 1, e p,q (k) ∈ Es (k).
Iterations are then alternating between the coordinator and waterborne AGVs until con-
sensus constraints (5.6) are satisfied according to certain criteria, implying that collision
avoidance couplings (5.5) are also satisfied on local solutions ∆rr (i|k).
5.3.2 Convergence analysis

To proceed with convergence analysis of the above ADMM iterations, we make the follow-
ing assumption:
Assumption 5.1 The centralized subgraph problem (5.2) – (5.5) is feasible.
With Assumption 5.1, the ADMM iterations by Steps 1 – 3 have the following conver-
gence properties:
Proposition 5.1 (Convergence of iterations by Steps 1 – 3) Under Assumption 5.1, the fol-
lowing convergence is achieved as iteration j → ∞:
j
1. Primal feasibility, i.e., for each waterborne AGV p ∈ Vs (k), ∆r̂r pj (i|k) → ∆rr p (i|k).

2. Objective convergence, i.e., primal objective value JGs X Gs (k),U U Gs (k) in (5.2) ap-
(
proaches the centralized optimal value JG∗ s (X
X Gs k),U
U Gs (k)).
5.3 Cooperative distributed waterborne AGVs 69
Algorithm 5.1 ADMM based distributed PPF-ATA: processed in parallel by each water-
borne AGV p ∈ Vs (k)
j
1: Initialize λ p,r (i|k) and ∆r̂r pj (i|k) at j = 0;
2: loop
j+1
3: Computes ∆rr p (i|k) solving the problem at Step 1;
j+1 j
4: Sends ∆rr p (i|k) and λ p,r (i|k) to the coordinator;
5: repeat
6: Wait;
7: until ∆r̂r pj+1 (i|k) arrives;
j+1
8: Computes λ p,r (i|k) as (5.10) at Step 3;
9: j + 1 → j;
10: end loop
j
3. Dual variable convergence, i.e., ∀p ∈ Vs (k), λ p,r (i|k) approaches the optimal dual
∗
value λ p,r (i|k).
Proof : The above proposition follows directly from general ADMM convergence proper-
ties in [12] where the proof is established under two mild assumptions: 1) The (extended-
real-valued) separable two functions are closed, proper, and convex. 2) The unaugmented
Lagrangian L0 has a saddle point. We prove Proposition 5.1 by showing that the two as-
sumptions hold also in our case. Firstly, for each waterborne AGV p ∈ Vs (k), define an in-
dicator function IC p (k) of the local convex constraint sets Cx p (k) and Cu p (k), and IC p (k) = 0
when (5.3) – (5.3) are satisfied and ∞ otherwise. Since local costs J p are convex, Cx p (k) and
Cu p (k) are non-empty convex, the problem at Step 1 is solvable. Likewise, indicator function
IC p,q (k) is also closed, proper, and convex, and the problem at Step 2 is also solvable. The
first assumption of [12] is satisfied. Secondly, since centralized problem (5.2) – (5.5) is fea-
sible by Assumption 5.1, problem (5.2) – (5.6) is also feasible. Let ∆rr ∗p (i|k), ∆r̂r ∗p (i|k) be
a feasible solution. As analyzed before, local cost functions J p are convex and constrained
sets Cx p (k), Cu p (k), and C p,q (k) are with non-empty relative interior(Slater’s condition holds
[11]), there exists λ ∗p,r (i|k) such that ∆rr ∗p (i|k), ∆r̂r ∗p (i|k), λ ∗p,r (i|k) is a saddle point of the
unaugmented Lagrangian L0 (k) (set ρ = 0 for (5.7)), i.e.,
L0 ∆rr ∗p (i|k), ∆r̂r ∗p (i|k), λ p,r (k) 6 L0 ∆rr ∗p (i|k), ∆r̂r ∗p (i|k), λ ∗p,r (k)

6 L0 ∆rr p (i|k), ∆r̂r p (i|k), λ ∗p,r (k) .

(5.12)
Therefore, the second assumption of [12] is also satisfied. 2
The ADMM based iterations are then implemented in a distributed way alternating be-
tween waterborne AGVs ∀p ∈ Vs (k) processing in parallel as Algorithm 5.1 and a coordi-
nator waterborne AGV for Gs (k) processing as Algorithm 5.2. The inputs are initialized at
j
j = 0. Outputs ∆uu p (i|k) from Algorithm 5.1 are returned after executing Line 7 in Algorithm
5.2.
Algorithm 5.2 ADMM based distributed PPF-ATA: processed by the coordinator water-
borne AGV of Gs (k)
1: repeat
2: repeat
3: Waits;
j+1 j
4: until ∆rr p (i|k) and λ p,r (i|k) arrive;
j+1
5: Computes ∆r̂r p (i|k) as (5.9) at Step 2;
6: Broadcasts ∆r̂r pj+1 (i|k) to ∀p ∈ Vs (k);
7: until Stopping criteria are met.
5.3.3 Stopping criteria

Convergence is achieved by primal and dual feasibility which, in practice, are indicated by
small primal and dual residuals, respectively, i.e.,
r j (k) = ∑ ∆rr pj (i|k) − ∆r̂r pj (i|k) 2

6 εpri , (5.13)
p∈Vs (k)
s j (k) = ∑ ∆r̂r pj+1 (i|k) − ∆r̂r pj (i|k) 2

6 εdual , (5.14)
p∈Vs (k)
where εpri and εdual are primal and dual feasibility tolerances specified using an absolute
and relative criterion following [12] as:
 
 
εpri = 2ns N p εabs + εrel max ∆rr pj (i|k) 2 , ∑ ∆r̂r pj (i|k) 2 ,
p
∑ (5.15)
 p∈V (k) p∈V (k)

s s
dual j
= 2ns N p εabs + εrel
p
ε ∑ λ p,r (k) . (5.16)
2
p∈Vs (k)
As suggested in [12], a varying step size ρ j is implemented as follows to improve prac-

tical convergence and make it less dependent on the initial value of ρ:

j j j
2ρ , for r (k) > 10s (k)

ρ j+1 = ρ j /2, for s j (k) > 10r j (k) (5.17)

 j
ρ, otherwise.
Tens of iterations are usually required to achieve convergence to a modest accuracy

for conventional ADMM [12]. A variant, fast ADMM, is proposed next to improve the
convergence rate in practice.
5.4 Fast ADMM for cooperative distributed waterborne

AGVs
Since waterborne AGVs in ADMM are solving local problems selfishly to achieve local
control goals and are only coordinated by a penalty term on position variables, the global
5.4 Fast ADMM for cooperative distributed waterborne AGVs 71
Waterborne AGV #5
and graph coordinator
)
|k
Drˆ j +1
(i
1
j+
8
Drˆ j +1 (i | k
Dr
2
5
(i | )
)
|k
k)
(i
Dr
)
j+
|k
j +1
8
(i |
Drˆ
2
(i
5 j +1
k)
Drˆ
Wa GV #
ne
or
te r 8
b
A
er #2
at
bo
W AGV
rne
Sub-network #2
Figure 5.1: Comparison of information flow in fast ADMM.
agreement on the coupling collision avoidance constraints could be slow. The proposed fast
ADMM involves adding, iteratively, approximated collision avoidance constraints to local
problems based on safe trajectories ∆r̂r pj+1 (k), p ∈ Vs (k) from the coordinator as:
d 0p,q (i|k) +C
C (i|k)∆r p (i|k) + D (i|k)∆r̂qj (i|k) > Ds , for p, q ∈ Vs (k), d p,q (k) 6 Dc . (5.18)
Approximated local collision avoidance constraints (5.18) are different from global col-
lision avoidance constraints (3.19) in two aspects:
1. Waterborne AGV p in (3.19) only cares about the pairwise collision constraints with
waterborne AGV q for p, q ∈ Vs (k), e p,q = 1 with p < q, while in (5.18), waterborne
AGV p cares about pairwise collision constraints with waterborne AGV q as along as
they are in the communication range of each other according to d p,q (k) 6 Dc ; and
2. In (5.18), coupling waterborne AGV q’s trajectory is treated as known and fixed while
in (3.19), trajectories of both waterborne AGV p and q are variables.
Besides, extra information in fast ADMM needs to be communicated from the coordinator
to waterborne AGVs to formulate (5.18). Taking subgraph 2, G2 (k) = (V2 (k), E2 (k)) where
V2 (k) = {2, 5, 8} and e2,5 = 1, e5,8 = 1 in Figure 3.2b as an example, the extra communi-
cated information to corresponding waterborne AGVs (waterborne AGV #5 is acting as the
coordinator) is highlighted in green in Figure 5.1.
Step 1 in fast ADMM then solves the following problem by each waterborne AGV
j
p ∈ Vs (k) parallely updating ∆rr p (k):
∆uu pj+1 (i|k), ∆rr pj+1 (i|k) = argmin J p (xx p (k), ∆uu p (i|k)) +

(5.19)
j 2
λ p,r (i|k)T (∆rr p (i|k) − ∆r̂r p (i|k)) + ρ/2 ∆rr p (i|k) − ∆r̂r pj (i|k) 2
subject to (5.3) – (5.4), and (5.18).

Similar with the definitions of r j (k) and s j (k) in (5.13) – (5.14), we denote the devia-
j+1 j+1 j+1
tions of ∆rr p (i|k) from ∆r̂r pj (i|k) for all waterborne AGVs in Gs as σs (k). Then σs (k)
inherently has two features along with iterations:
j+1
1. Deviations σs (k) cannot be too large because of the augmented penalty term (ρ/2)
2
∆rr p (i|k) − ∆r̂r pj (i|k) 2 in (5.8) or (5.19), which guarantees local problems at Step 1
of fast ADMM approximate well the original problem at Step 1 before convergence;
2. When both primal and dual residual convergence are achieved satisfying (5.13) –
j j+1
(5.14), i.e., ∆rr p (i|k) → ∆r̂r pj (i|k) and ∆r̂r pj+1 (i|k) → ∆r̂r pj (i|k), deviations σs (k) → 0
since
σsj+1 (k) = ∑ ∆rr pj+1 (i|k) − ∆r̂r pj (i|k)

(5.20)
p∈Vs (k)
∆rr pj+1 (i|k) − ∆r̂r pj+1 (i|k) + ∆r̂r pj+1 (i|k) − ∆r̂r pj (i|k) ,

= ∑
p∈Vs (k)
which implies that problem at Step 1 of fast ADMM at the terminal iteration finally
recovers the original problem at Step 1; feasibility and optimality are satisfied when
stopping criteria (5.13) – (5.14) are met.
Remark : Convergence is achieved as Proposition 5.1 for conventional ADMM based dis-
tributed PPF-ATA. When ρ is adaptive as (5.17) with iterative information, convergence is
difficult to prove. But the varying ρ technique is effective in practice, and convergence can
be achieved if ρ becomes fixed after a prior unknown but finite number of iterations [12]. In
the case of fast ADMM based distributed PPF-ATA, the approximated local collision avoid-
ance constraints (5.18) are also adaptive with respect to the results from a previous iteration.
If the constraint sets formulated by (5.18) become fixed after a finite number of iterations,
convergence can also be studied as in [12]. Practical improved convergence rates of fast
ADMM are further illustrated and analyzed in Section 5.5.2. 2

Theoretically, the proposed cooperative distributed approaches apply to any size of water-
borne AGV fleets. However, for maritime applications, typical encounters usually involve
two or three vehicles. As a representative ITT case study in the port of Rotterdam, we
consider a scenario shown as Figure 5.2 with five ITT tasks in Table 5.1 to illustrate the
effectiveness of the proposed control approaches. Five waterborne AGVs are denoted as
Vi , i = 1, 2, ..., 5 each assigned one ITT task1 and are positioned at their corresponding de-
parture terminals. There are two potential conflicting areas where waterborne AGVs may
encounter one another, indicated by the two circles in Figure 5.2. We still use the small
scaled marine vehicle model, CSII [121], to represent waterborne AGV dynamics (3.1),
(3.5). Simulation results based on CSII √ are then scaled-up according to Froude scaling law
[80], e.g., 1 : 70 for length (m) and 1 : 70 for time (s), for the real scale quantities. For
simulations in this chapter, sampling time is set as Ts = 0.5s and prediction horizon Np = 20.
1 The case that more than one ITT tasks are assigned to one waterborne AGV will be considered in Chapter 7.
Table 5.1: ITT tasks for waterborne AGVs.
ITT Routes Origin-Destination Departure-arrival

No. Berth → Berth (m) times (s)
1 → V1 1→2 (0, 0) → (411, 417) 0 → 460
2 → V2 2→1 (411, 417) → (0, 0) 0 → 418
3 → V3 3→4 (-50, 359) → (420, 52) 0 → 376
4 → V4 5→6 (1665, 236) → (441, 732) 0 → 920
5 → V5 7→8 (1121, 1144) → (1095, 149) 0 → 627
Figure 5.2: ITT scenario for waterborne AGVs.
Parameters for ADMM iterations are: maximum iteration jmax = 1000, absolute tolerance
εabs = 10−2 and relative tolerance εrel = 5 × 10−4 . Algorithms are implemented in MAT-
LAB 2014b [75] with optimization problems solved by Cplex [46]. Simulations are run on
a platform with Intel(R) Xeon(R) CPU E5-1620 v2 @3.70 GHz.
5.5.1 Safely carrying out ITT tasks

Optimal trajectories from fast ADMM based distributed PPF-ATA control schemes of the
five waterborne AGVs are shown in Figure 5.3. All waterborne AGVs are able to track their
assigned reference paths well, except for the conflicting areas where some deviations arise
to maintain a safety distance away from others. The trade-off between deviations from ref-
erence paths and safety has been optimized since all online optimizations are successfully
solved. Although trajectories overlay spatially, they do not overlay temporally at the same
time, which is demonstrated by positions of waterborne AGVs V1 , V2 , and V3 at t = 230s
and waterborne AGVs V4 , V5 at t = 418s plotted in different shapes in Figure 5.3. Colli-
sions are thus successfully avoided. Safety by the fast ADMM based distributed PPF-ATA
control approach is further confirmed by Figure 5.4 where Euclidean distances between
waterborne AGVs are above the minimum safety distance all the time. Furthermore, water-
borne AGVs arrive at their specified destinations punctually at 460s, 423s, 410s, 920s, and
636s, respectively, despite the possibly time consuming behavior for collision avoidance in
the conflicting areas.
V1
1200 V2
V3
V4
1000
V5
V1 at 230s
800
V2 at 230s
V3 at 230s
y (m)
600 V4 at 418s
V5 at 418s
400
200
0 200 400 600 800 1000 1200 1400 1600

x (m)
Figure 5.3: Trajectories of five waterborne AGVs.
1200
V1 and V2
V1 and V3
1000
V2 and V3
V4 and V5
800 safety distance
distances (m)
600
400
200
0
0 100 200 300 400 500 600 700 800 900
time (s)
Figure 5.4: Distances between waterborne AGVs.

5.5.2 Convergence behavior in distributed computations

Both the proposed ADMM and fast ADMM based distributed PPF-ATA control approaches
are based on the time varying graphs modeled as in Chapter 3. Figure 5.5 shows the graph
evolution modeling all working waterborne AGVs controlled by the fast ADMM based dis-
tributed PPF-ATA control approach. Circles with numbers represent the numbered water-
borne AGVs. Red dashed lines indicate there exists a collision avoidance coupling between
the corresponding two waterborne AGVs. Circles with the same color are waterborne AGVs
coupled together and thus in one subgraph. Time-varying subgraphs in Figure 5.5 indicate
the time-varying couplings among waterborne AGVs.
Based on the above subgraphs, both control approaches can achieve the above over-
all system performance and safety via iterative decomposition-coordination. However, the
fast ADMM based distributed PPF-ATA control scheme has a much faster convergence rate
than the ADMM based distributed PPF-ATA controller, as shown in Figure 5.6. The num-
ber of iterations with corresponding computation times for ADMM and fast ADMM based
controllers are reported in subplot (a) and (b), respectively. Number of iterations are the
maximum iterations over all subgraphs. Computation times are the summation of com-
puting times of all iterations of the corresponding subgraph and the computation time per
iteration is the summation of the maximum time of all vehicles solving the problem at Step
1, the time solving the problems at Step 2, and Step 3. ADMM involves large numbers of
iterations with long computational times in the two conflicting areas. A maximum of over
200 iterations with over 50s is required for ADMM to reach convergence. The computa-
tional advantage of fast ADMM over ADMM is obvious. Fast ADMM converges with a
maximum of six iterations and 0.27s which is within the system sampling time Ts = 0.5s.
Also, both subplots show that the number of iterations and the computation times are ap-
proximately positively related. Note that in practice, timely feedback within the sampling
time is generally critical for real-time control systems. Fast ADMM exploits the algorith-
mic structure of conventional ADMM computing timely for waterborne AGVs. In more
complicated and computational demanding scenarios, besides resorting to tuning controller
parameters, improving optimization model structure, and using more powerful computing
platforms etc., a reliable decision recovery mechanism is always necessary when real-time
decisions cannot be updated timely.
For further comparison, we applied ADMM based distributed PPF-ATA, fast ADMM
based distributed PPF-ATA, and centralized controllers to a same subgraph Gs (k) with
Vs (k) = {V1 ,V2 ,V3 } and Es (k) = {(V1 ,V2 ) , (V1 ,V3 ) , (V2 ,V3 )} at a particular time t = 142s.
ADMM based distributed PPF-ATA requires 108 iterations with a total solver time of 4.65s
before convergence while fast ADMM based distributed PPF-ATA requires only six iter-
ations with 0.27s. Subplot (a), (b), and (c) in Figure 5.7 show the detailed primal resid-
ual, dual residual and overall cost convergence behaviors, respectively, of ADMM and fast
ADMM based approaches. Primal residuals being small means that trajectories updated in
parallel at Step 1 and coordinated trajectories considering collision avoidance at Step 2 are
driven close to each other. Dual residuals being small means that trajectories updated by the
sub-coordinator do not change much from their previous iteration any more. Global safety
and optimality are then achieved when primal and dual residuals satisfy (5.13) – (5.14).
Global optimality is further illustrated in subplot (c) where the centralized cost acts as a
baseline and cost differences of ADMM and fast ADMM based controllers iterate approach-
Subgraphs at time step 1 Subgraphs at time step 15

5
1000 1000 5
y (m)
y (m)
500 500
3 2 2
3 4
4
0 1 0 1
0 500
1000 1500 0 500 1000 1500
x (m) x (m)
1000 5 1000
y (m)
y (m)
500 500 5
4
3 2 4 1
0
1 0 2 3
0 500 1000 1500 0 500 1000 1500

x (m) x (m)
1000 1000
y (m)
y (m)
500 54 500 45
1 1
0 2 0
0 5001000 1500 0 500 1000 1500

x (m) x (m)
1000 1000
4
y (m)
y (m)
500 4 500
5
0 0
0 500 1000 1500 0 500 1000 1500

x (m) x (m)
Figure 5.5: Time-varying subgraphs of waterborne AGVs

(a) ADMM
100
computation time (s)

iterations
computation time
200
iterations
50
0 0
100 200 300 400 500 600 700 800 900
Time (s)
(b) fast ADMM
0.4
computation time (s)

iterations
5 computation time
iterations
0.2
0 0
100 200 300 400 500 600 700 800 900
Time (s)
Figure 5.6: Computation time comparison.
(a)
primal residual
80 12
10 fast ADMM ADMM tolerances
60 8
40 6
4
20 2
0 2 4 6
10 20 30 40 50 60 70 80 90 100
(b)
dual residual
80 20
15
60 10
40 5
20 5 10
0
10 20 30 40 50 60 70 80 90 100
(c)
10
Cost difference
−10
−20
10 20 30 40 50 60 70 80 90 100
Iterations
Figure 5.7: Convergence behavior at time t = 142s.

14
x 10
2.5
ADMM
fast ADMM
2
Objective value
1.5
0.5
0
100 200 300 400 500 600 700 800 900
Time (s)
Figure 5.8: Objective comparison.
fast ADMM ADMM

160 160
iteration 1 iteration 1
140 140
120 iteration 4 120 iteration 40
safe distance safe distance
100 100
distances
distances
80 80
60 60
40 40
20 20
5 10 15 20 5 10 15 20
prediction step prediction step
Figure 5.9: Predicted distance evolution at time t = 142s.

5.6 Conclusions 79
ing zero. Overall costs of all time steps of the two distributed approaches are further shown
as Figure 5.8. Large costs are observed around the initial and two conflicting areas. The two
cost trajectories concur although fast ADMM based distributed PPF-ATA approximates the
original problem and converges within much fewer iterations. Figure 5.9 further shows how
global safety at time t = 142s is achieved and compares iterative distances between V2 and
V3 over the prediction horizon of the two controllers. ADMM based control approach sees
distances at several prediction steps below the minimum distance during the first iterations;
but by iterative communication and coordination, initial infeasible trajectories are adjusted
and driven above the safety line. For fast ADMM based control approach, since waterborne
AGVs are not solving local problems selfishly but also considering approximated collision
avoidance constraints which could be conservative during first iterations, the initial dis-
tances are actually above the safety line. Along with the convergence of primal and dual
residuals, however, distances converge to the real collision avoidance constraints yielding
global optimality as well.
5.6 Conclusions
In this chapter, we propose a distributed PPF-ATA control approach for multiple waterborne
AGVs carrying out ITT tasks. Computing in parallel within time-varying subgraphs as mod-
eled in Chapter 3 is achieved following the ADMM decomposition-coordination procedures.
Furthermore, possible poor convergence rates of the conventional ADMM are improved by
the proposed fast ADMM based distributed PPF-ATA. In our ITT case study in the port of
Rotterdam, comparing ADMM and fast ADMM at the same time step, ADMM requires 108
iterations with a total solver time of 4.65s before convergence while fast ADMM requires
only six iterations with 0.27s. Therefore, fast ADMM offers a more practical cooperative
distributed approach considering the short sampling time (0.5s in our simulations) of water-
borne AGVs. The proposed distributed PPF-ATA algorithm based on fast ADMM achieves
cooperative distributed waterborne AGVs and answers the third Key Research Question as
listed in Chapter 1.
The distributed control approach in this chapter is applicable to multiple waterborne
AGVs when environmental disturbances are perfectly known and the assignment of ITT
tasks to waterborne AGVs is given. We will consider cases when environmental distur-
bances are not perfectly known in Chapter 6 and a closed-loop scheduling and control design
in Chapter 7 based on the distributed algorithm proposed in this chapter.
Chapter 6
Cost-effective robust distributed

control of waterborne AGVs
In both Chapters 4 and 5, environmental disturbances are assumed to be known perfectly,

and waterborne AGV dynamics are modeled as (3.1) and (3.5). In this chapter, we consider
multiple waterborne AGVs modeled as (3.1) – (3.2) and (3.6) maneuvering in uncertain
environments with not perfectly known disturbances.
The research discussed in this chapter is based on [151, 152]
6.1 Introduction
The problems of a single waterborne AGV carrying out an assigned ITT task and multiple
waterborne AGVs with possible collision avoidance couplings are considered in Chapters 4
and 5, respectively. Control goals such as smooth path tracking, timing, energy efficiency,
and distributed cooperativeness have been achieved in deterministic scenarios, i.e., perfectly
known environmental disturbances. However, in practice, we can only roughly predict the
environmental influences by, e.g., weather forecast. For cases in which uncertainties in such
predictions exist, maintaining overall safety and energy efficiency in executing ITT tasks
with multiple waterborne AGVs remains an issue.
For waterborne AGVs that face uncertainties, besides the control goals in deterministic
scenarios, robust satisfaction of system constraints due to physical limitations and collision
avoidance is critical. Ideally, it is desirable to have constraint satisfaction for all possible
realizations of uncertainties modeled as (3.6) with stochastic characteristics; in practice,
however, waterborne AGVs may still fail at those worst cases of the sea which are, though
possible, very rare. Moreover, system performance could be degraded dramatically if the
system needs to be robust to those cases that rarely happen. Therefore, a practical design
should increase safety levels at a cost-effective price.
In this chapter, we propose a cost-effective robust distributed MPC (RDMPC) approach
for multiple waterborne AGVs facing uncertain scenarios. The approach is cost-effective in
the sense that the overall system robustness level and the associated price of robustness are
explicitly optimized considering system and uncertainty characteristics. In particular, we
81
82 6 Cost-effective robust distributed control of waterborne AGVs
maximize the robust probability of uncertainties while minimizing the nominal cost with
tightened constraints dependent on uncertainty bounds as in tube-based MPC [78]. The
problem is still decomposed and coordinated following the Alternating Direction Method
of Multipliers (ADMM) achieving cooperative parallel distributed control of coupled water-
borne AGVs as in Chapter 5. However, since probabilistic distributions are approximated
by introducing binary variables, the convexity assumptions for ADMM convergence do not
necessarily hold. Therefore, we propose an efficient integrated branch & bound (B&B)
and ADMM algorithm that solves the cost-effective RDMPC problem. The algorithm ex-
ploits the special ordered probability sets conducting smart search in B&B and integrates
branching criteria with intermediate ADMM results for early termination.
The remainder of this chapter is organized as follows. In Section 6.2, a cost-effective
centralized approach is proposed to solve the problem for multiple waterborne AGVs with
uncertainties. A linear programming (LP) relaxed RDMPC problem is first formulated and
solved in Section 6.3, and the exact RDMPC solutions are proposed in Section 6.4. In
Section 6.5, simulation experiments and results are presented and discussed, followed by
conclusions of this chapter in Section 6.6.
6.2 Cost-effective robust centralized formulation

In this section, we propose a cost-effective robust centralized approach for coupled wa-
terborne AGVs in one subgraph Gs (k) = (Vs (k), Es (k)) based on tube-based MPC [78].
For notational simplicity, we consider Gs (k) as the only subgraph in G (k), and thus use
G (k) = (V (k), E (k)) for the group of coupled waterborne AGVs considered in this chapter.
The centralized control goals are:
• individual ITT task achievement;
• overall safety;
• overall minimal energy consumption; and
• overall cost-effective robustness.
Tube-based MPC [78], as reviewed in Chapter 2, solves closed-loop optimization prob-

lems by parameterizing the control policy with an open-loop control sequence and a local
feedback. Accordingly, system dynamics with uncertainties are partitioned into a nominal
part and an uncertain part. The nominal dynamics generate the tube center and the uncertain
dynamics use the local feedback steering uncertain trajectories towards the center formu-
lating tube cross sections. Controlled system trajectories for all possible realizations of
uncertainties then evolve around the tube center within the tube. One key advantage of tube
MPC is that system goals are achieved by solving a nominal MPC problem with tightened
constraints while actual system constraints are still satisfied.
6.2.1 Parameterized uncertainty bounds

One of the assumptions for tube-based MPC is that uncertainties are bounded so that con-
straints can be properly tightened. However, as modeled by (3.1) – (3.2) and (3.6) in Chap-
6.2 Cost-effective robust centralized formulation 83
3.5
Bounds on uncertainties 3
2.5
1.5
0.5
0
0 0.2 0.4 0.6 0.8 0.95 1
Probability
Figure 6.1: Probability-bound of uncertainties following the standard normal distribution.
ter 3, waterborne AGVs moving in open waters experience stochastic environmental un-
certainties with infinite support. Obviously, robustness against 100% of such uncertain-
ties is impossible. As shown in Figure 6.1 for a standard normal distribution, the bound
on uncertainties increases exponentially as the probability approaches one, and thus could
degrade system performance dramatically. A natural way that handles uncertainties with
infinite support is to be robust to only a certain probability of uncertainties. For uncertain-
ties b ∼ N(b̄, Σ) acting upon waterborne AGVs, uncertainty bounds are determined by the
inverse Gauss error function erf [1] as
√
z = b̄ + 2Σerf−1 (p), (6.1)
so that the probability of b ∈ [−z, z] is p. Figure 6.1 plots (6.1) with b̄ = 0 and Σ = 1.
System performance and safety depend on the uncertainty bounds z and thus probability
p. With large probability p, system performance or even feasibility in finding solutions
within physical limits is lost; with small probability p, a certain level of system safety is not
guaranteed. We hence define system robustness level based on probability p as follows:
Definition 6.1 (Robustness level) The system with uncertain stochastic uncertainties b ∼
N(b̄, Σ) is said to have robustness level p if the system is robust to uncertainties in a compact
set [−z, z], where z is defined as (6.1).
A practical design increases the system robustness level at a cost-effective price of being
robust. The cost-effective robust approach proposed in [151] uses flexible bounds in tight-
ening constraints and penalizes deviations of the corresponding probability from a desired
robustness level. The idea is to integrate and make explicit use of the known probabilistic
distributions relating probability and uncertainty bounds in online optimizations. Online
optimizations can become intractable using complex distribution functions as Gauss error
functions. We, therefore, approximate the distribution functions by:
p = a P, (6.2)
√
z = b̄ + Σ (aaZ ) , (6.3)
nb
∑ a = 1, a ∈ {0, 1} (6.4)
where Z ∈ √ R nb and P ∈ R nb are bound and probability vectors, respectively, and satisfy
P = erf(ZZ / 2); a is a binary vector that parametrizes the distribution function by relating
the probability and bound via (6.2) and (6.3), and guaranteeing that exactly one probability
or bound is selected via (6.4). The set of red dots in Figure 6.1 is an example of P and
corresponding Z with nb = 40; the selected (p, z) pair with p = 0.95 and z = 2 is an example
of the 21st element of a being one and all others being zero. For parametrized uncertainties
(6.2) – (6.4) with unknown bounds, we are now ready to design a tube-based robust MPC
control strategy for multiple waterborne AGVs.
6.2.2 Tube-based MPC for successively linearized models with para-

menterized bounded uncertainties
We consider the successively linearized uncertain waterborne AGV models (3.13) – (3.15),
(3.18). The uncertain system states of each waterborne AGV p ∈ V (k) are partitioned as
x p (i|k) = x0p (i|k) + ∆x̄x p (i|k) + ∆x̃x p (i|k), (6.5)
where ∆x̄x p is the nominal perturbation state and ∆x̃x p the deviation of the actual perturbation
state ∆xx p (defined as (3.13)) from ∆x̄x p . Likewise, the control input is partitioned as:
u p (i|k) = u 0p (i|k) + ∆ūu p (i|k) + ∆ũu p (i|k). (6.6)
The nominal system dynamics then evolve as:
x̄x p (i + 1|k) = x 0p (i|k) + A p (i|k)∆x̄x p (i|k) + B p (i|k)∆ūu p (i|k)
and the uncertain dynamics evolve as:
∆x̃x p (i + 1|k) = A p (i|k)∆x̃x p (i|k) + B p (i|k)∆ũu p (i|k) + E p (i|k)∆b(i|k), (6.7)
with ∆x̃x p (0|k) = 0 since x̄x p (0|k) = x 0p (0|k) = x p (k). The certain part of b, i.e., b̄ has been
incorporated in x 0p (i|k) in calculating x 0p and thus in the nominal dynamics. By (6.3), the
uncertain part is then bounded as ∆b(i|k) ∈ W (i|k) where W (i|k) is a time-varying compact
set with origin in its interior explicitly dependent on a (i|k) as:
h √ √ i
W (i|k) =: − Σ (aa(i|k)ZZ ) , Σ (aa(i|k)ZZ ) . (6.8)
Incorporating feedback in predictions and optimizing over control policies rather than
control sequences is necessary when uncertainties exist to reduce conservativeness. How-
ever, optimizing over arbitrary feedback policies is practically intractable. Similarly as
tube-based MPC [78], we employ an affine feedback control policy as:
∆ũu p (i|k) = K p (i|k)∆x̃x p (i|k), (6.9)

where K p (i|k) is a time-varying feedback gain that needs to be calculated online. Par-
ticularly, for time-varying uncertain dynamics (6.7), we solve K p (i|k) as a finite horizon
unconstrained time-varying LQR controller [111] as:
• set P p (Np |k) := Q f ;
• for i = Np , Np − 1..., 1
Q + A Tp (i|k)P
P p (i − 1|k) =Q B p (i|k) R + B Tp (i|k)
P p (i|k)B
B p (i|k))−1 B Tp (i|k)P
P p (i|k)B P p (i|k)A
A p (i|k); (6.10)
• for i = 0, 1, ..., Np − 1
−1
B p (i|k) B p (i|k)T P p (i + 1|k)A
K p (i|k) = − R + B p (i|k)T P p (i|k)B A p (i|k), (6.11)
where Q , Q f and R are state cost, terminal state cost, and input cost matrices, respectively,
of the time-varying LQR controller, being the same for all waterborne AGVs. Then (6.7) in
closed-loop is:
∆x̃x p (i + 1|k) = A K p (i|k)∆x̃x p (i|k) + E p (i|k)∆b(i|k) (6.12)
with A K p ( j|k) = A p (i|k) + B p (i|k)K K p (i|k). Denote the set for uncertain perturbation states
as X̃ p , i.e., ∆x̃x p (i|k) ∈ X̃ p (i|k), we further have:
X̃ p (i + 1|k) := A K p (i|k)X̃ p (i|k) ⊕ E p (i|k)W (a(i|k)) (6.13)
with X̃ p (0|k) = {00}. The operator ⊕ defines the Minkowski set sum: A ⊕ B := {a + b|a ∈
A , b ∈ B }. The sizes of sets {X̃ p (i|k)} are expected to be smaller than those calculated from
unstable pairs of (AA(i|k), B (i|k)) since A K (i|k) is now stable by design. Moreover, {X̃ p (i|k)}
explicitly depends on the bounds of uncertainty sets W (i|k), and thus is also parametrized
by a(i|k). On the one hand, system performance desires small uncertainty sets; on the other
hand, it is necessary the system robustness level approaches one.
Following (6.5) and (6.13), the state tube with {xx0p (i|k)+∆x̄x p (i|k)} as centre and {X̃ p (i|k)}
as cross sections is defined as:
X p (i|k) := x 0p (i|k) + ∆x̄x p (i|k) ⊕ X̃ p (i|k),

(6.14)
and likewise control input tube is defined as:
U p (i|k) := u 0p (i|k) + ∆ūu p (i|k) ⊕ K p (i|k)X̃ p (i|k).

(6.15)
System constraints imposed by (3.3) are local convex constraints on states and control in-
puts, indicated by Cx p and Cu p , respectively, for waterborne AGV p. Coupling collision
avoidance constraints (3.8) between waterborne AGV p and q are non-convex and are later
convexified as (3.19) indicated by Cx p,q . Then, the centralized problem, termed Problem
1, for multiple waterborne AGVs is readily formulated to achieve the goals listed at the
beginning of this section as:
min ∑ J p (xx p (k), ∆ūu p (i|k)) + Qa k11 − p(i|k)k1 (6.16)

∆ūu,aa
p∈V (k)
subject to
X p (i|k) ⊆ Cx p ∀p ∈ V (k), (6.17)

U p (i|k) ⊆ Cu p ∀p ∈ V (k), (6.18)
(X p (i|k), Xq (i|k)) ⊆ Cx p,q ∀e p,q = 1, (6.19)
P,
p(i|k) = a (i|k)P (6.20)
nb
∑ a (i|k) = 1, a (i|k) ∈ {0, 1} (6.21)
2 2
where J p (xx p (k), ∆ūu p (i|k)) = x ep (k) w + k∆uu p (k)k2w 2 + x νp (k) w is the nominal local
1 3
convex cost function. The three terms in J p are formulated so that path following errors,
control input changes, and kinetic energy consumption are minimized as in Chapter 4 to
achieve smooth path tracking, energy efficiency, and arrival time awareness for ITT. Cost
function (6.16) also penalizes the deviation of the overall robustness level from one in the
last term where Q a is a weight parameter.
6.2.3 Implementations
Solving Problem 1 requires set computations ⊕ which are time consuming, and due to the
time-varying nature of the system (3.13) – (3.15), (3.18) and constraint (3.19), set computa-
tions are necessarily conducted online. Therefore, implemented tube-based MPC treats each
constraint separately. Suppose that each waterborne AGV p ∈ V (k) has Ix state constraints
and Iu control input constraints. For j = 1, ..., Ix ,
C pj (i|k)xx p (i|k) = C pj (i|k) x 0p (i|k) + ∆x̄x p (i|k) + ∆x̃x p (i|k) 6 d pj (i|k).

It is sufficient to tighten the nominal term via an offset defined as the out-bounding of the
uncertainty term:
Bxj p (i|k) = max C pj (i|k)∆x̃x p (i|k) ∆x̃x p (i|k) ∈ X̃ p (i|k)

a (i|k)
= max C pj (i|k)Θ

Φ p (i|k)|∆b(i|k) ∈ W (i|k)
Θ p (i|k)Φ (6.22)
a(i|k)
with
" #
i−1 i−1
Θ p (i|k) = ∏ A Kp (n|k)EE p (0|k) ··· ∏ E p (i − 2|k) E p (i − 1|k)
A Kp (n|k)E
n=1 n=i−1
T
and Φ (i|k) = ∆bT (0|k) ∆bT (1|k) · · · ∆bT (i − 1|k) . For general compact sets defined as

√
(6.8) with bounds Σ (aa(i|k)Z Z ), a solution to (6.22) is guaranteed to exist [67]. Moreover,
for the structured norm bounded uncertainties in our case, explicit maximization based on
the duality norm [67] is applicable as:
Bxj p (i|k) = C pj (i|k)Θ T (i) 1

Θ p (i|k)T (6.23)
j
= max C p (i|k)Θ T (i)∆bbt (i|k)|k∆bbt (i|k)k∞ 6 1 ,
Θ p (i|k)T
∆bbt (i|k)
where ∆bbt (i|k) = T −1 (i)∆bb(i|k) and T (i) is a diagonal translation matrix

T (i) √ 0
T (i + 1) = (6.24)
0 Σ (aa(i|k)Z
Z)
The possibility of explicit maximization avoids solving the programming problem (6.22)
for each constraint online which is then tightened as:
C pj (i|k)∆x̄x p (i|k) 6 d pj (i|k) −C

C pj (i|k)xx0p (i|k) − Bxj p (i|k). (6.25)
Control input constraints (6.18) are dealt with in a similar way. For j = 1, ..., Iu :
H pj (i|k)uu p (i|k) = H pj (i|k) u 0p (i|k) + ∆ūu p (i|k) + ∆ũu p (i|k) 6 h pj

Tightened constraints on nominal control inputs can be obtained as:
H pj (i|k)∆ūu p (i|k) 6 h pj (i|k) − H pj (i|k)uu0p (i|k) − Buj p (i|k). (6.26)
with tightening offsets
Buj p (i|k) = H pj (i|k)K

K p (i|k)Θ T (i) 1 .
Θ p (i|k)T (6.27)
Suppose there are Ic coupling constraints in (6.19), and for j = 1, ..., Ic ,

j j
r p (i|k)
E p (i|k) E q (i|k) 6 f p,q (i|k), (6.28)
r q (i|k)
where p ∈ V (k), q ∈ V (k) and e p,q = 1. Then (6.28) can also be tightened as in (6.25) but
involving tube information (nominal trajectory and bounds) from both coupled waterborne
AGVs p and q. Tightened constraints are on nominal perturbation position states ∆r̄r p (i|k)
and ∆r̄r q (i|k) as:

j j
∆r̄r p (i|k)
E p (i|k) E q (i|k) 6 f p,q (i|k)− (6.29)
∆r̄r q (i|k)
r 0p (i|k)

− B pj (i|k) − Bqj (i|k),
j
E p (i|k) E qj (i|k)
r 0q (i|k)
with tightening offsets

B pj (i|k) = E pj (i|k)Θ T (i)
Θ p (i|k)T 1
(6.30)
and
Bqj (i|k) = E qj (i|k)Θ T (i) 1 .
Θq (i|k)T (6.31)
Define the tightened state, control input, and collision avoidance constraint sets as
Cx p (aa(k)), Cu p (aa(k)), and Cr p,q (aa(k)), respectively, then constraints (6.17), (6.18), and
(6.19) are implemented as
∆x̄x p (i|k) ∈ Cx p (aa(k)) ∀p ∈ V (k), (6.32)

∆ūu p (i|k) ∈ Cu p (aa(k)) ∀p ∈ V (k), (6.33)
(∆r̄r p (i|k), ∆r̄r q (i|k)) ∈ Cr p,q (aa(k)) ∀e p,q = 1, (6.34)
Note that compared with (5.3) – (5.5) for the deterministic case in Chapter 5, (6.32) – (6.34)
depend explicitly on a (k). Problem 1 is a mixed integer programming (MIP) problem that
involves Np nb binary variables. A sensible simplification is to fix the probability over Np
with nb binary variables and to replace (6.21) by
∑ a (k) = 1, a (k) ∈ {0, 1}nb . (6.35)
The implemented centralized problem with objective (6.16) and constraints (6.32), (6.33),
(6.34), and (6.35) is termed Problem 2. Problem 2 is coupled due to
• pairwise collision avoidance constraints (6.34);
• coupling cost for overall robustness level.
Chapter 5 proposes a distributed approach based on ADMM for multiple waterborne

AGVs in deterministic cases. However, due to the introduction of binary variables a (k), the
convexity assumptions for ADMM convergence do not necessarily hold. A relaxed cost-
effective RDMPC algorithm is proposed first in the next section for the centralized Problem
2; exact solutions are proposed in Section 6.4.
Remark : The cost-effective robustness in Problem 2 could be interpreted as a type of

“relax-penalize” soft-constrained approach with constraint violations being parametrized
explicitly by uncertainty distributions. This is important since constraint violations are pe-
nalized considering the intrinsic interactions between system and uncertainty properties in-
stead of being penalized equally. For stochastically distributed uncertainties, penalizing
already very small deviations of probability to the desirable one could dramatically de-
grade system performance and even cause controller failures. In other words, it is not cost-
effective to sacrifice unnecessarily large system performance to increase little robustness
level. 2
6.3 Relaxed cost-effective RDMPC with ADMM

The main advantage of ADMM is that parallel distributed decision making still achieves
overall feasibility and optimality. However, the cost-effective robust centralized Problem
2 involves binary variables in both the cost function and constraints, which does not nec-
essarily guarantee convergence of ADMM iterations. In this section, we propose a relaxed
cost-effective RDMPC problem using ADMM. This problem will act as the node problem
in an efficient B&B search procedure to be proposed in Section 6.4.
6.3.1 Derivation of ADMM for the relaxed RDMPC problem

The binary variable constraints (6.35) of the original centralized Problem 2 are first relaxed
as:
∑ a (k) = 1, 0 6 a (k) 6 1. (6.36)
The relaxed cost-effective problem with (6.36) is termed Problem 3. Besides the collision
avoidance constraints (6.34) as in the deterministic case, waterborne AGVs with uncertain-
ties need to reach consensus also on an overall robustness level depending on a (k).
6.3 Relaxed cost-effective RDMPC with ADMM 89
The same as that in Chapter 5, one of the waterborne AGVs in graph G (k) is assigned
as the graph coordinator. For each waterborne AGV p ∈ V (k), we introduce a copy of
the bound variables as âa p (k) and a copy of the perturbation nominal position variables as
∆r̂r p (i|k). Then Problem 3 can be rewritten as:

min ∑ Jp (xx p (k), ∆ūu p (i|k)) + Qa /nv k1 − âa p (k)PPk1
∆ūu(i|k),aa0p (i|k)
(6.37)
p∈V (k)
subject to ∀p ∈ V (k)
∆x̄x p (i|k) ∈ Cx p (âa p (k)), (6.38)
∆ūu p (i|k) ∈ Cu p (âa p (k)), (6.39)
∑ âa p (k) = 1, 0 6 âa p (k) 6 1, (6.40)
(∆r̂r p (i|k), ∆r̂r q (i|k)) ∈ Cr p,q (âa p (k), âaq (k)), (6.41)
âa p (k) = a (k), (6.42)
∆r̂r p (i|k) = ∆r̄r p (i|k), (6.43)
where convex constraint sets Cx p (âa p (k)), Cu p (âa p (k)), and Cr p,q (âa p (k), âaq (k)) now explicitly
depend on local copies âa p (k) of a (k) due to local translation matrices instead of (6.24):

T p (i) √ 0
T p (i + 1) = .
0 Σ (âa p (i|k)Z
Z)
The augmented Lagrangian [12] that relaxes (6.42) and (6.43) is:
Pk1 + λ p,a (k)T λ p,r (i|k)T

Lρ (k) = ∑ Jp (xx p (k), ∆ūu p (i|k)) + Qa /nv k1 − âa p (k)P
p∈V (k)
(6.44)
2!
a (k) − âa p (k) a (k) − âa p (k)
+ ρ/2 + IC (k),
∆r̄r p (i|k) − ∆r̂r p (i|k) ∆r̄r p (i|k) − ∆r̂r p (i|k) 2
where λ p,a (k) ∈ R nb and λ p,r (i|k) ∈ R 2 are dual variables with respect to (6.42) and (6.43),
respectively; ρ is the augmented Lagrangian parameter. The coupling collision avoid-
ance constraint (6.41) has been replaced by the non-differential indicator function IC for
Cr p,q (âa p (k), âaq (k)) as: (
0, for (6.41)
IC (k) =
∞, otherwise.
j j
Given initial values λ p,a (k), λ p,r (i|k), a j (k), and ∆r̂r pj (i|k) at j = 1, the ADMM decomposition-
coordination at each iteration j consists of the following three steps:
Step 1: each waterborne AGV p ∈ V (k) solves the following local problem updating copied
uncertainty bounds âa p (k) and original perturbation position states ∆r̄r p (k):
∆ūu pj+1 (i|k), âa pj+1 (k), ∆rr pj+1 (i|k) = argmin J p + Qa /nv k1 − âa p (k)P

Pk1 +
2
a j (k) − âa p (k)
h
i −âa (k)
j T j T p
λ p,a (k) λ p,r (i|k) + ρ/2 (6.45)
∆r̄r p (i|k) ∆r̄r p (i|k) − ∆r̂r pj (i|k) 2
subject to (6.38) – (6.40).
Step 2: the graph coordinator solves the following problem updating original uncertainty
bound variables a (k) and copied perturbation position states ∆r̂r p (i|k):
h i
j j a (k)
a j+1 (k), ∆r̂r pj+1 (i|k) = argmin IC (k) + ∑
T T
λ p,a (k) λ p,r (i|k)
−∆r̂r p (i|k)
p∈V (k)
(6.46)
" # 2
a (k) − âa pj+1 (k)
+ρ/2 j+1
.
∆r̄r p (i|k) − ∆r̂r p (i|k)
2
Step 3: each waterborne AGV p ∈ V (k) updates local dual variables λ p,a (k) and λ p,r (i|k) as:
" # " # " #
j+1 j
λ p,a (i|k) λ p,a (k) a j+1 (k) − âa pj+1 (k)
j+1 = j +ρ j+1 . (6.47)
λ p,r (i|k) λ p,r (k) ∆r̄r p (i|k) − ∆r̂r pj+1 (i|k)
Step 1 and Step 3 are both carried out in parallel on board of each waterborne AGV
p ∈ V (k). The coordinator problem at Step 2 can be split further into two problems:
1. A global variable consensus problem [12] implemented as:

j
a j+1 (k) := 1/nv ∑ âa pj+1 (k) + λ p,a (k)/ρ , (6.48)
p∈V (k)
2. A problem of determining the Euclidean projections onto Cr p,q (âa p (k), âaq (k)) imple-
mented as:
2
j+1
∆r̂r pj+1 (i|k) := argmin ∑ ∆r̂r p (i|k) − ∆r̄r pj+1 (i|k) + λ p,r (i|k)/ρ (6.49)
2
p∈V (k)
subject to (6.41).
Note that in addition to the updated information on bound uncertainties and position states,
local tightening offsets (6.30) and (6.31) are necessary for the coordinator to formulate
Cr p,q (âa p (k), âaq (k)) at Step 2. Iterations are then alternating between the coordinator and
waterborne AGVs until consensus constraints (6.42) and (6.43) are satisfied, illustrated as
Figure 6.2, implying that
• all waterborne AGVs have agreed on the uncertainty bound parametrized by a (k);
and
• local solutions are also satisfying collision avoidance coupling (6.34).
6.3.2 Convergence analysis and stopping criteria

Similar to the convergence analysis in Chapter 5 for deterministic scenarios, convergence
analysis of the ADMM-based distributed algorithm for the relaxed cost-effective RDMPC
problem is based on an assumption for the relaxed centralized problem:
6.3 Relaxed cost-effective RDMPC with ADMM 91
Waterborne AGVs Coordinator
aˆ p
rp
p a
aˆq rˆp
rq
rˆq
q
rˆr
aˆr
rr
Figure 6.2: Agreement between waterborne AGVs and the coordinator.
Assumption 6.1 The relaxed cost-effective robust centralized problem Problem 3 is feasi-
ble.
Then the relaxed cost-effective RDMPC problem has the following convergence prop-
erties:
Proposition 6.1 (Convergence of the relaxed cost-effective RDMPC) Under Assumption

6.1, the following convergence is achieved as iteration j → ∞:
1. Primal feasibility, i.e., for each waterborne AGV p ∈ V (k), âa pj (k) → a j (k) and ∆r̂r pj (i|k)
j
→ ∆r̄r p (i|k).
2. Objective convergence, i.e., primal objective value of (6.37) approaches optimality.

j j
3. Dual variable convergence, i.e., ∀p ∈ V (k), λ p,a (k) and λ p,r (k) approach dual opti-
∗ ∗
mal points λ p,a (k) and λ p,r (k), respectively.
The proof of Proposition 6.1 follows the proof for Proposition 5.1, and is skipped in this
chapter.
Stopping criteria can also be extended from those in Chapter 5 as:
a j (k) − âa pj (k)

j
r (k) = ∑ j 6 εpri , (6.50)
p∈V (k)
∆r̄r p (i|k) − ∆r̂r pj (i|k) 2
a j+1 (k) − a j (k)

s j (k) = ∑ ρ 6 εdual , (6.51)
∆r̂r pj+1 (i|k) − ∆r̂r pj (i|k)
p∈V (k) 2
Algorithm 6.1 Relaxed cost-effective RDMPC: processed in parallel by each waterborne

AGV p ∈ G (k)
j j
1: initializes λ p,a (k), λ p,r (i|k), a j (k), and ∆r̂r pj (i|k) at j = 1;
2: loop
j+1
3: computes âa pj+1 (k), ∆rr p (i|k) as (6.45);
j+1 j j
4: sends âa pj+1 (k), ∆rr p (i|k), λ p,a (k) and λ p,r (i|k) to the coordinator;
5: repeat
6: wait;
7: until a j+1 (k), ∆r̂r pj+1 (k) arrive;
j+1 j+1
8: computes λ p,a (k) and λ p,r (i|k) as (6.47);
9: j + 1 → j;
10: end loop
Algorithm 6.2 Relaxed cost-effective RDMPC: processed by the coordinator waterborne

AGV of G (k)
1: repeat
2: repeat
3: wait;
j+1 j j
4: until âa pj+1 (k), ∆rr p (i|k), λ p,a (k) and λ p,r (i|k) arrive;
j+1 j+1
5: computes a (k), ∆r̂r p (i|k) as (6.46);
6: broadcasts a j+1 (k), ∆r̂r pj+1 (i|k) to ∀p ∈ V (k);
7: until stopping criteria (6.50) – (6.51) are met.
with εpri and εdual defined as:
εpri = 2nv N p + nb εabs + εrel max

p
 

a j j

 â p (k) a (k)
, ∑ ,
 p∈∑
V (k)
∆r̄r j
p (i|k) 2 p∈V (k)
r pj (i|k)
∆r̂ 2

" #
j
dual
p abs rel λ p,a (k)
ε = 2nv N p + nb ε + ε ∑ j .
p∈V (k)
λ p,r (i|k)
2
Algorithms 5.1 and 5.2 are extended as Algorithms 6.1 and 6.2 processed by AGVs
and the graph coordinator, respectively. However, output a j (k) from the relaxed problem
generally contains fractional elements, which is infeasible for the original Problem 2. We
next exploit the problem structure and propose an efficient integrated B&B and ADMM
based cost-effective RDMPC method to retrieve exact solutions.
6.4 Efficient exact cost-effective RDMPC

In this section, we derive an efficient integrated B&B and ADMM algorithm that retrieves
exact solutions from the Relaxed problem as presented in the previous section. Cooperative
6.4 Efficient exact cost-effective RDMPC 93
distributed computations as well as overall feasibility, optimality and cost-effective robust-

ness are still achieved. The approach exploits special ordered sets (SOS) [6] conducting
smart search in B&B, and integrating branching criteria with intermediate ADMM results
for early termination of iterations.
6.4.1 B&B with SOS1

Generic B&B is an exact solution paradigm that is the core for virtually any modern soft-
ware solving MIP problems [141]. B&B uses a search tree to implicitly enumerate possible
solutions by recursively partitioning the solution space into subspaces and pruning solution
subspaces that preclude the optimal solution. Active tree nodes are subspace problems that
are relaxed and solved as Algorithm 6.1 and 6.2 to provide lower bounds (for minimiza-
tion problems) that are then compared with the current upper bound, i.e., the incumbent
objective. Denote the incumbent objective as J ∗ , the current tree node objective JLP
∗ with
∗ ∗

an optimal solution ∆ūu p (i|k), a (k) , then three cases arise after solving the relaxed cost-
effective RDMPC problem:
∗ > J ∗ , then prune the node;
• Case 1: JLP
∗ 6 J ∗ and a ∗ (k) is integer, then set J ∗ = J ∗ , ∆ū
u∗p (i|k), a ∗ (k) as the

• Case 2: JLP LP
incumbent solution, and prune the node;
∗ 6 J ∗ and a ∗ (k) is fractional, then generate descendants of the current
• Case 3: JLP
node.
Branches in first two cases are both pruned either by bound or by optimality. The third
case requires partitioning the current solution space or subspace further into subspaces gen-
erating descendant nodes. Generic B&B branches on fractional variables ai , i ∈ I with index
set I = {1, 2, ..., nb } by adding branching constraints as ai = 0 and ai = 1 to the two descen-
dant nodes, respectively. Search strategies such as depth first, breadth first or best first etc.
[141] then decides the order in which the active nodes are to be processed by Algorithm 6.1
and 6.2 and matching the three cases to either obtain a better solution or verify the optimal-
ity of the current feasible solution. In such procedures, each variable may be branched on
multiple times during the search. Moreover, since the constraints on binary variables are
in the form of (6.35), the branch with ai = 0 makes little progress in improving the bounds
while the branch with ai = 1 immediately returns an integer solution; the search tree is un-
balanced. We next exploit the structure of the model and branch on set of variables to have
balanced search trees.
Constraints as (6.35) are called generalized upper bound (GUB) constraints [141] where
exactly one variable takes value 1 and all others are 0. Instead of branching on variables,
branching on the GUB constraints with ∑ ai = 1 for one subspace and ∑ ai = 0 the other
i∈I 0 i∈I 0
results in a more balanced search tree; I 0 ⊆ I is a subset of binary variable indices. In
Problem 2, constraints (6.35) enforce that exactly one bound or one probability is selected
for the uncertainties. Furthermore, since ai with increasing index i corresponds to a larger
probability as (6.2) or larger bound as (6.3), there is a logical ordering of variables ai , i ∈ I
which are called special ordered sets of type 1 (SOS1) first proposed in [6]. The vectors de-
ciding on the logical ordering, i.e., P or Z , are termed the reference row. The corresponding
branching method is then called SOS1 branching. Since we put heavy penalty on deviations
of the robustness level p to one, the probability vector P is set as the reference row in our
case. The idea is to take advantage of the conventionally overlooked fractional solutions in
the third case of B&B and combine them with the reference row to branch in a smarter way.
Suppose the current solution a ∗ (k) to the relaxed cost-effective RDMPC problem is
fractional and satisfies the SOS1 constraint (6.35):
a∗1 + a∗2 + ... + a∗nb = 1 (6.52)
Since the reference row P is ordered as p1 6 p2 6 ... 6 pnb ,the SOS1 branching suggests
the branch point as:
i∗ := argmin {pi }, (6.53)
i∈I I 0
where the index set I0 is computed as:
I 0 := {i|pi 6 ∑ a∗i pi }. (6.54)

i∈I
The solution space can now be partitioned into two subspaces by adding ∑ ai = 1 for
i∈I 0
one subspace and ∑ ai = 0 or equivalently, ∑ ai = 1 for the other subspace. Gener-
i∈I 0 i∈I I 0
ally, when the relaxed problems at the two subspaces are solved, integer feasible solutions
are obtained either for the branch with ∑ ai = 1 as amax{I 0 } = 1 or for the branch with
i∈I 0 i
∑ ai = 0 as amin{I I 0 } = 1. If not, the subspace with fractional solutions can apply the
0
i∈I I i
above SOS1 branching procedure recursively.
6.4.2 Integrated B&B with SOS1 and ADMM

Besides the special structure SOS1 constraints that facilitate efficient branching, two obser-
vations are noteworthy for further improving the efficiency of the B&B search procedure
with intermediate results of ADMM iterations.
First, Algorithms 6.1 and 6.2 need not necessarily run to convergence for the lower
bound JLP∗ . In B&B, J ∗ is compared with the incumbent objective J ∗ in three branching
LP
cases. However, we also save primal and dual objectives before convergence that can be
used for the comparisons to achieve early termination of ADMM as well as early pruning
j j
of a B&B node. Define JP as the primal overall objective of (6.37) and JD the dual overall
j j
Lagrangian objective L0 (k). Then, we track an intermediate objective value JLP at iteration
j as:
j j
• JLP = JP if primal feasibility (6.50) and dual feasibility (6.51) are not satisfied;
j j j
• JLP = max{JP , JD } if dual feasibility (6.51) is satisfied but primal feasibility (6.50) is
not satisfied;
j ∗ if both primal (6.50) and dual feasibility (6.51) are satisfied.
• JLP = JLP
The following proposition is then established for safe early termination of ADMM and
pruning of the node.
6.4 Efficient exact cost-effective RDMPC 95
Proposition 6.2 (Early termination and pruning) At iteration j of a node problem, i.e.,
j
Problem 3 solved by Algorithm 6.1 and 6.2, if JLP > J ∗ , then we can safely terminate ADMM
iterations and prune the node.
j
∗ where J ∗ is a lower bound of
Proof : Proposition 6.2 follows from the fact that JLP 6 JLP LP
j ∗ ∗ ∗
Problem 2. Whenever JLP > J , we also have JLP > J , therefore, iterations of ADMM can
be safely stopped and the current node can be pruned as Case 1 of B&B. If primal feasibility
j
is not achieved, the primal objective JP of a minimization problem with feasible solutions
by solving partially constrained problems at Step 1 is smaller than the optimal objective,
∗ . If dual feasibility is achieved, since the dual problem is a maximization problem
i.e., JLP
j
of which the optimal dual value JD∗ > JD provides an lower bound for JLP ∗ , i.e., J ∗ 6 J ∗ , we
D LP
j
have JD 6 JLP ∗ . Proposition 6.2 is established. 2
Secondly, we observe that if the solution for binary variables are integer, the consensus
constraints (6.42) are satisfied within a small number of iterations; otherwise if fractional,
they stay fractional for possible large number of iterations before convergence. This moti-
vates a special treatment of the stopping conditions in ADMM. Specifically, for the second
j
case of B&B, if âa p are integer and the stopping criteria for consensus constraint (6.42) are
j
satisfied, the local variables âa p are set as ãa j during all the following iterations until the
j
stopping criteria for consensus constraint (6.42) are also satisfied. If, however, a p are frac-
tional, since the solution space of this node is to be further partitioned, an optimal objective
∗ and an optimal solution ∆ū
value JLP u∗p (i|k), a ∗ (k) are not candidates for the ultimate exact
optimal solution anyway; instead, we are only interested in fractional solutions that could
provide collective information together with the reference row P to suggest a branching
point as (6.53) and (6.54). Therefore, we can safely terminate the ADMM iterations with
0 . The third case of B&B then follows the SOS1 branching procedure using
a small jmax
fractional solutions ãa j which are not necessarily optimal.
For waterborne AGVs coupled as G (k) = (V (k), E (k)), we then provide an efficient
exact cooperative distributed solution for the centralized cost-effective robust Problem 2.
Waterborne AGVs ∀p ∈ V (k) still process in parallel as Algorithm 6.1. The waterborne
j
AGV coordinator of G (k) processes the computations of JLP and three branching rules in
B&B using intermediate ADMM results. Algorithm 6.2 in this way is extended into Algo-
rithm 6.3.
Algorithm 6.3 Exact cost-effective RDMPC by the coordinator waterborne AGV of G (k)
1: Initialize J ∗ = ∞; mark the root node as active;
2: repeat
3: repeat
4: repeat
5: Wait;
j+1 j j
6: until âa pj+1 (k), ∆rr p (i|k), λ p,a (k) and λ p,r (i|k) arrive;
j+1 j+1
7: Computes a (k), ∆r̂r p (i|k) as (6.46);
8: Broadcasts a j+1 (k), ∆r̂r pj+1 (i|k) to ∀p ∈ V (k);
j+1
9: if JLP is computed according to (6.50) – (6.51) then
j+1
10: if Case 1: JLP > J ∗ thenw
11: prune the node and go to Line 1 of Algorithm 6.1;
12: end if
j+1 j
13: if Case 2: JLP 6 J ∗ and a p for p ∈ V (k) are integer and equal then
j+1
14: set a p = âa j+1
15: end if
j+1
16: if Case 3: JLP 6 J ∗ and âa j+1 is fractional and j = jmax
0 then
17: Mark the two descendants as active and go to Line 1 of Algorithm 6.1;
18: end if
19: end if
20: until (6.50) – (6.51)
are met.
∗ j j
21: Set J = JLP and ∆ūu p (i|k), ãa j (k) as the incumbent solution;
22: until no active nodes
Table 6.1: ITT tasks for waterborne AGVs
ITT Routes Origin-Destination Departure-arrival

tasks Berth → Berth (m) times (s)
1 → V1 1→2 (0, 0) → (411, 417) 0 → 401
2 → V2 2→1 (411, 417) → (0, 0) 0 → 401
3 → V3 3→4 (-50, 359) → (420, 52) 0 → 401

In this section, simulations are run to demonstrate the effectiveness of the proposed cost-
effective RDMPC strategy for waterborne AGVs that carry out ITT tasks while facing with
uncertainties. The proposed algorithm is applicable to any size of waterborne AGV fleets.
For simplicity, we consider scenarios that could illustrate how the algorithm works. Figure
6.3 shows the ITT scenario in the port of Rotterdam with three ITT tasks in Table 6.1 as-
signed to three waterborne AGVs which are denoted as V1 , V2 , and V3 . The proposed cost-
effective RDMPC algorithm is also applicable to scenarios with more waterborne AGVs
handling more ITT tasks. In Figure 6.3, there is one potential conflicting area where wa-
terborne AGVs may encounter one another, as the yellow circle in Figure 6.3 shows. The
port authority has predicted environmental disturbances but stochastic uncertainties exist in
the predictions. Waterborne AGVs are required to track the shortest reference path between
the origin and destination smoothly and arrive at the destination as punctual as possible.
Moreover, each waterborne AGV computes locally while achieves overall energy efficiency
and avoids collisions with others despite of environmental disturbances due to wind, waves,
and currents.
We assume homogeneous dynamics (3.1)–(3.2), (3.6) for three waterborne AGVs with
identical values for hydrodynamic parameters based on a small-scaled vessel model, CSII
[121]. Our simulation uses a sampling time of Ts = 0.6 s with a prediction horizon Np = 7.
Other controller parameters are set as follows: weight parameters w1 = diag([100, 50, 10]),
w 2 = I 3×3 , w 3 = 1, and Q p = 107 . Time-varying LQR parameters in (6.10) – (6.11) are set
Figure 6.3: ITT scenario for waterborne AGVs.
as: Q = diag( 100 20 20 0 0 0 ), Q f = diag( 104 104 104 0 0 0 ),

R = I 3×3 . Stopping criteria for ADMM convergence are set with absolute tolerance εabs =
2 × 10−3 and relative tolerance εrel = 10−4 . Stochastic uncertain predictions of environmen-
tal disturbances from the port authority are b ∼ N(b̄, Σ) with mean value b̄b = 343000N and
variances Σ = 219520N2 . Physical system constraints are: −0.5 m/s 6 u 6 1.8 m/s and
T
τ max = −ττmin = 1372000N 1372000N 72030000Nm .
The bound vector Z ∈ R n with nb = 21 is sampled with interval 0.2 from [0, 3.9] since for
3.9 3.9
standard normal distributions, erf( √ 2
)=0.9999. Then the probability vector P = erf( √ 2
).
We further assign a big value M = 104 and a probability 1 to Z and P , respectively, to
approximate the real function achieving ∞ by p = 1, z = M.
Algorithms are implemented in MATLAB 2014b [75] with optimization problems solved
by Cplex 12.51 [46] on a platform with Intel(R) Xeon(R) CPU E5-1620 v2 @3.70 GHz.
Simulations are run for carrying out the given ITT tasks using waterborne AGVs controlled
by the proposed cost-effective RDMPC.
6.5.1 Safely fulfilling ITT tasks

Optimal trajectories of the three waterborne AGVs carrying out their assigned ITT tasks are
shown in Figure 6.4a. Generally, all waterborne AGVs are able to track the shortest refer-
ence path between origin and destination pairs well. We observe two types of deviations
along the paths. One is the relatively small deviations outside of conflict zone due to the
existence of environmental uncertainties. The other is due to the encountering of water-
borne AGVs in the conflict zone where deviations from the reference path are necessary to
maintain a safety distance away from others. The trade-off among deviations, robustness
level, and safety has been optimized since all online optimizations are successfully solved.
Furthermore, waterborne AGVs arrive at their specified destinations punctually at 397s,
401s, and 401s, respectively, despite the possibly time consuming behaviors for collision
avoidance and environmental uncertainties.
The proposed cost-effective RDMPC contains a bunch of uncertain trajectories in a tube

that achieves both reference tracking and safety. The trajectory tubes of three waterborne
AGVs are shown in Figure 6.4b which zooms in the conflict area of Figure 6.4a. Actual
trajectories that are steered to nominal trajectories are guaranteed to stay within tubes if real
uncertainties are within bounded sets. Note that although trajectory tubes overlay spatially,
they do not overlay temporally at the same time. This is demonstrated by different posi-
tions of V1 , V2 , and V3 at the a same time t = 209s plotted in different shapes as shown in
Figure 6.4. Collisions are thus successfully avoided. Figure 6.5 further confirms overall
safety by showing that Euclidean distances for all pairwise waterborne AGVs are above the
safety distance line all the time. However, unlike deterministic scenarios where waterborne
AGVs approach to each other until the minimal possible distance, i.e., the safety distance,
is reached, the minimal distances in uncertain scenarios are a clear space above the safety
distance line. The spaces intuitively act as buffers accounting for uncertainties as to be
demonstrated later.
6.5.2 Convergence of distributed decision making

Waterborne AGVs with stochastic uncertainties compute in a distributed way and commu-
nicate to achieve overall satisfactions of coupling constraints and an overall cost-effective
robustness level by Algorithms 6.1 and 6.3. The efficiency of the algorithms lies in the appli-
cation of SOS1 branch and the integration of B&B and ADMM. Iterations with fractional
solutions are terminated before convergence, and iterations with integer solutions follow
ADMM stopping criteria (6.50) – (6.51). We show simulation results with integer solu-
tions to demonstrate how feasibility and optimality convergence in a relaxed cost-effective
RDMPC problem are achieved in this section, and illustrate the efficient B&B with SOS1
search procedure in Section 6.5.4.
At time step k = 35, three waterborne AGVs are coupled as graph G (k) = {V (k), E (k)}
with V (k) = {V1 ,V2 ,V3 } and E (k) = {(V1 ,V2 ), (V1 ,V3 ), (V2 ,V3 )}. The second relaxed RDMPC
problem, i.e., the second B&B node problem of this time step reaches integer solution agree-
ment (Line 14 in Algorithm 6.3) in the first iteration. Therefore, the remaining iterations
follow standard ADMM steps until convergence. The iterative satisfaction of coupling con-
straints is illustrated in Figure 6.6 which shows nominal distances over N p between V1 and
V2 . The safety distance line defines the actual minimal separation distance between two
waterborne AGVs. Nominal distances, however, are tightened due to the existence of un-
certainties by appending a distance buffer as the red dashed lines shows to the actual safety
distance. Waterborne AGVs with uncertainties controlled by the nominal control law are
then still stay a safety distance away from each other, as demonstrated in Figure 6.6. Dur-
ing the first iterations, nominal distances at several prediction steps are below the distance
buffer or even the safety distance; but by iterative communication and coordination, both V1
and V2 adjust their initial calculated trajectories which are driven above the distance buffer
in the end. The distances for (V1 ,V3 ), and (V2 ,V3 ) show similar convergence behaviors.
Figure 6.7 shows objective convergence along with iterations. In Figure 6.7(a), the
overall primal objective which is a sum of distributed calculated objectives from three wa-
terborne AGVs as (6.37) converges to the centralized objective at the same time step. Figure
j
6.7(b) demonstrates that the intermediate objective JLP defined in Section 6.4.2 converges
to the optimal distributed objective JLP ∗ from the below, which confirms the possibility of
D1
400
350
300
actual
250 nominal
tube bounds
y (m)
200 reference
V1 at t=201s
150 V2 at t=201s
V3 at t=201s
100
50 D3
0 D2
−50 0 50 100 150 200 250 300 350 400
x (m)
(a) Trajectories of three waterborne AGVs.
230
225
220
215
210
y (m)
205
actual
200
nominal
195 tube bounds
reference
190
V1 at t=201s
185 V2 at t=201s
V3 at t=201s
180
180 190 200 210 220 230
x (m)
(b) Trajectory tubes
Figure 6.4: Trajectories of waterborne AVGs for ITT

600
V1 and V2
V1 and V3
500
V2 and V3
60 safety distance
400 50
distances (m)
40
300 30
20
180 190 200 210
200
100
0
0 50 100 150 200 250 300 350
time (s)
Figure 6.5: Distances between waterborne AGVs.
90
80 iteration 1
iteration 10
iteration 30
70
iteration 60
iteration 136
60 safety distance
distances (m) V1 &V2
distance buffer
50
40
30
20
10
0
0 1 2 3 4 5 6 7
prediction step
Figure 6.6: Robust distance iterations between waterborne AGVs 1 and 2 at time step k =
35.
7 (a)
x 10
−2.9508
distributed
objectives
−2.9508
centralized
−2.9508
−2.9508
20 40 60 80 100 120
iteration
7 (b)
x 10
−2.9508
objectives
−2.9508 j
JLP
∗
JLP
−2.9508
−2.9508
20 40 60 80 100 120
Figure 6.7: Objective iterations at time step k = 35.
early termination as in Proposition 6.2. Overall feasibility and optimality are further indi-
cated by primal and dual residual convergence to specified tolerances in Figure 6.8. Primal
j
residuals being small means that the consensus (6.42) is achieved and local solutions ∆r̄r p (k)
are also satisfying coupling constraints (6.34). Dual residuals being small means that the
iterative coordinator solutions do not change much any more. Overall safety and optimality
are achieved when primal and dual residuals satisfy (6.50) and (6.51) at the final iteration
136.
6.5.3 Cost-effective robust tubes

This section explores the cost-effective robustness aspect of the proposed algorithm for co-
operative distributed waterborne AGVs. The main idea is to explicitly consider system and
uncertainty characteristics so that overall control performance including problem solvabil-
ity and overall safety are guaranteed. This is achieved by using varying robustness levels
with varying probabilities and uncertainty bounds as shown in Figure 6.9. Figure 6.9(a)
shows the optimal robustness levels, i.e., probabilities of environmental disturbances V1 can
be robust to. The deviations from 100% are necessary for the feasibility of online optimiza-
tions and deviations are as small as possible. All robustness levels are higher than 95% to
ensure high safety level. Note that when waterborne AGVs encounter in close proximity
between 100s – 300s, robustness levels slightly drop to account for possible large tracking
errors due to collision avoidance. This also showcases the optimal trade-off property of the
proposed algorithm. Subplot (b) shows the uncertainty bounds corresponding to the robust-
ness levels and uncertainties that actually impose on waterborne AGVs. Real uncertainties
are all contained in the optimally bounded sets, which demonstrates safety of high robust-
ness levels. Those extreme uncertainties beyond the bounds, though possible, are very rare
(a)
0.4
primal residuals primal tolerances
primal residual 0.3
0.2
0.1
0
20 40 60 80 100 120
(b)
1
dual residuals dual tolerances
0.8
dual residual
0.6
0.4
0.2
0
20 40 60 80 100 120
Iterations
Figure 6.8: Primal and dual residuals at time step k = 35.
to occur. Robustness against them could degrade system performance dramatically and is
thus deemed as not cost-effective. Throughout the simulation, infeasibility is not observed.
This might be because a large magnitude of the uncertainty has been taken care of by the
designed robustness. When uncertainties that do cause infeasibility issues in cost-effective
RDMPC, rarely though, safety/recovery procedures need to be activated.
Cost-effective robust satisfaction of coupling constraints has been shown in Figure 6.6
with nominal distances tightened by an extra distance buffer. Cost-effective robust satisfac-
tions of local constraints on speed and control inputs are shown in Figure 6.10 and Figure
6.11, respectively. The common patterns for Figure 6.10 and Figure 6.11 are that the actual
speed and control input trajectories are both well constrained within system limits denoted
by the red lines and are both a clear distance away from their corresponding limits. This
is the saved space for uncertainties, i.e., the price of robustness. Figure 6.10 and Figure
6.11 differ in that the actual and nominal trajectories for control input coincide while for
speed not. This is because MPC applies, in a calculated optimal sequence, only the first
control input which by (6.6) and (6.9) has no uncertainty involved. However, due to the
introduction of uncertainties via (6.9), predicted control inputs over 1, 2, ..., N p − 1 are nec-
essarily tightened. Figure 6.12 shows the nominal control input trajectories and tubes over
N p at one time step. The zero initial control input tightening offsets can also be clearly seen
in Figure 6.12. Similarly, Figure 6.13 shows the speed tubes over N p . Speed and control
input tubes evolve according to (6.14) and (6.15), respectively. The speed tube sizes ap-
proximately increase over N p due to set additions narrowing the feasible region for nominal
speeds. Therefore, nominal speeds decrease as the prediction step proceeds. However, on-
line optimization problems remain feasible due to the explicit integration of the uncertainty
bounds. Tightly fit tubes in Figure 6.10, Figure 6.13, and Figure 6.12 but still with a feasible
nominal solution illustrate this.
Robustness level (a)
99%
98%
97%
96%
95%
0 50 100 150 200 250 300 350
6 (b)
x 10
1
Uncertainties (N)
0.5
−0.5
uncertainty bounds real uncertainty
−1
0 50 100 150 200 250 300 350
time (s)
Figure 6.9: Varying robustness level: probability and uncertainty bounds.
2.5
1.5
1
u (m/s)
0.5
actual
nominal
0
limits
tube
−0.5
−1
−1.5
0 50 100 150 200 250 300 350 400
t (s)
Figure 6.10: Speed trajectories with tubes of the whole simulation.

6
Surge force (N) x 10
2
−2
0 50 100 150 200 250 300 350 400
6
x 10
2
Sway force (N)
−2
0 50 100 150 200 250 300 350 400
8
Yaw moment (Nm)
x 10
1
−1
0 50 100 150 200 250 300 350 400
time (s)
Figure 6.11: Control inputs of the whole simulation.
6
x 10
Surge force (N)
−2
1 2 3 4 5 6 7
6
x 10
2
Sway force (N)
−2
1 2 3 4 5 6 7
8
x 10
yaw moment (Nm)
−1
1 2 3 4 5 6 7
prediction steps
Figure 6.12: Predicted control inputs with tubes at one time step.
2.5
1.5
1
u (m/s)
0.5
−0.5
−1
−1.5
0 1 2 3 4 5 6 7 8
prediction steps
Figure 6.13: Predicted speed with tubes at one time step.
6.5.4 Efficient B&B search

Efficient B&B search retrieves exact solutions from relaxed RDMPC problems by exploit-
ing the special structure of SOS1 constraints (6.35) and the intermediate results of ADMM.
With the proposed cost-effective RDMPC algorithm, we observe at most two branching op-
erations throughout the simulation. Figure 6.14 and Figure 6.15 show the B&B search trees
with height one and two, respectively. In both figures, the tree nodes are attached with SOS1
constraints on subset of variables and the upper (U) and lower (L) bounds by solving this
node problem; the numbers left to the nodes indicate the sequence by which node problems
are dealt with. Note that the sum of the subset of variables in SOS1 constraints should equal
one, e.g., ∑21 21
i=1 ai = 1 for the top node in Figure 6.14, but ∑i=1 ai has been used for simplic-
ity. The division of the variable set at a parent node to the sets at two child nodes are based
on fractional solutions from the parent node problem by deciding the branching point as
(6.53) – (6.54). The upper bound, i.e., incumbent objective, is either updated with the better
bound produced by the current problem with integer solutions, or otherwise inherited from
the parent node. The lower bound is either the objective solving the relaxed cost-effective
RDMPC problem, or ∞ if the relaxed problem is infeasible.
Overall, the algorithm first searches a feasible solution, and then either improve this
solution or verify the optimality of this solution by implicitly enumerating all possible solu-
tions. E.g., for the tree in Figure 6.15, node 1 solves the relaxed Problem 3 and the fractional
solutions after 20 iterations suggest a branching point i∗ = 15. Therefore, ∑14 i=1 ai = 1 and
21
∑i=15 i a = 1 are added to the two child nodes, node 2 and node 5, respectively. In a similar
way as node 1, node 2 generates two child nodes of which node 3 has an integer feasible
solution. Since the optimal objective of node 3 is smaller than the incumbent objective,
i.e., its parent upper bound ∞, the incumbent objective and incumbent solution are updated.
X
21
U: Inf
1 ai
i=1 L: −2.98e+007
X
14
U: −2.97e+007 X
21
U: −2.97e+007
2 ai 3 ai
i=1 L: −2.97e+007 i=15 L: Inf
height = 1
Figure 6.14: B&B search tree with height 1.
X
21
U: Inf
1 ai
i=1 L: −2.98e+007
X
14
U: Inf
2 ai
i=1 L: −2.97e+007
X
13
U: −2.95e+007 U: −2.95e+007 X
21
U: −2.95e+007
3 ai 4 a14 5 ai
i=1 L: −2.95e+007 L: Inf i=15 L: Inf
height = 2
Figure 6.15: B&B search tree with height 2.
Problems at node 4 and node 5 are infeasible and are pruned immediately. Therefore, gen-
erally only one node needs to iterate to convergence in a search tree and thus making the
algorithm efficient.
6.6 Conclusions 107
6.6 Conclusions
In this chapter, we have proposed a cost-effective robust distributed MPC (RDMPC) control
approach for multiple waterborne AGVs that carry out ITT tasks in scenarios with uncer-
tainties. Environmental disturbances due to wind, waves, and currents are assumed to be
predicted by the port authority. However, there still exist stochastic uncertainties in the pre-
dictions, as modeled in Chapter 3. Each waterborne AGV computes locally those actions
that fulfill its assigned ITT task by tracking a reference path with arrival time awareness in
an energy-efficient way with a cost-effective price of being robust. In our simulations, all
robustness levels are higher than 95% to ensure high safety level. Moreover, overall mini-
mal cost, robustness level, and coupling collision avoidance constraints have been satisfied
by communicating within a neighborhood. The cost-effective robustness extends the tube-
based robust approach for bounded uncertainties to stochastic uncertainties with infinite
support by explicitly considering system and uncertainty characteristics. The cost-effective
RDMPC problem for coupled waterborne AGVs has been solved exactly and efficiently
by integrating a special type of B&B and ADMM. With the proposed efficient searching
strategy, at most two branching operations at each time step are required throughout the
simulation. Simulation results of an ITT case study in the port of Rotterdam illustrate that
the proposed cost-effective RDMPC algorithm is effective for controlling multiple water-
borne AGVs with uncertainties in carrying out ITT tasks, which answers the fourth Key
Research Question in Chapter 1.
So far, control problems for waterborne AGVs carrying out ITT tasks have been con-
sidered in Chapters 4 – 6. In all these problems, schedules regarding the assignment of
ITT tasks to waterborne AGVs and the routing of waterborne AGVs are assumed given. In
the next chapter, coordination considering both the control and a higher level scheduling
problems of waterborne AGVs will be discussed.
Chapter 7
Closed-loop scheduling and

control for autonomous Inter
Terminal Transport
In Chapters 4 – 6, effective controllers are developed given one ITT request per waterborne
AGV. The scheduling level decisions are assumed made and passed to the controllers acting
as the references that waterborne AGVs should track. The scheduling and control prob-
lems are solved in an open-loop way. This chapter considers the scheduling of ITT tasks for
multiple waterborne AGVs and proposes a closed-loop energy-efficient scheduling and con-
trol framework to realize an autonomous waterborne ITT system. The scheduling problem
is formulated as a pick-up and delivery problem (PDP) that considers speed optimization
and safe intervals between berthing time slots of different waterborne AGVs at the same
berth. Waterborne AGVs are controlled in a cooperative distributed way to fulfill assigned
schedules as in Chapter 5.
The research discussed in this chapter has been published in [153].
7.1 Introduction
In the literature, multi-vehicle scheduling problems are largely formulated as Vehicle Rout-
ing Problems (VRPs) [130] determining the assignment of vehicles to transport tasks and
the sequence of points possibly with temporal requirements that a vehicle should visit. How-
ever, assumptions are implicitly made that vehicles are dimensionless mass points finishing
assigned tasks as scheduled perfectly without consideration safety intervals between vehi-
cles. Moreover, the scheduling and control level problems are typically solved separately,
as illustrated in Figure 1.2. Sharing the common aims of making economical and environ-
mentally friendly decisions, scheduling and controlling waterborne AGVs given multiple
ITT requests are expected to achieve further benefits via a tighter interaction of these two.
In this chapter, we propose a closed-loop scheduling and control approach for a fleet
of waterborne AGVs to realize an autonomous ITT system. Closed-loop means that both
scheduling and control levels make decisions online based on system states measured at a
109
110 7 Closed-loop scheduling and control for autonomous Inter Terminal Transport
fast sampling rate. Decisions are still made hierarchically to guarantee tractability. More-
over, a new PDP scheduling model considering necessary time intervals between different
waterborne AGVs visiting a particular berth is proposed. Furthermore, we propose a par-
tial scheduling problem that is efficient to solve, and an interaction model that integrates
the scheduling and control problems. Solving the scheduling problem generates for each
waterborne AGV a sequence of terminals to visit to load or unload certain amount of con-
tainers sailing at an energy optimal speed arriving/departing at coordinated berthing times,
while still satisfying service time windows. Cooperative distributed model predictive con-
trol based on the fast ADMM algorithm of Chapter 5 is then adopted by the group of in-
volved waterborne AGVs to accomplish the schedules safely and accurately. The main
advantage of using a closed-loop scheme over an open-loop scheme is that real-time factors
such as unconsidered physical system limits, disturbances, and collision avoidance that are
difficult, if not impossible, to be integrated in a scheduling problem can be reflected timely
by the online updated schedules.
The remainder of this chapter is organized as follows. The overall problem statement for
an autonomous ITT system using waterborne AGVs is first introduced in Section 7.2. Then
in Section 7.3, the energy efficient scheduling problem with coordinated berthing times is
formulated. The closed-loop scheduling and control based on a real-time coupling speed
assignment problem and an interaction model are proposed in Section 7.4. In Section 7.5,
simulation experiments and results are presented, followed by conclusions of this chapter in
Section 7.6.
7.2 Problem statement

We consider an autonomous ITT system: a fleet of waterborne AGVs that handles a set of
emerging ITT requests to transport specified amounts of containers between specified ori-
gins and destinations within specified time windows autonomously in an energy efficient
way. Without loss of generality, two simplifications are made: 1) Each terminal has one
waterborne AGV berth. In practice, a terminal can have multiple berths, which can, how-
ever, be viewed as multiple pickup/delivery locations. Therefore, the problem formulation
is essentially not changed for one berth per terminal; and 2) ITT requests are decoupled
between different planning horizons so that requests arising within each planning horizon
are completed within that horizon.
For the scheduling problem, each waterborne AGV has a finite capacity which can ac-
commodate mixing containers from different requests. Besides, waterborne AGVs use “en-
vironmentally friendly” engines and perform “slow steaming” by cruising at lower speeds
if possible. There is no central depot for waterborne AGVs and they stay at the park lot
of the final service berth. Finally, waterborne AGVs are with certain dimensions and need
to keep a certain safety distance from others to avoid collisions. Collision avoidance is
achieved among moving waterborne AGVs by cooperative distributed control as presented
in Chapter 5 while waterborne AGVs visiting the same berth to perform loading or unload-
ing operations are yet to be coordinated. Figure 7.1 shows an illustration of an ITT system
with a fleet of three waterborne AGVs, six waterborne AGV berths and 12 routes1 .
1 These routes are all with much shorter distances by water than by land while routes connecting, e.g., berth 5
and 6, with short land distances are considered by other ITT modes.
7.2 Problem statement 111
Figure 7.1: Illustration of an ITT system with six berths for three waterborne AGVs.
In light of the above ITT network and available fleet of waterborne AGVs, the list of
ITT requests should be available upon making decisions. In particular, each ITT request is
associated with information on seven aspects:
1. Request ID which is sorted by all requests’ release times;
2. Origin berth ID corresponding to the pick-up location;
3. Destination berth ID corresponding to the delivery location;
4. Release time defining when a set of containers are ready to be shipped, being the
earliest time that the loading service can start;
5. Due time defining when the set of containers are ready for subsequent operations,
being the latest time of completing this request including the unloading time at the
destination berth; and
6. Volume of the set of containers to be shipped in TEUs;
7. Service time for loading/unloading the set of containers.
In addition, since delays or waiting times do occur in reality and meeting hard time win-
dows may fail in finding a feasible solution, requests are allowed to be serviced within soft
time windows, but customer inconvenience cost will incur if not within hard time windows.
Note that trade-offs can also be made by using more waterborne AGVs to reduce delays.
Containers with the same request ID cannot be shipped by different waterborne AGVs as
assumed in Chapter 1. Finally, containers have to be transported without transhipment, i.e.,
loading and unloading operations happen exactly once for each request.
The autonomous ITT system runs in a closed-loop fashion, i.e., both scheduling and
control level problems in Figure 1.2 are solved in real-time using updated system states.
Control of a fleet of waterborne AGVs is realized as in Chapter 5 while the energy efficient
scheduling problem as well as the closed-loop scheduling and control design are presented
in the following two sections.
7.3 Energy efficient scheduling of ITT using waterborne

AGVs
Traditionally, the scheduling problem of a fleet of ships traveling back and forth among
terminals to transport goods relies on human dispatchers necessarily with high competence
and experience. Complex decisions need to be made satisfying various possibly conflicting
objectives (e.g., saving energy by sailing at low speeds while meeting time windows at high
speeds) considering transport requests and available vehicle lists. Assignment of waterborne
AGVs to routes and timing can break such an operator. In face of real-time operational
delays and uncertainties, the problem can frustrate human dispatchers even more. We next
present a scheduling model based on mixed integer programming for ITT using waterborne
AGVs to ease the workload of human dispatchers. We first introduce relevant notations
including input parameters to the model and decision variables to be solved from this model.
Then the mathematical model is presented which is further transformed into a mixed integer
linear programming (MILP) problem to reduce required computation times.
7.3.1 Notations
The planning horizon considered within which a set of ITT requests R among the set of
berths B arise is 0, Tp . For each request i ∈ R , we denote a 7-element tuple hi, pi , di ,ti,min ,
ti,max , qi , si i to represent the information associated with request i as described in Section 7.2,
i.e., request ID, pick-up berth, delivery berth, release time, due time, volume, and service
time. For each pick-up location pi , a positive load +qi is attached, and each delivery location
di , a negative load −qi attached. The set of nv waterborne AGVs is V and homogenous.
The set for start locations for all waterborne AGVs is defined as Vo = {1, ..., nv } and the set
for end locations as Ve = {nv + 2n + 1, ..., 2nv + 2n} with n = |R |. All waterborne AGVs
have the same capacity Q in TEUs, curb weight m and cruising speed range [umin , umax ]. All
TEU of containers are assumed to have the same weight mc .
As has been discussed before, ITT scenarios inevitably involve waterborne AGVs shut-
tling back and forth, and thus pick-up and delivery locations of different requests might
actually be the same physical berths. This is one of the main differences of our schedul-
ing problem with land-based VRPs [130] or PDPs [115] based on assumptions of dis-
tinct pick-up and delivery locations and vehicles visit each location exactly once. We,
hence, define virtual pick-up and delivery node sets Pn = {nv + 1, nv + 2, ..., nv + n} and
Dn = {nv + n + 1, nv + n + 2, ..., nv + 2n}, respectively. Then, our scheduling problem is de-
finedover the virtual graph Gs = (N , A ) with node set N = Pn ∪ Dn ∪ Vo ∪ Ve and arc set
A = (i, j) (i, j) ∈ (Pn ∪ Dn ) × ((Pn ∪ Dn )) ∪ (i, j) i ∈ Vo , j ∈ Pn ∪ Dn ∪ {(i, j) |
i ∈ Pn ∪ Dn , j ∈ Ve , i 6= j . The physical locations of nodes in virtual graph Gs are
mapped as a vector L corresponding to N . Denote di j as the travel distance between nodes i
and j for all (i, j) ∈ A . Note that since waterborne AGVs stay at their final service berth, the
locations for virtual end nodes Vo vanish and distance di j = 0 if i ∈ Pn ∪ Dn ∪ Vo , j ∈ Ve .
For duplicated elements in L (same berths), we further cluster the corresponding nodes as
set Cb = {i |Li = b, b ∈ B }. For the nodes in a same set Cb , if they are visited by different
waterborne AGVs, a time interval T is imposed to the service time slots of the waterborne
AGVs to keep safety considering waterborne AGV dimensions.
The following decision variables are introduced to solve the scheduling problem:
7.3 Energy efficient scheduling of ITT using waterborne AGVs 113
• Binary variables: xi jv for (i, j) ∈ A and v ∈ V equals to 1 if waterborne AGV v travels

from node i → j and 0 otherwise;
• Binary variables: ziv for i ∈ N and v ∈ V equals to 1 if node i is visited by waterborne
AGV v and 0 otherwise;
• Binary variables: Ii j for i, j ∈ Cb , b ∈ B , i 6= j equals to 1 if node i is visited before
node j and 0 otherwise;
• Binary variables: Si j for i, j ∈ Cb , b ∈ B equals to 1 if nodes i, j are visited by different
waterborne AGVs and 0 otherwise;
• Integer variables: yi for i ∈ N denotes the load on board the waterborne AGV upon
arriving node i;
• Continuous variables: ai for i ∈ N specifies the arrival time at node i;
• Continuous variables: wi for i ∈ N is the waiting time at node i;
• Continuous variables: di for i ∈ N is the delay time at node i;
• Continuous variables: ui j for (i, j) ∈ A is the speed a waterborne AGV travels at on
leg i → j.
Additional auxiliary variables for the transformation into an MILP problem will be intro-
duced in Section 7.3.3.
7.3.2 Mixed integer programming problem

The overall goal is to compute a set of schedules that minimize the cost of fulfilling all
requests in R according to some cost metrics while satisfying various constraints. For our
case, a mixed integer programming problem is formulated as follows:
min c1 ∑ ∑ ∑ xi jv + c2 ∑ ∑ xi jv (mc y j + m)di j + c3 ∑ u2i j di j

v∈V j∈Pn ∪Dn i∈Vo v∈V (i, j)∈A (i, j)∈A (7.1)
+ c4 AVe − AVo 1
+ c5 kwk1 + c6 kdk1
subject to
∑ ziv = 1 ∀i ∈ N , (7.2)
v∈V
ziv = z(i+nr )v , ∀i ∈ Pn , v ∈ V , (7.3)
zii = 1, ∀i ∈ Vo , (7.4)
∑ xi jv = ∑ x jiv = ziv , ∀i ∈ N , v ∈ V , (7.5)
j∈N j∈N
∑ xvo jv = 1, ∀v ∈ V , (7.6)
j∈N /Ve
∑ xivd v = 1, ∀v ∈ V , (7.7)
i∈N /Vo
ai 6 Ai+nr , ∀i ∈ Pn , (7.8)
xi jv = 1 ⇒ max(Ai ,ti,min ) + si + di j /ui j = A j , ∀(i, j) ∈ A , v ∈ V , (7.9)
ti,min − wi 6 ai 6 ti,max − si + di , ∀i ∈ N , (7.10)
0 6 wi 6 wmax , ∀i ∈ N , (7.11)
0 6 di 6 dmax , ∀i ∈ N , (7.12)
Ii j + I ji = 1, ∀i, j ∈ Cb , b ∈ B , (7.13)
Si j = 1 − ∑ ziv z jv , ∀i, j ∈ Cb , b ∈ B , (7.14)
v∈V
Ii j Si j = 1 ⇒ max(ai ,ti,min ) + si + T 6 A j , ∀i, j ∈ Cb , b ∈ B , (7.15)
yvo = yve = 0, ∀v ∈ V , (7.16)
xi jv = 1 ⇒ yi + qi = y j , ∀i ∈ N , v ∈ V , (7.17)
0 6 yi 6 Q, ∀i ∈ N , (7.18)
umin 6 ui j 6 umax , ∀(i, j) ∈ A , (7.19)
xi jv , ziv , Ii j , Si j ∈ {0, 1} ∀(i, j) ∈ A , v ∈ V , (7.20)
where the objective (7.1) contains six cost terms that are related to energy efficient schedules
for waterborne AGVs. The first term counts the number of waterborne AGVs deployed
for the set of requests R . The fleet of deployed waterborne AGVs is not necessarily the
same with the fleet of available waterborne AGVs; we always minimize the number of
deployed waterborne AGVs considering high fixed deployment cost. Both the second and
third terms measure the cost of energy consumption and emissions traveling from node
i → j. The pollution-routing problem [7] employed similar emission measurement terms.
Cost term 2 is incurred due to the weight including waterborne AGV curb weight and the
weight of containers on board of the waterborne AGV. Cost term 3 reflects the nonlinear
dependence of energy consumption on cruising speed and distance. “Slow steaming” is
imposed by minimizing this term if possible. The fourth term considers the total sojourn
time of all waterborne AGVs. Departure times from starting locations are also optimized
with this formulation. The last two terms account for customer inconvenience measured by
waiting and delay times, respectively. The trade-off among these cost penalties is balanced
by weight parameters c1 , c2 , · · · , c6 .
Constraint (7.2) represents that each node is visited exactly by one waterborne AGV.
By constraints (7.3) and (7.4), we ensure that pick-up and delivery nodes of a particular
request are visited by the same waterborne AGV and all waterborne AGVs visit their own
starting nodes, respectively. Constraint (7.5) restricts that a waterborne AGV only enters and
leaves a node if it visits that node. Constraints (7.6) and (7.7) impose that each waterborne
AGV starts and ends at the right locations, respectively. Constraints (7.8) – (7.15) together
impose time constraints. Specifically, inequality (7.8) guarantees that pick-up nodes are
visited before delivery nodes. Constraint (7.9) enforces time consistency where the max
operation indicates that loading/unloading services cannot start earlier than the release times
of requests. Time window constraints are specified by (7.10) - (7.12). The coordinated
berthing times taking waterborne AGV dimensions and safety distances into consideration
are realized with constraints (7.13) – (7.15). The logic in (7.15) implies that if node i, j
relate to the same physical location (i, j ∈ Cb ) and are visited by different waterborne AGVs
(Si j = 1) and node i is visited before node j (Ii j = 1), that then the arrival time of the
waterborne AGV behind should be later than the departure time of the earlier waterborne
AGV at least for a time T . This is a novel feature of our waterborne AGV scheduling
problem. VRPs and variants have typically assume vehicles as dimensionless mass points
without consideration of safety distances. Load consistence and capacity constraints are
introduced via (7.16) – (7.18). Lastly, cruising speed is bounded by (7.19), and (7.20)
defines binary variables.
7.3 Energy efficient scheduling of ITT using waterborne AGVs 115
The above mixed integer programming problem (7.1) – (7.20) involves several nonlin-
earities:
• the multiplication of binary variable xi jv and integer load variable y j in the second
cost term of (7.1);
• the quadratic energy function of speed variable ui j in the third cost term of (7.1);
• the reciprocal of speed ui j in (7.9);
• logic implications in constraints (7.9), (7.15) and (7.17); and
• the multiplications of binary variables ziv z jv in (7.14) and Ii j Si j in (7.15).
All these nonlinearities bring about even more challenges to finding an optimal solution to
the already notorious NP-hard routing problem. We next present transformations of these
nonlinearities to obtain an easier to solve MILP problem.
7.3.3 Transformations into linearity

Linearizations of the above reported nonlinearities rely mainly on two techniques: dis-
cretization [7, 25], and logic and integer formulations [140]. We first deal with the non-
linearities caused by nonlinear functions of speed by discretization.
Generally, two discretization approaches are proposed: discrete speeds as in [7] and
discrete travel times as in [25]. Essentially, these two approaches are the same since speeds
and travel times are related by a constant travel distance. We apply discrete speeds as in [7]
due to its more intuitive formulation in modeling the “slow steaming” effect. The continuous
cruising speed range [umin , umax ] is discretized by equal intervals (umax − umin )/nu into a set
of nu speed levels [ur,min , ur,max ] for r = 1, ..., nu . An average speed is then calculated as
ūr = (ur,min + ur,max ) /2 and assigned to that level. Therefore, the speed optimization in
the continuous range [umin , umax ] becomes the optimal speed selection in the discrete speed
set {ūr |r = 1, ..., nu }. Correspondingly, we introduce binary variables bi jrv equal to 1 if
waterborne AGV v travels from node i → j at speed ūr . Then, the third cost term, which is
a quadratic speed function, is rewritten as:
!
nu
∑ u2i j di j = ∑ ∑ ū2r bi jrv di j . (7.21)
(i, j)∈A (i, j)∈A r=1
Similarly, the reciprocal speed term in (7.9) is rewritten as:

nu
di j /ui j = ∑ (di j /ūr )bi jrv . (7.22)
r=1
Note that different from the VRP in [7] with distinct visiting locations, we have dupli-
cated pick-up and delivery locations, which leads to di j = 0 when nodes i, j are actually the
same physical berths. Therefore, the relation between bi jvv and xi jv is constrained as:
nu
∑ bi jrv = δi j xi jv , ∀(i, j) ∈ A , v ∈ V , (7.23)
r=1
where δi j is a binary constant equal to 1 if di j > 0 and 0 if di j . This formulation enforces

a zero speed on arc (i, j) if nodes i, j are the same physical berths and a non-zero speed
otherwise.
To deal with the nonlinearities due to multiplications of binary and integer variables
xi jv y j in (7.1), multiplications of binary and binary variables ziv z jv and Ii j Si j in (7.14) and
(7.15), respectively, and logic implications in (7.9), (7.15) and (7.17), the following lin-
earizing approaches based on logic and integer formulations [140] are implemented.
Introduce auxiliary real variables Xi jv = xi jv y j , then we are able to replace the nonlinear
term in xi jv y j in (7.1) by Xi jv subject to the following set of linear constraints:
Xi jv 6 Qxi jv , ∀(i, j) ∈ A , v ∈ V ,
Xi jv > 0, ∀(i, j) ∈ A , v ∈ V ,
(7.24)
Xi jv 6 y j , ∀(i, j) ∈ A , v ∈ V ,
Xi jv > y j − Q (1 − xi jv ) ∀(i, j) ∈ A , v ∈ V .
The equivalence is due to the bound on variables y j as constraints (7.18).

Slightly differently, we replace the nonlinear terms in (7.14) and (7.15) with auxiliary
binary variables Zi jv = ziv z jv and Yi j = Ii j Si j along with the following two sets of linear
constraints, respectively:
−ziv + Zi jv 6 0, ∀i, j ∈ Cb , b ∈ B , v ∈ V ,
−z jv + Zi jv 6 0, ∀i, j ∈ Cb , b ∈ B , v ∈ V , (7.25)
ziv + z jv − Zi jv 6 1, ∀i, j ∈ Cb , b ∈ B , v ∈ V ,
−Ii j +Yi j 6 0, ∀i, j ∈ Cb , b ∈ B ,

−Si j +Yi j 6 0, ∀i, j ∈ Cb , b ∈ B , (7.26)
Ii j + Si j −Yi j 6 1, ∀i, j ∈ Cb , b ∈ B .
Finally, logic implications in (7.9) are transformed as:
nu
max(ai ,ti,min ) + si + ∑ (di j /ūr )bi jrv 6 A j + Mi1j (1 − xi jv ) , ∀(i, j) ∈ A , v ∈ V ,
r=1
!
nu
− max(ai ,ti,min ) + si + ∑ (di j /ur )bi jrv 6 −A j − m1i j (1 − xi jv ) , ∀(i, j) ∈ A , v ∈ V ,
r=1
(7.27)
with Mi1j = ti,max + wmax + si + di j /umin − (ti,min − wmax ), m1i j = ti,min + si − (ti,max + wmax ),
and (7.17) as:
yi + qi 6 y j + Mi2j (1 − xi jv ) , ∀(i, j) ∈ A , v ∈ V ,
(7.28)
− (yi + qi ) 6 −y j − m2i j (1 − xi jv ) , ∀(i, j) ∈ A , v ∈ V ,
with Mi2j = Q + qi , m2i j = qi − Q, and (7.15) simply as:
max(ai ,ti,min ) + si + T 6 A j + Mi3j (1 −Yi j ) , ∀i, j ∈ Cb , b ∈ B , , (7.29)
with Mi3j = t j,max + wmax + si + T − (ti − wmax ).

7.4 Real-time closed-loop scheduling and control 117
So far, we have transformed the nonlinear mixed integer programming problem (7.1) –
(7.20) into an MILP problem by replacing nonlinear terms in the cost function and con-
straints with auxiliary variables and linear constraints as formulated as (7.21) - (7.29).
Schedules generated by solving
the MILP problem are, for each waterborne AGV v ∈ V ,
sequences
of nodes Nv = i i ∈ N , ziv = 1, v ∈ V to visit, the corresponding arrival times
Av = ai ziv = 1, v ∈ V in ascending order, load/unloading volumes Qv = {qi |ziv = 1,
v ∈ V , as well as traveling speeds Uv = ui j xi jv = 1, (i, j) ∈ A , v ∈ V on each leg.
7.4 Real-time closed-loop scheduling and control

Ideally, a seamless integration of scheduling and control problem requires that all decisions
are made simultaneously achieving objectives and satisfying various constraints at both lev-
els, which still remains an open, though important, issue. The main technical challenges lie
in the different time horizons and different nature of decisions (continuous time and discrete
events) which result in a highly complex problem to solve considering current computing
power. A descriptive integrated problem P could be defined as:
(u∗s , u∗c ) := arg min J (Js , Jc )

Subject to Scheduling constraints (7.2) − −(7.29),
Waterborne AGV dynamics (3.1) − −(3.5), (7.30)
Disturbance dynamics (3.4),
Control constraints (3.3), (3.8),
where the total cost J depends on scheduling cost Js and control cost Jc . The optimal
scheduling and control decisions u∗s , u∗c are simultaneously made by solving problem P.
Apparently, the decision frequency of problem P should be the same as the control problem
which has faster decision frequencies. However, the already complex scheduling problem
(7.1)–(7.29) coupled still with lower level motion and disturbance models will typically not
be easily solved to optimality at a high frequency. Even if we decompose problem P hi-
erarchically and solve scheduling and control problems sequentially every time a control
decision is implemented and new system states are available, solving the scheduling prob-
lem at the control frequency could still preclude it from practical applications. We propose
an interaction model and a real-time scheduling problem that enable solving scheduling and
control problems both hierarchically and real-time in a closed-loop way.
7.4.1 Modeling interactions and real-time speed assignment

Literature dealing with scheduling or control problems independently can actually be viewed
as considering a simplified problem P with assumptions. From a scheduling perspective, im-
plicit assumptions are made that schedules are all executed perfectly, i.e., waterborne AGVs
depart and arrive at berths following the exact scheduled order and timing irrespective of
operational disturbances. This is implicitly achieved by assuming a constant speed and thus
constant travel times on all routes if speed is not an optimization variable, e.g., in [96],
or variable speeds as functions of discrete events if a combined route-speed optimization
problem is solved, e.g., in [7, 25, 92] as well as our scheduling problem (7.1)–(7.29). No
Yn
j
Xb
v
sl
u
Yb
v y
y Ob
yr r
i
Disturbances (wind,
waves, currents, etc.)
w
Xn
On x
Figure 7.2: Waterborne AGV v and pose projection in route (i, j).
operational disturbances and physical system limitations can further be incorporated. There-
fore, waterborne AGV motions are simply modeled as a first order integrator with constant
speeds on one arc in a combined routing-speed optimization problem, i.e.,
sv (k + 1) = sv (k) + ui(i+1)v Ts , ∀i ∈ Nv (k), v ∈ V (k), kTs ∈ [max(ai (k),ti,min ), Ai+1 (k)],

(7.31)
where sv is the distance waterborne AGV v has traveled on route i.
We take advantage of the implicit motion model in the scheduling problem and propose
an interaction model based on a two-level parameterization of reference paths similar as in
Chapter 4. The lower level is embedded in the online control problem which takes care of
the waterborne AGV dynamics, system limitations, operational disturbances, and collision
avoidance with other traffic, and modeled as:
slv (k + 1) = slv (k) + ulv (k)Ts , ∀v ∈ V (k), (7.32)
where slv (k), ulv (k) are a lower level path parameter and its speed, respectively. Both slv (k)
and ulv (k) are decision variables in online control optimization problems, as to be introduced
further in Section 7.4.2. The path parameter determines the reference orthogonal projections
of waterborne AGV v onto its current route (i, i + 1), i ∈ Nv by
x p (k) = slv (k) sin (ψi ) + xi ,

(7.33)
y p (k) = slv (k) cos (ψi ) + yi ,
as shown in Figure 7.2, with (x p (k), y p (k)), (xi , yi ) the inertial frame coordinates of the
reference projection and node i, respectively.
The upper level of the two-level parameterization scheme is a partial scheduling problem
of problem (7.1)–(7.29) and updates schedules based on the lower level states which in turn
reflect waterborne AGV operational details including real-time control performances, delays
caused by environmental disturbances or collision avoidance, etc. In this way, the two-level
interaction model connects scheduling and control problems but make them decomposable
7.4 Real-time closed-loop scheduling and control 119
from each other while still allowing both to be solved online using real-time feedback. The
upper level problem is formulated as a real-time speed assignment problem as follows.
At each time step k, we collect feedback information for the set of waterborne AGVs
V (k) that are scheduled with tasks but have not arrived at the scheduled last node each
with load y0v (k), v ∈ V (k) on board. For each waterborne AGVnv ∈ V (k), theolist of yet to
v
visit nodes are Nv (k), the corresponding list for time windows [tmin,i v
,tmax,i ] , ∀i ∈ Nv (k)
and the list of service times {svi } , ∀i ∈ Nv (k). Besides time window constraints, different
waterborne AGVs still need to coordinate their service time slots at a particular berth by
guaranteeing a safe time interval among them. Variables related to the real-time speed
assignment problem are:
• Binary variables: Ii j (k) for i, j ∈ Cb , i ∈ N p , j ∈ Nq , p 6= q equal to 1 if node i is visited
by waterborne AGV p before node j by waterborne AGV q and 0 otherwise;
• Continuous variables: avi (k) for i ∈ Nv , v ∈ V (k) specifies the arrival time of water-
borne AGV v at node i;
• Continuous variables: wvi (k) for i ∈ Nv , v ∈ V (k) is the waiting time of waterborne
AGV v at node i;
• Continuous variables: div (k) for i ∈ Nv , v ∈ V (k) is the delay time of waterborne AGV
v at node i;
• Continuous variables: uvi(i+1) (k) for i ∈ Nv , v ∈ V (k) is the speed waterborne AGV v
travels at on leg i → i + 1.
The overall goal is to compute schedules that still minimize the overall cost of fulfilling
all remaining requests while satisfying time window and coordinated berthing constraints.
The mixed integer programming problem is formulated as:
min c3 ∑ u2i j (k)di j (k) + c4 AVe (k) − AVo (k) 1
+ c5 kw(k)k1 + c6 kd(k)k1 (7.34)
(i, j)∈A
subject to
v
max(ai (k),ti,min ) + svi + di(i+1) (k)/ui(i+1) (k) = Ai+1 , ∀i ∈ Nv (k), v ∈ V (k), (7.35)
v
ti,min − wvi (k) 6 ai (k) 6 ti,max
v
− svi + div (k), ∀i ∈ Nv (k), v ∈ V (k), (7.36)
0 6 wvi (k) 6 wmax , ∀i ∈ Nv (k), v ∈ V (k), (7.37)
0 6 div (k) 6 dmax , ∀i ∈ Nv (k), v ∈ V (k), (7.38)
Ii j (k) + I ji (k) = 1, ∀i, j ∈ Cb , b ∈ B ,
i ∈ N p , j ∈ Nq , p 6= q, (7.39)
v
Ii j (k) = 1 ⇒ max(ai ,ti,min )(k) + svi (k) + T 6 A j, ∀i, j ∈ Cb , b ∈ B ,
i ∈ N p , j ∈ Nq , p 6= q, (7.40)
u− 6 uvi(i+1) (k) 6 ū, ∀i ∈ Nv , v ∈ V (k), (7.41)
Ii j (k) ∈ {0, 1} ∀(i, j) ∈ A , v ∈ V , (7.42)
where objective (7.34) contains four terms that are the same as the last four terms in (7.1),
minimizing energy consumption due to variable speed, total sojourn times, waiting times,
and delay times, respectively. Constraints (7.35) indicates the time consistency between
two successive nodes in the node list of each waterborne AGV v ∈ V (k). Soft time win-
dows are imposed by constraints (7.36) with constraints (7.37) and (7.38) specifying the
maximum waiting and delay times, respectively. Constraints (7.39) and (7.40) together for-
mulate the coordinated berthing times between different waterborne AGV visiting a same
berth. Variable speeds are bounded by constraint (7.41) and constraint (7.42) define the
only binary variable in this problem. By solving this problem, we obtain updated sched-
ules Nv (k), Av (k), Qv (k), Uv (k) for waterborne AGVs v ∈ V (k) as well as parameterized
reference paths at the upper level, defined as:
suv (k + 1) = suv (k) + ui(i+1) (k)Ts , ∀v ∈ V (k), i ∈ Nv (k), (7.43)
where suv (k), ui(i+1) (k) are references for slv (k), ulv (k), respectively.
7.4.2 Closing the real-time loop

With the interaction model, we are now ready to decompose problem P hierarchically into
a scheduling problem Ps
u∗s := arg min Js

(7.44)
Subject to Scheduling constraints (7.2) − −(7.29) or (7.35) − −(7.42),
and a control problem Pc
u∗c := arg min Jc

Subject to Waterborne AGV dynamics (3.1) − −(3.5),
(7.45)
Disturbance dynamics (3.4),
Control constraints (3.3), (3.8).
In particular, at each control time step k, we first solve scheduling problem Ps based on
updated feedback information: the set of waterborne AGVs V (k) that are scheduled with
0
tasks but have not arrived at the scheduled last node, each
v v
with load yv (k), v ∈ V (k) on
board, projected waterborne AGV positions x p (k), y p (k) ∀v ∈ V (k) that are determined
by measured waterborne AGV positions xv k, yv (k)∀v ∈ V (k). The projected positions are
utilized to 1) update the Euclidean distances di(i+1) (k)∀i ∈ Nv (k), v ∈ V (k); and 2) initialize
xvp (k) − xi
both levels in the interaction model by slv (k) = suv (k) = . Note that only
yvp (k) − yi 2
at the beginning of the planning horizon, problem (7.1) –(7.29) needs to be solved as the
scheduling problem Ps. Real-time scheduling is achieved by solving problem (7.34) – (7.42)
with a smaller set of integer variables and is efficient to solve to optimality.
At each time step k, the control problem Pc is solved after receiving references from
scheduling problem Ps. The overall control goals are to 1) execute schedules to fulfill ITT
requests in an economical way; 2) maintain safety by satisfying system physical limitations
and avoiding collisions with other traffic in the presence of disturbances; and 3) maneuver
in a distributed way. We present the formulation of Jc here which involves the lower level
of the interaction model in the closed-loop scheduling and control:

2
Jc (k) = ∑ η(k) − η r (k)k22 + c8 slv (k) − suv (k) + c9 /2kνν(k)k2M (k) ,
c7 kη (7.46)
2
v∈V (k)
7.5 Experiments and discussion 121
Real-time scheduling and control
Scheduling problem Ps
Interaction model load
Control problem Pc
Forces/moments
Real-time states
Physical system & environment
Figure 7.3: Closed-loop scheduling and control of energy-efficient waterborne AGVs.
T
where ηr is reference pose determined by (7.32) and (7.33) as ηr = x p y p ψi . Con-
trol performance is also affected by scheduling results reflected in the mass matrix as:
m + mc y0v (k)
 
0 0
M (k) =  0 m + mc y0v (k) m + mc y0v (k)xg  + M A , (7.47)
0 m + mc y0v (k)xg Iz
The first goal of executing schedules in an energy efficient way is then achieved by mini-
mizing Jc (k) as in Chapter 5. The closed-loop schedule and control of waterborne AGVs is
shown in Figure 7.3.
7.5 Experiments and discussion

Simulations are run to illustrate the effectiveness of the proposed closed-loop scheduling
and control of waterborne AGVs for ITT. Note that in practice, the number of ITT requests
and the fleet of waterborne AGVs could be large. However, as simplified in Section 7.2,
i.e., requests in different planning horizons are decoupled so that small sets of requests can
be considered independently. Scheduling problems for each set can be solved repetitively
using the proposed approach. The idea is to demonstrate how the algorithm works as well
as its effectiveness. Therefore, an ITT scenario as shown in Figure 7.1 with six berths and
a fleet of three waterborne AGVs (Vi , i = 1, 2, 3) is considered in simulations in this chapter.
Assume that there are seven ITT requests arising within the scheduling horizon 0 s – 2100
s as detailed in Table 7.1 of which the available request information structure has been
designed according to [118]. Positions of the six berths are determined in latitude/longitude
and subsequently converted to coordinates in inertial frame with Berth 1 as the origin. The
fleet of available waterborne AGVs for this set of ITT requests are initially all positioned at
Berth 1.
Speed limits of waterborne AGVs are umin = 2.57 m/s, umax = 6.68 m/s which are corre-
sponding to 5 knots and 13 knots, respectively. Each waterborne AGV can carry a maximum
Table 7.1: ITT requests to be carried out.
Request ID Origin Destination Release time (s) Due time (s) Volume (TEU)
1 5 2 125 1865 2
2 1 3 690 1155 2
3 1 4 700 1485 1
4 6 2 725 1535 2
5 6 1 1230 1755 2
6 2 3 1345 1750 2
7 1 4 1640 2085 1
of four TEUs, i.e., Q = 4, and each TEU of container weighs mc = 24000kg. Each move
of a quay crane can load/unload one or two TEUs and requires 120s [129]. Therefore,
for all ITT requests in Table 7.1, service times are the same as ts = 120s. The neces-
sary safety time interval between different waterborne AGVs visiting a same berth is set to
T = 60s based waterborne AGV lengths and sailing speeds. Other parameters concerned
with waterborne AGV dynamics are implemented as in [150]. The weight parameters in
cost functions (7.1), (7.34) and (7.46) for trade-offs of different performance metrics are set
as: c1 = 104 , c2 = 10−2 , c3 = 102 , c4 = 103 , c5 = 108 , c6 = 108 , c7 = 100, c8 = 100, c9 = 1.
Algorithms are implemented in MATLAB 2011b [75]. Optimization problems are solved
by Cplex [46]. All the simulations are run on a platform with Intel (R) Core (TM) i5-3470
CPU @3.20 GHz.
The closed-loop schedule and control algorithm as shown in Figure 7.3 needs to replace
human operators to make “smart” decisions:
• For the fleet of waterborne AGVs, energy efficient schedules as well as actuator inputs
to execute these schedules should be autonomously generated satisfying waterborne
AGV physical limitations, e.g., maximum capacity, rudder force range etc. and guar-
anteeing safety;
• For the list of ITT requests, certain amount of containers should be transported from
specified origins to destinations after the release time while before the due time; and
• Each berth can accommodate at most one waterborne AGV, and service time slots of
different waterborne AGVs should keep a buffer time interval.
Simulation results from these three perspectives are presented next to demonstrate the ef-
fectiveness of the proposed closed-loop scheduling and control algorithm.
7.5.1 From the waterborne AGV perspective

The set of seven ITT requests as shown in Table 7.1 calls for all three waterborne AGVs of
which initial schedules by solving problem (7.1) – (7.28) are shown in Figure 7.4 as green-
circle line, magenta-hexagram dotted line, and green-square dashed line, respectively. All
three waterborne AGVs start from Berth 1 but at different times. The small rectangles are
one TEU containers and the numbers attached identify IDs of requests that the containers
belong to. We display the set and mix of containers on board waterborne AGVs departing
4
4 V1
1 5
6 1 5
V2
V3
1
1 one TEU
5
4
Berth ID
4 6
2 4 6
2
1 2 3 7
0 500 1000 1500 2000

Time (s)
Figure 7.4: Initial energy efficient schedules.
berths. Each schedule contains information on the sequence of berths to visit, the corre-
sponding arrival and departure times as well as the load/unload operations at each berth.
Take the schedule of waterborne AGV V2 as an example, we place a hexagram upon wa-
terborne AGV V2 ’s arrival at a berth. There are three hexagrams at Berth 2 between 1000s
and 1500s because waterborne AGV V2 performs three load/unload operations at Berth 2.
From the set and mix of containers on board when departing, we can derive that waterborne
AGV V2 first unloads the two containers from request 1 taking 120s, then unloads the two
containers from request 4 taking another 120s and finally loads the two containers from re-
quest 6 before departing from Berth 2 to Berth 3 which is its final destination. Note that the
time for solving the problem (7.1) – (7.28) depends on the numbers of waterborne AGVs
and requests, and tightness of imposed time windows. It takes 127 s in our case, and could
be even more time consuming for larger problems.
Travel speeds along all route segments are also explicitly optimized. In fact, with berth
IDs, arrival and departure times known, travel speeds can easily be derived from 7.4. The
travel speed profile for waterborne AGV V2 along its route is shown as Figure 7.5. As can
also be observed in Figure 7.4, all three waterborne AGVs carry no more than four TEU
containers. Figure 7.6 further shows the total number of containers on board throughout the
simulation which are all within the maximum capacity of four TEU containers.
Waterborne AGVs receiving schedules as shown in Figure 7.4 are then controlled with
a first goal of guaranteeing operational safety and a secondary goal of executing those
schedules. Since complex system dynamics, physical limitations, disturbances and collision
avoidance between moving waterborne AGVs are not considered in the scheduling problem,
real-time waterborne AGVs do not necessarily behave safely and as scheduled: following
the scheduled route at specified speed and arriving at scheduled berths at specified times.
Figure 7.7 shows the evolutions of velocities of three scheduled waterborne AGVs, respec-
tively. Velocities in three degrees of freedom: surge, sway and yaw are all within safe
5
distance
speed
Travel distance (m)\berth ID
3398\3
Speed (m/s)
2784\2
1965\6
1252\5
0\1 0
0 500 1000 1500 2000
Time (s)
Figure 7.5: Travel speed profile of waterborne AGV V2 .
4.5
4
V1
3.5
V2
3 V3
Load on board (TEU)
max capacity
2.5
1.5
0.5
−0.5
0 500 1000 1500 2000
Time (s)
Figure 7.6: Containers on board of the waterborne AGVs.

Sway velocity (m/s) Surge velocity (m/s)

10 V1 V2 V3
0
0 500 1000 1500 2000
2
−2
0 500 1000 1500 2000
0.05
Yaw rate (rad/s)
−0.05
0 500 1000 1500 2000
Time (s)
Figure 7.7: Velocities in surge, sway and yaw of three waterborne AGVs.
maneuvering ranges as the red lines show. Surge velocity is a function of time and sees
fluctuations which is different with piecewise constant speeds determined in scheduling
problems which are functions of events as shown in Figure 7.5. This is due to the neces-
sary accelerations and decelerations when operating in real environment. Sway velocity and
yaw rate are not considered in scheduling problems at all. Likewise, control inputs in surge,
sway and yaw interact with complex system dynamics and environment to achieve control
goals and are all within the safety limitations as shown in Figure 7.8.
Waterborne AGVs are treated equally as agents that are controlled in a distributed way
and make control decisions parallelly using a fast distributed control method proposed in
[148]. During the execution of assigned requests, waterborne AGVs might encounter con-
flicts with each other. The distributed control algorithm guarantees ITT request fulfillment
locally for each waterborne AGV while achieves an overall minimal cost and safety for all
waterborne AGVs. The route following performance with small tracking errors of water-
borne AGV V3 is demonstrated as in Figure 7.9(a). Fluctuations are seen at the beginning
and around 1200s due to the start up and encountering with waterborne AGVs V2 and V1 .
However, a safety distance away from them is ensured by the control level when waterborne
AGVs are in close range. Figure 7.9(b) shows the distances between waterborne AGV V2
and the other two waterborne AGVs which are all above the minimum safety distances.
7.5.2 From the ITT request perspective

Assignment of requests in Table 7.1 to three waterborne AGVs can be derived from the re-
quest IDs attached with containers in Figure 7.4. This assignment is produced being aware
of constraints as waterborne AGV capacities and request time windows as well as travel cost
specified in (7.1). For each request, Figure 7.10 shows the time window (red bar) specified
by the release and due times of the request, the initial scheduled duration time (yellow bar)
6
Surge force (N) x 10
5 V1 V2 V3
−5
0 500 1000 1500 2000
6
x 10
Sway force (N)
−2
0 500 1000 1500 2000
Yaw moment (Nm)
8
x 10
2
−2
0 500 1000 1500 2000
Time (s)
Figure 7.8: Control input evolution in surge, sway and yaw of three waterborne AGVs.
(a)
1.5
Tracking error (m)
0.5
0
1050 1100 1150 1200 1250 1300 1350
Time (s)
(b)
1000
V1 and V3
800
Distances (m)
V2 and V3
600 minimum
400
200
0
1050 1100 1150 1200 1250 1300 1350
Time (s)
Figure 7.9: Tracking performance and collision avoidance.

Time windows
R7 (1−>4)
Scheduled V1
Actual
R6 (2−>3)
V2
R5 (6−>1)
V1
ITT requests
R4 (6−>2)
V2
R3 (1−>4)
V3
R2 (1−>3)
V1
R1 (5−>2)
V2
0 500 1000 1500 2000 2500
time (s)
Figure 7.10: Satisfaction of request time windows in the closed-loop.
Time windows
R7 (1−>4)
Scheduled V1
Actual
R6 (2−>3)
V2
R5 (6−>1)
V1
ITT requests
R4 (6−>2)
V2
R3 (1−>4)
V3
R2 (1−>3)
V1
R1 (5−>2)
V2
0 500 1000 1500 2000 2500
time (s)
Figure 7.11: Requests executed in the open-loop.

solving problem (7.1) – (7.29) and the actual duration time (green bar) specified by the wa-
terborne AGV’s arrival at the origin berth and departure from the destination berth. All the
scheduled and actual duration times are within required time windows, i.e., time windows
of all requests are satisfied by the scheduling problem and the control problem succeeds in
operating waterborne AGVs with timing window awareness except for request 3 with some
delays. Some green bars, e.g., request 1, however, are not within the corresponding yellow
bars. This is due to the real-time update of schedules by solving closed-loop scheduling
problem (7.34) – (7.42). The updated scheduled duration times are not necessarily the same
with the initial schedules, but still are guaranteed to satisfy time window constraints as
(7.36). The satisfaction of time windows by actual duration times proves this.
For comparison, Figure 7.11 shows the request completion times of schedules executed
in open-loop, i.e., the initial schedule by solving (7.1) – (7.28) is not updated. Some requests
with relatively tight time windows, e.g, request 6 and 7, see delays. This is because lower
level system details that may cause inaccurate execution of schedules are neither considered
in scheduling problem (7.1) – (7.28) nor reflected in updated schedules in solving real-
time problem (7.34) – (7.42). The inaccuracies accumulate along routes and lead to delays.
Waterborne AGV 1, for example, serves request 2, 5, and 7 in sequence (see Figure 7.4).
Delays in finishing request 2 lead to later start of request 5 compared to the scheduled
start time, and give rise to the violation of the tight time window of request 7 in the end.
Since “non-performance”, which happens when there are containers that are delivered with
delays, is the most important criterion for ITT, we define the “non-performance” rate as the
percentage of delayed number (in TEUs) of ITT containers with respect to the total number
of ITT containers. Therefore, for the seven ITT requests with 12 TEUs in Table 7.1, the
closed-loop approach has a “non-performance” rate of 0% while the open-loop approach is
41.67%. This illustrates the advantage of closed-loop over open-loop.
7.5.3 From the berth perspective

Each of the six waterborne AGV berths as shown in Figure 7.1 is designed to handle max-
imum one waterborne AGV one time and there should be a time interval between different
waterborne AGV’s visits to a same berth for berth practice and guaranteeing safety consid-
ering dimensional waterborne AGVs. In other words, the departure time of one waterborne
AGV should be at least earlier the specified time interval T = 60s than the arrival time of the
next waterborne AGV as constrained by (7.13) - (7.14). From the berth perspective, Figure
7.12, 7.13, and 7.14 illustrate berth occupations over time as initially scheduled, real-time
scheduled at t = 750s, and the actually executed by waterborne AGVs, respectively.
For all berths, there are intervals between bars in different colors along one horizontal
line representing different waterborne AGVs visiting a same berth in all three figures. Par-
ticularly in Figure 7.14 the for actual berthing time slots, Berths 1, 3, 4, and 6 see visits
from different waterborne AGVs. The minimum actual time interval is 60s> T (minimum
safety time interval) which happens between waterborne AGVs 1 and 3 at Berth 1. For one
waterborne AGV performing more than one load/unload operation at one berth, however,
there is no time interval. This is shown by the bars in same colors linked together, e.g.,
the three magenta bars at Berth 2 in Figure 7.12. Actually, waterborne AGV V2 performs
three load/unload operations at Berth 2 if we recall Figure 7.4 as analyzed in Section 7.5.1.
Note that the vertical lines at Berth 1 in Figure 7.12 and Figure 7.14 indicates that water-
6 V1
V2
V3
5
4
Berth ID
0 500 1000 1500 2000

Time (s)
Figure 7.12: Berthing time slots from the initial schedule.
6 V1
V2
V3
5
4
Berth ID
0 500 1000 1500 2000

Time (s)
Figure 7.13: Berthing time slots from the schedule at t = 750s.

6 V1
V2
V3
5
4
Berth ID
0 500 1000 1500 2000

Time (s)
Figure 7.14: Actual berthing time slots.
borne AGV V2 travels from Berth 1 to Berth 5 directly without any load/unload operations
at Berth 1. Lines instead of bars, i.e., no load/unload operations, may only arise at water-
borne AGVs’ initial positions (Berth 1 in our scenario) since any other berths waterborne
AGVs to visit are involved an ITT request with certain amount of containers to load/unload
and thus requires certain service times. Generally, all time slots (bars) should be not shorter
than ts = 120s. The shorter time slots in Figure 7.13, e.g., as shown by the relatively shorter
green bar at Berth 1 and the first magenta bar at Berth 6, are because certain amount of ser-
vice has been conducted at t = 750s. A bar longer than ts = 120s simply means waterborne
AGV waits for some time before the load service can start (release time).
7.6 Conclusions
In this chapter, we have proposed a real-time closed-loop scheduling and control scheme
for waterborne AGVs applied to ITT. The contributions are twofold. Firstly, we propose a
new pick-up and delivery scheduling model for ITT using waterborne AGVs by consider-
ing safety time intervals between their service time slots at one berth. For all the berths,
safety time intervals are guaranteed for different waterborne AGVs. Secondly, we propose
a partial scheduling problem that is efficient to solve. By integrating this partial scheduling
problem with the control problem of multiple waterborne AGVs in Chapter 5, we realize
real-time closed-loop scheduling and control of an autonomous ITT system. In our simula-
tion experiments based on a potential ITT scenario in the port of Rotterdam, time windows
of all ITT requests are satisfied by the closed-loop approach with 0% “non-performance”
rate compared to 41.67% from the open-loop approach. The proposed algorithm provides
an effective way realizing autonomous ITT systems using waterborne AGVs.
Chapter 8
Conclusions and future research
In this thesis, a new type of container transporter, waterborne AGVs is proposed, and ad-
vanced control and scheduling strategies for efficiently operating such waterborne AGVs
are discussed. Waterborne AGVs are coordinated for carrying out Inter Terminal Transport
(ITT) tasks in port areas. This final chapter presents the conclusions and the main con-
tributions of this thesis in Section 8.1, and gives recommendations for future research on
waterborne AGVs in Section 8.2.
8.1 Conclusions and contributions

In this section, we first conclude the thesis by addressing the main research goal and an-
swering the five Key Research Questions raised in Chapter 1 using the proposed strategies
from the previous chapters. Contributions of this PhD research with respect to coordination
methodologies and logistics applications are then explicitly discussed.
8.1.1 Answering the research questions

As presented in Chapter 1, the main research goal of this thesis is to develop advanced con-
trol and scheduling strategies for coordinated waterborne AGVs applied to ITT. In Chapters
2 – 7, relevant literature has been reviewed and novel systematic methods have been pro-
posed to achieve the aforementioned goal. Simulation results of ITT case studies using
waterborne AGVs have demonstrated the effectiveness of the proposed approaches. More
specifically, the five Key Research Questions for addressing the main research goal are an-
swered as follows:
1. Which technique is suitable for the control of waterborne AGVs?

We propose that the control of waterborne AGVs carrying out ITT tasks, as in Chapters 4
– 7, uses model predictive control (MPC). MPC can look into the future and take actions
at an earlier stage by predicting system trajectories over a future horizon. Advantages
of MPC include handling system constraints explicitly and optimizing multiple objec-
tives quantitatively. Different ITT scenarios with multiple possibly conflicting control
goals are investigated. In Chapter 4, one single waterborne AGV with an assigned ITT
131
132 8 Conclusions and future research
task is controlled to track reference paths smoothly and arrive at the specified terminal
at the specified time in an energy-efficient way. Overshoots during the switching of ref-
erence line segments are avoided. Various constraints such as physical limitations and
obstacle avoidance are also satisfied. Chapter 5 considers multiple waterborne AGVs,
each of them assigned with an ITT task. In the proposed cooperative distributed MPC
(DMPC) framework, waterborne AGVs are able to solve local problems in parallel while
possible collisions between waterborne AGVs are avoided and the overall cost is min-
imized. A cost-effective robust DMPC (RDMPC) algorithm is proposed in Chapter 6
for waterborne AGVs maneuvering in uncertain environments. Cost-effective robustness
optimizes the trade-off between system performance and robustness level considering
system and uncertainty characteristics. In Chapter 7, the DMPC approach of Chapter 5 is
utilized in a closed-loop scheduling and control scheme for multiple waterborne AGVs.
All the control problems consider detailed waterborne AGVs dynamics as modeled in
Chapter 3. In MPC, successively linearized models are used for prediction to reduce the
possible computational complexity. Various simulation experiments have been carried
out using the proposed MPC-based algorithms in different ITT case studies. Simulation
results show that control goals specified are all successfully achieved, which validates
the suitability of MPC in controlling waterborne AGVs for ITT.
2. What performance criteria should be considered in optimizing the process of one water-
borne AGV carrying out one ITT task and how can the optimal performance be achieved?
For the most fundamental scenario of one waterborne AGV carrying out one ITT task,
smooth reference path tracking, arriving at the destination as punctually as possible, and
energy-efficiency should be considered as the performance criteria. Geometric shortest
reference paths are given as connected straight-line segments from which derivations of
waterborne AGV trajectories are minimized. Waterborne AGVs arrive at the destinations
punctually or with a minimal delay to lower the “non-performance” rate of ITT. Energy
consumption related to the waterborne AGV speed has also been minimized. These per-
formance criteria are in alignment with the practice of ITT and the requirements in port
areas. A predictive path following with arrival time awareness (PPF-ATA) controller has
been proposed to satisfy those criteria in Chapter 4. Waterborne AGV kinematics are
remodeled in a connected path coordinate in which cross-track and along-track errors
are defined. Cross-track errors are minimized to achieve the reference path convergence,
and along-track errors that facilitate a path switching logic for avoiding overshoots are
minimized to achieve arrival time awareness. Smooth tracking and timing have been
guaranteed by a two-level double integrator scheme. Simulations have shown that wa-
terborne AGVs achieve the specified performance criteria with the proposed PPF-ATA
controller in two ITT scenarios.
3. How can multiple waterborne AGVs be coordinated for multiple ITT tasks with water-
borne AGVs making decisions locally while minimizing the overall cost in a cooperative
and distributed way?
In Chapter 5, cooperative distributed PPF-ATA controllers have been proposed for mul-
tiple waterborne AGVs carrying out multiple ITT tasks. In such a scenario, couplings
between waterborne AGVs in a neighborhood arise as collision avoidance constraints as
modeled in Chapter 3. An iterative decomposition-coordination technique, the Alternat-
8.1 Conclusions and contributions 133
ing Direction Method of Multipliers (ADMM), has been proposed to achieve coopera-
tive distributed control of waterborne AGVs in the framework of DMPC. Furthermore,
the possible poor convergence rates of conventional ADMM are improved by a pro-
posed fast ADMM approach by iteratively approximating global information in local
problems. In our ITT case study in the port of Rotterdam, comparing ADMM and fast
ADMM, ADMM requires 108 iterations with a total solver time of 4.65s before conver-
gence while fast ADMM requires only six iterations with 0.27s. Therefore, fast ADMM
offers a more practical cooperative distributed approach considering the short sampling
times of waterborne AGVs. For both ADMM and fast ADMM distributed control ap-
proaches, all waterborne AGVs can fulfill assigned ITT tasks successfully. Cooperative
and safe behaviors have been observed in conflicting areas. Therefore, the proposed co-
operative distributed controllers are effective in coordinating multiple waterborne AGVs
for multiple ITT tasks.
4. How can environmental disturbances due to wind, waves, and current be systematically
handled by cooperative and distributed waterborne AGVs?
The influences of environmental disturbances have been considered in two scenarios: one
with perfectly known disturbances, and the other with not perfectly known but roughly
predicted disturbances, as modeled in Chapter 3. If environmental disturbances are per-
fectly known, the cooperative DPPF-ATA controllers proposed in Chapter 5 are readily
applicable. For roughly predicted stochastic disturbances, a cost-effective RDMPC ap-
proach for multiple waterborne AGVs has been proposed in Chapter 6. The approach
is cost-effective in the sense that the overall system robustness level and the associated
price of robustness are explicitly optimized considering system and uncertainty charac-
teristics. Since probabilistic distributions of uncertainties are parameterized by introduc-
ing binary variables, the convexity assumptions for the cooperative distributed control
in Chapter 5 do not necessarily hold. An efficient integrated branch & bound (B&B)
and ADMM algorithm has been proposed to solve the cost-effective RDMPC problem.
The algorithm exploits the special ordered probability sets conducting smart search in
B&B and integrates branching criteria with intermediate ADMM results for early termi-
nation. With the proposed efficient searching strategy, at most two branching operations
at each time step are required throughout the simulation. Simulation results show that,
similar with the deterministic cases in Chapter 5, all waterborne AGVs fulfill assigned
ITT tasks successfully and safely in uncertain scenarios as well. Cooperative distributed
decision making has also been achieved following ADMM iterations. The trade-off be-
tween robustness level and the price of being robust has been optimized systematically.
In our simulations, all robustness levels are higher than 95% to ensure high safety level.
Therefore, the cooperative DPPF-ATA control strategy proposed in Chapter 5 and the
cost-effective RDMPC approach proposed in Chapter 6 have provided effective ways for
handling environmental disturbances for multiple waterborne AGVs.
5. In what way can the scheduling and control loop for waterborne AGVs be closed in order
to obtain an energy-efficient autonomous ITT system?
An energy-efficient autonomous ITT system using waterborne AGVs has been realized
in Chapter 7 with a closed-loop scheduling and control approach. Both the scheduling
and control problems have been solved in real-time with updated system states. Deci-
sions are made hierarchically to guarantee tractability. However, factors such as uncon-
sidered physical system limits, disturbances, and collision avoidance that are difficult,
if not impossible, to be integrated in a scheduling problem can be reflected timely in
the online updated schedules. A new coordinated berthing time schedule model consid-
ering necessary time intervals between different waterborne AGVs visiting a particular
berth has been proposed. For all the berths, safety time intervals are guaranteed for dif-
ferent waterborne AGVs. Furthermore, a partial scheduling problem that is efficient to
solve, and an interaction model that integrates the scheduling and control problems have
been proposed. Cooperative distributed waterborne AGVs are controlled using the fast
ADMM algorithm proposed in Chapter 5. Given a set of ITT requests and a fleet of
waterborne AGVs, the proposed closed-loop schedule and control framework works in
an autonomous way. In our simulation experiments based on a potential ITT scenario in
the port of Rotterdam, time windows of all ITT requests are satisfied by the closed-loop
approach with 0% “non-performance” rate compared to 41.67% from the open-loop ap-
proach. The proposed autonomous ITT system using waterborne AGVs is demonstrated
to be effective and contributes to smarter port logistics.
8.1.2 Contributions
Having answered the above Key Research Questions, the research presented in this PhD
thesis contributes to the state-of-the-art in the following aspects:
• A novel type of container transporter, waterborne AGVs, has been proposed for improv-
ing the port level automation and efficiency, which also stimulates a new research stream
in the fields of control and logistics, see, e.g., [148, 150, 151, 153].
• Systematic advanced control and scheduling strategies have been proposed and demon-
strated to be effective for coordinating waterborne AGVs with applications in ITT.
Control and scheduling

Particularly, the contributions of this PhD thesis to the state-of-the-art with respect to control
and scheduling approaches are as follows:
• Complex networked nonlinear systems in both certain and uncertain cases have been
approximated using a successive linearization approach in the framework of MPC, and
trade-offs among computational complexity, optimality, and ease of controller design
have been achieved in Chapter 3 – 6 (see also [148, 150, 151]);
• The problem of predictive path following with arrival time awareness has been proposed
and solved with a new reference path coordinate system and a two-level parameterization
scheme in Chapter 4 (see also [150]);
• Parallel computations in cooperative distributed MPC for time-varying networked sys-
tems has been achieved via ADMM and a fast ADMM algorithm with faster convergence
rate in Chapter 5 (see also [148, 149]);
• The robustness level of uncertain systems has been proposed, and a cost-effective RDMPC
scheme has been proposed for time-varying networked systems with cooperative dis-
tributed decision making in Chapter 6 (see also [151, 152]);
8.2 Future research 135
• A novel scheduling approach considering safe intervals between berthing time slots of
different vehicles visiting the same berth has been proposed in Chapter 7 (see also [153]);
• Real-time scheduling and control loop has been closed by a partial scheduling model
and an interaction model with feedback reflecting neglected factors from lower levels in
Chapter 7 (see also [153]).
Transport and logistics

With respect to transport and logistics, several typical ITT case studies have been investi-
gated using waterborne AGVs coordinated by the proposed strategies. More specifically,
• A singe ITT task has been achieved using one waterborne AGV with smoothly path
tracking, arrival time awareness, and energy efficiency in Chapter 4 (see also [150]);
• A strategy for cooperatively accomplishing multiple ITT tasks is proposed based on

coordinating multiple waterborne AGVs in a parallel distributed way in Chapter 5 (see
also [148, 149]);
• Cost-effective robust performance has been achieved in fulfilling ITT tasks by multiple
waterborne AGVs even when environmental disturbances are present in Chapter 6 (see
also [152]); and
• An energy efficient autonomous ITT system using waterborne AGVs has been realized
by the proposed closed-loop real-time scheduling and control framework in Chapter 7
(see also [153]).
8.2 Future research

As discussed in Chapter 1, a typical transport system consists of various decision problems
at different levels. In principle, mutual interactions exist between different problems and
an integrated design is desirable. However, an integrated problem size could be intractable.
For the proposed new type of transporter, waterborne AGVs, this thesis has provided several
effective coordination strategies for the involved scheduling and control problems. Recom-
mendations for future research are given in this section with respect to the coordination and
further development of waterborne AGVs. The aim is to stimulate more awareness of water-
borne AGVs in the community, inspire further work on this topic, and advance waterborne
AGVs in practical port logistics systems.
8.2.1 Directions for coordinating waterborne AGVs

This section briefly discusses several technically open issues relevant with coordination of
waterborne AGVs, and gives recommendations for future work on these aspects.
• Modeling accuracy.
This thesis models the waterborne AGV behavior by the dynamics of marine surface ve-
hicles with three DOFs (surge, sway, and yaw) as presented in Chapter 3. This model
considers detailed hydrostatics, hydrodynamics, and kinematics with environmental in-

fluences in the inertial and body-fixed coordinate systems, and involves highly nonlinear
terms. The main maneuvering characteristics of waterborne AGVs can be captured by
this model. However, in practice, the hydrodynamics can be varying in different occa-
sions. Examples include: a) the load, i.e., the number or position of containers, on board
is altered. In Chapter 7, the rigid body mass is adapted to the number of containers on
board of waterborne AGVs. More accurate hydrodynamic influences should be consid-
ered. b) the hydrodynamic interactions between waterborne AGVs in the proximity are
not modeled. For networked waterborne AGV models, collision avoidance couplings
with independent waterborne AGV dynamics are considered. However, hydrodynamic
interactions could give rise to couplings also in dynamics. Model-based controller de-
sign, e.g., MPC, is expected to achieve better control performance using more accurately
modeled waterborne AGVs. In particular, for varying and uncertain parameters in wa-
terborne AGV models, more research work could be done using, e.g., online parameter
identification, robust, and adaptive methods.
• Computational efficiency.
In this thesis, scheduling and control problems have been handled hierarchically for a
tractable solution. Nonetheless, computational efficiency for both the scheduling and
control problems of waterborne AGVs could still be improved.
The scheduling problems proposed in Chapter 7 for ITT using waterborne AGVs are
originally mixed integer nonlinear problems, and transformed into mixed integer linear
problems which are still NP-hard. The computational complexity increases exponentially
with increasing sizes of the waterborne AGV fleet and ITT requests. Therefore, the full
scheduling problem is solved offline handling seven ITT requests within the considered
horizon. More dynamic scenarios dealing with larger numbers of ITT requests are ex-
pecting either more advanced computational hardware, faster exact optimization solvers,
or a fast heuristic solution approach accepting certain optimality gaps.
For the control problems, since MPC requires solving optimization problems online
repetitively, computational efficiency is also critical. Efforts that have been made in
this thesis to relieve possibly heavy computational burdens include successive lineariza-
tions to convexify optimization problems (Chapter 3), distributed control (Chapter 5),
a proposed fast ADMM based distributed control approach (Chapter 5), analytical tube
bounds (Chapter 6), and the efficient search in an integrated branch & bound and ADMM
approach in robust control (Chapter 6). However, for more complicated scenarios, it can
still currently not be guaranteed that decisions are updated timely even with the aforemen-
tioned designs. For example, when a large number of waterborne AGVs are involved, it
could take a long time before an agreement is reached within the group. In general, shar-
ing more information within the coupling group helps in terms of convergence rates for
distributed control algorithms. However, privacy issues might occur. For cases in which
timely update and feedback are critical, a reliable decision recovery mechanism, e.g.,
making safe but not necessarily optimal decisions based on the previous solution, could
be necessary. Furthermore, comparisons between the proposed controllers and other ad-
vanced or practically implemented controllers with respect to system performance and
computational efficiency could also be done in the future work.
8.2 Future research 137
• Coordination between waterborne AGVs with different owners and between waterborne
AGVs and manned marine vehicles.
Waterborne AGVs considered in this thesis are designed to be cooperative to optimize an
overall cost and no priorities are assigned. However, in practice, waterborne AGVs could
belong to different terminals and operated by different companies. A more complicated
scenario involves manned marine vehicles which cannot be controlled to be cooperative.
In these cases, non-cooperative coordination approaches should be investigated. One op-
tion is to assign priorities to vehicles according to certain rules, e.g., the International
Regulations for Preventing Collisions at Sea (COLREGs) [97]. COLREGs are the nav-
igation rules followed by manned ships to prevent collisions. Flexibility and optimality
could be the concerns for the rule-based coordination. More work could be done on inte-
grating rule-based coordination into the proposed coordination strategies in this thesis to
handle mixed traffic situations.
8.2.2 Additional directions for future research

Real applications of waterborne AGVs to ITT in port areas still call for more technological,
methodological, and legal advances. Key challenges regarding the development of water-
borne AGVs are as follows:
• Cost-benefit analysis.
Research on a systematic cost-benefit analysis of applying waterborne AGVs to ITT could
be carried out. Data in terms of labor, time, and cost savings, etc., could be collected from
port authorities for the cost-benefit analysis. Moreover, analytical and simulation models
of an ITT system with waterborne AGVs involved could be built to predict and analyze
the potential strengths and weaknesses of the system.
• Design and construction.

Waterborne AGV design and construction could be seen as a naval architecture problem
and solved by ship building experts. However, knowledge on transport and logistics
should be given as the input. For example, size and capacity of waterborne AGVs can be
related with the ITT demand scenarios. In the context of green port logistics, clean power
systems should be considered in the design and construction of waterborne AGVs.
• Supporting infrastructures.
Parking, maintenance, charging or refueling infrastructures should be designed and built
for Waterborne AGVs. Besides, communication infrastructures including waterborne
AGV-to-waterborne AGV and waterborne AGV-to-infrastructure are important in real-
izing the coordination framework for waterborne AGVs.
• Interactions with other port logistics processes.

As discussed in Chapter 2, currently, ITT is mainly realized by multi-trailer systems. The
assignment problem of ITT requests to different transport modes should be extended to
consider waterborne AGVs. Interactions with terminal operations, such as the scheduling
of quay cranes for loading/unloading waterborne AGVs and the scheduling of land-side
AGVs for transporting unloaded containers could also be investigated.
• Macroscopic marine traffic flow analysis.

At the macroscopic level, the impacts of introducing waterborne AGVs to the existing
maritime traffic could also be identified. On the one hand, since waterborne AGVs take
advantage of advanced information and communication technologies and are operated in
an optimal way, the introduction of them might improve maritime traffic efficiency. On
the other hand, as an extra type of vehicles with possibly lower capacities than manned
barges, the introduce of them might increase the maritime traffic flow. Therefore, it is
suggested that the current maritime traffic flow analysis incorporates waterborne AGVs
in the models and simulations as well.
• Industrial awareness and acceptance.

More work should be done to raise the awareness of waterborne AGVs in relevant indus-
tries. Pilot or experimental platforms should be built for tests and demonstrations.
Bibliography
[1] L. C. Andrews. Special Functions for Engineers and Applied Mathematicians.

Macmillan, London, UK, 1985.
[2] D. Angeli. Economic model predictive control. In J. Baillieul and T. Samad, editors,
Encyclopedia of Systems and Control, pages 1–9. Springer London, 2014.
[3] APM Terminals. APM terminals maasvlakte II terminal officially

opens. http : / / www . apmterminals . com / news / press-releases / 2014 / 12 /
apmt-terminals-maasvlakte2-officially-opens, 2015. Accessed: May 2016].
[4] K. J. Aström and R. M. Murray. Feedback Systems: An Introduction For Scientists

and Engineers. Princeton university press, New Jersey, USA, 2nd edition, 2015.
[5] M. Baldea and I. Harjunkoski. Integrated production scheduling and process control:
A systematic review. Computers & Chemical Engineering, 71:377–390, 2014.
[6] E. M. L. Beale and J. A. Tomlin. Special facilities in a general mathematical program-

ming system for non-convex problems using ordered sets of variables. In Operational
Research, volume 69, pages 447–454. Tavistock Publishing, London, 1970.
[7] T. Bektaş and G. Laporte. The pollution-routing problem. Transportation Research

Part B: Methodological, 45(8):1232–1250, 2011.
[8] A. Bemporad, M. Morari, V. Dua, and E. N. Pistikopoulos. The explicit linear

quadratic regulator for constrained systems. Automatica, 38(1):3–20, 2002.
[9] E. A. Boroujeni and H. R. Momeni. Adaptive sliding mode control for roll motions
of ships. In Proceedings of International Conference on Control, Automation and
Systems, pages 1622–1625, Seoul, South Korea, 2008.
[10] J. R. Boyd. Patterns of conflict. Technical report, Washington, DC: U.S. Air Force,
1986. Accessed: October 2016.
[11] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge university press,

Cambridge, UK, 2004.
[12] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein. Distributed optimization and
statistical learning via the alternating direction method of multipliers. Foundations
and Trends in Machine Learning, 3(1):1–122, 2011.
139
140 Bibliography
[13] G. Bruzzone, G. Bruzzone, M. Bibuli, and M. Caccia. Autonomous mine hunting

mission for the charlie USV. In Proceedings of 2011 OCEANS, pages 1–6, Santander,
Spain, 2011.
[14] S. Campbell, W. Naeem, and G. W. Irwin. A review on improving the autonomy

of unmanned surface vehicles through intelligent collision avoidance manoeuvres.
Annual Reviews in Control, 36(2):267–283, 2012.
[15] C. Chen, C.-C. Liao, and C. Chiang. Adaptive neuro-wavelet control for the ship
trajectory tracking problem. In Proceedings of the International Joint Conference on
Neural Networks, pages 1–6, Brisbane, Australia, 2012.
[16] M. Christiansen, K. Fagerholt, B. Nygreen, and D. Ronen. Maritime transportation.

In C. Barnhart and G. Laporte, editors, Handbooks in Operations Research and Man-
agement Science: Transportation, volume 14, pages 189–284. Elsevier, Amsterdam,
2007.
[17] J. Curcio, J. Leonard, and A. Patrikalakis. SCOUT – a low cost autonomous surface
platform for research in cooperative autonomy. In Proceedings of 2005 OCEANS,
pages 725–729, Washington, USA, 2005.
[18] R. Diestel. Graph Theory. Springer-Verlag, Heidelberg, Germany, 4th edition, 2010.
[19] K. D. Do and J. Pan. Robust path-following of underactuated ships: Theory and

experiments on a model ship. Ocean Engineering, 33(10):1354–1372, 2006.
[20] J. Du, J. Park, I. Harjunkoski, and M. Baldea. A time scale-bridging approach for
integrating production scheduling and process control. Computers & Chemical En-
gineering, 79:59–69, 2015.
[21] M. B. Duinkerken, R. Dekker, S. T. Kurstjens, J. A. Ottjes, and N. P. Dellaert. Com-

paring transportation systems for inter-terminal transport at the maasvlakte container
terminals. OR Spektrum, 28(4):469–493, 2006.
[22] W. B. Dunbar and R. M. Murray. Distributed receding horizon control for multi-
vehicle formation stabilization. Automatica, 42(4):549–558, 2006.
[23] Egemin Automation Inc. Automated guided vehicles (AGVs). http : / / www.
egeminusa.com/pages/agvs/agvs general.html, 2015. Accessed May 2016.
[24] Erasmus Smart Port Rotterdam. SmartPort/Port research center Rotterdam

Delft poster session 2014. http :/ /www.erim. eur. nl/centres /smartporterasmus /
news / detail / 3585-smartportport-research-center-rotterdam-delft-poster- \
session-20131-posters-were-presented-in/, February, 2015. Accessed: June,
2016.
[25] K. Fagerholt, G. Laporte, and I. Norstad. Reducing fuel emissions by optimizing

speed on shipping routes. Journal of the Operational Research Society, 61(3):523–
529, 2010.
Bibliography 141
[26] F. Farokhi, I. Shames, and K. H. Johansson. Distributed MPC via dual decompo-
sition and alternative direction method of multipliers. In J. M. Maestre and R. R.
Negenborn, editors, Distributed Model Predictive Control Made Easy, pages 115–
131. Springer, 2014.
[27] M. A. Figliozzi. The time dependent vehicle routing problem with time windows:
Benchmark problems, an efficient solution algorithm, and solution characteristics.
Transportation Research Part E: Logistics and Transportation Review, 48(3):616–
636, 2012.
[28] T. I. Fossen, M. Breivik, and R. Skjetne. Line-of-sight path following of underactu-

ated marine craft. In Proceedings of the 6th IFAC on Manoeuvring and Control of
Marine Craft, pages 244–249, Girona, Spain, 2003.
[29] T. Fossen and T. Perez. Kalman filtering for positioning and heading control of ships
and offshore rigs. IEEE Control Systems Magazine, 29(6):32–46, 2009.
[30] T. I. Fossen. Handbook of Marine Craft Hydrodynamics and Motion Control. John
Wiley and Sons Ltd., West Sussex, U.K., 2011.
[31] H. P. Geering. Optimal Control with Engineering Applications. Springer, Berlin,

German, 2007.
[32] E. Gerritse. Analysis for inter terminal transportation demand scenarios for the
Maasvlakte I and II in 2030. Technical report, Delft University of Technology, Delft,
The Netherlands, 2014.
[33] P. Giselsson, M. D. Doan, T. Keviczky, B. De Schutter, and A. Rantzer. Accelerated

gradient methods and dual decomposition in distributed model predictive control.
Automatica, 49(3):829–833, 2013.
[34] P. Gomes, C. Silvestre, A. Pascoal, and R. Cunha. A path-following controller for

the DELFIMx autonomous surface craft. In Proceedings of 7th IFAC Conference on
Manoeuvring and Control of Marine Craft, Lisbon, Portugal, 2006.
[35] Google. Maasvlakte Rotterdam, 51.962398◦ N and 4.056800◦ E, Google Earth,

August 7, 2013. Accessed: 2014-10-17.
[36] L. Grüne and J. Pannek. Nonlinear Model Predictive Control. Springer-Verlag Lon-
don, 2011.
[37] Gurobi Optimization. Gurobi optimizer reference manual, 2012. URL http://www.
gurobi.com. Accessed: June, 2016.
[38] N. Harl and S. Balakrishnan. Impact time and angle guidance with sliding mode
control. IEEE Transactions on Control Systems Technology, 20(6):1436–1449, 2012.
[39] B. He, H. Yang, and S. Wang. Alternating direction method with self-adaptive penalty
parameters for monotone variational inequalities. Journal of Optimization Theory
and applications, 106(2):337–356, 2000.
142 Bibliography
[40] L. Heilig and S. Voß. A cloud-based SOA for enhancing information exchange and
decision support in ITT operations. In R. G. Gonzlez-Ramrez, F. Schulte, S. Voß, and
J. A. Ceroni Daz, editors, Computational Logistics, pages 112–131. Springer, New
York, USA, 2014.
[41] L. Heilig and S. Voß. Inter-terminal transportation: an annotated bibliography and

research agenda. Flexible Services and Manufacturing Journal, pages 1–29, 2016.
doi: 10.1007/s10696-016-9237-7.
[42] C. Hetherington, R. Flin, and K. Mearns. Safety in shipping: The human element.
Journal of safety research, 37(4):401–411, 2006.
[43] S. Hill and M. Cannon. A potential feedback approach to ecosystem-based man-

agement: model predictive control of the antarctic krill fishery. Commission for the
Conservation of Antarctic Marine Living Resources Science, 20:119–137, 2013.
[44] Y. Ho and T. Liao. Zone design and control for vehicle collision prevention and load
balancing in a zone control agv system. Computers & Industrial Engineering, 56(1):
417–432, 2009.
[45] J. E. Hopcroft and R. E. Tarjan. Efficient algorithms for graph manipulation. Com-
munications of the ACM, 16(6):372378, 1971.
[46] ILOG. IBM ILOG CPLEX Optimizer, 2010. URL http://www-01.ibm.com/software/

integration/optimization/cplex-optimizer/. Accessed: June 2016.
[47] D. Jia and B. Krogh. Min-max feedback model predictive control for distributed con-
trol with communication. In Proceedings of the 2002 American Control Conference,
volume 6, pages 4507–4512, Anchorage, USA, 2002.
[48] S. J. Julier, J. K. Uhlmann, and H. F. Durrant-Whyte. A new approach for filter-

ing nonlinear systems. In Proceedings of the 1995 American Control Conference,
volume 3, pages 1628–1632, Seattle, USA, 1995.
[49] C. M. Kellett and A. R. Teel. Smooth lyapunov functions and robustness of stability
for difference inclusions. Systems & Control Letters, 52(5):395–405, 2004.
[50] T. Keviczky, F. Borrelli, K. Fregene, D. Godbole, and G. J. Balas. Decentralized

receding horizon control and coordination of autonomous vehicle formations. IEEE
Transactions on Control Systems Technology, 16(1):19–33, 2008.
[51] B. Kouvaritakis, M. Cannon, and J. Rossiter. Non-linear model based predictive

control. International Journal of Control, 72(10):919–928, 1999.
[52] B. Kouvaritakis, M. Cannon, S. V. Raković, and Q. Cheng. Explicit use of prob-

abilistic distributions in linear predictive control. Automatica, 46(10):1719–1724,
2010.
[53] Y. Kuwata and J. P. How. Cooperative distributed robust trajectory optimization using
receding horizon MILP. IEEE Transactions on Control Systems Technology, 19(2):
423–431, 2011.
Bibliography 143
[54] Y. Kuwata, M. T. Wolf, D. Zarzhitsky, and T. L. Huntsberger. Safe maritime au-

tonomous navigation with COLREGs using velocity obstacles. IEEE Journal of
Ocean Engineering, 39(1):110–119, 2014.
[55] Laboratory for Autonomous Marine Sensing Systems. MOOS-IvP home page.
http://oceanai.mit.edu/moos-ivp/pmwiki/pmwiki.php?n=Main.HomePage, 2015.
Accessed: October 2016.
[56] S. G. Lachmann and R. G. McKinstry. Software engineering of a navigation and
guidane system for commercial aircraft. In Proceedings of the AIAA Digital Avionics
System Conference, pages 1–15, Boston, US, 1975.
[57] G. Laporte, S. Ropke, and T. Vidal. Heuristics for the vehicle routing problem. In
P. Toth and D. Vigo, editors, Vehicle Routing: Problems, Methods, and Applications,
pages 87–116. SIAM, Philadephia, 2nd edition, 2014.
[58] M. Lazar, W. Heemels, S. Weiland, and A. Bemporad. Stabilizing model predictive
control of hybrid systems. IEEE Transactions on Automatic Control, 51(11):1813–
1818, 2006.
[59] H. Li and Y. Shi. Robust distributed model predictive control of constrained
continuous-time nonlinear systems: a robustness constraint approach. IEEE Trans-
actions on Automatic Control, 59(6):1673–1678, 2014.
[60] L. Li, R. R. Negenborn, and B. De Schutter. Intermodal freight transport planning–
a receding horizon control approach. Transportation Research Part C: Emerging
Technologies, 60:77–95, 2015.
[61] S. Li, H. Zheng, R. R. Negenborn, and G. Lodewijks. Coordination for efficient trans-
port over water. In Proceedings of the 19th International Conference on Computer
Supported Cooperative Work in Design, pages 389–394, Calabria, Italy, 2015.
[62] Z. Li and J. Sun. Disturbance compensating model predictive control with application
to ship heading control. IEEE Transactions on Control Systems Technology, 20(1):
257–265, 2012.
[63] Z. Li, J. Sun, and S. Oh. Path following for marine surface vessels with rudder and
roll constraints: an MPC approach. In Proceedings of American Control Conference,
pages 3611–3616, St. Louis, Missouri, 2009.
[64] S. Lin, B. De Schutter, Y. Xi, and H. Hellendoorn. Efficient network-wide model-
based predictive control for urban traffic networks. Transportation Research Part C:
Emerging Technologies, 24:122–140, 2012.
[65] X. Liu, Y. Shi, and D. Constantinescu. Distributed model predictive control of con-
strained weakly coupled nonlinear systems. Systems & Control Letters, 74:41–49,
2014.
[66] J. Lofberg. YALMIP: A toolbox for modeling and optimization in matlab. In Proc-
cedings of 2004 IEEE International Symposium on Computer Aided Control Systems
Design, pages 284–289, Taipei, Taiwan, 2004.
144 Bibliography
[67] J. Löfberg. Automatic robust convex programming. Optimization Methods and Soft-
ware, 27(1):115–129, 2012.
[68] Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, and P. Haves. Model predictive
control for the operation of building cooling systems. IEEE Transactions on Control
Systems Technology, 20(3):796–803, 2012.
[69] Maersk. Big ship means big pollution, right? http://www.maersk.com/en/
test-section/wls-test/stories/big-ship-means-big-pollution-right, 2013. Accessed
January 2016.
[70] J. Maestre, D. Muñoz De La Peña, and E. Camacho. Distributed model predictive
control based on a cooperative game. Optimal Control Applications and Methods, 32
(2):153–176, 2011.
[71] J. M. Maestre and R. R. Negenborn. Distributed model predictive control made easy.
Springer, Dordrecht, 2014.
[72] J. Majohr and T. Buch. Modelling, simulation and control of an autonomous surface
marine vehicle for surveying applications Measuring Dolphin MESSIN. In G. N.
Roberts and R. Sutton, editors, Advances in Unmanned Marine Vehicles, volume 69,
pages 329 – 352. MPG Books Ltd, Cornwall, UK, 2006.
[73] M. Maloni, J. A. Paul, and D. M. Gligor. Slow steaming impacts on ocean carriers
and shippers. Maritime Economics & Logistics, 15(2):151–171, 2013.
[74] K. Margellos, P. Goulart, and J. Lygeros. On the road between robust optimiza-
tion and the scenario approach for chance constrained optimization problems. IEEE
Transactions on Automatic Control, 59(8):2258–2263, 2014.
[75] MATLAB. R2014b. The MathWorks Inc., Natick, Massachusetts, 2014.
[76] D. Q. Mayne. Model predictive control: Recent developments and future promise.
Automatica, 50(12):2967–2986, 2014.
[77] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. Scokaert. Constrained model
predictive control: Stability and optimality. Automatica, 36(6):789–814, 2000.
[78] D. Q. Mayne, M. M. Seron, and S. Raković. Robust model predictive control of
constrained linear systems with bounded disturbances. Automatica, 41(2):219–224,
2005.
[79] N. Minorski. Directional stability of automatically steered bodies. Journal of Amer-
ican Society of Naval Engineers, 42(2):280–309, 1922.
[80] L. Moreira, T. I. Fossen, and C. G. Soares. Path following control system for a tanker
ship model. Ocean Engineering, 34(14):2074–2085, 2007.
[81] J. F. Mota, J. M. Xavier, P. M. Aguiar, and M. Püschel. D-ADMM: A distributed
algorithm for compressed sensing and other separable optimization problems. In
Proceedings of International Conference on Acoustics, Speech and Signal Process-
ing, pages 2869–2872, Kyoto, Japan, 2012.
Bibliography 145
[82] J. F. Mota, J. M. Xavier, P. M. Aguiar, and M. Puschel. Distributed ADMM for model
predictive control and congestion control. In Proceedings of 51st Annual Conference
on Decision and Control, pages 5110–5115, Maui, USA, 2012.
[83] T. Müller. Automated guided vehicles. IFS (Publications) Ltd./Springer-Verlag,

UK/Berlin, 1983.
[84] W. Naeem, R. Sutton, J. Chudley, F. Dalgleish, and S. Tetlow. An online genetic algo-
rithm based model predictive control autopilot design with experimental verification.
International Journal of Control, 78(14):1076–1090, 2005.
[85] W. Naeem, T. Xu, R. Sutton, and A. Tiano. The design of a navigation, guidance, and
control system for an unmanned surface vehicle for environmental monitoring. Pro-
ceedings of the Institution of Mechanical Engineers, Part M: Journal of Engineering
for the Maritime Environment, 222(2):67–79, 2008.
[86] W. Naeem, G. W. Irwin, and A. Yang. COLREGs-based collision avoidance strate-

gies for unmanned surface vehicles. Mechatronics, 22(6):669–678, 2012.
[87] NATO Industrial Advisory Group (NIAG). Pre-feasibility study on UAV autonomous
operations. Technical report, NIAG Special Group 75, 2004.
[88] R. R. Negenborn. Delfia-1 autonomous vessel facility. http://www.negenborn.net/,

2016. Accessed: June 2016.
[89] R. R. Negenborn and J. M. Maestre. On 35 approaches for distributed MPC made

easy. In J. M. Maestre and R. R. Negenborn, editors, Distributed Model Predictive
Control Made Easy, pages 1–37. Springer, 2014.
[90] R. R. Negenborn, B. De Schutter, and J. Hellendoorn. Multi-agent model predictive

control for transportation networks: Serial versus parallel schemes. Engineering
Applications of Artificial Intelligence, 21(3):353–366, 2008.
[91] L. A. Nguyen, M. D. Le, S. H. Nguyen, T. H. H. Nghiem, et al. A new and effective

fuzzy PID autopilot for ships. In SICE 2003 Annual Conference, volume 3, pages
2647–2650, 2003.
[92] I. Norstad, K. Fagerholt, and G. Laporte. Tramp ship routing and scheduling with
speed optimization. Transportation Research Part C: Emerging Technologies, 19(5):
853–865, 2011.
[93] S. Oh and J. Sun. Path following of underactuated marine surface vessels using line-
of-sight based model predictive control. Ocean Engineering, 37(2):289–295, 2010.
[94] F. Oldewurtel, C. N. Jones, A. Parisio, and M. Morari. Stochastic model predictive

control for building climate control. IEEE Transactions on Control Systems Technol-
ogy, 22(3):1198–1205, 2014.
[95] H. Y. Ong and J. C. Gerdes. Cooperative collision avoidance via proximal message
passing. In Proceedings of 2015 American Control Conference, pages 4124–4130,
Chicago, USA, 2015.
146 Bibliography
[96] K. W. Pang, Z. Xu, and C. L. Li. Ship routing problem with berthing time clash
avoidance constraints. International Journal of Production Economics, 131(2):752–
762, 2011.
[97] Parliament of the United Kingdom. International regulations for preventing collisions
at sea. Technical report, IMO (The International Maritime Organisation), London,
UK, 1972.
[98] A. Pavlov, H. Nordahl, and M. Breivik. MPC-based optimal path following for un-
deractuated vessels. In Proceedings of the 8th IFAC International Conference on
Manoeuvring and Control of Marine Craft, pages 340–345, Guaruj, Brazil, 2009.
[99] T. Perez, O. Smogeli, T. I. Fossen, and A. Sorensen. An overview of the marine

systems simulator (MSS): a simulink toolbox for marine control systems. Modeling,
Identification and Control, 27(4):259–275, 2006.
[100] Z. Pietrzykowski and J. Uriasz. The ship domain–a criterion of navigational safety
assessment in an open sea area. Journal of Navigation, 62(01):93–108, 2009.
[101] Port of Rotterdam. Port Vision 2030. http://www.portofrotterdam.com/en/Port/

port-in-general/port-vision-2030/Pages/default.aspx, 2011. Accessed: November
2014.
[102] Port of Rotterdam. Maasvlakte II. https://www.maasvlakte2.com/nl/index/, 2013.

Accessed: May 2016.
[103] Port of Rotterdam. Expected weather conditions in the port of rotterdam,

2015. URL https://www.portofrotterdam.com/en/shipping/up-to-date-information/
current-water-levels-flow-wind-and-visibility. Accessed: July 2016.
[104] J. A. Primbs and C. H. Sung. Stochastic receding horizon control of constrained

linear systems with state and control multiplicative noise. IEEE Transactions on
Automatic Control, 54(2):221–230, 2009.
[105] H. N. Psaraftis and C. A. Kontovas. Ship speed optimization: Concepts, models

and combined speed-routing scenarios. Transportation Research Part C: Emerging
Technologies, 44:52–69, 2014.
[106] H. N. Psaraftis, M. Wen, and C. A. Kontovas. Dynamic vehicle routing problems:

Three decades and counting. Networks, 67(1):3–31, 2016.
[107] S. J. Qin and T. A. Badgwell. A survey of industrial model predictive control tech-
nology. Control engineering practice, 11(7):733–764, 2003.
[108] Rafael Advanced Defense Systems Ltd. Unmanned naval patrol vehicle – PRO-
TECTOR, December 2010. URL http://www.rafael.co.il/Marketing/351-1037-en/
Marketing.aspx. Accessed: June 2016.
[109] J. B. Rawlings, D. Angeli, and C. N. Bates. Fundamentals of economic model predic-

tive control. In Proceedings of the 51st Annual Conference on Decision and Control,
pages 3851–3861, Maui, USA, 2012.
Bibliography 147
[110] J. B. Rawlings and D. Q. Mayne. Model predictive control: Theory and design. Nob
Hill Pub., Madison, USA, 2009.
[111] A. Richards and J. P. How. Robust distributed model predictive control. International
Journal of Control, 80(9):1517–1531, 2007.
[112] M. P. Richter. Analysis of operational manning requirements and deployment pro-

cedures for unmanned surface vehicles aboard US Navy ships. PhD thesis, Naval
Postgraduate School, Monterey, United States, 2006.
[113] S. Riverso and G. Ferrari-Trecate. Tube-based distributed control of linear con-

strained systems. Automatica, 48(11):2860–2865, 2012.
[114] G. Roussos and K. J. Kyriakopoulos. Decentralized navigation and conflict avoidance

for aircraft in 3-d space. IEEE Transactions on Control Systems Technology, 20(6):
1622–1629, 2012.
[115] M. W. Savelsbergh and M. Sol. The general pickup and delivery problem. Trans-
portation Science, 29(1):17–29, 1995.
[116] R. Scattolini. Architectures for distributed and hierarchical model predictive control–
a review. Journal of Process Control, 19(5):723–731, 2009.
[117] T. Schouwenaars, B. De Moor, E. Feron, and J. How. Mixed integer programming

for multi-vehicle path planning. In Proceedings of 6th European Control Conference,
pages 2603–2608, Porto, Portugal, 2001.
[118] H. J. Schroër, F. Corman, M. B. Duinkerken, R. R. Negenborn, and G. Lodewijks.

Evaluation of inter terminal transport configurations at Rotterdam Maasvlakte using
discrete event simulation. In Proceedings of Winter Simulation Conference, pages
1771–1782, Savannah, USA, 2014.
[119] M. S. Sedehi and R. Z. Farahani. An integrated approach to determine the block lay-
out, agv flow path and the location of pick-up/delivery points in single-loop systems.
International Journal of Production Research, 47(11):3041–3061, 2009.
[120] F. Semet, P. Toth, and D. Vigo. Classical exact algorithms for the capacitated ve-
hicle routing problem. In P. Toth and D. Vigo, editors, Vehicle Routing: Problems,
Methods, and Applications, pages 37–57. SIAM, Philadephia, 2nd edition, 2014.
[121] R. Skjetne. The maneuvering problem. PhD thesis, Norwegian University of Science
and Technology, Trondheim, Norway, 2005.
[122] R. Skjetne, T. I. Fossen, and P. V. Kokotović. Adaptive maneuvering with experi-

ments for a model ship in a marine control laboratory. Automatica, 41(2):289–298,
2005.
[123] SNAME: The Society of Naval Architects and Marine Engineers. Nomenclature for
treating the motion of a submerged body through a fluid. Technical and Research
Bulletin, pages 1–5, 1952.
148 Bibliography
[124] A. J. Sørensen. A survey of dynamic positioning control systems. Annual Reviews

in Control, 35(1):123–136, 2011.
[125] T. H. Summers and J. Lygeros. Distributed model predictive consensus via the al-
ternating direction method of multipliers. In in Proceedings of 50th Annual Allerton
Conference on Communication, Control, and Computing, pages 79–84, Urbana IL,
USA, 2012.
[126] R. Sutton. Design of the multi-role springer unmanned surface vehicle. Technical
report, Plymouth University, Plymouth, UK, 2007.
[127] N. Tarcai, C. Virágh, D. Ábel, M. Nagy, P. L. Várkonyi, G. Vásárhelyi, and T. Vicsek.

Patterns, transitions and the role of leaders in the collective dynamics of a simple
robotic flock. Journal of Statistical Mechanics: Theory and Experiment, 2011(04):
1–26, 2011.
[128] Terex Corporation. Terex Gottwald automated guided vehicles. http://www.terex.

com/port-solutions/en/products/automated-guided-vehicles/index.htm, 2016. Ac-
cessed: May 2016.
[129] K. Tierney, S. Voß, and R. Stahlbock. A mathematical model of inter-terminal trans-

portation. European Journal of Operational Research, 235(2):448–460, 2014.
[130] P. Toth and D. Vigo. Vehicle Routing: Problems, Methods, and Applications, vol-
ume 18. SIAM, Philadelphia, 2nd edition, 2014.
[131] P. Trodden. Feasible parallel-update distributed MPC for uncertain linear systems
sharing convex constraints. Systems & Control Letters, 74:98–107, 2014.
[132] P. Trodden and A. Richards. Distributed model predictive control of linear systems
with persistent disturbances. International Journal of Control, 83(8):1653–1663,
2010.
[133] P. Trodden and A. Richards. Cooperative distributed MPC of linear systems with
coupled constraints. Automatica, 49(2):479–487, 2013.
[134] R. Vilanova and A. Visioli. PID Control in the Third Millennium. Springer, Springer-
Verlag London, UK, 2012.
[135] I. F. Vis. Survey of research in the design and control of automated guided vehicle
systems. European Journal of Operational Research, 170(3):677–709, 2006.
[136] I. F. Vis, R. M. B. de Koster, and M. W. Savelsbergh. Minimum vehicle fleet size

under time-window constraints at a container terminal. Transportation Science, 39
(2):249–260, 2005.
[137] A. Wahl and E. Gilles. Track-keeping on waterways using model predictive control.
In Proceedings of the IFAC Conference on Control Applications in Marine Systems,
pages 149–154, Fukuoka, Japan, 1998.
Bibliography 149
[138] M. Wang, W. Daamen, S. P. Hoogendoorn, and B. van Arem. Cooperative car-

following control: Distributed algorithm and impact on moving jam features. IEEE
Transactions on Intelligent Transportation Systems, 17(5):1459–1471, 2016.
[139] X. Wang, Z. Zou, T. Li, and W. Luo. Path following control of underactuated ships
based on nonswitch analytic model predictive control. Journal of Control Theory and
Applications, 8(4):429–434, 2010.
[140] H. P. Williams. Logical problems and integer programming. Bulletin of the Institute
of Mathematics and its Applications, 13:18–20, 1977.
[141] L. A. Wolsey. Integer programming, volume 42. Wiley, New York, 1998.
[142] J. Xin, R. R. Negenborn, and G. Lodewijks. Energy-aware control for automated con-
tainer terminals using integrated flow shop scheduling and optimal control. Trans-
portation Research Part C: Emerging Technologies, 44:214–230, 2014.
[143] J. Xin, R. R. Negenborn, F. Corman, and G. Lodewijks. Control of interacting ma-

chines in automated container terminals using a sequential planning approach for
collision avoidance. Transportation Research Part C: Emerging Technologies, 60:
377–396, 2015.
[144] M. N. Zeilinger, M. Morari, and C. N. Jones. Soft constrained model predictive

control with robust stability guarantees. IEEE Transactions on Automatic Control,
59(5):1190–1202, 2014.
[145] H. Zheng, Z. Huang, C. Wu, and R. R. Negenborn. Model predictive control for
intelligent vehicle lane change. In Proceedings of the 2nd Internation Conference on
Transportation Information and Safety, pages 265–276, Wuhan, China, 2013.
[146] H. Zheng, R. R. Negenborn, and G. Lodewijks. Trajectory tracking of autonomous

vessels using model predictive control. In Proceedings of the 19th IFAC World
Congress, pages 8812–8818, Cape Town, South Africa, August 2014.
[147] H. Zheng, R. R. Negenborn, and G. Lodewijks. Survey of approaches for improving

the intelligence of marine surface vehicles. In Proceedings of the 16th International
Conference on Intelligent Transportation Systems, pages 1217–1223, Den Haag, The
Netherlands, 2013.
[148] H. Zheng, R. R. Negenborn, and G. Lodewijks. Cooperative distributed collision

avoidance based on ADMM for waterborne AGVs. In F. Corman, S. Voß, and R. R.
Negenborn, editors, Computational Logistics, pages 181–194. Cham, Switzerland,
2015.
[149] H. Zheng, R. R. Negenborn, and G. Lodewijks. Fast ADMM for distributed

model predictive control of cooperative waterborne AGVs. Accepted for publica-
tion in IEEE Transactions on Control Systems Technology, In press, 2016. doi:
http://dx.doi.org/10.1109/TCST.2016.2599485.
150 Bibliography
[150] H. Zheng, R. R. Negenborn, and G. Lodewijks. Predictive path following with arrival
time awareness for waterborne AGVs. Transportation Research Part C: Emerging
Technologies, 70:214 – 237, 2016.
[151] H. Zheng, R. R. Negenborn, and G. Lodewijks. Explicit use of probabilistic distribu-

tions in robust predictive control of waterborne AGVs - a cost-effective approach. In
Proceedings of the 15th European Control Conference (ECC16), pages 1278 – 1283,
Aalborg, Denmark, 2016.
[152] H. Zheng, R. R. Negenborn, and G. Lodewijks. Robust distributed predictive control

of waterborne AGVs – a cooperative and cost-effective approach. Submitted to a
journal, 2016.
[153] H. Zheng, R. R. Negenborn, and G. Lodewijks. Closed-loop scheduling and control

of waterborne AGVs for energy-efficient inter terminal transport. Transportation
Research Part E: Logistics and Transportation Review, In press, 2016. doi: http:
//dx.doi.org/10.1016/j.tre.2016.07.010.
Glossary
List of symbols and notations

Below follows a list of the most frequently used symbols and notations in this thesis.
ai arrival time at node i

a binary vector parameterizing the distribution function
a∗ LP-relaxed node solution
âa p copied binary parameterizing vector by waterborne AGV p
A Jacobian state matrix
AK p state matrix with feedback
A virtual arc set
b environmental disturbances
bp binary switching logic decision variable
{b p } the body-fixed coordinate system of waterborne AGV p
b binary arrival time decision variable
b obs,i , i = 1, 2, 3, 4 binary obstacle avoidance decision variable
b̄ predicted disturbance force
j j
B p , Bq outboundings of uncertainty perturbation positions of waterborne AGV p and q
j
Bu p outbounding of uncertain perturbation control inputs of waterborne AGV p
j
Bx p outbounding of uncertain perturbation states of waterborne AGV p
B Jacobian control input matrix
B berth set
c1 , · · · , c6 weight parameters
C Jacobian position matrix of waterborne AGV p
Cb set of nodes representing the same physical berths
Cx p convex set constraints on perturbation states of waterborne AGV p
Cu p convex set constraints on perturbation control inputs of waterborne AGV p
Cr p,q convex set constraints on perturbation positions variables of waterborne AGVs p and q
di delivery berth of request i or delay time at node i

di j distance between nodes i and j
ds safety distance from the obstacle
D Jacobian position matrix of waterborne AGV q
151
152 Glossary
Dc communication range of waterborne AGVs

Ds minimum safety distance between waterborne AGVs
Dn delivery node set
e p,q edge between waterborne AGVs p and q

E Jacobian disturbance matrix
E , Es set of couplings between waterborne AGVs
G waterborne AGV graphs

Gs virtual scheduling graph
i request or node ID
i∗ branch point by SOS1 branching
IC indicator function for Cr p,q
Ii j binary variable for whether node i visited before node j, i, j ∈ Cb
Ix , Iu number of state and control input constraints
I index set for a
I0 subset of I
j iteration number in ADMM

J∗ incumbent objective
∗
JLP LP-relaxed node objective
j j
JP , JD primal and dual objectives at iteration j
k discrete time step

Kp time-varying feedback gain of waterborne AGV p
lj length of path segment j

L vector of berths corresponding to N
Lρ ρ-augmented Lagrangian function
mc mass of one TEU container

M a big value
n, ns number of waterborne AGVs

nu number of speed levels
nb dimension of a
nG number of subgraphs
{n} the inertial coordinate system
N discrete preferable arrival times
Nmax discrete latest arrival times
Np prediction horizon
N virtual node set
p waterborne AGV ID or probability of b ∈ [−z, z]

pi pick-up berth of request i
{p j } the path coordinate j
P discrete probability vector
Glossary 153
Pn pick-up node set
qi container number of request i

Q state cost weight matrix
Qf terminal state cost weight matrix
r discrete speed levels

rp yaw rate of waterborne AGV p
rj primal residual at iteration j of ADMM
rp position vector of waterborne AGV p
R control input weight matrix
R request set
s subgraph ID
sj dual residual at iteration j of ADMM
si service time of request i
sv traveled distance of waterborne AGV v on a route
Si j binary variable for whether node i, j ∈ Cb visited by different waterborne AGVs
t continuous time
ti,min release time of request i
ti,max due time of request i
tf continuous and discrete preferable arrival times
tf continuous and discrete latest arrival times
T time interval between service time slots of different waterborne AGVs at the same berth
Ts sampling time
T diagonal translation matrix
ui j speed a waterborne AGV travels at on leg i → j

up surge speed of waterborne AGV p
ur,min , ur,max minimum and maximum speeds of speed level r
ūr average speed of speed level r
up system control input vector of waterborne AGV p
∆ūu p nominal perturbation control inputs of waterborne AGV p
∆ũu p uncertain perturbation control inputs of waterborne AGV p
us,l lower level control input vector
u s,h higher level control input vector
U Gs control input vector of subgraph s
Up control input tube of waterborne AGV p
v waterborne AGV ID
vp sway speed of waterborne AGV p
νp velocity vector of waterborne AGV p
ν p,r relative velocity vector of waterborne AGV p with respect to current speed
ν p min minimum velocity vector of waterborne AGV p
ν p max maximum velocity vector of waterborne AGV p
ν (tf ) final velocity vector
154 Glossary
Vc current speed
Vi , i = 1, 2, · · · numbered waterborne AGVs
V , Vs set of waterborne AGVs
Vo , Ve start and end location sets of waterborne AGVs
wi waiting time at node i

wmax , dmax maximum waiting and delay times at all nodes
W compact set for uncertainty bounds
xi jv binary variable for whether waterborne AGV v traveling from node i → j

xp coordinate along Xn axis of waterborne AGV p
x s,l lower level state vector
x s,h higher level state vector
xp system state vector of waterborne AGV p
x ep j tracking error vector in the path coordinate {p j }
∆x̄x p nominal perturbation states of waterborne AGV p
∆x̃x p uncertain perturbation states of waterborne AGV p
X Gs state vector of subgraph s
Xp state tube of waterborne AGV p
∆X̃ p convex sets of uncertain perturbation states of waterborne AGV p
yp coordinate along Yn axis of waterborne AGV p

yi load on board waterborne AGVs upon arriving node i
z uncertainty bounds
ziv binary variable for whether node i visited by waterborne AGV v
Z discrete bound vector
βc current angle
j+1 j+1
δs deviations of ∆rr p from ∆r̂r pj of subgraph s
εpri primal tolerance

εdual dual tolerance
ηp pose vector of waterborne AGV p

ηpj pose vector in the path coordinate j
η (tf ) final pose vector
λ p,a dual variable for consensus constraints on binary parametering vectors

λ p,r dual variable for position variable consensus constraints of waterborne AGV p
τu,p surge force of waterborne AGV p

τv,p sway force of waterborne AGV p
τr,p yaw moment of waterborne AGV p
τ control input vector of waterborne AGV p
τ p min minimum control input vector of waterborne AGV p
τ p max maximum control input vector of waterborne AGV p
Glossary 155
ψp heading angle of waterborne AGV p

ψj angle of path segment j with Xn axis
Σ covariance of disturbance prediction distribution
List of abbreviations
The following abbreviations are used in this thesis:
waterborne AGVs waterborne Autonomous Guided Vessels

AGVs Automated Guided Vehicles
ITT Inter Terminal Transport
TEU Twenty-foot Equivalent Unit
MPC Model Predictive Control
NGC navigation, guidance, and control
LOS line-of-sight guidance
DOF degree-of-freedom
PPF-ATA predictive path following with arrival time awareness
MIQP mixed integer quadratic programming
DPPF-ATA distributed predictive path following with arrival time awareness
ADMM the Alternating Direction Method of Multipliers
DMPC distributed Model Predictive Control
RDMPC robust distributed Model Predictive Control
B&B branch and bound
LP linear programming
MILP mixed integer linear programming
GUB generalized upper bound
SOS1 special ordered sets of type I
PDP the pick-up and delivery problem
VRP the vehicle routing problem
TRAIL Thesis Series
The following list contains the most recent dissertations in the TRAIL Thesis Series. For a
complete overview of more than 150 titles see the TRAIL website: www.rsTRAIL.nl.
The TRAIL Thesis Series is a series of the Netherlands TRAIL Research School on
transport, infrastructure and logistics.
Zheng, H., Coordination of Waterborn AGVs, T2016/25, December 2016, TRAIL Thesis
Series, the Netherlands
Yuan, K., Capacity Drop on Freeways: Traffic Dynamics, Theory and Modeling, T2016/24,
December 2016, TRAIL Thesis Series, the Netherlands
Li, S., Coordinated Planning of Inland Vessels for Large Seaports, T2016/23, December
2016, TRAIL Thesis Series, the Netherlands
Berg, M. van den, The Influence of Herding on Departure Choice in Case of Evacuation:
Design and Analysis of a Serious Gaming Experimental Set-up, T2016/22, December 2016,
TRAIL Thesis Series, the Netherlands
Luo, R., Multi-Agent Control of Urban Transportation Networks and of Hybrid Systems
with Limited Information Sharing, T2016/21, November 2016, TRAIL Thesis Series, the
Netherlands
Campanella, M., Microscopic Modelling of Walking Behavior, T2016/20, November 2016,
Horst, M. van der, Coordination in Hinterland Chains: An Institutional Analysis of Port-
related Transport, T2016/19, November 2016, TRAIL Thesis Series, the Netherlands
Beukenkamp, Securing Safety: Resilience Time as a Hidden Critical Factor, T2016/18,
October 2016, TRAIL Thesis Series, the Netherlands
Mingardo, G., Articles on Parking Policy, T2016/17, October 2016, TRAIL Thesis Series,
the Netherlands
Duives, D.C., Analysis and Modelling of Pedestrian Movement Dynamics at Large-scale
Events, T2016/16, October 2016, TRAIL Thesis Series, the Netherlands
Wan Ahmad, W.N.K., Contextual Factors of Sustainable Supply Chain Management Prac-
tices in the Oil and Gas Industry, T2016/15, September 2016, TRAIL Thesis Series, the
Netherlands
Liu, X., Prediction of Belt Conveyor Idler Performance, T2016/14, September 2016, TRAIL
157
158 TRAIL Thesis Series
Thesis Series, the Netherlands

Gaast, J.P. van der, Stochastic Models for Order Picking Systems, T2016/13, September
2016, TRAIL Thesis Series, the Netherlands
Wagenaar, J.C., Practice Oriented Algorithmic Disruption Management in Passenger Rail-
ways, T2016/12, September 2016, TRAIL Thesis Series, the Netherlands
Psarra, I., A Bounded Rationality Model of Short and Long-Term Dynamics of Activity-
Travel Behavior, T2016/11, June 2016, TRAIL Thesis Series, the Netherlands
Ma, Y., The Use of Advanced Transportation Monitoring Data for Official Statistics, T2016/10,
June 2016, TRAIL Thesis Series, the Netherlands
Li, L., Coordinated Model Predictive Control of Synchromodal Freight Transport Systems,
T2016/9, June 2016, TRAIL Thesis Series, the Netherlands
Vonk Noordegraaf, D.M., Road Pricing Policy Implementation, T2016/8, June 2016, TRAIL
Thesis Series, the Netherlands
Liu, S., Modeling, Robust and Distributed Model Predictive Control for Freeway Networks,
T2016/7, May 2016, TRAIL Thesis Series, the Netherlands
Calvert, S.C., Stochastic Macroscopic Analysis and Modelling for Traffic Management,
T2016/6, May 2016, TRAIL Thesis Series, the Netherlands
Sparing, D., Reliable Timetable Design for Railways and Connecting Public Transport Ser-
vices, T2016/5, May 2016, TRAIL Thesis Series, the Netherlands
Rasouli, S, Uncertainty in Modeling Activity-Travel Demand in Complex Urban Systems,
T2016/4, March 2016, TRAIL Thesis Series, the Netherlands
Vries, J. de, Behavioral Operations in Logistics, T2016/3, February 2016, TRAIL Thesis
Series, the Netherlands
Goñi-Ros, B., Traffic Flow at Sags: Theory, Modeling and Control, T2016/2, March 2016,
Khademi, E., Effects of Pricing Strategies on Dynamic Repertoires of Activity-Travel Be-
haviour, T2016/1, February 2016, TRAIL Thesis Series, the Netherlands
Cong, Z., Efficient Optimization Methods for Freeway Management and Control, T2015/17,
November 2015, TRAIL Thesis Series, the Netherlands
Kersbergen, B., Modeling and Control of Switching Max-Plus-Linear Systems: Reschedul-
ing of railway traffic and changing gaits in legged locomotion, T2015/16, October 2015,
Brands, T., Multi-Objective Optimisation of Multimodal Passenger Transportation Net-
works, T2015/15, October 2015, TRAIL Thesis Series, the Netherlands
Ardi, Ö., Road Pricing Policy Process: The interplay between policy actors, the media and
public, T2015/14, September 2015, TRAIL Thesis Series, the Netherlands
Samenvatting
Coördinatie van Water AGVs

De verwachte grotere containeroverslag en de beperkte behandelingscapaciteit van bestaande
haveninfrastructuren leggen een steeds grotere druk op havens om het concurrentievermo-
gen te verbeteren. Binnen containerterminals worden aan de landzijde al decennia lang op
grote schaal automatisch geleide voertuigen gebruikt om operationele efficiëntie en duur-
zaamheid te verbeteren. Vervoer tussen de terminals, zogenaamd Inter Terminal Transport
(ITT), wordt op dit moment vooral gerealiseerd met behulp van bemande vrachtwagens. Het
wegverkeer in havengebieden met beperkte wegcapaciteit is echter al zwaar. Voor geogra-
fisch complexe havens, zoals de haven van Rotterdam, kunnen reisafstanden tussen termi-
nals over land veel langer zijn dan de afstanden tussen die terminals over water. Uitbreiding
van de bestaande fysieke infrastructuur voor transport brengt extreem hoge kosten met zich
mee. Als alternatief moeten meer efficiënte en duurzame manieren om havenlogistiek te
verbeteren worden onderzocht.
Dit promotieonderzoek stelt een nieuw vervoerstype voor containers voor: Water AGVs
(“Waterborne Autonomous Guided Vessels”), met name voor de het uitvoeren van ITT. De
nadruk ligt in dit onderzoek op geavanceerde besturings- en planningsstrategieën voor het
coördineren van Water AGVs. Besturingstechnieken gebaseerd op modelgebaseerd voor-
spellend regelen (“model predictive control” (MPC)) van gedetailleerde Water AGV dyna-
miek worden onderzocht en coördinatietechnieken voor het oplossen van planningsproble-
men tussen meerdere Water AGVs voor het uitvoeren van ITT worden voorgesteld.
De waarde van Water AGVs wordt geanalyseerd voor vier belangrijke ITT scenario’s
met toenemende mate van complexiteit:
• Het individuele Water AGV scenario.

In dit scenario wordt een enkele Water AGV een ITT taak toegewezen, bestaande
uit het verplaatsen van containers op bepaalde tijden tussen gespecificeerde herkomst
en bestemmingsterminals. De Water AGV streeft er hierbij naar om een referentie-
pad nauwkeurig te volgen en op het juiste moment, op een energie-efficiënte manier
en zonder aanvaringen aan te komen op de bestemming. Om deze doelen te berei-
ken wordt een besturingstechniek voor een Water AGV voorgesteld gebaseerd op het
idee van voorspellend padvolgen met aankomsttijd bewustzijn (“predictive path follo-
wing with arrival time awareness” (PPF-ATA)). De PPF-ATA besturing is gebaseerd
op een zogenaamde aaneengesloten-pad assenstelsel, een referentieschakellogica, en
een twee-level dubbele-integrator schema voor referentiepadparameterisatie. Simula-
159
160 Samenvatting
tieresultaten van relevante ITT case studies in de haven van Rotterdam laten zien dat
Water AGVs de toegewezen ITT taken succesvol uit kunnen voeren met de voorge-
stelde aanpak.
• Meerdere coöperatieve Water AGVs.

In het scenario met meerdere coöperatieve Water AGVs wordt elke Water AGV een
ITT taak toegewezen en moeten Water AGVs een veilige afstand van elkaar hou-
den. Bij de uitvoering van ITT taken lossen de Water AGVs lokale problemen in
parallel op, terwijl zij een overkoepelende doelstelling nastreven en antibotsingsbe-
perkingen respecteren. De betrokken Water AGVs worden hierbij gerepresenteerd
met tijdsvariërende grafen. Een parallelle coöperatieve controller (“distributed PPF-
ATA”) wordt voorgesteld om acties te bepalen op basis van gedistribueerde MPC en
de wisselende richting van multipliers methode (ADMM). Een variant, snelle ADMM
(“fast ADMM”) wordt voorgesteld om mogelijk trage convergentie van conventionele
ADMM te verbeteren. Simulatieresultaten tonen aan dat alle Water AGVs succesvol
en veilig hun toegewezen ITT taken uitvoeren met de voorgestelde besturing.
• Meerdere Water AGVs in een omgeving met onzekerheid.

Wanneer er onzekerheden bestaan in omgevingen met verstoringen moeten prestaties
en veiligheid van Water AGVs worden gerealiseerd op een robuuste gedistribueerde
manier. Voor onzekerheden met stochastische eigenschappen en met oneindige basis,
is het onmogelijk en onnodig om tegen alle mogelijke onzekerheidsrealisaties be-
stand te zijn. Het concept van kosteneffectieve robuustheid wordt voorgesteld en de
trade-off tussen de waarschijnlijkheid van robuustbare onzekerheden en prestaties van
het systeem worden geoptimaliseerd met een robuust gedistribueerd MPC (RDMPC)
raamwerk.
Naast het voorkomen van botsingen voor meerdere Water AGVs zoals in het deter-
ministische geval moet overeenstemming worden bereikt over het niveau van over-
koepelende systeemrobuustheid wanneer er onzekerheden bestaan. Een efficiënte
geı̈ntegreerde “branch and bound” (B&B) methode en het ADMM algoritme worden
voorgesteld om het RDMPC probleem op te lossen. Het algoritme maakt gebruik van
speciale gesorteerde waarschijnlijkheidssets die slimme zoekopdrachten in de B&B
doen en integreert vertakkingscriteria met tussentijdse ADMM resultaten voor vroeg-
tijdige beëindiging. Simulatieresultaten tonen aan dat, ondanks de onzekerheden, alle
Water AGVs nog steeds de toegewezen ITT taken succesvol en veilig uitvoeren. De
trade-off tussen het robuustheidsniveau en de prijs van robuustheid worden systema-
tisch geoptimaliseerd voor meerdere Water AGVs met onzekerheden.
• Gesloten-lus planning en besturing van Water AGVs voor een autonomoon ITT sys-
teem.
Gegeven een set van ITT aanvragen en een vloot van Water AGVs is een gesloten-lus
plannings- en besturingsaanpak voorgesteld om beslissingen voor het ITT-systeem op
een autonome manier te maken. Gesloten-lus betekent hierbij dat zowel planning en
besturingsproblemen continu worden opgelost. Op een hiërarchische manier omvat
het planningsprobleem het bepalen van de volgorde en de tijden waarop terminals
worden bezocht voor het laden/lossen van voorgeschreven hoeveelheden containers
Samenvatting 161
door elke Water AGV. Het besturingsprobleem omvat het op een coöperatieve gedis-
tribueerde manier bepalen van de optimale acties die genomen moeten worden om
de gedetailleerde Water AGV dynamiek de toegewezen plannen uit te laten voeren.
Daarnaast beschouwt het planningsprobleem gecoördineerd afmeren rekening hou-
dend met de benodigde minimale tijdsintervallen tussen afmeren van verschillende
Water AGVs op dezelfde ligplaats. De gesloten-lus planning en besturingsbenadering
maakt het mogelijk dat lagere niveau factoren zoals onvoorziene fysieke systeembe-
perkingen, verstoringen, en de moeilijke botsingsvermijding kunnen worden gente-
greerd in tijdig vernieuwde hogere niveau plannen. Simulatieresultaten tonen aan dat
energie-efficiënte plannen worden gegenereerd die voldoen aan de plannings- en be-
sturingsbeperkingen van de Water AGVs. Gecoördineerde tijden voor het afmeren
worden ook bereikt voor alle Water AGV ligplaatsen. Daarnaast is de zogenaamde
“non-performance” lager met behulp van de voorgestelde gesloten-lus aanpak dan
een open-lus aanpak. Het voorgestelde autonome ITT systeem kan met behulp van
Water AGVs een effectieve bijdrage aan slimmere havenlogistiek leveren.
In het kort, dit proefschrift stelt Water AGVs voor als een innovatieve en effectieve ma-
nier voor het uitvoeren van ITT. Hoewel de voorgestelde besturing- en planningsmethodes
zijn ontworpen voor Water AGVs en zijn geoptimaliseerd voor ITT, kan hun toepassing
worden gegeneraliseerd naar andere soortgelijke coördinatiescenarios.
Summary
Coordination of waterborne AGVs

The possible larger amount of container throughput and the limited handling capacities of
existing infrastructures impose increasingly high pressure on large ports in improving com-
petitiveness. Inside container terminals, land-side automated guided vehicles have been
used extensively for decades to improve terminal operational efficiency and sustainability.
Transport between terminals, i.e., Inter Terminal Transport (ITT), is currently mainly re-
alized by manned trucks. However, road traffic has already been heavy in port areas with
limited land. For geographically complex ports, e.g., the port of Rotterdam, travel distances
by land can be much longer than by water between some terminals. Expanding the existing
physical transportation infrastructure might be an option, at extremely high costs nonethe-
less. As an alternative, more efficient and sustainable ways for port logistics need to be
investigated.
This PhD research proposes a new type of container transporter, waterborne Autonomous
Guided Vessels (waterborne AGVs), for ITT. The focus is on advanced control and schedul-
ing strategies for coordinating waterborne AGVs. Particularly, control problems based
on model predictive control (MPC) considering detailed waterborne AGV dynamics and
scheduling problems of multiple waterborne AGVs given ITT requests are investigated com-
prehensively. Four representative ITT scenarios with increasing levels of complexity using
waterborne AGVs are studied:
• A single waterborne AGV scenario.

In this scenario, a waterborne AGV is assigned an ITT task to move containers be-
tween specified origin and destination terminals at specified times. Performance cri-
teria are that the waterborne AGV needs to track a reference path smoothly and arrive
at the destination punctually in an energy-efficient way. To achieve these goals, a
predictive path following with arrival time awareness (PPF-ATA) controller for one
waterborne AGV is proposed. The PPF-ATA controller consists of a connected path
coordinate system in which the inertial waterborne AGV kinematics are remodeled,
a reference switching logic to avoid overshoots, and a two-level double-integrator
scheme for parameterizing reference paths. Moreover, obstacle avoidance can also
be addressed by this controller. Simulation results of relevant ITT case studies in the
port of Rotterdam show that waterborne AGVs accomplish assigned ITT tasks suc-
cessfully according to the defined performance criteria with the proposed approach.
163
164 Summary
• Multiple cooperative waterborne AGVs.

For multiple cooperative waterborne AGVs, each waterborne AGV is assigned an
ITT task and waterborne AGVs should keep a safe distance away from others. While
carrying out ITT tasks, waterborne AGVs solve local problems in parallel while min-
imizing an overall objective and satisfying coupling collision avoidance constraints.
The involved waterborne AGVs are modeled as time-varying graphs and a parallel
cooperative controller, distributed PPF-ATA, is proposed based on distributed MPC
and the alternating direction method of multipliers (ADMM). Collision avoidance
constraints are satisfied in the framework of ADMM via iterative decomposition and
coordination. Furthermore, a variant, fast ADMM, is proposed to improve the pos-
sible poor convergence rates of the conventional ADMM. Both the ADMM and fast
ADMM based controllers achieve cooperative distributed computations through it-
erations. Simulation results show that all waterborne AGVs successfully and safely
accomplish their assigned ITT tasks with local decision making.
• Multiple waterborne AGVs facing uncertain environmental disturbances.

When there exist uncertainties in environmental disturbances, performance and safety
of waterborne AGVs still need to be achieved in a robust distributed way. For uncer-
tainties with stochastic characteristics and with infinite support, being robust to all
possible realizations is impossible and not necessary. The concept of cost-effective
robustness is proposed and the trade-off between the probability of robustable un-
certainties and system performance is optimized in a cost-effective robust distributed
MPC (RDMPC) framework. Besides the agreement on collision avoidance couplings
for multiple waterborne AGVs as in the deterministic case, further agreement on an
overall system robustness level has to be reached when uncertainties exist. An effi-
cient integrated branch and bound (B&B) and ADMM algorithm is proposed to solve
the cost-effective RDMPC problem. The algorithm exploits the special ordered prob-
ability sets conducting smart search in B&B and integrates branching criteria with
intermediate ADMM results for early termination. Simulation results show that, in
spite of uncertainties, all waterborne AGVs still fulfill assigned ITT tasks successfully
and safely. Cooperative distributed decision making is also iterativelly achieved fol-
lowing ADMM. The trade-off between robustness level and the price of being robust
is optimized systematically for multiple waterborne AGVs with uncertainties.
• Closed-loop scheduling and control of waterborne AGVs for an autonomous ITT sys-
tem.
Given a set of ITT requests and a fleet of waterborne AGVs, a closed-loop schedul-
ing and control approach is proposed to make decisions for the ITT system in an
autonomous way. Closed-loop means that both scheduling and control problems
are solved in real-time. In a hierarchical way, the scheduling problem decides, for
each waterborne AGV, the sequence of terminals to visit at specified times with load-
ing/unloading specified amount of containers. The control problem decides, in a co-
operative distributed way, the optimal forces and moment considering detailed wa-
terborne AGV dynamics in order to carry out the assigned schedules. Moreover,
the scheduling problem considers coordinated berthing with necessary time intervals
between different waterborne AGVs at the same berth. The real-time closed-loop
Summary 165
scheduling and control approach enables that lower level factors such as unconsid-
ered physical system limits, disturbances, and collision avoidance that are difficult, if
not impossible, to be integrated in a scheduling problem can be reflected timely in the
online updated schedules. Simulation results show that energy-efficient schedules are
generated and fulfilled successfully by waterborne AGVs satisfying both scheduling
and control constraints. Coordinated berthing time slots are also achieved for all wa-
terborne AGV berths. The so-called “non-performance” rate is lower using the pro-
posed closed-loop approach than an open-loop approach. The proposed autonomous
ITT system using waterborne AGVs is demonstrated to be effective contributing to
smarter port logistics.
Briefly, this PhD thesis proposes an innovative and effective way, waterborne AGVs,
for conducting ITT. Although the proposed control and scheduling methodologies are de-
signed for waterborne AGVs with the specifications of ITT, their applications can also be
generalized to other similar coordination scenarios.
Curriculum vitae
Huarong Zheng was born in December, 1988 in Jingshan, Hubei, China. She obtained the
B.Sc degree on Logistics Engineering in 2011 and the M.Sc degree on Intelligent Trans-
portation Engineering in 2013, both from Wuhan University of Technology in Wuhan,
China.
Starting from 2013, Huarong is sponsored by China Scholarship Council as a PhD
candidate at the Department of Maritime and Transport Technology, Delft University of
Technology, Delft, the Netherlands. In her PhD project, Huarong proposes advanced co-
ordination strategies for a new type of container transport vehicles, waterborne AGVs, for
Inter Terminal Transport. Her research interests include model predictive control, intelligent
transportation systems, and autonomous vehicles.
List of my publications during PhD

1. H. Zheng, R.R. Negenborn, G. Lodewijks. Closed-loop scheduling and control of
waterborne AGVs for energy-efficient Inter Terminal Transport. Transportation Re-
search Part E: Logistics and Transportation Review, In press, 2016, http://dx.doi.org/
10.1016/j.tre.2016.07.010.
2. H. Zheng, R.R. Negenborn, G. Lodewijks. Fast ADMM for Distributed Model Pre-
dictive Control of Cooperative Waterborne AGVs. IEEE Transactions on Control Sys-
tems Technology, 2016. In press, doi: http://dx.doi.org/10.1109/TCST.2016.2599485.
3. H. Zheng, R.R. Negenborn, G. Lodewijks. Predictive path following with arrival time
awareness for waterborne AGVs. Transportation Research Part C: Emerging Tech-
nologies, vol. 70, pp. 214 – 237, 2016. http://dx.doi.org/10.1016/j.trc.2015.11.004.
4. H. Zheng, R.R. Negenborn, G. Lodewijks. Explicit use of probabilistic distributions

in robust predictive control of waterborne AGVs - a cost-effective approach. In pro-
ceedings of the 15th European Control Conference (ECC16), Aalborg, Denmark, pp.
1278 - 1283, June - July 2016.
5. H. Zheng, R.R. Negenborn, G. Lodewijks. Cooperative distributed collision avoid-

ance based on ADMM for waterborne AGVs. In Proceedings of the 6th International
Conference on Computational Logistics (ICCL’15), Delft, The Netherlands, pp. 181
- 194, September 2015.
167
168 Curriculum vitae
6. C. Liu, H. Zheng, R.R. Negenborn, X. Chu, L. Wang. Trajectory tracking control for
underactuated surface vessels based on nonlinear Model Predictive Control. In Pro-
ceedings of the 6th International Conference on Computational Logistics (ICCL’15),
Delft, The Netherlands, pp. 160 - 180, September 2015.
7. S. Li, H. Zheng, R.R. Negenborn, G. Lodewijks. Coordination for efficient transport

over water. In Proceedings of the 19th IEEE International Conference on Computer
Supported Cooperative Work in Design (IEEE CSCWD’15), Calabria, Italy, pp. 389
- 394, May 2015.
8. H. Zheng, R.R. Negenborn, G. Lodewijks. Model predictive control of a waterborne

AGV at the operational level. In Proceedings of the International Maritime and Port
Technology and Development Conference (MTEC 2014), Trondheim, Norway, pp. 99
- 108, October 2014.
9. H. Zheng, R.R. Negenborn, G. Lodewijks. Trajectory tracking of autonomous vessels

using model predictive control. In Proceedings of the 19th IFAC World Congress
(IFAC WC’14), Cape Town, South Africa, pp. 8812 - 8818, August 2014.
10. C. Liu, R.R. Negenborn, F. Ma, X. Chu, H. Zheng. Towards an Experimental Platform
for Inland Waterway Ship Navigation. In Proceedings of the 2014 IEEE International
Conference on Networking, Sensing, and Control (ICNSC 2014), Miami, Florida, pp.
707 - 712, April 2014.
11. H. Zheng, R.R. Negenborn, G. Lodewijks. Survey of approaches for improving the
intelligence of marine surface vehicles. In Proceedings of the 16th International IEEE
Conference on Intelligent Transportation Systems (IEEE ITSC 2013), The Hague, The
Netherlands, pp. 1217 - 1223, October 2013.
12. H. Zheng, R.R. Negenborn, G. Lodewijks. Robust distributed predictive control of

waterborne AGVs - a cooperative and cost-effective approach. Submitted to a journal,
2016.

Thesis Huarong Zheng

Uploaded by

Copyright:

Available Formats

Thesis Huarong Zheng

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Huarong Zheng

Uploaded by

Copyright:

Available Formats

Delft University of Technology

Coordination of waterborne AGVs

This work is downloaded from Delft University of Technology.

ter verkrijging van de graad van doctor

Master of Science in Intelligent Transportation Engineering,

Composition of the doctoral committee:

Rector Magnificus chairman

TRAIL Thesis Series T2016/25, the Netherlands TRAIL Research School

Published and distributed by: Huarong Zheng

Keywords: Waterborne AGVs, model predictive control, cooperative distributed control,

Copyright c 2016 by Huarong Zheng

Printed in the Netherlands

2 Literature review and background 9

3 Dynamic models of waterborne AGVs 27

4 Predictive path following with arrival time awareness 39

5 Cooperative distributed waterborne AGVs 65

6 Cost-effective robust distributed control of waterborne AGVs 81

7.5.1 From the waterborne AGV perspective . . . . . . . . . . . . . . . 122

8 Conclusions and future research 131

TRAIL Thesis Series publications 157

Curriculum vitae 167

1.1 Waterborne AGVs for Inter Terminal Transport

• Waterborne AGVs, comparable to land-side AGVs, could be optimally operated 24/7

To develop a transport system using waterborne AGVs involves multi-discipline tech-

1.2 Scope and assumptions

Physical system level & environment

deflections/revolutions or spatial no-sailing zones, etc., need to be satisfied; d) complex

1.3 Research questions

1.4 Thesis outline

• Chapter 6 considers the influences of environmental disturbances and proposes a

5. Cooperative 6. Waterborne AGVs

Figure 1.3: Road map. Arrows indicate read before relations.

Literature review and

2.1 Relevant intelligent vehicles

2.1.1 Automated guided vehicles

Figure 2.1: The ECT Delta Terminal Rotterdam (Courtesy of ECT).

Figure 2.2: Terex Gottwald AGVs [128].

2.1.2 Intelligent marine surface vehicles

Environment map Control

Software Navigation System Hardware

Table 2.1: Prototypes of intelligent surface vehicles.

Organization Vehicle Purpose Year

Figure 2.4: Cybership II at Norwegian University of Science and Technology [121].

Figure 2.5: Delfia-1 at Delft University of Technology [88].

2.2 Motion control literature

within a specified time window otherwise;

• The aforementioned two design requirements could be achieved in an energy eco-

• A certain level of robustness could be achieved when uncertainties are involved.

2.2.1 General control techniques

• Trajectory tracking: Explicit time parameterized references (e.g., positions, veloci-

2.2.2 Model predictive control

Control technique Advantage Disadvantage

Algorithm 2.1 Basic MPC algorithm

Figure 2.6: Illustration of MPC.

Systems with uncertainties

2.3 Scheduling approaches

2.4 Inter terminal transport

waterborne AGVs with environmental disturbances. Different scheduling approaches have

Dynamic models of waterborne