Professional Documents
Culture Documents
Signature Redacted: Tiignature
Signature Redacted: Tiignature
Marine Vehicles
by
Joseph William Leavitt
B.S. Electrical Engineering, United States Naval Academy (2008)
Submitted to the Department of Mechanical Engineering and the
Department of Electrical Engineering and Computer Science
in partial fulfillment of the requirements for the degrees of
Naval Engineer
and
Master of Science in Electrical Engineering and Computer Science
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2017
Massachusetts Institute of Technology 2017. All rights reserved.
Certified by...
Signature
Si n tMay redacted-
(/ Michael R. Benjamin
Research Scientist, Department of Mechanical Engineering
Thesis Supervisor
A~
LN 21R
LIBRARIES
2
Intent-Aware Collision Avoidance for Autonomous Marine
Vehicles
by
Joseph William Leavitt
Abstract
Applications of autonomous marine vehicles in dynamic and uncertain environments
continuously grow as research unveils new enabling technology and academic, com-
mercial, and government entities pursue new marine autonomy concepts. The safe
operation of these vehicles in the marine domain, which is currently dominated
by human-operated vehicles, demands compliance with collision avoidance protocol,
namely the International Regulations for Preventing Collisions at Sea (COLREGS).
Strict application of this protocol can lead to a highly constrained motion planning
problem, in which it is difficult for a vehicle to identify a safe and efficient motion
plan.
This thesis proposes a multi-objective optimization-based method for COLREGS-
compliant autonomous surface vehicle collision avoidance in which vehicles use shared
intent information, in addition to vehicle state information, to identify safe and ef-
ficient collision avoidance maneuvers. The proposed method uses intent information
to relax certain COLREGS-specified constraints with the goal of providing sufficient
maneuvering flexibility to enable improvements in safety and efficiency over a non-
intent-aware system. In order to arrive at an intent-aware solution, this thesis ex-
plores the concept of intent, including intent formulations for the marine domain,
intent communications, and the application of intent to the COLREGS-compliant
motion planning problem. Two types of intent information are specifically evaluated:
COLREGS mode intent, in which the give-way vessel in an overtaking or crossing
scenario communicates its intent to maneuver in a certain direction with respect
the stand-on vessel, and discrete trajectory intent in which vehicles communicate
projected future positions. Simulations and on-water experiments demonstrate the
feasibility of the proposed intent-aware method, as well as improvements in perfor-
mance, in terms of both vehicle safety and mission efficiency, over a non-intent-aware,
COLREGS-compliant collision avoidance method.
3
Thesis Supervisor: Michael R. Benjamin
Title: Research Scientist, Department of Mechanical Engineering
4
Acknowledgments
The most important words in this document are those I leave here. I am forever
grateful to all of the people in my life, my friends, my family, my peers, and my
mentors, who have supported me, guided me, and rode along with me through this
To Mike Benjamin, your guidance, support, and encouragement are the reason I
was able to do all that I have done this past year. I am incredibly lucky to have had
the opportunity to work with you and I hope that we can continue our pursuit of fun
To John Leonard, thank you for believing in me and giving me the opportunity
to pursue this work in such a broad, interesting, and fast moving area of research. It
has been an honor to work with you; you exemplify the spirit of MIT and all that we
To Misha Novitzky, from our days together in Duckietown to our early mornings
wrangling robots on the water, you have made it possible for me to accomplish every-
thing that I've accomplished. It's been a real adventure and I can never thank you
To Kyle Woerner, Paul Robinette, Hugh Dougherty, and everyone else in the lab
who has supported me somewhere along the way, through your expertise, guidance,
robot wizardry, and nautical skills, you were absolutely instrumental to my success.
To my classmates, not only have you been there to prop me up along the way,
but your friendship is something that I will value for the rest of my days. I owe a
very special thank you to Tom Finley, whose hard work, dedication, leadership, and
friendship enabled me to get absolutely everything I could out of my time at MIT.
To my family and friends who have supported me and been patient with me as I
have pursued this education, you are a constant source of strength in my life and I
look forward to nothing more than making up for lost time together.
To my parents, who have given me the world and the belief that, with a little hard
5
work, I can be anything in it, your endless love and support through all the years
have made it possible for me to go after my dreams. You inspire everyone around you
through your example as parents and as simply wonderful people. The whole world
Finally, and without words to express my gratitude, to my wife Jessie, who has
been there through every early morning and every late night to lift me up and keep
me going through it all. It would be enough in a lifetime to have given me what you
have over these last three years, but it comes at the tail end of a decade in which you
have carried us through far more than that. You are the key to everything in my life
and there's no telling where I would be without you. While I will not miss time spent
working instead of time spent together, I will definitely miss all of our spontaneous
adventures in this great city, especially hanging out together at our baseball home. I
am so proud of all that you have accomplished and I look forward now to the next
6
Contents
1 Introduction 23
1.1 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
mization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4 O verview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2 Literature Review 39
7
2.3.1 Background of Intent in Motion Planning . . . . . . . 52
.
2.3.2 Intent Formulations . . . . . . . . . . . . . . . . . . . 55
.
2.3.3 Intent Inference . . . . . . . . . . . . . . . . . . . . 62
.
2.3.4 Intent Communication . . . . . . . . . . . . . . . . 62
.
2.4 Intent-Aware Motion Planning . . . . . . . . . . . . . . . . 66
.
2.4.1 Intent-Aware Motion Planning Frameworks . . . . . 66
.
2.4.2 Intent-Based, COLREGS-Compliant Motion Planning 68
2.5 Summary of Literature Review . . . . . . . . . . . . . . . 69
.
3 Intent Formulation 71
3.1 Intent Formalization . . . . . . . . . . . . . . . . . . . . . 71
.
3.2 From Intent to Vehicle Motion . . . . . . . . . . . . . . . . 73
.
3.3 From Contact Intent to Motion Planning . . . . . . . . . . 74
.
3.4 Vehicle Trajectory . . . . . . . . . . . . . . . . . . . . . . . 75
.
3.5 Vehicle Intent in the Marine Domain . . . . . . . . . . . . 75
.
3.5.1 Marine Domain Intent Abstractions . . . . . . . . . 76
.
3.5.2 Selected Marine Domain Intent Formulations . . . . 80
.
4 COLREGS-Compliant Motion Planning 85
4.1 COLREGS Collision Avoidance Scenarios . . . . . 85
.
4.1.1 Overtaking . . . . . . . . . . . . . . . . . 89
.
4.1.2 Head-On . . . . . . . . . . . . . . . . . . . 91
.
4.1.3 Crossing . . . . . . . . . . . . . . . . . . . 92
.
8
4.5 Summary of COLREGS-Compliant Motion Planning ... . . . . . . 113
.
5.4 Intent-Aware COLREGS Constraints . . . . . . . . . . . . . . . . . 128
.
5.4.1 Constraints for COLREGS Mode Intent . . . . . . . . . . . 128
.
5.4.2 Constraints for Trajectory Intent . . . . . . . . . . . . . . . 131
.
5.5 Intent-Aware COLREGS Behavior . . . . . . . . . . . . . . . . . . . 133
.
5.5.1 Processing Intent Data . . . . . . . . . . . . . . . . . . . . . 134
.
5.5.2 Intent-Aware Mode Checking Algorithm . . . . . . . . . . . 138
.
5.5.3 Intent-Aware Objective Functions . . . . . . . . . . . . . . . 141
.
5.6 Intent Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
.
9
6.3 Canonical COLREGS Scenario Assessment . . . . . . 163
6.3.1 Overtaking Scenario . . . . . . . . . . . . . 166
.
6.3.2 Crossing Scenario . . . . . . . . . . . . . . . 181
.
6.3.3 Head-On Scenario . . . . . . . . . . . . . . . 199
.
6.4 Multi-Vehicle Evaluation . . . . . . . . . . . . . . . 215
.
6.4.1 Traffic Pattern Simulations . . . . . . . . . . 215
.
6.4.2 On-Water Validation . . . . . . . . . . . . . 221
.
6.5 General Results . . . . . . . . . . . . . . . . . . . . 229
.
6.6 Evaluation Summary . . . . . . . . . . . . . . . . . 233
.
7 Conclusions 235
7.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
.
7.2 Recommendations for Future Work . . . . . . . . . . . . . . . . . 237
.
A Intent-Aware Software System 241
A.1 Payload Autonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
10
B.3 Head-on Scenario Simulations . . . . . . . . . . . . . . . . . . . . . . 259
B.4 Traffic Pattern Simulations . . . . . . . . . . . . . . . . . . . . . . . . 262
D Notation 267
11
12
List of Figures
(180, 360) are disallowed. Blue represents actions of low utility and
13
4-6 Collision avoidance objective function. . . . . . . . . . . . . . . . . . 102
tive function that accounts for intent in addition to vehicle states. . . 133
m ode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5-10 A notional speed objective function for the CommunicatedBow submode. 147
14
6-4 Overtaking scenario - range at CPA (rcpa) distributions by intent type. 167
6-6 Overtaking scenario - time efficiency (?It) distributions by intent type. 168
6-10 Overtaking scenario - sequence leading to collision using mode intent. 171
6-13 Overtaking scenario - improved efficiency due to shared mode intent. . 176
ulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6-19 Overtaking scenario - fpa violation rate by intent type with updated mode
15
6-21 Overtaking scenario - range at CPA (rpa) distributions by intent type
showing comparison with kinematic trajectory intent. . . . . . . . . . 183
6-23 Overtaking scenario - cpa violation rate by intent type with kinematic
6-24 Randomized crossing scenario in which the stand-on vessel has an in-
6-25 Crossing scenario - range at CPA (rca) distributions by intent type. . 186
6-26 Crossing scenario - distance efficiency (rd) distributions by intent type. 187
6-27 Crossing scenario - time efficiency (rqt) distributions by intent type. . 187
6-31 Crossing scenario - unsafe encounter with none or mode intent. ..... 191
6-32 Crossing scenario - unsafe encounter with none or mode intent. ..... 192
tent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6-35 Crossing scenario - inefficient encounter without use of intent information. 196
16
6-39 Crossing scenario - fipa violation rate by intent type with kinematic
6-41 Head-on scenario - range at CPA (rpa) distributions by intent type. . 200
6-42 Head-on scenario - distance efficiency (rd) distributions by intent type. 201
6-43 Head-on scenario - time efficiency (rt) distributions by intent type. . . 201
intent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
6-55 Head-on scenario - fcp, violation rate by intent with kinematic trajec-
6-57 Traffic pattern - range at CPA (rca) distributions by intent type. . . 217
6-58 Traffic pattern - trajectory efficiency (r) distributions by intent type. 218
17
6-59 Traffic pattern - ANOM comparison of means between none and mode
intent for 2-contact encounters. . . . . . . . . . . . . . . . . . . . . . 218
6-60 Traffic pattern - ANOM comparison of means between none and non-
kinematic trajectory intent for 2-contact encounters. . . . . . . . . . . 219
6-61 Traffic pattern - fpa violation rate by intent type and contact count. 220
6-62 Traffic pattern - three-vehicle encounter without intent use resulting
in a collision. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
6-63 Traffic pattern - three-vehicle encounter with mode intent resulting in
a collision. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
6-64 Traffic pattern - three-vehicle encounter with non-kinematic trajectory
intent resulting in a collision. . . . . . . . . . . . . . . . . . . . . . . 224
6-65 Traffic pattern - three-vehicle encounter with improved efficiency due
to non-kinematic trajectory intent. . . . . . . . . . . . . . . . . . . . 225
6-66 Traffic pattern - three-vehicle encounter with improved efficiency due
to non-kinematic trajectory intent use. . . . . . . . . . . . . . . . . . 226
6-67 Traffic pattern - ANOM comparison of means between none and updated mode
intent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
6-68 Traffic pattern - fpa violation rate by intent type with updated mode
intent objective functions. . . . . . . . . . . . . . . . . . . . . . . . . 227
6-69 Traffic pattern - ANOM comparison of means between none and kine-
matic trajectory intent. . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6-70 Traffic pattern - fpa violation rate by intent type with kinematic tra-
jectory intent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6-71 Traffic pattern - on-water efficiency and rcpa in comparison to simulated
encounters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
18
B-2 Overtaking scenario - range at CPA (rcpa) mean and standard deviation
by experim ent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
B-3 Overtaking scenario - distance efficiency (1d) mean and standard devi-
ation by experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
B-4 Overtaking scenario - time efficiency (qt) mean and standard deviation
B-7 Crossing scenario - range at CPA (rqa) mean and standard deviation
B-8 Crossing scenario - distance efficiency (ad) mean and standard devia-
B-9 Crossing scenario - time efficiency (qt) mean and standard deviation
B-12 Head-on scenario - range at CPA (rpa) mean and standard deviation
B-13 Head-on scenario - distance efficiency (rd) mean and standard deviation
B-17 Traffic pattern - range at CPA (rca) mean and standard deviation by
B-18 Traffic pattern - trajectory efficiency (Yq) mean and standard deviation
19
20
List of Tables
21
22
Chapter 1
Introduction
Autonomous ships are on the horizon, both literally [30] and figuratively [29, 68],
institutional, that autonomous vehicles will adhere to the rules of the nautical road,
which promote the safety of all vehicles. In the marine domain, this means vehi-
cle compliance with the International Regulations for Preventing Collisions at Sea
(COLREGS) [54].
COLREGS is a protocol consisting of rules that are intended to facilitate colli-
sion avoidance' between surface vehicles interacting at sea. When two vehicles are
interacting such that a "risk of collision" 2 exists, COLREGS requires the application
straints give some assurance to both vehicles that the actions taken by each to avoid
a collision will in fact do so, and not result in some more dangerous situation. For
example, when two vehicles are crossing, as shown in Figure 1-1, COLREGS dictates
that vehicle 1, as the vehicle to the left, or port, of the other, is the "give-way" vessel
and should maneuver around vehicle 2, while vehicle 2 is the "stand-on" vessel and
'The term collision avoidance is used to specifically refer to avoidance of other surface vehicles
and excludes avoidance of static obstacles or dynamic obstacles that are not autonomously or human-
controlled.
2
While the literature suggests various means of defining risk of collision, this thesis defines risk
of collision based on the expected range, or distance between two vehicles, at their closest point of
approach (CPA): rcpa.
23
Vehicle 1
t4
CD) Vehicle 1, Giveway Vessel
Vehicle 2, Stand-on Vessel
Giveway Planned Path
- Giveway Passage Ahead
Giveway Passage Astern
Stand-on Passage
Vehicle 2
24
Desired path
Path due to1
*...
maintaining course
and speed through
encounter -------
Give-way S
2
Stand-on
Consider the scenario depicted in Figure 1-2. This is another basic crossing sce-
nario in which vehicle 1 is the give-way vessel and vehicle 2 is the stand-on vessel,
except in this case vehicle 2 has a plan to make an immediate maneuver to the east.
By strictly applying the constraint to maintain course and speed, vehicle 2 drives
past its intended turning point and finally turns once vehicle 1 has traveled further to
the southeast. The resulting maneuver is inefficient with respect to vehicle 2's origi-
nal plan and could be considered unnecessary given information about the intent of
vehicle 1, such as if vehicle 1 intends to continue to the southeast and not maneuver
Furthermore, consider the scenario depicted in Figure 1-3, which adds a third
vehicle to the scenario in Figure 1-2. Now vehicle 2 is in a crossing scenario with
vehicle 3 as the give-way vessel in addition to being the stand-on vessel in the crossing
stand-on constraints with respect to vehicle 1 may now result in an unsafe encounter
with vehicle 3 as it could delay vehicle 2's ability to execute an appropriate give-
way maneuver. One could continue to add vehicles to this scenario until vehicle 2 is
25
3
Stand-on
------ Desired path (ii"O"" t.
S2
Stand-on AND Give-way
Figure 1-3: A crossing scenario resulting in an unsafe and inefficient stand-on vessel
maneuver. The stand-on vessel plans to maneuver to the east and immediate execu-
tion of the planned maneuver and an appropriate collision avoidance maneuver are
prevented by strict adherence to the COLREGS constraint to maintain course and
speed.
ment for vehicle 2 to maintain course and speed with respect to vehicle 1 does not
imply the need to maintain exact course and speed in light of other navigational
constraints, such as the need to maneuver around vehicle 3. This flexibility of in-
for the give-way vessel to account for the intent of the stand-on vessel when consid-
In the above scenario, we call the plan of vehicle 1 to pass ahead or astern of vehicle
2 a representation of vehicle l's intent. The plan of vehicle 2 to maneuver to the east
to meet some objective, likewise, is a representation of vehicle 2's intent. Whether or
not and when vehicle 2 can maneuver to the east, so as to meet this objective, is a
This thesis explores the concept of using vehicle motion intent information to
better inform the process of risk of collision assessment and COLREGS constraint
and execution.
26
1.1 Approach
The goal of this thesis is to demonstrate that the application of certain types of
problem can improve the safety3 and efficiency 4 of vehicle encounters5 . There is no
states and vehicle motion models for trajectory prediction. In order to arrive at an
following approach:
" Development of the concept of intent through exploration of related work (Chap-
27
1.1.1 Intent in the Marine Domain
Humans use shared motion intent information every day to navigate sidewalks, roads,
particular area on the road due to observation of a car with an active turn blinker
or stepping in a certain direction on the sidewalk because another pedestrian used a
hand gesture to tell you she plans to stop and let you go ahead. Sometimes human
reasoning of intent information can be more complex, such as inferring the trajectory
humans use this intent information to better inform the process of predicting future
ambiguity in assessing vehicle intentions, so long as all interacting vehicles are com-
plying with the rules. However, inherent flexibility in rule interpretation and options
for vehicles to take certain actions according to the rules, such as a crossing vessel
passing ahead or astern, beg the question of whether or not knowledge of this in-
tent can be applied to achieve safer and more efficient collision avoidance. Several
projects have already demonstrated the safety and efficiency benefits of exchanging
applied to perform collision avoidance in the marine domain based on information that
motion planning, two specific types of intent information are selected for development
and evaluation:
COLREGS Mode Intent: The COLREGS rule by which a vehicle intends to ex-
ecute collision avoidance and any other information associated with executing
Discrete Trajectory Intent: A discrete set of positions and the times at which the
28
Application of this intent information is considered in the case where vehicles
establish their own intent and then share it via broadcast to other vehicles. In the
case of discrete trajectory intent, vehicles establish a collision avoidance priority and
"negotiate" this priority to establish which vehicle has responsibility to take collision
to maneuver in some manner or do not agree and simply apply the basic COLREGS
constraints.
Given the COLREGS protocol and intent information that might be used to execute
this protocol, autonomous marine vehicles must employ a motion planning and execu-
general autonomy system design considerations for the use of shared-intent informa-
tion in the motion planning process and then proposes specific algorithms for using
solves for the optimal vehicle maneuvering action based on a linear combination of
29
optimization "behavior"6 employs algorithms specific to the objective to develop ob-
jective functions.
For the purposes of evaluation, the behavior discussed in the previous section is
the baseline, non-intent aware system through simulations and on-water experiments
counters.
Using specified measures of safety and efficiency, it was found that application
over the non-intent-aware baseline. These findings come with the recognition that
additional research should be conducted in which the assumptions made in this work
intent inference, intent-based COLREGS constraints, and other related areas such as
human-robot interaction.
30
and efficiency over the non-intent-aware system. These improvements are associated
with the added flexibility provided by the proposed constraint relaxations. Significant
findings include the need to address the case in which the give-way vessel changes
or does not execute the originally communicated intent and the potential benefit of
constraint relaxations.
efficiency in all cases and improvements in safety in some, but not all, cases. The most
significant improvements in both safety and efficiency are achieved through the appli-
cation of trajectory intent generated with the use of a vehicle motion model. Improve-
ments in efficiency are based on the proposed system of collision avoidance priority
based risk of collision exists and are assigned a "give-way" priority in the negotiation
process. Specific shortcomings of this work, which result in some observed degrada-
tions of safety, include the need for a more robust method of assigning and planning
collision avoidance maneuvering priority when more than two vehicles are involved
in an encounter, such that non-conflicting 7 actions are executed by any vehicle with
tion approach only plans the next maneuvering action, it cannot be used to develop
may not represent actual future vehicle motion due to deviations for collision avoid-
the aviation intent literature [12,37,45,47, 80,95, 106,109], should investigate plan-
ning methods that produce intended vehicle trajectories that account for collision
avoidance maneuvers.
7
1n the context of collision avoidance, a conflicting action, or maneuver, is defined as an action
that results in an unsafe encounter with respect to the actions of another vehicle, such as if one
vehicle turns right and the other turns left when meeting head-on.
31
1.2 Formal Problem Statement
[0, 360) are position and heading in the Earth-fixed reference frame and v E R is speed.
and Hansman in 183]. The subscript "os" is used to indicate a variable associated
with ownship8 and the subscript "cn" is used to indicate a variable associated with a
contact9 where an index may also be used to describe a particular contact in a group
of contacts under consideration. Let a = (A, Vd) E [0, 360) x R be a potential ownship
maneuvering action, where 6 d and Vd are desired heading and speed, respectively.
Given ownship state x0,, contact states xn,i, and some ownship or contact intents
I, and Ii, the problem is to establish the utility of potential maneuvering actions
a = (Od, Vd) for all motion objectives such that execution of the globally optimal
improves upon the safety and efficiency of the resulting motion with respect to the
motion planning will result in the improvement of vehicle safety and efficiency in
COLREGS-governed vehicle encounters through intent-based application of COL-
REGS constraints.
avoidance system, this thesis seeks to identify appropriate formulations of the intent
I for use in the marine domain and associated procedures to process and apply intent
8
The
term ownship is used to refer to the vehicle that is being controlled.
9
The term contact is used to refer to a vehicle not under control, about which ownship may need
to reason for the purposes of motion planning.
32
1.3 Scope and Assumptions
As an initial exploration into the concept of applying shared intent information to the
tate development and analysis. The following list defines the scope and assumptions
" COLREGS Rules for Power-Driven Vessels: This thesis examines the applica-
tion of COLREGs rules for collision avoidance between power-driven vessels [54].
Other rules, or considerations thereof, are discussed as they may illuminate the
discussion of intent, however these rules are not specifically addressed or eval-
uated in the proposed intent-aware algorithms. In the event of a COLREGS
a sailing vessel, COLREGS provides priority rules that dictate vehicle respon-
these rules through some method of vehicle classification, the resulting priority
" Homogeneous Autonomy: All vehicle COLREGS encounters are between au-
cles or with human and autonomous vehicles would provide more information
on the general usefulness of intent information and its ability to support the
" Intent Communicated, Not Inferred: The use of inferred intent information is
33
lision avoidance system, however intent inference is not specifically developed
are considered.
" Assumed Vehicle Motion Model: This thesis employs the concept of a vehicle
experimentation.
holonomic. This assumption has not prevented the ability to execute safe
COLREGS-compliant collision avoidance in most vehicle encounters as seen
in [18,110, 111] and this thesis. However, certain cases do arise where this as-
sumption does result in unsafe vehicle encounters. A key area for future work is
" No State Uncertainty: Ownship and contact states, in terms of position (x, y),
heading 6, and speed v are known such that any state uncertainty is considered
and state estimation algorithms are considered to provide state estimates that
" All Contact State and Intent Information is Available to Ownship: State and
34
any generated intent information of all contacts is available to ownship through
cess of contact state estimation is abstracted into this use of AIS data, such
that contact state estimation is separated from the collision avoidance problem.
represent actual vehicle intentions, such that the contact will not perform ma-
neuvering actions inconsistent with the communicated intent. This same as-
lision avoidance distances are such that vehicle geometry may be considered
circular with radius of the minimum distance centered at the body-fixed locus
distances between vehicles during a particular encounter, such that safety vio-
10AIS is a standard marine communication system that broadcasts vessel information such as
position, heading, and speed, and can include many other types of information such as vessel iden-
tification and destination [52].
35
lations or severe safety violations are considered to occur if the distance between
vehicle positions is, at any point, less than a defined threshold distance.
1.4 Overview
36
Finally, Chapter 7 provides overall thesis conclusions and recommendations for
future work in the area of intent-aware autonomous marine vehicle collision avoidance.
eral system requirements identified in Chapter 5 with the purpose of identifying the
rithms for the proposed intent-aware behavior. Appendix D identifies notation used
throughout this thesis along with the section in which the notation is defined.
37
38
Chapter 2
Literature Review
intent, and COLREGS compliance. This chapter begins with a broad overview of mo-
tion planning, as the overarching concept behind the process of collision avoidance,
and then more closely examines related work in COLREGS-compliant motion plan-
ning and the use of intent in motion planning. The chapter concludes with a discussion
planning. As the focus of this thesis is the marine domain and COLREGS-compliance,
review of related work in this area was performed in detail, while contributions from
research in other fields is explored to the extent that it illuminates general concepts
The field of motion planning is an incredibly rich field of study reaching across the
virtually every physical domain imaginable, including land vehicles, air vehicles, space
solutions to the motion-planning problem exist, there still remain many interesting
39
using algorithmic or computational advances, and exploration of problems specific to
a E A in a state x E X.
" A set of criteria by which the goal is to be achieved, e.g. optimality in terms of
some metric or feasibility in terms of obstacle avoidance.
The goal of a motion planning algorithm is then to determine the series of actions
that will allow the agent to move from the initial state to a state in the goal region
while satisfying the criteria set. The resulting motion plan typically consists of a
path r : [to, tf] -+ X, such that T(to) = xo and r(tf) = xf E X where xf is the
final state in the path, for some to E R and tf E R defined by the planner such that
t o 5 tf [65]. In the case where a planning algorithm associates a time with each r(t)
for all t E [to, tf], r is then referred to as a trajectory. There are a broad range of
approaches to motion planning, each of which may use different formulations of and
Motion planning algorithms may be classified as either global, local, or a hybrid of the
two [41,48]. With some subtle differences, the literature sometimes refers to global
path planning algorithms as deliberative and local algorithms as reactive [6]. This
40
thesis will use the terms global and local unless otherwise more appropriate to use
the terms deliberative and reactive, in which case any differences will be explained.
Global motion planning algorithms establish a desired vehicle path or trajectory from
the initial state to the goal region using all available knowledge of the environment and
some assumptions or inference on how the environment will evolve up to the planning
horizon [48]. These methods provide the advantages of allowing the agent to account
computational cost and the need to acquire more complete environmental knowledge
Local motion planning algorithms, in contrast to global methods, only use knowledge
of obstacles in the vicinity of the vehicle to develop a motion plan that applies to a
short planning horizon or nearby goal region [48], which is generally not the same as
the global goal region. These methods are generally used for local obstacle avoidance
and do not plan paths or trajectories beyond the results of a single action from the
current state; in other words, local methods typically use instantaneous environmental
state information to select only the next action. Due to generally lower computational
complexity than global methods, local methods provide the advantage of allowing
continuous and rapid re-planning as the local environment evolves. This capability
the vehicle must continuously react to new information, including new obstacles.
The main disadvantage is that local methods cannot, on their own, account for global
objectives and may fail to account for all available and relevant environmental data.
41
For this reason, there is generally no guarantee of path optimality when using a local
planner [48].
Due to the limited scope of local motion planning, the literature sometimes sep-
arates local planning methods from global methods by referring to local methods as
collision or obstacle avoidance techniques, while reserving the term of path planning
for global motion planning methods [61,94]. This thesis will use the term local to
refer to the planning method and the term collision avoidance to refer to the general
act of planning with respect to dynamic obstacles, since collision avoidance may be
As Kunchev observes in [61], a vehicle that only applies a global planning method that
does not account for a dynamic environment runs the risk of colliding with unknown
or dynamic obstacles and a vehicle that only applies a local planning method may
never reach its goal. One solution might be to use a global method that supports
rapid replanning, such as incremental planning [58,93], or that attempts to predict
and account for future states of dynamic obstacles [11, 91], but these methods may
still suffer from the disadvantages discussed above, namely computational complexity
in which a global method is used to establish a path from the initial state to the
goal region and a local method is used for collision avoidance. The local planner
may be employed either in conjunction with the global planner to guide the vehicle
along the nominal trajectory and plan deviations as necessary to avoid dynamic or
unforeseen obstacles or independently to guide the vehicle when triggered by newly
sensed information [48]. In this manner, hybrid methods attempt to draw from the
42
2.1.2 Motion Planning Challenges
In the world of the simple motion planning problem, in which the robot is the only
moving object, all obstacles are static, the robot is holonomic and can move without
regard for kinematic and dynamic properties, and there is no uncertainty in robot
relatively easy. However, in the real world, none of these conditions are typically
satisfied, resulting in the following widely recognized complications to the motion
" The presence of dynamic, or moving obstacles, whose own motion patterns are
" Uncertainty in every aspect of the problem, including observations and asso-
ciated state estimation, kinematic and dynamic models, and all aspects of the
This thesis is directly concerned with the issue of managing dynamic obstacles
in the form of other marine vehicles. While this problem is not decoupled from the
other issues, namely kinematics, dynamics, and uncertainty, these issues are handled
As discussed in the previous sections, one of the major difficulties in robotic mo-
achieve collision free motion, an agent must be able to predict the future states of
dynamic obstacles with sufficient accuracy such that the developed plan will actually
be collision free or must replan at a sufficient rate to account for deviations of the
dynamic obstacle from its expected motion [48]. Techniques to predict dynamic ob-
stacle motion include both deterministic trajectory prediction methods [36,67] and
43
probabilistic methods [7,11, 57,92,104], such as Markov Decision Processes (MDPs)
and Gaussian processes. Both deterministic or probabilistic methods may incorpo-
rate assumed or learned vehicle models and motion planning patterns to apply to the
trajectory prediction process.
In [67], Lef6vre proposes a classification of trajectory prediction based on the level
of vehicle intent for which the prediction accounts. The three proposed classifications,
starting with the lowest level of motion abstraction, are:
" Physics-based prediction: projects vehicle motion based on some vehicle model;
this is the simplest prediction technique but is typically only accurate for a very
short time horizon.
Each level of prediction provides a greater degree of future vehicle state prediction
accuracy with an associated increase in complexity of the prediction problem.
Since the goal of motion planning is ultimately to plan motion that a mobile robot
will execute, due consideration must be given to the interface of the motion planning
process with other robot functions and the environment. The boundaries of this in-
terface may even be dependent on the particular approach, however, in general, one
can consider the problem as shown in Figure 2-1 where the robot uses sensed infor-
mation to update a model of the real world, that model is used to plan actions, and
actions are executed through which the robot interacts with the world [87]. One can
consider hierarchical verses behavioral or subsumption architectures [8], or a hybrid
of the two, however the general idea of taking information from the environment and
applying it towards selecting actions is maintained.
44
World
sense act
Model
plan
Figure 2-1: Benjamin's original figure from [15], a basic robotic system paradigm in
which the robot uses sensed information to update a model of the real world, that
model is used to plan actions, and actions are executed through which the robot
interacts with the world.
Figure 2-2, from [15], is a behavior-based alternative to Figure 2-1 where individual
robot behaviors that represent different robot objectives are used to process sensed
While many solutions have been successfully applied to motion planning problems in
general, the problem of COLREGS-compliant motion planning is particularly difficult
due to the need to apply protocol-based constraints to the planning problem, where
the agent must first decide, in-situ, what constraints to apply [26,94]. As an added
maker, wherein it provides flexibility to rule interpretation based on all of the potential
situational variables that a human can observe and reason about given experience and
tion of the rules has been called into question [28,81], this is not a trivial task. This
45
__World * --
senseat
-- +( Behavior
- (Behavior) ato
Behavior_ selection
Figure 2-2: Benjamin's original figure From [15], a behavior-based robotic system
paradigm in which sensed information is passed to individual behaviors, the output
of which is used to select robot actions.
avoidance scenarios as well as the required actions for each vessel in the scenario, so
as to maximize the probability that vessels will maneuver safely with respect to one
another.
While COLREGS requirements are applicable "to all vessels upon the high seas
and in all waters connected therewith navigable by seagoing vessels," [54] individual
governments may enact additional or alternative rules for inland waterways, as in the
Inland Rules promulgated by the U.S. Coast Guard [96]. Where the application of
international rules may differ from that of inland rules as it pertains to the collision
avoidance discussions herein, a note will be made of the differences and appropriate
consideration thereof.
due to the fact that many investigations, including this one, only address a portion
46
of COLREGS. Additionally, no universally accepted definition or test exists for au-
erature focuses on work aimed at COLREGS compliance in this same area. Other
areas of COLREGS compliance, such as the interpretation of visual and sound signals
through visual and audio sensors or the visual differentiation between different types
of vessels, are related to collision avoidance, but may be considered distinct problems.
driven vessel collision avoidance. Rule 2 places general responsibility on all individual
vessels to comply with the rules and to depart from the rules in the case that it is
readily apparent" to the other vessel involved in the maneuver [54]. Rules 13 through
17 specify actions required by vessels in the cases where risk of collision does exist.
the following steps [26], which follow from the COLREGS requirements in Rule 7 to
3. If no risk of collision exists, continue to observe the contact but take no maneu-
(a) Use own and contact state data (position, heading, and speed) to determine
47
(b) Apply the constraints associated with the applicable COLREGS rule to
the motion planning algorithm.
pertain to the application and resolution of the COLREGS protocol, and step 4. (b) is
from steps 2 through 4. (a). The following sections, therefore, focus on work related to
steps 2 through 4 in addition to the motion planning approaches taken for COLREGS-
assessment of the "risk of collision" [54]. Given the existence of a risk of collision,
COLREGS demands that the vehicles at risk take certain actions in accordance with
particular rules. Therefore the ability to assess risk of collision is an essential capa-
Many approaches in the literature use the ownship-contact closest point of ap-
to evaluate risk of collision. If rcPA is below some threshold, then it may be deemed
that the contact will approach ownship within an unsafe distance and risk of collision
therefore exists. If tCPA or rcPA are above some threshold, it may be deemed that the
contact is too far away or may change its planned trajectory some time before it closes
ownship within an unacceptable range and therefore does not pose an immediate risk
of collision.
Other approaches take additional factors into account including vehicle dimen-
sions and geometry or other attributes of the interaction geometry, where it may be
considered that certain arrangements between ownship and the contact may represent
48
greater risk despite having similar r'CPA and tCPA 1941.
Accurately assessing risk of collision through the use of CPA or other geometrical
In the marine domain, per COLREGS, "Every vessel shall use all available means
collision exists" and "If there is any doubt such risk shall be deemed to exist" [54].
sion include visual observations, visual cues (including lights and day shapes), sound
specifies requirements for the use of visual cues and sound signals, it does not con-
radio, for the assessment of risk of collision. The inland rules for the U.S. permit
Means that might be available to an autonomous marine vehicle for state estima-
tion may include, but are not limited to, the following [35]:
" Radar.
49
The COLREGS-compliant motion planning literature generally takes one of the
following approaches when applying one or more of the above data sources to contact
future state estimation for risk of collision assessment and motion planning:
" Assume contact will maintain current velocity, and project future states using
this velocity [5,18, 62, 63, 72, 1101. In most cases, this means constant linear
contact and use the model to infer future contact states based on current and
to establish a notion of vehicle intent that can subsequently be used to evaluate risk
COLREGS compliance.
All methods that infer a vehicle trajectory do so based on an assumed motion
goal and use a predictive motion model that captures some deterministic or stochas-
tic contact behavior [90, 91]. The motion model is used in a forward simulation of
the current vehicle state to produce a predicted future state, set of future sets, or
distribution over potential future states. So far, these methods only account for fu-
ture contact maneuvers based on collision avoidance and do not account for potential
maneuvers based on objectives other than the goal [90,91]. This concept is discussed
further in Section 2.4.2.
REGS rules and their design for use by a human mariner [26,94].
The prevailing technique for COLREGS rule determination is a rule-based system
50
knowledge of ownship and contact states, associated geometrical relationships, and
linear velocity projections [5, 18, 62, 63, 72, 90, 91, 101, 110]. Greater detail on the
Other rule determination methods include the use of fuzzy set theory and neural
networks [46,66,71].
general motion planning techniques with varying assumptions and consideration for
the motion planning challenges discussed in Section 2.1.2. Like general motion plan-
ning approaches, the COLREGS-compliant approaches may be examined as global,
Global approaches, including use of state-space search over grid or lattice struc-
tures, are employed in [5,90,91]. Local approaches, including the use of velocity obsta-
methods, including the combination of RRTs with Dynamic Windows and lattice-
based state space search with a velocity obstacle extension, are explored in [72,101].
Section 2.1 illustrated challenges in predicting or inferring the future state of dynamic
obstacles in a stochastic environment in the context of establishing a collision-free
trajectory. In the case of any autonomous agent equipped with some capacity of
reasoning and interaction with the environment, including humans, future states can
be considered a manifestation of the agent's intent. Dictionary definitions of intent
51
0 "Something that is intended; purpose; design; intention" [33].
Using these definitions, it could be concluded that motion intent can be used to
infer a motion plan. Given an appropriate mathematical formulation of a dynamic
obstacle's motion intent, as some representation of how the obstacle plans to move, the
general goal of intent-aware motion planning techniques is to improve the prediction
of the obstacle's motion, and therefore plan its own motion with more certainty in
the safety or mission utility that will be achieved.
Intent in Aviation
In the 1990's, the aviation community began a substantial and ongoing effort to re-
search, develop, and employ Air Traffic Management (ATM) technology that leverages
the concept of aircraft intent [1,3,9, 12,44,85,99]. The following sections include an
exploration of aviation research and development projects that are relevant to the ap-
plication of intent to motion planning in the marine domain. Specific attention is paid
to early efforts in the development of intent-based aviation concepts, since there has
been limited exploration into these concepts for marine vehicles. Additionally, while
many results of the aviation efforts can be directly applied to marine applications, the
early efforts shed light on certain domain-specific considerations for an intent-aware
marine vehicle, such as the need for exploration of new COLREGS requirements.
While the aviation research literature is a useful resource for any exploration into
the concept of intent, when considering the application of aviation research-derived
concepts to the marine domain, the following differences must be noted:
* Aircraft are generally required to file flight plans prior to departure and these
flight plans are available to central, ground-based air traffic controllers (ATC).
Certain, but not all, marine vessels are required to develop navigation plans
prior to departure, but a substantial infrastructure to share these navigation
plans is still under development [50].
52
" Currently, in general, aircraft trajectory conflict detection and resolution is the
avoidance maneuver. The rules for power-driven vessels under COLREGS are
most similar to aviation right-of-way rules, which require aircraft of the same
type at the same altitude to maneuver in the same manner as marine surface
vessels under Rules 13 through 15 [2]. Much of the aviation research explores
arate set of rules in which qualifying aircraft in certain airspace are individually
responsible for collision avoidance and do so using many of the intent-based
concepts discussed in this thesis [109]. In the marine domain, all responsibility
for collision avoidance is with the mariner under one set of rules that applies to
all vessels and there are very limited instances of central control and communi-
" Much of the aviation research is performed in the context of developing au-
tomation tools to assist human operators, not necessarily to support fully au-
could be considered to possess some level of autonomy. While this thesis ex-
plores intent use cases for fully autonomous marine vehicles, research in human
interactions with intent-aware autonomy could be vital in bridging the gap be-
tween a marine environment with all human operators to one where human
" Aircraft may affect collision avoidance through the use of vertical separation
53
Intent in Autonomous Ground Vehicles and Cars
One area of particular interest to the autonomous car research community is the use
of pedestrian motion intent to predict pedestrian motion such that autonomous cars
can avoid hitting pedestrians. Work in this area has resulted in the development
of considerable insight into mathematical formulations of intent, including the use
probabilistic models for intent inference and application to motion planning [10,11,57].
Similar to the aviation research community, the maritime research community has
begun to undertake projects that investigate the benefits of exchanging intent in-
concept overseen by the International Maritime Organization (IMO) and the Interna-
tional Association of Marine Aids to Navigation and Lighthouse Authorities (IALA),
several projects have explored or are exploring the communication of intended vehicle
related aviation concepts, these projects in the marine domain are already demon-
strating the benefits of shared vehicle intent in the safe and efficient navigation of
vessels at sea [23, 70, 81, 821. These and other e-Navigation projects include inves-
tigations into the use of shore-based services, which could provide functionality or
As previously discussed, there have been limited applications of vehicle intent to au-
tonomous motion planning in the marine domain. Intent exploration to date consists
of both using intent to predict a target vehicle's motion with the goal of tracking the
54
vehicle [5, 101] and the use of assumed contact goals and motion models to predict
contact motion for COLREGS-compliant collision avoidance [90, 91, 101]. The use
2.4.2.
motion planning. These formulations range from high level abstract definitions, as
gories, down to precise formal, mathematical definitions, that may be used to resolve
a set of possible or exact motion plans. In an effort to arrive at appropriate formu-
lations of intent for the marine domain, each of these three levels of abstraction are
explored, understanding that these levels are notional and the degrees of abstraction
Definitions of Intent
Definitions of motion intent in the literature vary broadly, but all tend to define it in
terms related to the motion planning problem, including goals, action selection, and
In 2000, a meeting was held between the U.S. and European aviation authorities,
along with representatives of industry and academia, with the purpose of establishing
a unique definition of aircraft intent and developing a common vision for the devel-
opment of intent-based ATM decision support tools [85]. The conference concluded
that two separate definitions of intent were required, one with respect to the aircraft,
which can express precise aircraft motion with account for aircraft dynamics, and one
with respect to the flight, which expresses the plan of the pilot or ATC [85]:
Aircraft intent is "information on the planned future aircraft behaviour that can
be obtained from the aircraft systems (avionics), implying both:
55
" aircraft trajectory data e.g. way-point, route of flight, [top of descent],
[top of climb], and constraints such as weather and terrain, or
" aircraft control data e.g. heading, speed, flight level/altitude and con-
Flight intent is "the future aircraft trajectory expressed as a 4-D [three dimensional
position and time] profile until destination, calculated and "owned" by the air-
craft FMS agreed by the pilot communicated to ground and air systems,
path planned using a global planning method and the aircraft intent could be thought
of as the precise aircraft trajectory as computed using vehicle and environmental
models.
In the context of the overall notion of intent as a plan or purpose, these definitions
time. While these definitions are useful for the purpose of mathematically modeling
intent, they fail, as basic definitions, to capture the possibility that intent may be
described in some more abstract form, such as a plan to turn right, from which these
forms of intent must be inferred or derived. This may not be a problem in the case
doesn't require or support abstractions, which may be the case in aviation with the
strict standardization of flight plans and avionic system specifications, but would be
a problem in an environment, such as the marine domain as it stands today, that
permits the exchange of intent through modes such as sound signals or general radio
communications.
There are also several examples in the literature in which intent is represented
strictly in terms of vehicle states, including those that use a series of positions or
56
waypoints and several that simply characterize intent as a goal state, given which
other aspects of vehicle motion must be known or inferred [10,11, 91].
the manner in which the vehicle will be guided during the time interval over which
which an agent selects an action given current state and a goal state.
intent, they are actually consistent with one another if examined in the context of
light of vehicle state, vehicle dynamics, and environmental state is in fact the vehicle's
where intent fits into the process of the trajectory formulation. Since, in order to
plan a path or a trajectory, a vehicle must have both a planned goal and a means
" A policy describing how the vehicle plans to achieve a particular goal.
used with current vehicle state to predict the most likely vehicle trajectory over some
specified time interval [112]. The concept of intent parameters is discussed further
in Section 2.3.2, but the overarching principle is that intent may be described using
spatial and temporal attributes of current and future vehicle motion, such as planned
waypoints, as well as parameters that describe the manner in which, or behavior with
57
which, the vehicle will execute its plan, such as an objective to minimize distance
traveled.
Categories of Intent
Early aircraft intent research breaks intent down into categories that provide a struc-
ture to the spectrum of intent formulations discussed in the previous section, where
intent can be expressed with varying degrees of resolution to the trajectory produced
by exercising the intent [44,45,85,112]. While in some cases authors break intent down
by specific instances of intent representations [44,45], such as control inputs and way-
points, Zhao provides the following set of general categories to classify groups of intent
parameters [112]:
Motive intent: used to describe immediate motion plans, including changes in di-
rection and speed.
Objective intent: used to specify information about a target state, including target
heading or speed. This differs from motive intent in that motive intent may
specify a trend in motion that may result in any number of final states, while
an objective intent, by giving information about the final state, could be used
to arrive at a more precise trajectory through interpolation from the current
state.
Cost intent: used to describe the utility and constraints with which a vehicle com-
putes an optimal motion plan.
58
an obstacle, but may not have any specific objective for its speed or final direction,
therefore it could express a motive intent, but not necessarily objective or trajectory
intent without further resolving the motive intent through planning and modeling of
its own future motion.
Additionally, the above intent parameter categories fit into the broader intent
definition where any of the intent categories may be used to express the planned
motion behavior and objective or trajectory intent may be used to express the motion
goal.
The literature also classifies intent in terms of the time horizon over which it
applies [85,113], generally in terms of short term, or tactical, intent and longer term,
or strategic, intent. Zhao suggests the idea of specifying tactical intents over finite
time intervals and then specifying a strategic intent as a series of tactical intents,
where the composition of the strategic intent may change over time as different tactical
intents are selected to meet the strategic objective [113]. Additional levels of temporal
breakdown are possible, as in [85], where short, medium, and long term classifications
are used, with each category being assigned a specific time interval, i.e. 0 to 3 minutes
for short term.
59
and contact intent I, 7r gives the expected contact action, which can then be applied
to the contact transition model and successive states to resolve the predicted contact
trajectory. The pedestrian motion prediction literature assumes that the form of 7r
is known given I even when I may only represent some portion of the contact intent,
such as its goal state. This would mean that some portion of the vehicle intent, such
as the underlying motion planning optimality criterion, is assumed and accounted for
in ir. While it may be possible to assume and incorporate intent information into the
policy definition in certain applications, applying the general definition of intent to I,
which in turn influences 7r, allows one to apply this formulation more generally with
varying levels of intent information, as characterized by the intent categories in the
previous section.
The intent I, which represents any set of relevant intent information, then requires
some mathematical definition such that it can be applied to the computation of 7r, or
some other means of arriving at predicted contact motion. Reynolds and Hansman
introduce a state vector representation of intent in [831, in which vehicle intent is
considered as a state along with spatial and temporal state information. Such a
representation is given by Equation (2.1), where X.,, is a general ownship state vector,
inclusive of physical state and intent information.
XO (t)=) (2.1)
G(t)
Equation (2.1) gives I as a function of time, where I(t) may be any abstract repre-
sentation of intent, including some set of intent parameters. Reynolds and Hansman
suggest that I(t) consists of current target states and the subsequent planned trajec-
tory [83], but this would restrict the ability to express intent in terms of all possible
abstractions. Additionally, they separate goal information into G(t), which may be
done as a matter of convenience and doesn't impact the relationship of goal informa-
tion to overall intent information, so long as I accounts for both I(t) and G(t). By
60
constructing I(t) with appropriate intent parameters, the state vector representation
The advantage of the state vector representation is the ability to view intent
as a time varying quantity, which, just as other vehicle state information, may be
associated with some amount of uncertainty, can be compared to other state vectors,
and can be predicted at times in the future.
The intent vector consists of intent information represented by state variables. One
possibility for defining the state variables is through the use of intent parameters. The
aviation intent literature thoroughly examines the concept of intent parameters, pro-
intent parameters could be formulated as any variable that captures intent informa-
" Motive intent: mode of navigation (e.g. intent to cooperate with COLREGS),
speed change indication (in implicit terms, e.g. "right," or explicit terms, e.g.
" Objective intent: commanded heading, commanded speed, goal state, start
" Trajectory intent: vehicle trajectory (time indexed positions), trajectory change
In the context of motion planning, an agent might use any of the above parameters
61
2.3.3 Intent Inference
Since the ultimate goal of intent inference in motion planning is to obtain some
estimate of the future contact states, one approach is to assume the vehicle intent
and apply it to the problem of trajectory prediction, as presented in [67]. An ex-
ample is the assumption that the vehicle will maintain its current course and speed,
with account for environmental conditions, and therefore future states are simply an
extrapolation of the current state. Intent application where there is no online intent
determination is simply the case of non-intent aware motion planning, or intent-aware
motion planning with fixed intent.
The concept of intent communication has been broadly explored in the aviation intent
literature, since many potential instances of intent parameters are readily available
in digital form within existing aviation systems and, thus, may be readily commu-
nicated. The same is true of marine navigation systems, especially those onboard
autonomous vehicles, which generally possess some numerical representation of their
planned motion. Challenges associated with intent communication include the ability
of all agents involved to exchange, understand, and authenticate the communicated
intent. These challenges are addressed, respectively, through the establishment of
appropriate communication modes, content, and protocols, each of which is explored
in the context of motion intent by the literature.
62
Intent Communication Content
In order to overcome the issue of mutual understanding of intent, the literature gener-
ally establishes some formal specification for the content of an intent communication,
as derived from a particular set of intent parameters. The aviation literature takes
this specification a step further by exploring the definition and use of formal intent
The most fully developed aircraft intent description language is the Aircraft Intent
Description Language (AIDL), which is introduced in [74,75,1001 and further devel-
oped in [22,39,75], along with the Intent Composite Description Language (ICDL) and
the Flight Intent Description Language (FIDL). Together, AIDL, ICDL, and FIDL,
each consisting of an alphabet and a grammar, form a hierarchy that permits the ex-
pression of intent over particular time horizons and different levels of abstraction. In
a composite of AIDL instructions which may be used to represent more complex flight
behaviors and may not translate directly into a specific trajectory, such as a plan to
achieve a particular speed without precise definition of the control actions to achieve
the speed; and FIDL is the highest-level language which is used to represent overall
use of ICDL and AIDL instructions [39]. It is observed that this language definition
is consistent with the idea that intent may be represented through varying degrees
of resolution of the trajectory resulting from exercise of the intent. Additionally, the
intent, with which one could express any of the previously discussed intent categories.
Use of AIDL, ICDL, and FIDL to communicate aircraft intent and then predict
aircraft trajectory within acceptable degrees of accuracy is demonstrated in [40,59,
60,80, 95].
Development of such a language for the marine domain is outside the scope of this
63
require some standardization of communications as identified in [88]. The format of
waypoint exchange proposed in [88] has been adopted for the marine Electronic Chart
Display and Information System by the IEC [50]. Intent communication content in
addition to waypoint exchange is a potential area for future work. This thesis attempts
to establish the level of communication standardization necessary to test the proposed
intent-aware motion planning system.
64
and estimated time of arrival, however, like ADS-B, AIS guidelines do not require
the broadcast of this information [55, 97], and there is no further explicit provision
for intent communication. AIS does have a provision for Application-Specific Mes-
sages (ASMs), which are specially tailored messages for certain applications [51], but
these messages must be specifically defined and authorized by international and local
authorities as in [56] and [97]. While not used for the purposes of motion intent
communication, it is noted that [50,56] do provide for a route messages and waypoint
advice messages, which contain waypoints as data fields. Also like ADS-B, which
is only required in certain airspace, AIS is only required to be installed on ships of
300 gross tonnage or more engaged on international voyages, cargo ships of 500 gross
tonnage or more not engaged on international voyages, and passenger ships under
any circumstances [76].
Other, more general, methods of communication include sound signals, light sig-
nals, and voice communications, which require some added layer of processing to
arrive at an intent formulation that could be applied in an intent-aware motion plan-
ning algorithm.
With the development of inter-vehicle communication networks, there has also been
significant exploration into the communication protocols required to execute com-
munication with respect to the type of communication, including communication in
support of collision avoidance [1051. One example of protocol application is [42] in
which a "handshake" procedure is established to achieve vehicle agreement on actions
to avoid collision at an intersection.
65
2.4 Intent-Aware Motion Planning
The motion intent literature, whether considering intent as specific spatial informa-
tion, behavioral information, or some combination of the two, in all cases focuses
on the prediction of contact motion given the intent information. Intent-aware mo-
tion planning seeks to use contact intent information to inform contact trajectory
prediction in the process of planning ownship motion. Examples of general intent-
aware motion planning frameworks are explored, followed by exploration of intent
applications in COLREGS-compliant motion planning.
Use of intent in aviation research is often done so in the context of Conflict Detection
and Resolution (CDR), where a conflict between aircraft is synonymous to the exis-
tence of a risk of collision in the marine domain. Frameworks in the aviation intent
literature incorporate the following mechanisms or considerations for the application
of intent to the aircraft collision avoidance problem [12,37,45,47,80,95,106,109]:
" Distinction between tactical and strategic conflict detection and resolution [34,
107], where tactical conflicts represent near-term, immediate risk of collision,
and strategic conflicts represent less immediate risk, and different techniques
may be employed to handle each.
66
Similarities can be drawn between the above considerations and requirements for
COLREGS-compliant collision avoidance, where risk of collision determination and
rule determination are of fundamental concern.
Investigation into the above framework components reveals several additional con-
siderations for intent communication frameworks, including: appropriate amount of
intent information to share including time horizons of shared intent [45, 106]; the
need for special protocol rules, including specification of collision avoidance respon-
sibility or priority [12,45,106]; the use of priority verses cooperation in protocol rule
specification [37]; the possibility that communicated intent or protocol rules are not
followed [106]; and potential need to confirm intent and/or responsibility through
negotiation [45].
in the aviation domain found that, in general, use of intent information resulted in
The pedestrian motion prediction and avoidance literature, such as the IAMP pre-
motion planning through the application of probabilistic models, such as Markov De-
cision Processes. These models are used to estimate intent, associated contact motion,
and the resulting optimal sequence of maneuvering actions [10, 11, 57]. Limitations
limited account for actions taken by pedestrians to avoid one another or in response
67
2.4.2 Intent-Based, COLREGS-Compliant Motion Planning
The use of vehicle intent for the purposes of COLREGs-compliant collision avoid-
ance has been explored by Shah and $vec in [90,91, 101].
In [101], Svec employs a predictive motion model that computes the worst-case
contact maneuvering action given an assumed contact goal state. Potential maneu-
vering actions are then identified by forward simulating ownship motion using a model
of ownship dynamics and identifying the actions that do not result in collision with
known contacts. The planner then determines whether or not a COLREGS-defined
encounter exists with any contact. If in a COLREGS encounter, actions are that
are non-COLREGS compliant are penalized in the computation of an optimal ma-
neuvering action towards the next ownship waypoint. Svec demonstrates COLREGS
compliance in two-vehicle encounters for each of the basic COLREGS encounters
while executing a target-following mission.
In [90], Shah proposes an intention model that also assumes knowledge of contact
goal state, either through some other estimation process or through radio communi-
cations. The model includes a classification function that estimates whether or not a
vehicle intends to comply with COLREGS based on history of observed states of the
contact and other contacts. Future contact motion is predicted by forward simulating
contact motion using an assumed model that accounts for the COLREGS-compliance
classification. The result is a deterministic trajectory, which Shah transforms into
a distribution over future vehicle states by assuming variance in vessel position as
a function of distance from the contact current and goal locations, where the mean
position is on the deterministic trajectory. These state distributions are then applied
to risk-of-collision and COLREGS scenario determination in cost-function definition
for a state-space search over dynamically feasible ownship motion primitives. At
68
each state expansion in the state-space search, the planner evaluates all candidate
maneuvers for COLREGS-compliance with respect to all contacts.
using the proposed algorithm is sufficiently limited such that Shah employs a veloc-
ity obstacle-based local motion planner to follow the nominal trajectory generated
of accounting for vehicle dynamics, use of more environmental information than the
step in the motion plan. The proposed method does result in lower efficiency, in
terms of time and distance, than the velocity obstacle method due to limited replan-
ning capability and the need to return to the state-space search defined path upon
the motion planning problem, including findings that use of intent information im-
proves the ability of an autonomous vehicle to safely and efficiently execute collision
69
avoidance. Additionally, it is observed that the use of motion intent by autonomous
vehicles in the marine domain has been explored in the form of local vehicle goal
locations and assessment of intent to comply with COLREGS, but there is no known
work that explores other forms of intent and the explicit communication of this intent
information.
70
Chapter 3
Intent Formulation
For the purposes of this thesis, motion intent, referred to simply as intent, is defined
as follows, as derived from the literature review presented in Section 2.3.2:
As in the case of aircraft intent discussed in Section 2.3.2, in the marine domain
do not directly specify vehicle intent as a goal or planned behavior, but give some
information as to what its goal or planned behavior might be. These abstract intent
descriptions may or may not be enough to resolve some or all of the vehicle's planned
motion, but can be used to inform the motion planning process. The intent categories
in Section 2.3.2 are adopted to describe different intent abstractions for the marine
domain.
71
Perception/ raw sensor data
state Sens
Esdmaon
Environment
moioControuer
PManner a E - + Actuatorm
q Vehicle-
T
Mistion
PlanCeng
Figure 3-1: A general mobile robotic system model, where vehicle intent I,, is de-
veloped through some mission planning process and motion is executed through a
motion planner that uses intent and information in an onboard world model XE to
determine an action a that leads to actuator commands u, generalized forces q, and
ultimate evolution of vehicle state in the form of a trajectory r.
72
and I, refers to the intent of some contact under consideration. Intent I is captured
in the mobile robotic system as shown in Figure 3-1, where XE represents an onboard
environmental model that may include ownship state x,,, contact states x,,j, and
contact intents Ini, for n contacts such that i E {1, ...
,
Similar to physical states, intent of other agents may be observable or non-
observable, and as such may be processed using some state estimation or inference
method [11]. While there has been considerable work in the area of intent inference,
the intent of other agents is generally not observable. However, intent can be made
observable through direct communications from one agent to another. Intent is first
considered from the perspective of applying an agent's own intent to its motion plan-
ning process and then from the perspective of applying another agent's intent to the
motion planning process.
Considering the system represented by Figure 3-1, as an expansion of the model pre-
sented in Section 2.1.4, some mission planning process is used to develop ownship
intent I, which consists of any information required by the motion planner to es-
tablish motion goals and desired behavior for motion execution. The motion planner,
using some planning method that accounts for I, and some or all of the information
contained in the environmental model XE determines the next maneuvering action
a E A to be issued to the controller. The controller applies some control law that
maps an action a E A and feedback information contained in XE to a set of actuator
commands u E U, where U is the vehicle's actuator command space. The actuators
and environment produce a vector of generalized forces q that act on the robot such
that xos = fveh (x0s, q), where fveh is a dynamic model of the vehicle. Vehicle state
then evolves according to c, as the motion plan is executed, resulting in a vehi-
cle trajectory T. Through this process a vehicle intent is transformed into a vehicle
trajectory.
In a completely deterministic setting, the above relationships might be used to
73
predict a unique vehicle trajectory [67]. In general, many of aspects of the system
states must be estimated from noisy sensors, actuator and vehicle models may not
perfectly represent actual dynamics, and disturbance forces perturb actuator and ve-
estimates of system behavior may be achieved to execute safe and efficient motion
planning.
In the robotic system discussed in the previous section, the environmental model
XE = (xi, ... , Xn), consists of individual states xi, each of which consists of variables
xj that capture information about the environment. Some of these xi represent the
state of moving obstacles, or contacts, consisting of both the contact's physical state
xo,i and intent I.,. Given some knowledge of x.,i, Ici, and other relevant infor-
contact trajectory could be predicted with a perfect vehicle model, vehicle state, and
current and future environmental state information. This predicted trajectory could
be applied to the motion planning process, where predicted contact motion would
also need to account for reaction to ownship planned motion.
motion models, and other environmental variables, including the case of incomplete
intent information, some mechanism must be applied to estimate future contact mo-
tion such that it can be applied to ownship motion planning. In light of these consid-
erations, intent-aware motion planning is explored through the need to relate contact
74
3.4 Vehicle Trajectory
This thesis considers the use of discrete trajectories. Where a continuous trajectory,
and a time t E [to, t1 ] [65]. The time period over which the trajectory is defined
TO = (xo, to), rn = (xn, tf), ti+1 > ti Vi E {, 1, ... , n}, and length is defined as the
number of states represented by the trajectory.
States in the discrete trajectory between ti and ti+1 can be estimated through
the use of an interpolation function. In the case of linear interpolation, estimation
by which will suffice in most instances where the interval from ti to tj is sufficiently
small in relation to the dynamics of the system in question, intermediate states are
t - ti
x(t) = xi + (x3 - x) -- - (3.1)
Note, that in the general case, Equation 3.1 can be used to interpolate the state
x for any value of t, however to maintain the ordering, accuracy, and validity of the
estimated trajectory, the following should hold: tj > ti and t E [ti, t].
vehicle trajectory may terminate at some time horizon, tf, at which vehicle motion
may continue with some known velocity, but unknown future changes in velocity. In
this case, future states (t > tf) can be projected or inferred as a function of xn.
While all vehicles in the marine domain may possess a fully defined intent, including
a goal state and planned behavior for achieving that state, certain partial intent
75
terms of the intent categories defined in Section 2.3.2. The identified forms of intent
are not intended to be an exhaustive list of all forms of intent information in the
marine domain, but rather an initial exploration of possible forms that might be
The following paragraphs identify potential marine domain intent abstractions, which
implicitly capture vehicle motion intent information. These abstractions are proposed
on the basis of information typically available through the course of motion planning
A complete trajectory meets the definition of trajectory in the Section 3.4 and is
illustrated in Figure 3-2. It represents the complete vehicle motion plan, including
some account for the dynamics over the time interval covered by the trajectory for-
mulated for the current state and environmental conditions. This information may
be available to a marine vehicle based on its mission or navigation plan, its current
motion plan, and a model that captures the system depicted in Figure 3-1.
Since the underlying intent is used to formulate the trajectory but is not explicitly
expressed in the trajectory, the complete trajectory, without additional inference, can
only be used to express what the intent was at the time the trajectory was formulated.
In the case where future actions are updated by the underlying intent due to change
in state, goal, or the environment, an intent expressed as a complete trajectory must
also be updated.
goals or waypoints, which, at the very least consist of an ordered set of positions
(x, y) E R 2 . This information may be available to a marine vehicle based on its mission
76
CD) Ownship
CJD Contact
Ownship Trajectory
State Vector
T(tf)
Figure 3-2: Intent as a complete trajectory, where the underlying vehicle intent is
used to generate a predicted trajectory r that gives future vehicle states for all times
t E [to, tj.
or navigation plan, where it may not be possible to develop a complete trajectory due
to incomplete or insufficient system modeling to support trajectory generation.
including planned times for each waypoint, from which speed can be inferred. Like the
since the waypoint specification essentially provides intermediate vehicle goals and
allows estimation of vehicle action selection and execution through the interpolation
process, intent specification as waypoints is less volatile through the course of vehicle
motion than a complete trajectory. At the same time, a greater degree of inference or
prediction must be performed in order translate waypoints into future vehicle states.
77
-4
Ownship
C&Z Contact
-*Connected Waypoints
-- + State Vector
Figure 3-3: Intent as a set of waypoints, where the underlying vehicle intent is used
to generate a series of planned positions (xi, yi) beginning at (xo, yo) and ending at
(xf,Yf).
Collision avoidance mode intent indicates the behavior with which a vehicle intends
to avoid other vehicles. This information is readily available to a vehicle based on
decisions made with respect to execution of collision avoidance. For a vehicle that
plans to cooperate with COLREGS, this intent might be expressed as the COLREGS
scenario that the vehicle assesses that it is in with another vehicle and therefore
what COLREGS constraints it intends to apply to its motion. This intent might also
express that the vehicle does not intend to cooperate with COLREGS.
78
specification in relation to ownship state and contact state are described as follows:
the contact, as in a ship using a sound signal to indicate that it plans to pass
another on its port side.
they can be used to understand what portions of the physical state space may or
may not be occupied by the signaling vehicle during the encounter, in terms of some
Feature references are similar to directional specifications except that instead of sig-
naling an immediate plan, they signal an intent to maneuver with respect to some
feature in the environment. Examples include a ship planning to turn to some head-
ing once past a particular buoy or a car that will be turning at some upcoming
intersection.
mon understanding of what the feature is and where it is located. To ensure this
common understanding, the vehicle communicating the intended maneuver could also
specify the location of the maneuver, in which case the feature referenced intent be-
comes like a waypoint specification with an approximate waypoint location and the
The idea of a feature referenced intent is illustrated in Figure 3-4 for a planned
turn to starboard around a buoy.
79
CDOwnship
Contact
Potential Trajectory
State Vector
Buoy
Other potential marine domain intent abstractions include a destination or goal loca-
tion, a near-term target state, current controller input and any other intent-derived
information that can be applied to some model accounting for the processes in Figure
3-1 in order to reason about future vehicle motion [85].
Considering the intent abstractions presented in the previous section, the basic nature
of trajectories, and the importance COLREGS scenario determination within the
COLREGS-compliant motion planning problem, the following intent formulations are
proposed for initial exploration of surface vehicle intent communication and intent use
in COLREGS-compliant motion planning. The components of each formulation can
be considered as instances of intent parameters, as discussed in Section 2.3.2.
80
To
Ownship
Contact
* Interpolated Discete Trajactory
-+w State Vector
Figure 3-5: Discrete trajectory intent, where the underlying vehicle intent is expressed
as a series of vehicle state and time pairs, between which other states may be inferred
or interpolated.
81
neuvers associated with the mode in terms of a submode, Isubmode. COLREGS
mode and submode are discussed further in Chapter 4. For the give-way vessel,
submode may give the planned direction of passage, namely, port or starboard
for overtaking or head-on, and ahead or astern for crossing. Collision avoidance
mode and submode intent Iavoid is given by Equation (3.3).
Figure 3-6 shows a COLREGS-defined crossing scenario where the give-way vessel
recognizes the crossing scenario and intends to pass astern of the stand-on vessel, such
that it may follow any trajectory that causes it to pass behind the stand-on vessel.
This action is selected under the assumption that the stand-on vessel will follow
COLREGS and maintain its current course and speed. In the case of the give-way
vessel, therefore, the collision avoidance mode intent is compliance with a COLREGS
crossing scenario and the directional intent is passage astern of the stand-on vessel.
Application of these intent parameters to the COLREGS-compliant motion plan-
ning problem is discussed in Chapter 5 and elaboration on COLREGS modes and
constraints is provided in Chapter 4.
82
I
Figure 3-6: Collision avoidance mode and submode intent, where the underlying vehi-
cle intent is expressed through an intented method of collision avoidance and direction
of passage; this figure shows a COLREGS-defined crossing scenario in which the give-
way vessel intends to comply with COLREGS and pass astern of a stand-on vessel.
The stand-on vessel is expected to maintain heading and speed. The maneuvering
region for astern passage is notional based on current vehicle states.
83
84
Chapter 4
COLREGS-Compliant Motion
Planning
This chapter establishes the basic theoretical framework in which the COLREGS-
approach. The concepts in this chapter rely heavily upon previous work by Ben-
In all COLREGs encounters we consider interactions between two vehicles where one
vehicle has the responsibility to give-way, per [54] Rule 16, and the other vehicle the
responsibility to stand-on, per [54] Rule 17, or both vehicles must give-way, where
give-way and stand-on are defined as follows:
" Give-way vessel: vehicle required to keep out of the way of another vehicle and
" Stand-on vessel: vehicle about which a give-way vessel is maneuvering; this
vehicle shall keep course and speed unless it becomes apparent that the other
85
vehicle is not taking sufficient action to avoid collision, in which case it may
and shall take action to avoid collision.
a single contact is the other vessel in the encounter and each has a state x = (x, y, 6, v)
and speed v E R. The subscript "os" is used for ownship and the subscript "cn" is
used for a contact, i.e. ownship and contact states are x0 , and x., respectively.
venient to define several variables, as presented in [171, that describe the encounter
geometry based on vehicle states. Figure 4-1 depicts geometric references that are
used in computing these variables. The variables used in this thesis are given in
Table 4.1. Equations for each variable are provided in [17] and select variables are
illustrated in Figure 4-2. Additionally, equation (4.1), as given in [17], is defined for
convenience.
-+180)
>0360
0 '
[0]180 -- [60
-
Symbol Description
86
Symbol Description
means ownship will pass the contact on the contact's port side.
cross Boolean variable indicating that one vehicle will cross the other's
centerline if evaluated contact and ownship headings and speeds are
i.e. cross' means ownship will cross behind the stern of the
contact.
COLREGS encounters are defined in terms of current ownship and contact states,
where the encounter begins at time tencolnteT and ends at time tcomplete. This thesis
defines the time tencunte, as the time at which r' becomes less than an encounter
threshold range rencounter, and tcomplete as the time at which r" becomes greater than
COLREGS compliance is defined as the task of selecting and executing the ac-
tions a = (Od, Vd) consisting of desired heading Od and desired speed Vd that lead to
relative motion of the two vehicles that adheres to the constraints for the particular
87
Iport I starbo
bow
beam beam
stern
astern
centerline
North
North
bng contt
contact
ownhip
ownship
a'
Figure 4-2: Geometric encounter parameters as given in Table 4.1, based on Ben-
jamin's figure in [171.
88
Give-way Vessel
0
Stand-on Vessel
State Vector
4.1.1 Overtaking
Per [54], Rule 13, a vessel is overtaking another if it is approaching the other vessel
from a direction more than 22.5 degrees abaft' of its beam, and if there is any doubt as
to whether or not this is the case it shall be assumed so. The overtaking vessel is the
give-way vehicle and the vessel being overtaken is the stand-on vessel in this scenario.
Additionally, [54] directs that the give-way vehicle maintains give-way responsibility
until "finally past and clear" of the other vehicle. Once in an overtaking scenario any
alteration of bearing between the vehicles cannot change the scenario to a crossing
scenario.
Overtaking - Give-way
Entry into the overtaking scenario as the give-way vessel is defined mathematically,
as shown in [17], by the following conditions:
89
. pass" = true
follows, where the value of passos and pass", are evaluated at each time t and it is
recognized that the give-way vessel may decide to change the side of passage:
" ro > rcpa,min Vt E [tencounter, tcomplete] for some minimum safe range repa,min that
Overtaking - Stand-on
Entry into the overtaking scenario as the stand-on vessel is defined mathematically,
* pass' = true
608(t) =0"(to) E0 Vt E [to,0 tf] for some variation in heading eo that is very
small relative to ,, (to)
* v 08 (t) = vo8 (to) ,Ev Vt E [to, t1 ] for some variation in speed e, that is very small
relative to v0s(to)
* rg > rc1 a,min Vt E [to, tf] for some minimum safe range rqpa,min that may be
particular to the scenario or the vehicles (specification of this constraint is in
90
recognition of the fact that the stand-on vehicle is responsible for taking action
to avoid collision in the event that the overtaking vessel does not take sufficient
4.1.2 Head-On
Per [54], Rule 14, a vessel is in a head-on scenario with another if it is meeting
the other vessel on a reciprocal or "nearly" reciprocal course so as to involve risk-of-
collision, and if there is any doubt as to whether or not this is the case it shall be
In the international rules, the vehicles are required to alter course to starboard so
as to achieve a "port-to-port" passage to avoid collision in all cases, while in the U.S.
inland rules the vehicles vehicles are required to alter course to starboard unless they
the following conditions, where kheadon is used to define a nearly reciprocal course:
* I [C]18"0I 5 (head.o
S|[0]1801 Oheadon
Constraints for either vehicle in a head-on scenario can can be expressed as follows,
where the value of pass'" and pass" 8 are evaluated at each time t and it is recognized
that the give-way vessel may decide to change the side of passage:
" passnp = false for all t if port-to-port passage is not required and vehicles
" ro > rcpa,min Vt E [to, tf ] for some minimum safe range rpa,min that may be
particular to the scenario or the vehicles
91
4.1.3 Crossing
Per [54], Rule 15, a vessel is in a crossing scenario when meeting another vessel in such
a way that their courses intersect, they are not in an overtaking or head-on scenario,
and a risk-of-collision exists. The vessel with the other on its starboard side is the
give-way vessel and the other is the stand-on vessel. Additionally, [54] directs that
the give-way vessel should avoid crossing in front of the stand-on vessel. [54] Rule 17
gives the special provision for a stand-on vessel in a crossing scenario that it should
not alter course to port in the event that it must maneuver to avoid a collision.
Crossing - Give-Way
it is recognized that the give-way vessel may decide to change the side of passage:
" cross cb = true for all t in which ownship plans to pass ahead of the contact
* cross;" = true for all t in which ownship plans to pass astern of the contact
r > ropa,min Vt E [to, tf] for some minimum safe range rqpa,min that may be
Crossing - Stand-On
92
e ce < 112.5, and
Constraints for the stand-on vessel in a crossing scenario can be expressed in the
same manner as those for a stand-on vessel in an overtaking scenario given in Section
4.1.1.
plified two-dimensional trajectories are applied such that -rs(t) = (xos (t), yos (t)) E R 2
and -re(t)= (xen(t), ycn(t)) E R 2 allowing expression of rcpa using Equation 4.2.
If it is assumed that ownship and the contact will maintain constant course and
speed, i.e. 0 ,,(t) = 64,(O) = 9OS, vOS(t) = vos(O) = vo, Ocn(t) = Ocn(O) = Ocn, and
ven(t) = ven(O) = vn, where t = 0 represents the current time, and the vehicle
has zero lateral velocity, then the forms of Ts(t) and Tr.(t) are linear, as given by
Equations (4.3) and (4.4). This makes the form of the squared Euclidean norm in
Equation (4.2) quadratic such that one may compute rcPa by solving for the time tea
and range rc~a at which -rcpa = 0, as given by Equations (4.5) through (4.7) [17,201.
In Equations (4.3) through (4.7), the positions (xOS, YOS) and (xn, ycn) are ownship
93
and contact positions, respectively, at time to. Equation (4.6) recognizes that, for a
range rate r > 0, the vehicles will become no closer and rcPa = rm.
tcpa = {~
0 1>0
i (4.6)
ko = y2, - 2yosycn + y2 + X2 - 2x x + X2
(4.7)
In the general case, where r0 (t) and ren(t) are non-linear and not continuously
ren, where ros,i = ros(to,,) and rcn,i = -re(tcn,i). If discrete trajectories are available
but are not equally sampled such that t,,, = ten,i Vi E {0, 1, ... , n} or the trajectories
do not have the same length n, then interpolation as described in Section 3.4 can be
used to re-sample the trajectories at equal intervals over the overall interval of [to, tf].
Given the equally spaced, discrete representations one may apply Equation 4.2 and
the assumption of constant course and speed between each time ti and ti+1 to obtain
94
CPA range over ownship and contact trajectories, repa,,, as given by Equation (4.8),
where ros,i = (xos,i, yos,i) and -re,i = (xc ,i, e,,). The functions ro,i(t) and -rc,i(t)
are given by Equations (4.11) and (4.12), respectively, where the values of 0,,,i and
0e,,i are given by Equation (4.9), v.,,i and ven,i are given by Equation (4.10). The
function bng((xi, yi), (x 2 , Y2)), as defined in [17], gives the absolute bearing in the
interval [0, 360) from the position (x 1 , Yi) to the position (x 2 , y2).
_ I1i+1 - rill
=
-
ti+1 - ti (4.10)
Given a CPA, or other risk measure, a threshold value may be applied to determine
whether or not a risk of collision exists. Furthermore, if the current range is the CPA
range, meaning that vehicles will become no closer and the current range does not
represent any risk, then it is possible to conclude that a risk of collision does not
exist. Equation (4.13) defines a Boolean evaluation of risk of collision based on re~a
using a value rca,min as a threshold above which it is evaluated that risk of collision
discussed in [28], such that risk is not considered to exist if these values are so great
only activating a collision avoidance behavior in the case that r' is below a defined
95
threshold.
When evaluating risk of collision over two trajectories, it is recognized that multi-
ple local minima may exist in the evaluation of rpa over the legs of the two trajectories
such that each minimum is less than the risk threshold. Equation (4.8) assigns the
global minimum to repa,r since the existence of any risk of collision based on this
measure should trigger collision avoidance actions. These actions will be determined
either through local or global planning methods that account for the complete contact
trajectory and therefore resolve all minima that represent a risk of collision.
Given the basic COLREGS scenarios listed in Section 4.1, the goal of a COLREGs-
compliant motion planning system is to apply the constraints associated with those
scenarios in order to avoid collisions with other vehicles, while simultaneously meet-
ing mission objectives, such as transit to a goal location with minimum time or
to meet mission requirements and to avoid collisions with static and dynamic obsta-
cles. A multi-objective optimization approach, such as the IvP method introduced by
Benjamin in [21] and developed further for application ot marine autonomy in [20],
provides for tradeoff between these objectives by determining the optimal maneu-
objective function fi represents the utility of action a for a particular vehicle objective
i. Each objective function fi is weighted by some scalar function wi, which establishes
the relative importance of the objective i with respect to other objectives. Objective
functions and associated weights may be defined in terms of any number of variables
96
relevant to the particular objective. In the sense that the multi-objective optimization
approach computes only the next maneuvering action, it is a local motion planning
priate objective functions, such as with the waypoint behavior described in Section
4.3.3.
n
a* =argmax wifi(a) (4.14)
aEA
of the action space A that violates motion constraints, to some appropriately low
number, or zero, such that these actions will never be selected over allowable actions,
a E A \ A, [171. These allowable actions will generally have some higher, non-zero
utility. Formally, if for a single objective f, f(a,) < f(a) V a, E A, and a E A \ Ac,
and A \ A -L 0, optimization of f over A per Equation 4.15 will result in a* E A\ A.
cording to Equation 4.16, where the function g(a) gives the utility of action a and the
function h(a) E [0, 1] applies an effective discount to the utility value [17]. Motion
constraints can be enforced by defining h(a) such that it gives low values for actions
blue corresponding to actions of low utility. In this case, the actions a E (180,360)
97
-- I
270 90
180
h(a) = 1 Va E [0, 180]. All actions are equally desirable, therefore f(a) = c Va E A,
In the case of multi-objective optimization, the use of the above constraint for-
mulation and the definition of weighting functions wi, permits the tradeoff between
desirable actions and constraints amongst multiple motion objectives. In the case
where certain actions are not explicitly disallowed, but are less desirable than other
actions, the objective function can be appropriately constructed to produce some in-
termediate utility value, such that, in light of other objectives and their weights, the
action might be the optimal solution to satisfy both objectives. Since constraint for-
mulation is expressed in the objective function, a case may arise in which the utility
and weight of one objective is sufficiently large such that linear combination with
incorporated in one of the other objective functions. This behavior may be desired
if the system must have flexibility in constraint application, otherwise weighting and
98
Environment Inputs
utility function values must be established such that, when linearly combined with
other objective functions, the overall utility of disallowed actions is less than that of
allowed actions.
planning method, the problem is to develop weighting and objective functions that
produce the desired vehicle motion. This construct is illustrated in Benjamin's orig-
inal figure from [161, which is reproduced in Figure 4-5, where a particular vehicle
single objective, or IvP, function and weight, which are all processed by the solver to
99
desired heading Od and speed vd, respectively, where f(d, Vd) = wOfO(0d)
+
wVfA(vd), and wo and wv weight the importance between heading and speed
selection, or
A collision avoidance behavior seeks to identify actions that will prevent collision
with other objects, and therefore builds an objective function where expected safe,
collision-free actions have higher utility than those that are unsafe or result in a
collision. One such function that fits this profile is rca, as defined in 4.2 and use
of which is demonstrated in [15, 20]. The function rcpe,(O, v) is defined as the the
computation of rqa using Equations (4.5) through (4.7) for a potential action a =
(0, v), where ownship position (X 0 8, yos) and contact state xen = (Xc, yen, 6Oc, ven) are
treated as constants in the computation.
As shown in [20, 111], if a threshold is applied to the value of repa(6, v) such
that actions resulting in a rcpa below the threshold, ac E Ac, have zero utility and
actions resulting in a rcpa above the threshold, a E A \ Ac, have some constant non-
negative utility, the objective function presents a binary selection between A and Ac,
which can be equated to Fiorini's velocity obstacle approach [38]. The multi-objective
optimization formulation, however, allows for a non-binary selection by providing for
continuous variation of utility over a specified range of rcpa values.
As presented in [17,20, 110], a CPA-based collision avoidance utility function can
be constructed according to the following parameters:
" The rcpa value below which the objective function produces minimum utility:
rcpa
* The rca value above which the objective function produces maximum utility:
rcpa
100
The above parameters are used to define the utility function gavoid(a) = g(0, v)
in Equation (4.17), where gmi, is some minimum utility value and gmax is some
maximum utility value [17]. The value gintermediate is some function of the minimum
and maximum values and rcpa, which may simply be a linear interpolation between
gmin and 9max for rea on the interval [fcpa, cpal. This thesis employs the values
9min = 0, gmax = 100, and gintermediate as the linear interpolation given by Equation
(4.18).
For this generic collision avoidance function, there are no additional motion con-
straints, therefore h(a) = 1 Va E A and f(a) = g(a). The polar heat map of this
objective function for a particular (x,,, ys) and xe, is shown in Figure 4-6.
The weight associated with a collision avoidance behavior expresses the impor-
tance of selecting a collision avoidance action for a particular contact over other
motion objectives. By making this weight a function of the risk of collision with a
contact, collision avoidance with vehicles that present higher risk of collision is pri-
oritized over those that present lower risk. A collision avoidance weighting function
can be defined similar to the objective function, again as presented in [20,1101, using
* The contact distance below which the objective function weight takes a maxi-
" The contact distance above which the objective function weight takes a mini-
101
-4
0
Cc Ownship
Contact 0 100
State Vector
Utility
Figure 4-6: Collision avoidance objective function in which the red region corresponds
to actions resulting in rpa(G,v) > icpa, blue region corresponds to actions resulting
in rep,(0, v) < fp,,, and other colors represent some intermediate value of utility.
102
The above parameters are used to define the weighting function in Equation (4.19),
where wmin is some minimum weight value and wmax is some maximum weight value,
and wintermediate is some function of the minimum and maximum values and a contact
range r'. Like gintermediate, Wintermediate may simply be a linear interpolation between
wmin and Wmax for r" on the interval [i,,t, iWt]. This thesis employs the value wmin =
0 and wintermediate as the linear interpolation given by Equation (4.20). The value Wmax
min
08 >
rd> fiwt
The goal of a waypoint following behavior is to identify the best maneuvering actions
for a vehicle to traverse some specified set of waypoints. Using the method of combin-
ing separate objective functions over desired heading and speed, waypoint following
also be maximized towards the line segment from the previous waypoint to the next
waypoint such that the vehicle accounts for disturbances to maintain a track between
structed in this manner is shown in Figure 4-7, where fo and f, have equal weights,
wo = wV.
103
0
Ownship
State Vector 0~ 100
Path to Waypoint Utility
Figure 4-7: A waypoint following objective function, where the next waypoint is di-
rectly north of ownship. The utility values are produced through a linear combination
of separate utility functions over the heading and speed domains.
104
Varying the relative weight of the heading and speed objective functions effectively
causes the vehicle to favor deviations in speed over changes in heading, or vice versa,
depending on the relative weight. This behavior is observed in the case where the
waypoint objective function is combined with other objective functions such that the
waypoint-defined heading and speed do not maximize the overall objective function.
produce an objective function that results in collision avoidance while adhering to the
COLREGS protocol constraints. Returning to the general approach for COLREGS-
compliant motion planning presented in Section 2.2.1, the first problem is to identify
whether or not and what COLREGS scenario applies and the second problem is to
apply the applicable COLREGS constraints to the motion planning process. A general
behavior establishes an objective function f(a) using ownship state x,8 and contact
state xn obtained from a world model, and passes f(a) and a weight w to a solver
that combines f with other objective functions and solves to obtain the optimal
maneuvering action a*. This process is performed at every time step, where the step
105
size is user-defined, but is short enough to react to changes in ownship and contact
state. The behavior is initialized with ownship and contact state at the beginning of
an encounter and terminates at the completion of an encounter, as defined in Section
4.1, where rencounter = fpwt and rcompiete is user-defined.
Under COLREGS, the problem of setting a collision avoidance mode is the problem
of determining if a risk of collision exists with a contact and, if it does, determining
which COLREGS scenario applies. The term mode describes the applicable COL-
REGS scenario. Each mode may be associated with a submode used to describe any
additional scenario-dependent information associated with a particular mode, such as
whether not an overtaking vessel is overtaking to port or starboard. As discussed
in Section 2.2, mode determination is not always a straight forward problem due to
flexibility in the COLREGS language. However, given the mathematical relationships
between vehicle states presented in Section 4.1, it is possible to develop a rule-based
algorithm for setting the COLREGS mode. Algorithm 2 from [17] is a procedure
for systematically making the COLREGS rule determination based on vehicle states,
which results in establishing one of the following modes, each associated with the
given COLREGS rule:
106
World Model
Vehicle States
Tr
I Solvera
107
* StandOnX, or StandOn Crossing Mode (Rule 15,17)
" Null
The CPA mode is reserved for encounters in which a particular COLREGS scenario
does not apply but it is still desirable to exhibit some collision avoidance behavior
due to the range of the contact or some other aspect of the encounter. The Null mode
applies when the contact does not represent any risk to ownship, i.e. it is outside
some pre-specified range.
108
mode, i.e. in most cases relative vehicle geometry will not be the same at the be-
ginning of some collision avoidance encounter as it is when risk of collision no longer
exists. These considerations make it necessary to meet some mode exit criteria and
transition to Null or CPA before allowing the possibility of entering another mode.
Each mode check generally consists of the following processes, where [17] provides
specific instantiations for each COLREGS mode:
" A check for satisfaction of absolute release criteria from all COLREGS modes,
such as r" > rpt.
" A check for release from the particular COLREGS mode based on current vehicle
states. If release criteria are note met, then the mode is maintained, otherwise
mode is changed to CPA mode.
* A check for entry into the COLREGS mode if not currently in the mode.
" Setting or updating any submodes relevant to the mode if staying in or entering
the mode. Possible submodes for each mode are given by Equation (4.21).
0, otherwise
For the modes associated with a give-way vessel, Give WayO T and Give WayX, the
submodes are associated with the relative direction that the vehicle intends to pass
the stand-on vehicle. For the modes associated with a stand-on vessel, StandOnX
and StandOnOT, the submodes are associated with the relative direction that the
stand-on vessel assesses the give-way vessel will pass it. The InExtremis submode
provides for the COLREGS defined scenario in which the stand-on vessel assesses
109
that the give-way vessel has not taken sufficient action to avoid collision and the
stand-on vessel must take some action itself. There are no submodes associated with
with the COLREGS mode to optimization over a collision avoidance objective func-
tion. Combining the collision avoidance behavior and constraint application concepts
presented in Section 4.3, Benjamin [17] presents objective function formulations for
each of the COLREGS modes and submodes. Formulations for stand-on vessels in the
overtaking and crossing scenarios are discussed further here due to development of
method.
The assumed constraints for a vehicle with stand-on responsibility are to maintain
the course and speed of the vehicle at the time the encounter begins. Application of
this constraint to an objective function means high utility at the original heading and
speed and low utility at all other headings and speeds. Since it is not particularly im-
portant to maintain the exact original heading and speed, but something within some
small deviation, there may be some non-zero or intermediate utility value for head-
ings and speeds close to the original values. Since variations over heading and speed
of separate functions fo(O) and f,(v) over candidate headings 6 and speeds v, respec-
tively [17]. The overall objective function is given by Equation (4.22). Parametric
definitions of fo(6) and f,(v) are given by Equations (4.23) and (4.25), respectively,
where the definition for A6 is given in Equation (4.24) and a graphical definition of
the parameters used in the equations are given in Figures 4-9 and 4-10.
110
96,min + (gelow - go,min) ('0+180) -180 < AO < -A01ow
0 otherwise
An objective function that accounts for the stand-on constraints can then be built
by specifying appropriate values for the parameters identified in Figures 4-9 and 4-
10, where appropriately low values for all parameters except ge,med and gv,med ensure
favorability of 6 onig and Voig, or similar headings and speeds, over all other headings
and speeds. An example of a complete stand-on objective function constructed in
this manner is shown in Figure 4-11.
111
J
9
6
90,med
gema
96)ow - - - -- ..--.-
8i
-
I
I
i8
96,min
0 360
A01ow khirgh
g _______________________
9v,vned ~- -- mm~ -
-
_______________ S
-r
-
- -I.--------------------------------------------------------
9vJhnn
a I
gv~yln C I
9vhigh - --
112
Figure 4-11: A stand-on vehicle objective function, where the stand-on vehicle is
circled in yellow in the figure on the left. The peak utility is the dark red point, at
the values of stand-on vehicle heading and speed at the beginning of the encounter.
The relative size of the red region to the blue region illustrates how constraining this
formulation of stand-on responsibility is on vehicle motion.
ning
113
114
Chapter 5
Intent-Aware COLREGS-Compliant
Collision Avoidance
Since the concept of motion intent is being applied towards prediction of or assump-
tion on future vehicle motion for collision avoidance, it is important that this appli-
cation make appropriate considerations for the collision avoidance protocol, namely
COLREGS, and the ability to accurately predict vehicle motion using intent infor-
mation. This chapter seeks to account for these considerations in the introduction
We consider the general case, as shown in Figure 3-1, where intent information is
available to the mobile robotic system through processing of sensory data, which may
in work on intent inference, it may be the case that state estimation, intent inference,
and motion planning are combined into a single model that accounts for the related
or audible signals, the following discussion separates intent inference from motion
planning and state estimation, though this separation is not necessarily mandatory.
115
bilities, many of which are captured in the development of Autonomous Flight Rules
for the aviation domain, as documented in [109]:
" Intent Inference or Communication: an agent may obtain the intent of another
agent either by direct communication of intent information or through inference
of intent given observable data. In the absence of intent communication or
with updates from communicated intent information, an intent inference may
be performed on the basis of some intent model, e.g. a model that expresses
the likelihood that a vehicle intends to comply with COLREGS based on past
vehicle states [91].
116
planner that provides input necessary to develop motion plans, i.e. the compo-
nents of the motion planning problem presented in Section 2.1. Through the
communicate its intent to other agent's, either by broadcasting the intent in-
the receiving agent to understand and use the intent, the agent's must have a
common mode of communication and have mutual understanding of the intent
of transmitting vehicle intent, then there must exist some process by which
the received intent is "verified," such that "invalid" intent information is not
erroneously used by the motion planning system. This verification may involve
some examination of the communicated intent in light of past and current vehicle
requirements or in order to assure that vehicle actions will not create an unsafe
situation [54,109]. Similar to the case of intent communication, the process and
between vehicles.
117
Taking the above considerations into account, along with concepts of intent devel-
oped in Chapter 3 and COLREGS-compliant motion planning developed in Chapter
4, the following sections introduce concepts and algorithms for a multi-objective op-
intent-aware behavior is developed under the assumption that any processing of in-
In the context of the considerations presented at the beginning of this chapter and the
general mobile robotic system in Figure 3-1, Figure 5-1 depicts a general intent-aware
system by which an intent-aware marine vehicle may use communicated or inferred
the world model contains all information available to system in the form of sensed
and processed data, including ownship and contact states and intent information.
assess intent validity based on current and past vehicle states and may use informa-
tion in the world model and communicated intent information to update some inferred
intent. The concept of intent validity is discussed further in Section 5.6. Given that
the intent is understood and valid, the processed intent information is used to update
the world model.
intent that provides information required for developing motion plans, as discussed
in Section 3.2. This intent information may also be stored in the world model for use
by other processes.
118
model to develop a motion plan and determine the next maneuvering action. As
instances of intent information, the developed motion plan and the COLREGS mode
are updated in the world model. The next action determined by the motion planner
tion in the world model to transmit or broadcast to other agents by passing appro-
are handled collectively by behaviors for individual contacts and mission objectives.
These behaviors develop objective functions that are passed to a solver that solves
for a single optimal action to pass to the controller.
Since COLREGS mode and submode, as presented in Section 4.4.1, directly pro-
vide collision avoidance mode intent information discussed in Section 3.5.2, the com-
ponents of the collision avoidance mode intent Iavid = (Imode, Iubmode) are simply
equated to those of the mode and associated submodes, such that Imode = mode and
IsthmoIde = submode where mode and submode are determined through a COLREGS
mode determination algorithm and each has the same domain as given in Section
Trajectory intent Itaj is further refined from the definition in Section 3.5.2 to simplify
intepretation between vehicles. Itraj = (T, an) where T = (x1, ... , xn) is a discrete
trajectory consisting of a series of positions and times xi = (xi, yi, ti) where it is
assumed planned heading is the direction from one position to the next and speed
119
Communications Intent Intent Communications
Receiver Procosn/ I World Model coimuniatar Transmitter
-
OOLREGS Modk
120
is that required to achieve the time between positions. The vector an = (6, vn)
is the intended speed and heading following the final position. These simplifying
assumptions correspond to a vehicle that can instantaneously change heading and
speed, which is both used and evaluated for initial development. It is also noted that
simpler data structures for intent communication may be desired over more complex
in Figure 5-2:
" Non-kinematic trajectory intent: a set of waypoints for which vehicle kinematics
or dynamics are not specifically accounted and the associated time for each
" Kinematic trajectory intent: a trajectory generated through the use of a vehi-
cle model to account for low-level control and vehicle kinematics or dynamics.
accounts for vehicle kinematics verses one that does not, this thesis explores
simulating ownship motion given the current state, maneuvering action, and se-
waypoint once the forward simulated vehicle position is within the user-specified
waypoint "capture" radius. The desired heading applied to the model is the line-
of-sight heading to the next waypoint and the desired speed is the user-defined
waypoint traversal speed. The specific model is identified in Section 6.1.
In either case, the trajectory intent may be generated through processing of some
nation algorithm that accounts for contact intent information given in the form of
121
(a) Non-Kinematic Trajectory Intent (b) Kinematic Trajectory Intent
While there are many potential uses for COLREGS mode intent information, includ-
ing alerting ownship to a contact mode that is inconsistent with ownship mode, such
as a contact in Null mode while ownship is in HeadOn mode, the focus of this thesis
relaxation of COLREGS constraints. In most scenarios, the give-way vessel has the
flexibility to maneuver as needed around the stand-on vessel, with some specific cases
maintain the course and speed held at the beginning of the encounter, making stand-
on scenarios a priority for constraint evaluation. Since mode intent relates directly
the first attempt to apply COLREGS mode intent is to simply consider the intent
122
are added to the submode domain, and for the StandOnOT mode, the submodes
resulting in the update to Equation (4.21) given by Equation (5.1). When a vehicle
assesses that it is in one of the stand-on modes and it receives a communication from
a contact that it is in the associated give-way mode with a particular submode, then
0, otherwise
(5.1)
It is possible that mode intent information could be considered within the previ-
ously defined Bow, Stern, Port, or Starboard submodes, however the new submodes
provide for differentiating between the case where a contact has directly communi-
cated the mode intent verses some inference performed by ownship. This differen-
tiation provides for subsequent identification of appropriate COLREGS constraints.
Future work, where mode intent is inferred with some confidence level, where this
confidence level might be very high if the inference is updated with communicated
intent information, and is then provided to the mode determination and motion plan-
ning algorithms could consider the stand-on submode in light of the confidence level
Additionally, under the assumption that communicated intent represents the true
intent of the give-way vessel, the stand-on vessel accepts the communicated intent
123
is noted that COLREGS sound signaling rules provide for the ability or requirement
for the stand-on vessel to acknowledge give-way vessel intent. In this work the stand-
on acknowledgement is assumed.
In order to establish a measure of risk of collision over two discrete trajectory intents,
risk,, the concept of CPA range over ownship and contact trajectories, rcpa,r, from
Section 4.2 is applied to the trajectory intents. In order to apply Equation (4.8),
ownship and contact trajectories r are re-sampled through linear interpolation to
obtain equally sampled discrete trajectories with the same length k, as described in
Section 3.4. If the original final point in a trajectory of length n is at a time t" < tk,
an= (O,Onv) of the original trajectory is used to determine any trajectory points at
times greater than tn.
Risk of collision over intended trajectories, riskr, is then defined as in the case
of the linear CPA calculation, by applying some minimum threshold icpa,r to the
computed value of repa, as given by Equation (5.2). Equation (5.2) establishes a
hysteresis such that frequent switching between a state of risk and no risk does not
124
occur as trajectories are updated and the value of r,, is close to that of iej,,. Under
this formulation, risk is initialized such that risk,,O = false. This idea of trajectory-
based risk of collision is analogous to trajectory-based conflicts in the aviation domain
[109].
riskr,i otherwise
A Trajectory Mode
Given that a risk of collision does or does not exist over the intended contact trajecto-
ries may not be enough information to assure that some scenario will not develop that
ultimately places the vehicles at a risk of collision. Figure 5-3 shows the difference
between a state-only and intent-only risk of collision, as derived from the concept
implies that re,,a or some other risk measure, based only on current vehicle states is
below some threshold value, while the intent-only risk meets the definition of risk over
sure. Given these possibilities, collision avoidance mode determination must consider
e A state-only risk of collision exists, which one expects to trigger some COLREGS-
trajectory based risk of collision and continue to execute the planned trajecto-
ries. However if one vehicle chooses to follow the COLREGS maneuver and one
follows its originally planned trajectory then an unsafe encounter may result.
125
CD, Ownshlp N
CD Contact
Ct
Discrete Trajectory
State Vector
4
1'
7-
7-
I-u
0Th
+
TO
126
the equivalent to human operators agreeing to a particular passage arrangement
" An intent-only risk of collision exists, which, if a COLREGS rule does not cur-
trajectories such that the intent-only risk is eliminated. In the case of replan-
ning, if both vehicles simultaneously replan then there is no assurance that the
resulting plans will not still pose some risk of collision unless this replanning
for any state-based risk that arises in order to avoid a scenario in which the
" Both state-based and intent-based conflicts exist, in which case considerations
from both of the above two scenarios may be applied to arrive at a solution,
where at least one vehicle must maneuver or replan to avoid a collision and
both should apply COLREGS rules appropriately based on the level of available
ance requires establishment of some method to determine which vehicles will execute
collision avoidance maneuvers and under which conditions they will do so. In the case
where one vehicle has maneuvering priority' over another, as directed by COLREGS
Rule 18, such as in a sailing vessel over a power driven vessel, it is expected that
the vehicle with priority will maneuver according to its intended trajectory while the
'The term priority is taken from the aviation intent literature [109], in which ownship is required
to take collision avoidance action with respect to a contact if assigned a lower priority than the
contact. Priority can be likened to the term responsibility used in COLREGS Rule 18, in which
vessels axe assigned collision avoidance responsibility according to classification. Priority can also be
assigned to vessels of the same type if associated with the collision avoidance responsibility dictated
by Rules 13 through 17, as suggested with the similar aviation rules in [1091.
127
lower priority vehicle replans around the priority vehicle's trajectory. Where there
is not a specific Rule 18-type priority specification, it is still possible for vehicles to
In this case, priority can be established between vehicles through communication and
is introduced to handle the scenario in which two vehicles agree whether or not a
trajectory-based risk of collision exists and, if one does exist, they agree how the risk
will be resolved.
COLREGs constraints must be applied such that the resulting vehicle motion is
associated with the knowledge of intent through the communicated COLREGS mode
intent or discrete trajectory intent.
Constraints for COLREGS-compliant motion planning with the use of mode intent are
considered for the stand-on in overtaking and stand-on in crossing scenarios. Rule
17 requires that these stand-on vessels maintain course and speed so long as it is
apparent that action by the give-way vessel will result in safe passage.
If, however, conditions arise such that the stand-on vessel is the give-way ves-
sel in an interaction with another vehicle or conditions exist such that a particular
maneuver results in no risk of collision given the intended actions of the give-way
128
vessel, then strict adherence to the constraint to maintain course and speed may
not only be inefficient but also unsafe. In this context, relaxations to the stand-on
constraints are considered in light of give-way vehicle mode intent so as to obtain
COLREGS compliance with flexibility for maneuvering out of highly inefficient and
unsafe scenarios.
The two possibilities for contact mode intent in a stand-on in overtaking scenario
are (Give WayOT, Port) and (Give WayOT, Starboard). From this intent information,
it can be inferred that the contact will follow a trajectory composed of states that
fall on either the port side or starboard side of ownship for some unspecified time
horizon over which port and starboard remains in reference to ownship heading at
the beginning of the encounter. This scenario is illustrated in Figure 5-4 for the
give-way vessel overtaking the stand-on vessel on the port side.
Given the intent information, therefore, ownship, if necessary, could alter course
to the opposite side of intended contact passage which, assuming the contact carries
out its intended actions, will open range to the contact and thereby reduce the risk of
collision. This would allow the stand-on vessel to simultaneously meet the objective
for which it desires to maneuver. Since such action is being considered in the context
necessary, as allowed by COLREGS, to deviate for some other purpose such as the
presence of another vehicle [28]. Therefore, the proposed constraint relaxation is not
to simply allow any maneuver to the opposite side of overtaking vessel passage, but
to set the utility of these actions such that they may be preferred over maneuvers to
the side of overtaking vehicle passage but not over actions to maintain course and
speed. Additionally, the constraint on speed can be maintained while altering the
constraint on heading.
129
WI
C: Give-Way Vessel
Figure 5-4: Maneuvering region for an overtaking vessel with intent to maneuver to
the port of the stand-on vessel.
130
Stand-On in Crossing Constraint Relaxation
The two possibilities for contact mode intent in a stand-on in crossing scenario are
(Give WayX, Bow) and (Give WayX, Stern). From this intent information, it can be
inferred that the contact will follow a trajectory composed of states that fall alongside
and in front of or behind ownship. Unlike in the overtaking scenario where the give-
way vessel has a speed greater than ownship and maneuvers to a particular side can
vessel once it has crossed the bow or stern such that it could be guaranteed that any
change in course will not result in additional risk of collision. Additionally, maneuvers
to port by a stand-on vessel in a crossing scenario are explicitly prohibited. For this
reason, consideration is initially given only to modification of speed constraints in
Given the information that the give-way vessel intends to pass in front of or behind
will open range to the contact and thereby reduce the risk of collision. This would
allow the stand-on vessel to simultaneously meet the objective for which it desires
to change speed. Like the stand-on in overtaking case, since such action is being
considered in the context of stand-on responsibility, it is still desirable to maintain
speed unless necessary to deviate for some other purpose. Therefore, the proposed
constraint modification is to set the utility of speeds above or below original speed
such that they may be preferred over speed reductions or increases, respectively, but
Constraints for COLREGS-compliant motion planning with the use of trajectory in-
tent are considered for the case in which intent-based risk of collision exists and the
case in which it does not exist. In either case, a failure to establish agreement be-
131
trajectory-based collision avoidance should result in a default to application of the
appropriate state-based COLREGS constraints.
In the case where no intent-based risk of collision exists and ownship and contact agree
that no risk exists such that each can safely execute originally planned maneuvers, as
described in Section 5.3.2, then, assuming that each vehicle carries out the intended
maneuvers, the maneuvers may be safely executed without applying additional motion
constraints. An example of this scenario is the state-only risk of collision encounter
in Figure 5-3.
In the event that the vehicles unintentionally end up in an unsafe scenario with
respect to one another, such as closing within some pre-specified range, then, as a
safety feature, the vehicles can apply basic collision avoidance constraints as discussed
in Section 4.3.2.
132
World Model
I Solver -
*
Figure 5-5: A behavior model for producing a multi-objective optimization objective
function that accounts for intent in addition to vehicle states, updated from Figure
4-8.
5.5.
objective function for COLREGS-compliance given ownship and contact state data.
Figure 5-5 updates Figure 4-8 to represent an intent-aware COLREGS behavior that
accounts for ownship and vehicle intent information in addition to state data.
In order to account for the use of intent information, the top level behavior
133
onRunState 0 procedure in Algorithm 1 is updated by Algorithm 3, which adds
the process IntentData() procedure to update and perform any necessary intent
pre-processing prior to application to COLREGS mode determination and objective
function establishment. Lines 5 and 7 are updated to incorporate intent data in the
Algorithm 4 is proposed for processing intent data passed to the intent-aware COL-
priority. Algorithm 4 assumes that any intent data passed to the behavior repre-
sents the current intentions of ownship and the contact and that this data is available
upon behavior initialization; considerations for when this assumption is not valid are
Lines 2 and 3 of Algorithm 4 simply update stored values for Iavoid,cn, Itraj,os,
and Itra,cn and, in the case where a contact has communicated information regard-
ing evaluation of risk or priority, this information is updated as well. The variables
priority" and priority" are used to identify the contact's perceived priorities
for itself and ownship, where the subscript indicates the vehicle that has evaluated
the priority and the superscript indicates the vehicle to which the priority is as-
signed. Similarly, ownship evaluated priority for ownship and the contact are given
134
Algorithm 4 Process Ownship and Contact Intent Data
1: procedure PROCESSINTENTDATA()
2: update ownship intent data
3: update contact intent data
4: if 'traj,os available or Itrajcn available then
5: evaluate risk, > Equation (5.2)
6: setPriorityO > Algorithm 5
7: communicate perceived risk and priority > publish to external process
8: evaluate agree > Equation (5.3)
9: end if
10: end procedure
the trajectory intent. Line 5 evaluates risk, as discussed in Section 5.3.2 and given
by Equation (5.2).
sion avoidance as discussed in Section 5.3.2. In the case of different vessel types, e.g.
power verses sailing, this can be achieved through priority determination through
recognition of the different vessel types. In the case of power-driven vessels, how-
established so vehicles can communicate this responsibility to one another and sub-
uation of give-way and stand-on responsibility based on the entry criteria for the
three basic COLREGS scenarios, which is similar to a method for aircraft priority
{giveway, standon, no - risk}2 based on meeting the entry criteria for each of the
COLREGS modes where the mode entry criteria are checked based on the criteria
2
The priority values of giveway, standon, and no - risk are simply used to refer to agreed
upon collision avoidance responsibility between vehicles and are not necessarily always reflective of
the vehicle's state-based COLREGS responsibility.
135
provided in Section 4.1. Algorithms for each mode entry check, which are based on
applicable portions of mode check procedures given in [17], are provided in Appendix
C. The case where there is a trajectory-based risk of collision but no state-based risk,
which meets the condition of line 21 in Algorithm 5, is handled by assigning giveway
priority to both vehicles. The case of no perceived trajectory-based risk, which meets
the condition of line 25 in Algorithm 5 is handled by the no - risk value.
It is noted that the case where giveway responsibility is assigned to both vehicles
can result in an encounter in which vehicles make conflicting maneuvers as both try
to take action to avoid a collision. Such a case is observed in Section 6.3, where
simply applying the proposed objective functions for use of trajectory intent without
application of additional protocol constraints can result in vehicles turning towards
one another. Potential solutions to this problem include applying the basic protocol
constraints, such as the requirement to turn to starboard in a head-on situation, or
using a cooperative planning algorithm in which each vehicle must take some collision
avoidance responsibility and the replanning is executed in some sequence so as not
develop conflicting maneuvers. Such an approach is suggested for autonomous flight
rules, as discussed in [109].
136
Algorithm 5 Setting the Trajectory-Based Collision Avoidance Priority
1: procedure SETPRIORITY() > Called from within processIntentData()
2: if (priorityset) and -,(priority" == no - risk) then
3: return > do not change priority unless risk now exists
4: end if
5: if (risk,) then > trajectory-based risk exists
6: if checkOvertakingEntryo then > Appendix C, Algorithm 8
7: priority' +- giveway
8: priority' <- standon
9: else if checkHeadOnEntry() then > Appendix C, Algorithm 9
10: priorityg - giveway
11: priority' +- giveway
12: else if checkCrossingGivewayEntryo then > Appendix C, Algorithm 10
13: priority" +- giveway
14: priorityc +- standon
15: else if checkStandOnEntry() then > Appendix C, Algorithm 11
16: priority" +- standon
17: priority' +- giveway
18: else if checkStandOnOTEntry() the a > Appendix C, Algorithm 12
19: priority" +- standon
20: priorityg +- giveway
21: else > entry criteria n ot met for any COLREGS scenario
22: prioritys +- giveway
23: priority' <- giveway
24: end if
25: else trajectory-based risk does not exist
26: priority' +- no - risk
27: priority' <- no - risk
28: end if
29: priority set +- true
30: end procedure
137
tions, risk assessment, and priority assessment of the contact. Using the values of the
between vehicles.
false otherwise
This section presents an algorithm for performing collision avoidance mode checking
using collision avoidance mode intent or trajectory intent in addition to ownship and
Section 5.3.2. Next, the mode checking procedure for each COLREGS mode, as
Additionally the mode checks for the StandOnX and StandOnOT modes are up-
dated to allow for the "communicated" submodes introduced in Section 5.3.1. This is
dures in [17 to simply set the StandOnX or StandOnOT submode to the respective
"communicated" submode if the contact has communicated that it is in the associ-
ated give-way mode and submode. For example, if Iavoid,cn = (Give WayOT, Port) and
ownship has evaluated that it is in the StandOnOTmode, then ownship submode will
be set to CommunicatedPort.
check occurs in line 18 after all COLREGS mode checks (lines 2 through 16) and prior
to the CPA mode check (line 21) such that if the agreement release criteria added
138
Algorithm 6 Determining the COLREGS Major Mode - Updated
1: procedure SETAVOIDMODE() > Executed each iteration of the behavior
2: if mode E {Null, CPA, GiveWayOT} then
3: mode +- checkModeGiveWayOT(
4: end if
5: if mode E {Null, CPA, HeadOn} then
6: mode <- checkModeHeadOn()
7: end if
8: if mode E {Null, CPA, StandOnOT} then
9: mode +- checkModeStandOnOTO
10: end if
11: if mode E {Null, CPA, GiveWayX} then
12: mode <- checkModeGiveWayXO
13: end if
14: if mode E {Null, CPA, StandOnX} then
15: mode +- checkModeStandOnXO
16: end if
17: if mode E {Null, CPA, TrajectoryMode} then
18: mode +- checkModeTrajectory() > Trajectory mode check - Algorithm 7
19: end if
20: if mode E {Null, CPA} then
21: mode +- checkModeCPA()
22: end if
23: end procedure
139
occur. Additionally, if release from TrajectoryMode occurs then CPA mode is checked
as in release from any other mode.
The TrajectoryMode check given in Algorithm 7 follows the general mode check
pattern discussed in Section 4.4.1 using the same r" and f release criteria (lines 2 and
6) as the other mode checks in [17]. The logic for the TrajectoryMode release check be-
ginning on line 5 and TrajectoryMode entry check beginning on line 14, simply require
agreement between vehicles to remain in or enter TrajectoryMode, respectively, and
return TrajectoryMode as the mode and ownship evaluated priority as the submode.
Considering the processIntentData() procedure in Algorithm 4 and the
checkModeTrajectory() procedure in Algorithm 7 together, it is observed that, on
each iteration of the intent-aware behavior, the appropriate non-trajectory COLREGS
mode is entered or maintained unless all of the following conditions hold:
9 Each vehicle broadcasts updated projected trajectories, priorities, and risk as-
sessments on each iteration of the behavior.
140
" Trajectory intent information, which is assumed to be valid, is available.
By requiring these conditions to hold for entry into TrajectoryMode and applying
7 ensure that the behavior will default to applying state-based COLREGS constraints
This section proposes objective function formulations based on the constraint relax-
In the case where an overtaking contact has communicated that it intends to overtake
starboard by increasing the utility of starboard maneuvers. This naive relaxation can
be achieved using appropriate utility values to define the stand-on objective function
same as in a normal stand-on objective function while the value of 9o,high is set to some
value go,high,intent > j,iow 0 and go,max is set to some value go,max,intent > 90,min > 0.
This parameterization gives fo(6) in the form shown in Figure 5-6. An example of
6.3.1, revealed that simply allowing the overtaking vessel to maneuver to starboard
without account for overtaking vessel state or ownship kinematics resulted in unsafe
maneuvers towards the overtaking vessel in the event of another objective with high
141
go,med
9O,highintent
I
96,max,intent -44--
ge,371ar
g0,10
go,rn
0)
VLIJzL
I
t
I
I
I
360
1)
, I
AO,09~, ~JAUJh
Figure 5-6: A notional heading objective function for StandOnOT mode and
CommunicatedPortsubmode, where Oig is the vehicle heading at the beginning of
the encounter.
which it is combined with the basic collision avoidance objective function according
to Equation (5.4), where gavod (, v) is defined by Equation (4.17). The value of gmode
is established such that it is less than the maximum possible value of wefo + wf, and
maintaining course and speed is still favored. An example of this objective function
is shown in Figure 5-8.
we fe + w f, AO 6 [-180,0]
In the case where an overtaking contact has communicated that it intends to overtake
on the starboard side, the proposed constraint relaxation is to permit a maneuver
142
Figure 5-7: Overall objective function for StandOnOTmode and CommunicatedPort
submode, where the stand-on vehicle is circled in yellow in the figure on the left.
This initial, naive objective function formulation increases the utility of all maneu-
vers away from the side on which the overtaking vessel has communicated that it is
passing (starboard in this case). The increased utility is with respect to the baseline
formulation shown in Figure 4-11, which is maintained on the side of passage (port
in this case). Peak utility is maintained at the values of stand-on vehicle heading and
speed at the beginning of the encounter.
143
-1'
Figure 5-8: Modified overall objective function for StandOnOT mode and
CommunicatedPortsubmode, where the stand-on vehicle is circled in yellow in the
figure on the left. The CPA-based collision avoidance objective function is employed
in conjunction with the relaxation of stand-on constraints to allow the stand-on vessel
to safely maneuver to starboard. The peak heading and speed utility value is still
maintained at original heading and speed, as shown by the pink dot.
144
to port by increasing the utility of port maneuvers. This naive formulation can
be achieved in the same manner as in the CommunicatedPort case, except that the
objective function while the value of golow is set to some value gO,iow,intent > gO,high > 0
and gO,min is set to some value 9,min,intent > go,max > 0, giving fo(6) as a mirror of
In the case where a give-way crossing contact has communicated that it intends to
speed by increasing the utility of speeds greater than the speed at the beginning of
the encounter voig. Similar to the overtaking case, this relaxation can be achieved
using appropriate utility values to define the stand-on objective function according to
equations (4.22) through (4.25). Specifically, in Equation (4.25), the values of gv,min,
9vlow, gv,med, vi 0w, and Vhigh are kept the same as in a normal stand-on objective
function while the value of 9v,high is set to some value g9,high,intent > gvlow ;> 0 and
gv,max is set to some value gv,max,intent > gv,min > 0, giving f (v) in the form shown
in Figure 5-9. An example of the complete objective function can be seen in Figure
5-13.
In the case where a give-way crossing contact has communicated that it intends to
pass ahead of ownship, the proposed constraint relaxation is to permit a reduction in
145
.q
9v,high~fltent------7
7.. T----TM K
.......-- 1:
t
t'dmin w orig hh VI 1 01
Figure 5-9: A notional speed objective function for StandOnX mode and
CommunicatedStern submode, where vorig is the vehicle speed at the beginning of
the encounter.
speed by increasing the utility of speeds less than the speed at the beginning of the
encounter Vorig. This can be achieved in the same manner as in the CommunicatedBow
case, except that the values of g,max, gv,high, gv,med, Vto7, Vhigh are kept the same as
in a normal stand-on objective function while the value of gY,Io is set to some value
9vtow,intent > gv,high 0 and go,min is set to some value gv,min,intent > g>,ma ! 0,
giving fv(v) in the form shown in Figure 5-10. An example of the complete objective
function can be seen in Figure 5-13.
In the case where ownship and contact agree to navigate with respect to shared
trajectory intent and ownship has assumed the responsibility to avoid the contact
by taking a giveway priority, the proposed constraint is a basic collision avoidance
constraint where ownship must maintain some minimum range to the contact over
its trajectory. As in the basic collision avoidance behavior presented in Section 4.3.2,
this can be achieved through application of the utility function given by Equation
(4.17) except that ra, is replaced by rep,., evaluated as discussed in Section 5.3.2 and
146
W-- tWOM--- - - - - - - - - -
-
g9,jlwinterat
I S
9vhi h
yjv, max I -S I 1~-V
1
dmin Vlow lyoriq thigh Vdnax
Figure 5-10: A notional speed objective function for StandOnX mode and
CommunicatedBow submode, where voi is the vehicle speed at the beginning of
the encounter.
In the case where ownship and contact agree to navigate with respect to shared
trajectory intent and ownship has assumed the responsibility to traverse the intended
trajectory by taking the standon or no-risk priority, a collision avoidance objective
function is only produced for the case in which it becomes apparent that traversing
the trajectory is no longer safe based on contact state or an update to contact in-
tended trajectory. The case of evaluated risk based on updated trajectory intent is
handled by provisions in Algorithm 5 as discussed in Section 5.5.1. The case in which
vehicles close within some range without identifying a trajectory-based risk is handled
by applying the trajectory CPA-based objective function described in the previous
section; i.e. the objective function is applied if r' becomes less than some predefined
threshold, r,,inextremis, as in the InExtremis submode in the StandOnOT or standon
modes.
147
Ownship
Contact
Discrete Trajectory 0 100
Utility
Figure 5-11: A trajectory-based objective function, in which rpa over the contact
trajectory is used to specify utility values instead of rpa based only on current contact
state.
148
5.6 Intent Verification
As discussed at the beginning of this chapter, in the event that it cannot be assumed
that a vehicle will carry out an intent as communicated or when there is uncertainty
associated with the manner in which an intent will be translated to a trajectory, it
may be necessary for ownship to validate intent prior to use in the motion planning
process. A valid intent is defined as the case in which the vehicle actually executes
or is executing a communicated intent.
Under the assumption that vehicles will execute intended maneuvers exactly as
communicated to other vehicles, this thesis does not make further developments on
the concept of intent verification. However, as shown in Chapter 6, unless vehicles are
somehow restricted to maneuvering exactly in accordance with communicated intent,
there must be further consideration for vehicle deviation from communicated intent.
Such consideration may involve the development of measures that can be used to
assess the likelihood that a contact is following its communicated intent, which may
take advantage of previously developed intent inference mechanisms.
This chapter applied the concept of intent developed in Chapter 3 to the problem
of COLREGS-compliant collision avoidance presented in Chapter 4, including con-
sideration of intent in the processes of risk of collision assessment, COLREGS mode
determination, and collision avoidance objective function formulation. Figures 5-12
through 5-15 illustrate the functionality of the proposed intent-aware COLREGS colli-
sion avoidance behavior. Figure 5-12 illustrates the application of COLREGS collision
avoidance mode intent in an overtaking scenario. The figures shows the relaxation of
COLREGS constraints for a stand-on vessel, where the give-way vessel communicates
that it is going to overtake on the port or starboard side. Figure 5-13 illustrates the
application of mode intent in a crossing scenario. The figure shows the relaxation of
COLREGS constraints for a stand-on vessel, where the give-way vessel communicates
149
Communicates
Port
Communicates
Starboard
Stand-on in Overtaking I
that it is going to pass ahead or astern of the stand-on vessel. Finally, Figure 5-15
illustrates possible outcomes of discrete applying trajectory intent information to the
orities, and then choose the appropriate collision avoidance objective function based
150
Communcates
Bow I
Com[icates
Stand-on in Crossing
Figure 5-13: Example of mode intent application in a crossing encounter. On the left,
the give-way vessel has not decided or shared mode intent and therefore the stand-on
vessel (circled in yellow) applies the highly-constraining stand-on objective function.
On the right, the give-way vessel has communicated its intent to pass to ahead or
astern of the stand-on vessel, resulting in the stand-on constraint relaxation to allow
speed reduction or increase, respectively.
0 Vhide l
Vehide 2
State Vector
151
Generate Gvenerai*
Trajectory Trajectory
Communicate
No Constraint No Constraint
152
Chapter 6
Experimental Evaluation
is conducted in two stages, each consisting of both simulations and on-water demon-
strations. First, the impact of intent information on an instance of each of the three
to that of a baseline system that does not use intent information. Second, a sim-
analysis seeks to identify scenarios that illuminate the concept of intent information
6.1.1 Software
The behavior and general intent-aware system introduced in Chapter 5 were developed
in the MOOS-IvP software system. The system employs a payload autonomy concept
in which the autonomy software system is run on a "payload" computer that sends
vehicle actuator commands to and receives sensor data from a vehicle "front-seat"
153
computer [201. The front-seat computer directly interfaces with the vehicle propulsion
system and sensors. Additionally, a shoreside control station interfaces with the
Vehicle simulations were performed using the MOOS-IvP uSimMarine model as de-
vehicle front-seat computer, as described in Section A.1, such that the application
accepts vehicle actuator commands and provides vehicle state information, including
position, heading, and speed. The vehicle model is a simple kinematic model with
constraints on linear and rotational acceleration, where heading and speed are up-
dated according to rudder and thrust commands provided from a PID controller, as
described in Appendix A.
For experiments that employ kinematic trajectory intent generation, the same model
identified in the previous section, including the PID controller, is used for kinematic
trajectory intent generation by forward simulating ownship state using desired way-
point speed and desired headings to each waypoint. Desired heading inputs to the
PID controller are the line-of-sight heading from ownship current position to the next
waypoint. The next waypoint is updated once the forward simulated vehicle position
is within the waypoint capture radius defined for the experiment. By employing this
exact model of vehicle motion, the evaluation exposes benefits of accurate trajec-
154
tory communication over the case where trajectory is assumed to exactly follow the
planned waypoints.
On-water experiments were conducted on the Charles River at the MIT Marine Au-
tonomy Bay in the MIT Sailing Pavilion using the Clearpath Robotics Heron M200
and Heron M300 unmanned surface vehicle platforms [27J. The payload autonomy
faced with the vehicle front-seat computer through a TCP/IP connection. Through
this connection, the payload autonomy system provides left and right thruster com-
mands to the vehicle and receives vehicle state information from the vehicle's GPS
through 802.11 WiFi communication, where simulated vehicle AIS broadcasts are
relayed through the shoreside station and made available to all other vehicles.
The following independent variables were considered for assessing impact on COLREGS-
Section 5.5, such that intentuse E {none, mode, trajectory}. The use of the
155
to as mode intent. Additionally trajectory intent is identified as either non-
kinematic trajectory or kinematic trajectory intent, depending on the method
of trajectory generation according to Section 5.2.2.
" avd_ pwt: relative weight between collision avoidance behavior priority, avd_pwt,
and waypoint behavior priority, wpt _pwt, established by fixing waypoint weight
at wpt_pwt = 100 and varying collision avoidance weight.
" crs spd_ratio: relative weight between heading and speed objectives in the
waypoint behavior, 1, as defined by a course-speed weighting ratio.
156
Table 6.1: Parameter Values for Canonical COLREGS
Scenario Simulations
The values of crsspdratio and avdpwt provided in Table 6.2 were used
difference in safety or efficiency was observed for different avd pwt, therefore a con-
servative value is selected to ensure adequate on-water collision avoidance priority.
The value of crs_spd_ ratio was selected based on the balance between safety and
Parameter Value
avd-pwt 200
crsspdratio 50
Tables 6.3 and 6.4 provide parameters that were fixed in each of the experiments
discussed in this chapter with respect to mode and trajectory intent application, re-
collision avoidance (Table A.1) and supporting applications (Table A.2). These val-
ues are selected for initial evaluation based on observed system performance in early
experimentation and results contained in [110, 1111. Future work involving inves-
157
as discussed throughout and at the end of this chapter.
9mode 5.5.3 95
-
90,min,intent 5.5.3 95
-
ge,iow,intent 5.5.3 95
-
ge,high,intent 5.5.3 95
-
9g,max,intent 5.5.3 95
gv,min,intent 5.5.3 95
-
gv,low,intent 5.5.3 95
-
gv,high,intent 5.5.3 95
-
gv,max,intent 5.5.3 95
-
fcomms A.2.3 1.0 Hz
Intent
icpa,r 5.3.2 12 m
repa,, 5.3.2 18 m
r-r,inextremis 5.3.2 10 m
thorizon A.2.3 120 s
The primary metrics used for assessing the impact of intent on COLREGS-compliant
motion planning and execution are safety, as defined by range at CPA, rca, and
158
efficiency defined in terms of both distance traveled relative to planned distance and
time of travel relative to planned time of travel.
7
Distance efficiency, 7d, as defined by Equation (6.1) is used to measure distance
efficiency, where do is planned distance, da is actual distance traveled, and deff,o is a
constant value that establishes zero efficiency. Similarly, time efficiency, ?I, as defined
by Equation (6.2) is used to measure time efficiency, where To is defined by Equation
(6.3), Ta is the actual time of travel, and teff,O is a constant value that establishes zero
efficiency. The values di and vi are the planned distances and speeds, respectively,
over an interval i of the vehicle's motion plan where E di - do. The planned vehicle
state at the end of interval i is equal to the vehicle state at the beginning of interval
i + 1, and the speed vi is assumed to be constant over the interval i. Efficiency
measures are defined in this manner for ease of comparison between encounter sets
within each scenario, where subtraction of the planned time or distance from actual
time or distance and the constant standardization factors penalize all added distance
and time due to a collision avoidance maneuver equally.
7d = (i - d x 100 (6.1)
def f,O)
T=t -
tef f,O
)x 100 (6.2)
To d (6.3)
vi
For evaluating cases in which there is not a specific starting and stopping position
or time, a trajectory efficiency 17 is defined according to Equation (6.4), where d,
is given by Equation (6.5) and d,,o defines zero trajectory efficiency. The values of
(Xa(ti), ya(ti)) and (xo(ti), yo(ti)) are the vehicle actual and planned positions, respec-
tively, at discrete time ti where to = tencounter and tn = tcogplete. Planned positions
are based on the planned trajectory at tencounter. Trajectory efficiency measures the
degree to which the vehicle follows the originally planned trajectory using the average
159
Euclidean distance from the the planned trajectory position at each time step 11021.
= - x 100 (6.4)
d-r,O
values for deff,o, teff,o, and d,,o were selected as shown in Table 6.5. Note that the
selection of these values and the definition of efficiencies according to Equations (6.1)
and (6.2) do not restrict the efficiency values to the interval [0, 100]. For example, r/t
deff,o 200 m
tel ,o 200 s
dr,O 50 m
Besides evaluating raw rcpa, each encounter is also evaluated as to whether or not
a minimum CPA violation or severe minimum CPA violation occurred, the criteria
for which are defined in Table 6.6. The minimum CPA violation threshold is based
on the fixed value of fpa = 10 m, which is great enough to ensure that vehicles do
not collide given a 3-meter GPS accuracy, which gives a worst-case 6-meter error.
The 6 m severe minimum CPA violation threshold gives indication of whether or not
160
Table 6.6: Safety Violation Thresholds
Parameter Criterion
The summary statistics in Table 6.7 are evaluated for each experiment, as appli-
Experiments
161
Means of each metric from different experiments are compared using the Analysis
of Means method (ANOM) with a = 0.05 [79]. This method compares means to the
overall sample mean and is used to make pairwise comparisons to establish whether or
not statistically significant differences exist between metrics when intent information
is used and when it is not. While the standard ANOM test assumes the tested samples
are normally distributed, the test has been shown to be robust to non-normality [791.
Additionally, the results in this chapter are confirmed by performing an ANOM test
The following plot types are used to illustrate the results of each experiment:
" Distribution plot: An example is given in Figure 6-1. Each plot consists of a
histogram and a box plot. The histogram shows the relative number of data
points that fall in the interval covered by a particular histogram bar. The box
plot, as illustrated in Figure 6-1, includes a box that marks the interquartile
range, the median as the horizontal line in the middle of the box, the mean as
the center of the confidence diamond, whiskers that mark 1.5 times above and
below the interquartile range, and outliers as individual points that fall outside
the whiskers [89].
" ANOM plot: An example is given in Figure 6-2. The plot shows the sample
mean for the particular independent variable in relation to the upper decision
limit (UDL) and lower decision limit (LDL) for the test, as marked by the
blue box and the overall sample mean marked by the base of each black line.
A mean within the decision limits implies that the mean is not significantly
different from the overall sample mean and a mean outside the limits implies
" Bar graphs: standard bar graphs are employed to indicate the value of the
162
30
26 whisker
20 0 3rd quartile
15 shortest half
median sample value
fidence diamond
1st quartile
0 whiisker
Figure 6-1: Figure from [89] of an example distribution plot. The plot consists of a
histogram and a box plot. The histogram shows the relative number of data points
that fall in the interval covered by a particular histogram bar. The box plot box
marks the interquartile range. The sample median is the horizontal line in the middle
of the box. The mean is the center of the confidence diamond. The whiskers mark
1.5 times the interquartile range above and below the box. Outliers are shown as
individual points that fall outside the whiskers. The "shortest half" bracket is the
densest 50% of observations.
" On-water encounter plots: these plots are used to compare on-water encounters
to simulated encounters. Each encounter is a data point on a plot of efficiency
verses rqa. On-water data points are overlaid on simulated data points and
highlighted to show relative distribution.
163
EAfidsb of MU...
11 ---- UDL
10-
Avg = 8.53
LDL
A ' C
Figure 6-2: Figure from [89] of an example ANOM plot. The plot consists of a
histogram and a box plot. The plot shows the sample mean for the particular in-
dependent variable in relation to the upper decision limit (UDL) and lower decision
limit (LDL) for the test, as marked by the blue box. The base of each black line is
the overall sample mean. A mean within the decision limits, which implies that the
mean is not significantly different from the overall sample mean, is shown in green.
A mean outside the limits, which implies that it is significantly different from the
overall sample mean, is shown in red.
ment is carried out by comparing the performance, in terms of safety and efficiency, of
ing, crossing, and head-on scenarios where one of the vehicles has a mission-defined
the give-way vehicle, or one of the vehicles in the head-on case, is varied within the
limits of the relative vehicle geometry required for application of the given COLREGS
scenario as given in Section 4.1. An example for the overtaking scenario is shown in
the waypoint behavior-defined speed for the give-way vehicle. Each simulation is then
a set of random encounters for which the summary statistics listed in Table 6.7 are
generated.
164
N
---
----
-- ..... W~f
to~
CEO> - - S~
Giveway Vessel
CD Stand-on Vessel
Giveway Planned Path
Standon Planned Path
Limits of Overtaking Sector
Stand-on Passage
Figure 6-3: Randomized overtaking scenario in which the stand-on vessel has an
intended maneuver to port. The give-way vessel plans to travel from west to east
on a straight path with constant speed. On each experimental run, the overtaking
vessel speed and start position is uniformly sampled over intervals that establish an
overtaking scenario as given in Section 4.1.
165
6.3.1 Overtaking Scenario
The first scenario tested is the overtaking scenario shown in Figure 6-3 where the
give-way vessel plans to travel from west to east on a straight path with constant
speed and the stand-on vessel has a non-COLREGS-derived plan to maneuver to the
north. Approximately 3000 simulated encounters were executed for each combination
of independent variables where the specific encounter totals are given in Appendix B.
Figures 6-4 through 6-6 summarize the simulation results in terms of the distribution
of each metric by intent type. Qualitatively, use of non-kinematic trajectory intent
resulted in greatest efficiency in terms of both distance and time and the lowest
mean repa, while use of mode intent resulted in increased mean efficiencies over no
intent as well as increased mean rcpa. Table 6.8 provides a complete summary of the
overtaking encounter simulations. Figures 6-7 and 6-8 support these conclusions by
showing statistically significant differences in means using the ANOM method.
The increased efficiency under non-kinematic trajectory intent use is expected
based on the fact that collision avoidance constraints are only being placed on the
vehicles in the event that a trajectory-based risk of collision exists. The result of
lower fcpa follows from the fact that risk is based on a cpa,, value that, in this case, is
set lower than the rca value used in the collision avoidance behaviors under the none
and mode intent types. This setting allows for increased efficiency by not requiring
collision avoidance maneuvers unless fcpa,, is relatively low, but at the expense of
reduced safety in terms of rea. This result illustrates the need to find an appropriate
balance between safety and efficiency through collision avoidance behavior settings,
especially in the case where vehicles may deviate from communicated trajectories.
The improved efficiency and greater f,%, under mode intent over no intent follows
from the execution of stand-on vehicle maneuvers to port when the overtaking vessel is
passing on the starboard side. This maneuver means both that the stand-on vessel is
able to execute its planned motion and, in doing so, increases range to the overtaking
166
48-
40 40 40
88
30 30. 30
25 25 28,
20 20 20
18 15
10 10 10
5 8 5
0 0 0
Figure 6-4: Overtaking scenario range at CPA (repa) distributions by intent type.
-
90 90 90
80 80 80
70 70 70
80 60
80~
40 40
40-1
301
40
201 20' (0 1
10 10
10
0( 0 0'
Figure 6-5: Overtaking scenario distance efficiency ('?d) distributions by intent type.
-
vessel.
Appendix B contains plots of rva, qd, and qt mean and standard deviation verses
each of the intent types, crsspdratio values, and avd pwt values as given in
Table 6.1. From these plots it can be seen that the general relationship between the
responses based on intent type is maintained regardless of the crs_spd ratio and
avdpwt values, with some exceptions. Significant exceptions include the following:
* Low avdpwt and high crsspdratio: in this case, iqd is similar for all three
intent types due to the fact that the waypoint behavior has a high priority weight
167
110 110 - -- 110
-
90 90
80 80
70 70
80 60
7
7t 50 01
40 30
40 (0
30 30
20 20
0
20
101 10- 10
01 N
(a) No Intent (b) Mode Intent (c) Trajectory Intent
Figure 6-6: Overtaking scenario - time efficiency (qt) distributions by intent type.
V-z75
04.0
93.5 91.0
amo UDL. D
am-,
22 - 22. 7 m-- =93.=0
~9001.0
2M- 1
90.0
2a.W
IR.0 -- ----
MO&. Now Mode
-
N" MaO&
ITENT INTENT INTENiT
a -0.05 G =0.05 a =0.05
Figure 6-7: Overtaking scenario - ANOM comparison of means between none and
mode intent.
168
95 -
-
22.4
94
222
22.0
342A
CPR 214- =21.M7
M5J-
21.4
21.3 3,
21.2 1.5
21.0- 90
20.5 I2.0
Nam Thisaw
NW v
-
N m
RMW MW WN'ENT
a 0.06 a= 0.0 a = 0.06
Figure 6-8: Overtaking scenario - ANOM comparison of means between none and
non-kinematic trajectory intent.
compared to the collision avoidance behavior and the high crs_spd_ ratio
value causes the vehicle to favor changes in speed while maintaining the way-
point derived heading. Overall, this results in the vehicle maintaining the way-
point path more closely under the none and mode intent types than with other
crs_spd_ ratio and avdpwt values. However, the non-kinematic trajectory
intent type still exhibits greater time efficiency in these cases due to the speed
baseline results in higher fea as a result of the scenarios discussed in the fol-
Of specific concern are points in Figure 6-4 that meet the minimum CPA and severe
minimum CPA violation criteria given in Section 6.2.3, which indicate the possibility
of a collision. Figure 6-9 provides the minimum CPA violation rate per 1,000 en-
counters by intent type over all simulated encounters. The same plot is provided in
Appendix B showing the CPA violation rate for each set of independent variables,
where the overall relationship between the rates based on intent type is the same
across values of crsspdratio and avdpwt, except as noted below with respect
169
ovwtaadn Sceneolo - h~nbInmmn CFAs~om 1-c 10 ni) per 000 Overtakhkg Scenslto - Sewem Finh urn CPA Vtolmlone j< 6 m) per
100 Encountervs. we ham lpe 1,00 Encoutnters vs. mInWt lTp
mMO
im
I0.465
SA
-
NM.63
(a) repa < 10 m per 1,000 Encounters
-
Figure 6-9: Overtaking scenario - rpa violation rate by intent type. Violations under
mode intent are primarily due to attempted course reversals that occur as result of
waypoint utility combined with the proposed mode constraint relaxation, as illustrated
in Figure 6-10. Violations under non-kinematic trajectory intent only occur with high
crsspdratio values due to speed reduction and a failure to account for vehicle
kinematics in the communicated intent.
found that violations generally occurred as the result of the following contributing
" The combination of the relaxed constraint on all candidate headings away from
the overtaking side, including a heading in the opposite direction of the stand-
* The stand-on objective function does not account for the current overtaking
vessel state, for example, in the form of a CPA-based collision avoidance utility.
" The stand-on objective function places no restrictions on heading decision space,
such that a holonomic ownship is assumed and course reversals are expected to
The following insights are drawn from these safety violations under the use of
mode intent:
170
(1)
0
(2)
S
(3)
0
(4)
(a) Vehicles
0(b) Waypoint (c) Avoid (d) Collective
Figure 6-10: Overtaking scenario - sequence leading to collision using mode intent.
Objective functions are for the stand-on vessel. In (1), the overtaking vessel (left-most
vehicle) begins overtaking the stand-on vessel on its port side and communicates this
action to the stand-on vessel, resulting in increased utility for maneuvers to starboard.
In (2), the combined utility of the waypoint and collision avoidance objective functions
result in an optimal heading in the opposite direction of the original heading. In (3),
the stand-on vessel executes the course reversal by turning to port, towards the give-
way vessel and enters InExtremis mode. In (4), the vehicles collide due to the stand-on
vessel maneuver.
171
" Simply allowing the stand-on vehicle to maneuver in certain directions based on
communicated intent using an objective function that does not consider any in-
formation other than the communicated intent and basic stand-on responsibility
function domain modified for the use of communicated intent, and thereby
accounting for overtaking vessel state, it is likely that these safety violations
the end of the sequence in Figure 6-10, places the vehicles at greater risk of
generation.
Examining the violations occurring under the use of non-kinematic trajectory in-
tent it was found that all violations occurred with the high crsspdratio value,
generally as the result of the following contributing factors, as illustrated in Figure
6-11:
" A need for sufficient difference between rca,r and rr,inextremis such that in-
extremis actions are not executed in cases where vehicles deviate slightly from
The following insights are drawn from these safety violations under the use of
172
(1) -!5
(2)
(3)
(4)
173
" Lack of account for vehicle kinematics or dynamics in the generation of intended
trajectories results in inaccurate assessment of trajectory-based risk of collision,
leading to increased risk of collision when employing these trajectories for the
purpose of collision avoidance.
" The value of f must be sufficiently greater than fp such that trajectory-
based risk of collision is deemed to exist and associated objective functions
are employed prior to reaching a condition where in-extremis actions must be
executed. Alternatively, some other means of establishing in-extremis conditions
could be explored.
Figure 6-12 illustrates a common inefficient encounter without the use of intent,
while Figures 6-13 through 6-15 depict a similar encounter as executed under the mode
and non-kinematic trajectory intent types, respectively. Significant inefficiencies arise
without the use of intent because the stand-on vessel is not reasoning about which
side the overtaking vessel is overtaking and therefore conservatively continues along
the course and speed at the time the encounter began. Additionally, the overtaking
vessel is unaware of intended stand-on vehicle maneuvers and chooses to overtake
on the side that is most efficient in terms of the waypoint objective, instead of the
side that might make more sense or might be safer in terms of the intended stand-on
maneuver. For this reason, both in the no intent and mode intent case, the give-way
vessel chooses to overtake on the port side if initially on the port side and overtake on
the starboard side if initially on the starboard. This can cause the same inefficiencies
observed in the no intent case to occur in the mode intent case when the overtaking
vessel overtakes on the side of the intended stand-on maneuver, or as in the unsafe
174
(1)
(2)
scenario discussed in the previous section. When the overtaking vessel overtakes to
the side away from the intended maneuver, however, overall efficiency is improved as
shown in Figure 6-13.
Figure 6-14 illustrates how non-kinematic trajectory intent use in an encounter
involving both state-based and trajectory-based risk of collision leads to the safe over-
taking maneuver to the starboard side with an efficient stand-on maneuver to port.
Figure 6-15 illustrates non-kinematic trajectory intent use in an encounter involving
state-only risk of collision, where the vehicles identify the absence of trajectory-based
risk of collision and navigate with respect to their originally planned paths.
175
(1)
(2)
(3)
Figure 6-13: Overtaking scenario - with improved efficiency due to shared mode intent.
Objective functions are for the stand-on vessel, as shown in Figure 6-3. In (1),
the overtaking vessel (left-most vehicle) begins overtaking the stand-on vessel on
its starboard side and communicates this action to the stand-on vessel, resulting in
increased utility for maneuvers to port. In (2), the combined utility of the waypoint
and collision avoidance objective functions allow the stand-on vessel to make the
desired maneuver to starboard. In (3), the stand-on vessel's maneuver to port is
complete and the give-way vessel is able to return to its originally planned path.
176
(1)
(2)
(3)
177
(1)
(2)
(3)
178
mu...Effleincy vs. d-CPA - On-War Dama Pohts IUghdied im EIaney vs. d-CPA - Onftlw Dda Po&ts Highighbd
0 6 2 3S 46 ym 51 16 4 6 2mn.3 4
The previous simulations were performed using the updated mode intent objective
function formulation discussed in Section 5.5.3, where the stand-on vessel accounts
for the give-way vessel state when relaxing constraints to maneuver in one direction
or another based on overtaking vehicle mode intent. The results of this modification
is an elimination of all safety violations discussed in previous sections, as shown in
Figure 6-19. Figures 6-17 and 6-18 summarize the results of employing the updated
Figure 6-20 illustrates an encounter similar to the mode intent-related safety vio-
lation shown in Figure 6-10. The stand-on vessel still attempts a course reversal due
to the waypoint objective function, however this maneuver is delayed to the point
that it does not result in an unsafe maneuver due to the CPA-based portion of the
stand-on objective function. Additionally, the CPA-based stand-on objective function
formulation results in a maneuver away from the give-way vessel, instead of towards,
179
45
40-
40 40
25
30
25 25 25
rcpa M
20 20-- 20
15 15
m -~
10- 10
5 V 10
5
5a 0
Figure 6-17: Overtaking scenario - range at CPA (rep) distributions by intent type
showing comparison with updated mode intent objective function simulations.
mA
2.8 34.3
MAI0
UDL. 01.5L
W.124
=M
Nam.=
2.5 LDL
222=2%
62.0
2-43
Figure 6-18: Overtaking scenario - ANOM comparison of means between none and
updated mode intent.
in the attempt to reverse course. This stand-on maneuver is still relatively inefficient,
and could likely be improved through greater flexibility in speed variation as dictated
by the waypoint objective function. Since it is known that the overtaking vessel will
continue along the port side once adjacent to the stand-on vessel, the stand-on vessel
would ideally slow down and make the desired maneuver to port.
The previous simulations were performed using the kinematic trajectory intent gener-
ation discussed in Section 5.5.3, where each vessel communicates a discrete trajectory
180
Ovetaking Scenario w/Updated Mode Intent Objective Function Ovmlakhlg Scenario w/Updote Mode Intent Objective Function
-
Sewere Minimum CPA Violations J< Sm)r per 1.000 Encounters vs. Minimum CPA Violations (10 mn) per 1,000 Encounters vs. Intent
1.27 WTENt 2
257 u
I dot oW
1 2
i2DM
,AsM
jGM6
Figure 6-19: Overtaking scenario - rpa violation rate by intent type with updated mode
intent objective functions.
based on application of a vehicle model over the set of vehicle waypoints. The result
sections, as shown in Figure 6-23, with the exception of minimum CPA violations
when using the low value of crs_spdratio. Figures 6-21 and 6-22 summarize the
results of employing the kinematic trajectory intent, where the statistically significant
relationships between means are maintained with this updated method of trajectory
generation.
The minimum CPA violations under low crsspdratio values are a result of
the overtaking vessel continuously changing the desired side of passage until finally
committing to a side at a point that is too late to avoid an unsafe encounter as the
stand-on vessel maneuvers to the north. Encounters like this can be avoided by either
avoiding this low of crsspdratio value or by limiting the ability of the overtaking
vessel to change side of passage as in the baseline overtaking objective function given
in [17].
The next scenario tested is the overtaking scenario shown in Figure 6-24 where the
give-way vessel plans to travel from southwest to northeast on a straight path with
constant speed and the stand-on vessel has a non-COLREGS-derived plan to maneu-
181
(1)
(2)
(3)
(4)
Figure 6-20: Overtaking scenario - encounter with updated mode intent stand-on ob-
jective function, which accounts for give-way vehicle state. The give-way vessel is
the left-most vehicle in (1). Objective functions are for the stand-on vessel, as shown
in Figure 6-3. This encounter is similar to the one shown in Figure 6-10, except, as
seen in (1).(c), the stand-on constraint modification employs a CPA-based collision
avoidance objective function which limits the selection of headings towards the over-
taking vessel. The result is a safe, but somewhat inefficient encounter, where gains
in efficiency could be explored through adjustment to waypoint or collision avoidance
behavior settings.
182
41 45 - - - - 45---
-
40
35 35
30 30
25
20 m m
20 20
1M
a0 15 25
15
10 10-
5 5
Figure 6-21: Overtaking scenario - range at CPA (rp,) distributions by intent type
showing comparison with kinematic trajectory intent.
U6 -
-
22.4
.4
U.N
9.0
............ =94144
CM 21.8 -21.787 94.0 92
-
21.6-
80
21A go
68.0
-
212
INTENT2
a -0.05 a = 0.05 a=0.05
Figure 6-22: Overtaking scenario - ANOM comparison of means between none and
kinematic trajectory intent.
183
Overtulft Sceario wila mllW Tajectory Generauo - Severe ovaielng Scenario w/inemasic Tieactory Qeraflon - Savera
Midnm CPA Vilotm (-c 10 rn) per 1,000 Encounters v. intent IMninuom CPA Vcastlona (,c S rn) par 1,000 Enounters vs. Int
Type and Patience Pearmeter
INTENT
p0."
-000
1130
10 0j
0a'-2
DI
1.31 0.:
U2
0274
M-
f3 f
28 so n W 73
M_.M VOCWV Mr--Tnk--
INTIMAY60"
(a) r,,,a < 10 m per 1,000 Encounters (b) rpa < 6 m per 1,000 Encounters
Figure 6-23: Overtaking scenario - fea violation rate by intent type with kinematic
trajectory intent. It is observed that, using kinematic trajectory intent, all severe vio-
lations are eliminated and violations only occur under low crs_ spdratio values as
a result of allowing the give-way vessel to continuously change desired side of passage
by not applying protocol constraints to the trajectory-based objective function.
ver to the east. Approximately 3000 simulated encounters were executed for each
combination of independent variables where the specific encounter totals are given in
Appendix B.
Figures 6-25 through 6-27 summarize the crossing simulation results in terms of the
tory intent again resulted in the greatest efficiency in terms of both distance and time
and also the greatest ea, while use of mode intent resulted in a slightly increased
fpa and mean efficiencies over no intent. These conclusions are supported by the
ANOM plots in Figures 6-28 and 6-29. Table 6.9 provides a complete summary of
The result of greater fpa follows from the geometry of the particular crossing sce-
nario, where many encounters result in no trajectory-based risk of collision and more
low rcPa values occur when intent information is not used due to encounters like those
184
-A
4N
1-
-
-- > If
I
to
N'"
.... ..............................
Giveway Vessel
CD Stand-on Vessel
Giveway Planned Path
-- Standon Planned Path
- Umits of Stand-On Crossing Sector
Figure 6-24: Randomized crossing scenario in which the stand-on vessel has an in-
tended maneuver to starboard. The give-way vessel plans to travel from southwest
to northeast on a straight path with constant speed. On each experimental run, the
overtaking vessel speed and start position is uniformly sampled over intervals that
establish a crossing scenario as given in Section 4.1.
185
40 40 40
35 8 35
30 30 30
25 25
stt r
rcpa m m
20 20 20
15 15 1r
10 10 10
6. 55
0 0
Figure 6-25: Crossing scenario - range at CPA (rqpa) distributions by intent type.
The small difference between the none and mode intent types can be attributed
to the fact that the only difference between the two settings is the greater allowance
for speed changes under the mode intent type, meaning that when the give-way vessel
chooses to pass behind the stand-on vessel, the give-way vessel is still forced to travel
passed its desired turning point. Even with the greater speed variability allowed by
the waypoint objective function with a higher crsspdratio value, there is no
significant improvement achieved with the mode intent type. A slightly greater rpa is
achieved as a result of greater ability to slow down with the give-way vessel passing
ahead or ability to speed up with the give-way vessel passing astern. Future work
should investigate allowances for turning under communicated mode intent in the
crossing scenario.
Appendix B contains plots of rea, qd, and rt mean and standard deviation verses
each of the intent types, crsspdratio values, and avd pwt values as given in
Table 6.1. From these plots it can be seen that the general relationship between the
responses based on intent type is maintained regardless of the crs_spdratio and
avdpwt values, with some exceptions. Significant exceptions include the following:
186
1ao 100 -- 100
90 90
go
.
80 80
70 70 70
60 60
RTd 50 50 50
40}
40-
B
30- 30
20' 20
10
10] 10
0(N t
(a) No Intent (b) Mode Intent (c) Trajectory Intent
Figure 6-26: Crossing scenario - distance efficiency (rid) distributions by intent type.
I
-
-
100 100 100
90 90
80- 80 80
70 70 70
80 60-j
50~
4o 400
40 40
*
30-' W0 30
20i 20 20
10 N 10 10-
Figure 6-27: Crossing scenario - time efficiency (rqt) distributions by intent type.
187
93.10
25
UDL
25.0o6
mil
25.10 uD
2596 93.1155
C- - -Ag - 23.0297 Avg. - 93.44 jAug
mo- Amom
22.AS i.L
INTENT INTENT
0.05 a - 0.05 a -0.05
Figure 6-28: Crossing scenario - ANOM comparison of means between none and mode
intent.
eso 95.5
24.5- 96.0-
96.0
04.5
24.0
CPA -- ~ DL
J4.--23.722
96.
94.0 - - 94.115
23.5
04.0 96.5
22.0- 0.0
9"0
Nam 1Oy Nm Trewy
MOMn T"
INTENT INTENT INTENT
a -0.05 a =0.05 a =0.05
Figure 6-29: Crossing scenario - ANOM comparison of means between none and
non-kinematic trajectory intent.
188
* High crsspd_ratio: in this case, for the none and mode intent types, the
give-way vessel uses a speed change instead of course change to avoid the stand-
on vessel and maintain the waypoint path more closely, resulting in a greater
rcpa than with non-kinematic trajectory intent. This finding warrants further
investigation into combined effects of varying intent information and other col-
lision avoidance behavior settings.
As in the overtaking scenario, points in Figure 6-25 that fall low on the ra scale
were examined in detail, especially those that meet the minimum CPA and severe
minimum CPA violation criteria given in Section 6.2.3. Figure 6-30 provides the
minimum CPA violation rate per 1,000 encounters by intent type over all simulated
encounters. The same plot is provided in Appendix B showing the CPA violation rate
for each set of independent variables. It is observed that, while use of non-kinematic
trajectory intent results in the greatest number of minimum CPA violations in the
crossing scenario, all of these violations occurred with a high crsspdratio value
for the reasons discussed below. Additionally, use of non-kinematic trajectory intent
resulted in no severe minimum CPA violations. Severe violations under none and
mode intent only occur under high crsspdratio and high avdpwt due to greater
Examining the safety violations occurring under no intent use and mode intent
use, it was found that most violations occurred as the result of the following two
scenarios:
the stand-on vessel and once the crossing release conditions are reached both
vessels enter CPA mode resulting in a maneuver towards one another to achieve
respective waypoints. In this case, each vehicle effectively concludes that the
risk associated with the crossing encounter no longer exists without considering
189
Crashqg Sowtaoo- MInimum CPA VWialaa (c 10 mo pw 1,000 Crouuln Swomil - ftver a -LUnwm C1% Vkolamon (.cU m) per
21- Enoomba vs. Intent 1"me _ i,OM Swnter vs. Amen T~ip
1t TEMT
WENT.
NM
17.1 ~0.26
00::
oas
(a) r,a, < 10 mn per 1,000 Encounters (b) repa < 6 m per 1,000 Encounters
Figure 6-30: Crossing scenario - rcpa violation rate by intent type. Violations under
none and mode intent are primarily due to encounters illustrated in Figures 6-31
and 6-32. Violations under non-kinematic trajectory intent only occur with high
crsspd ratio values due to speed reduction and a failure to account for vehicle
kinematics in communicated non-kinematic trajectory intent.
e As illustrated in Figure 6-32: the give-way vessel chooses to pass astern of the
stand-on vessel and the stand-on vessel continues past the originally planned
path of the give-way vessel until crossing release conditions are reached. In CPA
mode, the stand-on vessel then maneuvers across the bow of the stand-on vessel
to achieve the next waypoint. Again, both vehicles effectively conclude that the
risk associated with the crossing encounter is complete without considering the
shared between vehicles. In this case, both vehicles assess that no trajectory-based
risk of collision exists and maneuver with respect to originally planned waypoints.
While this represents an efficient maneuver, Figure 6-34 illustrates the case where
assess the trajectory-based risk of collision. The vehicles employ a collision avoidance
objective function once within in-extremis range, however the strong preference for
maintaining course due to the high crs_ spd ratio value used to build the waypoint
objective function prevents a maneuver to reduce the severity of the unsafe encounter.
190
(1)
S4
(2)
0
(3)
(4)
Figure 6-31: Crossing scenario - unsafe encounter with none or mode intent. Ob-
jective functions are the collective objective functions for each vehicle. In (2), the
vehicles have entered a crossing encounter in which the stand-on vessel (right-most
vehicle) maintains course and speed and the give-way vessel chooses to cross astern
of the stand-on vessel. In (3), the Give WayX and StandOnX mode release criteria are
met for each vehicle, causing entry into CPA mode and application of collision avoid-
ance objective functions without protocol constraints. In (4), the combined waypoint
and collision avoidance objective functions result in maneuvers towards respective
waypoints resulting in a close encounter.
191
(1)
-
(2)
(3)
(4)
Figure 6-32: Crossing scenario - unsafe encounter with none or mode intent use.
Objective functions are the collective objective functions for each vehicle. In (1)
and (2), the vehicles have entered a crossing encounter in which the stand-on vessel
maintains course and speed and the give-way vessel chooses to cross astern of the
stand-on vessel. In (3), the stand-on vessel crosses the original give-way vessel path
and the Give WayX and StandOnX mode release criteria are met for each vehicle.
This results in entry into CPA mode and application of collision avoidance objective
functions without protocol constraints. In (4), waypoint utility causes the stand-on
vessel to cross back in front of the give-way vessel, resulting in a close encounter.
192
(1)-
(2)
(3)
Figure 6-33: Crossing scenario - encounter with use of shared non-kinematic trajectory
intent involving state-only risk of collision. The give-way vessel is the left-most vehicle
in (1). Objective functions are the collective objective functions for each vehicle. In
(1), the vehicles have communicated non-kinematic trajectory intent, as shown by the
paths in the vehicle graphics, and have agreed that no trajectory-based risk of collision
exists. As a result, the vehicles do not apply explicit collision avoidance constraints.
In (2) and (3), the vehicles continue efficient execution of originally planned paths
without applying collision avoidance constraints.
193
(1)
(2)
(3)
Figure 6-34: Crossing scenario - unsafe encounter with non-kinematic trajectory in-
tent use. The give-way vessel is the left-most vehicle in (1). Objective functions
are the collective objective functions for each vehicle. In (1), vehicles have agreed
that no trajectory-based risk of collision exists, and, therefore, do not apply collision
avoidance constraints. In (2), because communicated trajectories did not account for
vehicle kinematics or dynamics, vehicles close to the point that r" becomes less than
rT,inextremis, triggering the in-extremis collision avoidance objective function. In (3),
vehicles begin to maneuver away from one another due to combined waypoint and
collision avoidance utility.
194
AOL ---- . -- _-.4
were also examined. As in the overtaking case, these encounters generally occur with
In addition to the inefficiencies depicted in Figures 6-31 and 6-32, Figure 6-35
illustrates a common inefficient encounter under the none and mode intent types,
where the give-way vessel plans to cross the bow of the stand-on vessel. The stand-on
vessel, not knowing or reasoning about whether or not the give-way vessel will pass
ahead or astern, maintains course and speed until the give-way vessel is clear. This
Figure 6-33, based on either the absence of trajectory-based risk of collision or the
allowance for the stand-on vessel to maneuver with respect to its originally planned
path.
6.1 using the equipment described in Section 6.1.4. Figure 6-36 provides a comparison
The previous simulations were performed using the kinematic trajectory intent gen-
of almost all safety violations discussed in previous sections, as shown in Figure 6-39.
Figures 6-37 and 6-38 summarize the results of employing the kinematic trajectory
intent, where the statistically significant relationships between means are maintained
195
(1) m
(2)
(3)
Figure 6-35: Crossing scenario - inefficient encounter without use of intent informa-
tion. Objective functions are the collective objective functions for each vehicle. In
(2), the vehicles have entered a crossing encounter in which the give-way vessel (left-
most vehicle) chooses to cross ahead of the stand-on vessel and the stand-on vessel
maintains course and speed, without reasoning about give-way vessel intent. In (3),
the StandOnX mode release criteria are finally met and the stand-on vessel is able to
maneuver back in the direction of its next waypoint.
196
C0,01g arb On-Vwl Ranstit - DltUne EMlancy v. CromeVig Sasoemb On-Wlaf Resauts - Tim Efflelscyvs. r-CPA
rCPA Onused Dabt Pont In UackpmiQ Sokuaatd Da Points In Backuewa4
4 .1IUT
00 * IR
Aft* Umb
10.V
1WTEHT
To
0 4 s i 6 s 4
46
Figure 6-36: Crossing scenario - on-water efficiencies and repa in comparison to simu-
lated encounters, where on-water data points are overlaid on the grayed-out simulation
data points.
35 35 36
30 I 30 30
25 25 25
m
20 20 20
15 15 15
10 P 10 10
5 5 5
-
0 0
Figure 6-37: Crossing scenario - range at CPA (repa) distributions by intent type
showing comparison with kinematic trajectory intent.
197
99
98 95.0
1
25.0
97
24.5
94.5
-94.20
24.107 94.0
CPA 24.0
95
93.5
23.5
94
23.0 93.0
93
Figure 6-38: Crossing scenario - ANOM comparison of means between none and
kinematic trajectory intent.
Croski Sownmlo wAC(Innatc Tnjotiny GeneriOn - t.Urkmwn C1904011111 Scenino Wwiime Tralecimy Generatin - Seveos
CPAV~ol.ang(.cO n) per ,000EnOWAetrs v.AM
1
bTAM Iv Migniumn CPA Vhfltion. (< 10 us) pe 1,000 Encounter vs. nnt
2121 ENTBIT?2
NOW 2*
-I-C
MROOVl9Jlt9
1.222
~16
O
M
(a) rpa < 10 m per 1,000 Encounters (b) rpa < 6 m per 1,000 Encounters
Figure 6-39: Crossing scenario - frpa violation rate by intent type with kinematic
trajectory intent. It is observed that, using kinematic trajectory intent, all severe
violations are eliminated and all violations are reduced below the non-intent-aware
case.
198
4.
?t
. . .
.
CD Giveway Vessel
CD Stand-on Vessel
Giveway Planned Path
Standon Planned Path
Limits of Head-On Sector
Stand-on Passage
Figure 6-40: Randomized head-on scenario in which the east-bound vehicle has an
intended maneuver to port. The west-bound vessel plans to travel from east to west
with constant speed. On each experimental run, the west-bound vehicle speed and
start position is uniformly sampled over intervals that establish a head-on scenario as
given in Section 4.1.
The next scenario tested is the head-on scenario shown in Figure 6-40 where one ves-
sel plans to travel from east to west on a straight path with constant speed and the
other vessel has a non-COLREGS-derived plan to maneuver to the north. Approxi-
mately 3000 simulated encounters were executed for each combination of independent
variables where the specific encounter totals are given in Appendix B.
199
48
40 140
30 j
45 T
28
m
(
(aW oItn
10 10-
Figure 6-41: Head-on scenario - range at CPA (rqa) distributions by intent type.
Figure 6-41 and 6-43 summarize the head-on simulation results in terms of the distri-
butions of each metric by intent type. Qualitatively, use of non-kinematic trajectory
intent again resulted in greatest efficiency in terms of both distance and time and
lower fpa as in the overtaking scenario. Use of mode intent had no observable effect
in comparison to the non-intent-aware baseline, as expected due to the fact that there
under the use of mode intent in the head-on scenario. These conclusions are sup-
ported by the ANOM plots in Figures 6-44 and 6-45. Table 6.10 provides a complete
result of lower ?,p is attributed to the same explanation as in the overtaking scenario,
where the particular fcpa,-r, at the value of fi' results in more close-range vehicle
encounters. Particular scenarios that result in unsafe and inefficient encounters are
200
100 100 100
-
90
70
90 70 70
40-
90 50
0
20 40-
30
20j
10 10 10.
0 0
(a) No Intent (b) Mode Intent (c) Trajectory Intent
Figure 6-42: Head-on scenario - distance efficiency (7/d) distributions by intent type.
-
100 100 100
90 90
80_ 8o
70
ao 10
50 50 50
40~ 40(r
30 30
20~ 20 20
10- 10 10 -j
0 0 0
Figure 6-43: Head-on scenario time efficiency (rh) distributions by intent type.
-
201
95.18 UDL
!UDL
lam,-
1.40-
95.16
aen' un
CFR Avg = 95.1265 4- -0- Avg - 92.19M
Mu3 Avgo .18.3415 a 92.1
-
92.16-
16.0 65.00
LDL
66.06 - - -- - --- - LD
None Mods
INTENTf INTENT INTENT
a = 0.05 a- 0.06 a - 0.05
Figure 6-44: Head-on scenario - ANOM comparison of means between none and mode
intent.
94.0
67.0
12.5
16.2
UDL
16.0
-
17.6
02.0
Nam 110CMbY
INTENT INTENT INTENT
0.05 a=0.05
Figure 6-45: Head-on scenario - ANOM comparison of means between none and
non-kinematic trajectory intent.
202
Sensitivity to Independent Variables - Head-On
Appendix B contains plots of rpa, rd, and r7t mean and standard deviation verses
each of the intent types, crsspdratio values, and avdpwt values as given in
Table 6.1. From these plots it can be seen that the general relationship between the
e Low avdpwt and high crsspdratio: none and mode result in lower mean
rca due to waypoint utility giving strong preference for staying on the planned
path and flexibility for changing speed. This waypoint utility outweighs the
utility for collision avoidance maneuvers due to low avdpwt. Since vehicles
speed.
As for the other scenarios, points in Figure 6-41 that fall low on the ra scale were
examined in detail, especially those that meet the minimum CPA and severe mini-
mum CPA violation criteria given in Section 6.2.3. Figure 6-46 provides the minimum
CPA violation rate per 1,000 encounters by intent type over all simulated encoun-
ters. It is observed that non-kinematic trajectory intent use results in the greatest
minimum CPA violation rate but the smallest severe violation rate, both differences
being statistically significant from the no intent rate. Additionally the violation rates
are significantly greater under low avd_pwt and high crsspdratio due to the
Examining the collisions occurring under no intent use and mode intent use, it
was found that most safety violations occurred as the result of at least one of the
* The west-bound vehicle assesses that the vehicles are in a head-on scenario,
203
Heed-On Scenario - Mi m CPViolation (. 10 n0 per 1,000 Heed-On Scenario -SereM w C (C S mdper
Enom vsa<w. mnr 1pe E.000 Enonrste (b) rnmen 1,ooE
dtor6 f tst
1i
I:I
unermdeinet resmia totmeiohennitn-aaecs ineteei
while the east-bound vehicle does not. The result is the application of the
head-on objective function by the west-bound vehicle, while the east-bound
vehicle simply applies a CPA mode collision avoidance objective function that
allows a turn to port. As a result, the east-bound vehicle turns to port and the
west-bound vehicle turns to starboard.
* The east-bound vehicle begins the turn to port prior to initiating the collision
avoidance behavior based on the ria value. This turn may also occur in the
case where rg is sufficiently high, making collision avoidance priority weight
low enough based on Equation (4.19, or avd pwt is sufficiently low such that
the relative waypoint priority weight to collision avoidance priority weight is
high enough that the waypoint heading utility outweighs the collision avoidance
constraint.
204
(1)
(2)
(3)
(4)
Figure 6-47: Head-on scenario - unsafe encounter with no intent use. Objective
functions in (1) are waypoint objective functions and those in (2) through (4) are
COLREGS behavior objective functions. In (1), the vehicles have not yet initiated the
COLREGS behavior. In (2), the west-bound vehicle has assessed that it is in a head-
on encounter and the east-bound vehicle has assessed that enough separation exists
that no particular COLREGS scenario applies and goes into CPA mode. Additionally,
the east-bound vehicle has begun turning to port as the collision avoidance objective
function is activated. In (3), the west-bound vehicle, without accounting for the
intent of the east-bound vehicle, has executed a maneuver to starboard according
to its head-on responsibility, which conflicts with the east-bound vehicle maneuver.
In (4), a close encounter occurs as the west-bound vehicle tries to maneuver back
towards its next waypoint.
205
-W
violation as the vehicles did not account for kinematics or dynamics in the risk of
collision assessment. As in the crossing scenario case shown in Figure 6-34, the
vehicles employ a collision avoidance objective function once within in-extremis range,
however, based on the high rate of closure it is too late to improve the safety of the
encounter.
Examining the safety violations occurring under the use of non-kinematic trajec-
tory intent it was found that that most violations occurred as the result of at least
one of the following contributing factors:
" As illustrated in Figure 6-48: failure to account for vehicle kinematics and dy-
namics resulting in an inaccurate assessment of trajectory-based risk of collision
in a scenario that involves both state-based and trajectory-based risk of colli-
sion.
206
IAm I - __ - __ - ", - - ---
"
prevent switching side of passage once chosen. This objective function formu-
in the overtaking and crossing scenarios since only one vehicle is choosing the
The following insights are drawn from the observed safety violations:
" Any application of trajectory intent to the head-on scenario must be done from
a sufficient r' such that the vehicles are able to agree upon collision avoidance
about one another. This observation leads to the discussion of strategic verses
ature [106,109].
ance maneuvers.
" In the case where both vehicles are given collision avoidance responsibility, the
vehicles must either apply a protocol to avoid one another using non-conflicting
ment of any state-based risk of collision. Such replanning would need to occur
at a sufficient r" such that risk can be reduced prior to the development of an
unsafe scenario.
207
(1)
(2)
(3)
Figure 6-48: Head-on scenario - unsafe encounter with non-kinematic trajectory in-
tent. Objective functions are the collective objective functions for each vehicle. In
(1) and (2), vehicles have agreed that no trajectory-based risk of collision exists, and,
therefore, do not apply collision avoidance constraints. In (3), because communi-
cated trajectories did not account for vehicle kinematics or dynamics, vehicles close
to the point that r" becomes less than r,inextremis, triggering the in-extremis collision
avoidance objective function. However, at this point, it is too late to avoid a close
encounter.
208
(1)
(2)
(3)
Figure 6-49: Head-on scenario - unsafe encounter with non-kinematic trajectory in-
tent. Objective functions are the COLREGS behavior objective functions for each
vehicle. In (1), the west-bound vehicle has received the priority assessment commu-
nication from the east-bound vehicle and assessed that agreement has been achieved
and that a trajectory-based risk of collision does exist. However, the west-bound ve-
hicle has not yet processed the east-bound vehicle's priority assessment and goes into
HeadOn mode. In (2), the east-bound vehicle has now processed the priority commu-
nication and agrees with the west-bound vehicle, but has already begun a maneuver
to starboard. The west-bound vehicle maneuvers to port due to no protocol con-
straint application in the trajectory-based objective function. In (3), the conflicting
maneuvers result in an unsafe encounter.
209
Significant Inefficiencies - Head-On
that are low on the rd and rt scales in Figures 6-42 and 6-43, indicating low efficiency,
Figure 6-50 illustrates a typical encounter without intent use under this particular
head-on experimental setup. Under the use of non-kinematic trajectory intent, this
encounter is handled either as shown in Figure 6-48 where the vehicles assess there is
or as shown in Figure 6-51, where the vehicles gain some efficiency by accounting for
6.1 using the equipment described in Section 6.1.4. Figure 6-52 provides a comparison
of on-water encounters to the simulated encounters. It is observed that the on-water
The previous simulations were performed using the kinematic trajectory intent gen-
minimum CPA violation rate and roughly the same severe minimum CPA violation
rate as shown in Figure 6-39. These violation rates are attributed to the same fac-
tors discussed in the previous sections with respect to application of trajectory intent
to the head-on scenario. Figures 6-37 and 6-38 summarize the results of employing
210
(1) - W
(2)
(3)
(4)
Figure 6-50: Head-on scenario - inefficient encounter with no intent use. Objective
functions in (1) are waypoint objective functions and those in (2) through (4) are
COLREGS behavior objective functions. In (1), the vehicles have not yet initiated
the COLREGS behavior. In (2), both vehicles have assessed that they are in a head-
on encounter and begin maneuvering to starboard, despite the intended east-bound
vehicle maneuver to port. In (3) and (4), the vehicles begin maneuvering back towards
respective waypoints.
211
(1)-
(2)
(3)
212
Head-cn Suiuil On-Water Redtm - Distace Efficiey vs. Had-on Scear* On-NW Remsf -Thus cancy vs. rCM
rCPM(Ohnuisd Dal Pohe Bi
Bckgounc Wmu stod Data PoinSt In 3ackcou4
WTDrr W* N
9 Now
mod$ a mom
Figure 6-52: Head-on scenario - on-water efficiencies and r,,, in comparison to simu-
lated encounters, where on-water data points are overlaid on the grayed-out simulation
data points.
46 4 465
40 40 40
30 30
26- 25~ 25
m m
20 20 20
16. 15
10 10 10
6 j 5
0 a
-
Figure 6-53: Head-on scenario - range at CPA (rep,) distributions by intent type
showing comparison with kinematic trajectory intent.
213
4.0
M86 UDL
97.0 06.5
M6.4
W-_ 93.209
At
CPA
-T Avg =18.3042 W 9 0
1626 I
L26.0
18.24 L 0.0
Figure 6-54: Head-on scenario - ANOM comparison of means between none and
kinematic trajectory intent.
Hd-on Smm woo wAnMenafo lhulaay Ganermlf- MbInwmu H b-Mfmbf wAQMf Troatmy Ommafn - ftvom
naimwn CM Vkalallan (410.1 " W 1.00 EnM-6t- Ms Inten
$M VIcWons (410 nr per 1,00 nrAombra vs. hotM Two TWO6
WW2
MWW
MMWW-oO-"
IAI
I(
(a) rqpa < 10 m per 1,000 Encounters (b) repa < 6 m per 1,000 Encounters
Figure 6-55: Head-on scenario - rCpa violation rate by intent with kinematic trajectory
intent.
214
6.4 Multi-Vehicle Evaluation
aware system in basic two-vehicle encounters, however this evaluation does not assess
the relative performance in encounters involving more than two vehicles, a case which
simple traffic pattern where vehicles are merging with one another on common courses,
departing from this common course, and potentially overtaking one another due to
different speed objectives.
The evaluated pattern is depicted in Figure 6-56, including the objective speeds for
each vehicle. The same independent variable values were evaluated as given in Table
6.1 and the same fixed parameters were applied as given in Section 6.2. Since there
is not a specific starting and stopping point, in position or time, as there was in the
Additionally, only encounters that result in repa < 30 m and da < 150 m are
considered so as not to skew the results based on encounters that occur between
vehicles from across the traffic pattern and encounters that go on for multiple laps
of the pattern. Such encounters may result from an overtaking vehicle with a high
crsspdratio value simply sitting behind the stand-on vehicle indefinitely, which
While two-vehicle encounters are still of interest, the focus of the analysis in this
of the proposed intent-aware system in encounters involving more than two vehicles.
The simulations for the multi-vehicle traffic pattern were conducted using the same
software setup as used in the canonical encounter simulations. Over 1000 simulated
215
4N
CD: Vehicles
Planned Waypoints
State Vector
v4Q
* 1.4 m/s
vi
1.6 r/s M
O. m/s
v2
O.8 M/s
Figure 6-56: Evaluated multi-vehicle traffic pattern. While it is recognized that ve-
hicles would not continuously drive around loops such as these in the real-world, this
setup provides for random encounter generation in which vehicles are merging with
one another on common courses, departing from this common course, and poten-
tially overtaking one another due to different speed objectives. Additionally, a setup
such as this is particularly useful for continuously generating encounters in on-water
experiments.
216
30 30 30
25 25
20 20
-
S15 m 15 S15
10 10 10
5 :5 5
0 0 0
-
(a) No Intent (b) Mode Intent (c) Trajectory Intent
Figure 6-57: Traffic pattern - range at CPA (rp,) distributions by intent type.
encounters were executed for each combination of independent variables, with excep-
tion of the high crs_spd _ratio experiments due to the issue discussed in previous
section with respect v1 slowing down staying in an overtaking encounter with v2 for
General Results
Figures 6-57 and 6-58 summarize the simulation results in terms of the distribution
resulted in greatest efficiency and all intent types resulted in roughly the same fe,.
Table 6.11 provides a complete summary of the traffic pattern simulations. More
detailed analysis is explored by number of contacts involved in each encounter.
217
100 100
90 so 90-
-
90
s0
70 70 70
60
-
60 60
80-
71 50 50
40
40- 40 t
30 30 30
20
20 0 201
10 e 10 10
0 0 0
Figure 6-58: Traffic pattern - trajectory efficiency (,q.) distributions by intent type.
82.0
20.0
20.8
81.8 UDL
20.7 UDL
WA - AVg a 201M
20.8 I:Avg = 81.483
2U.
20A
LDL 81.4
1
81.2 LOL
20.8
20.2 81.0
NOM mom. Mode
INTENT
a - 0.0 a- 0.06
(a) fpa
Figure 6-59: Traffic pattern - ANOM comparison of means between none and mode
intent for 2-contact encounters.
Figures 6-59 and 6-60 show a comparison of means between none and mode or non-
kinematic trajectory, respectively, for encounters involving two contacts. It is observed
that mode intent results in a greater ipa and no significant difference in 7- and that
non-kinematic trajectory intent results in both greater fcpa and greater ?T in these
multi-contact encounters.
Figure 6-61 shows safety violation rates by intent type and the number contacts
involved in the encounter. Examination of single contact encounters reveals the same
issues observed in the canonical overtaking and crossing scenarios, where mode ap-
218
20A
20AU
88
20.82UDL
84- 84.32
LDL
2040
MNT NW
a -O.05 a .05
Figure 6-60: Traffic pattern - ANOM comparison of means between none and non-
kinematic trajectory intent for 2-contact encounters.
plication suffers from the simple nature of the initial formulation of the mode intent
objective function and non-kinematic trajectory application suffers primarily from a
failure to account for vehicle kinematic and differential constraints in trajectory in-
tent generation or the lack of protocol constraint application. Examples of specific
encounters of concern are addressed in the following sections.
Figure 6-62 illustrates a three-vehicle encounter without the use of intent information.
In this encounter, v1 chooses to overtake both v2 and v4 on the port side, despite
the fact that v1 has an upcoming maneuver to starboard and v4 has an upcoming
maneuver to port. The result is that when v1 and v4 open range to the point
that they enter CPA mode, they each choose maneuvers towards one another based
on waypoint utility. A failure to account for ownship kinematics or dynamics in
the collision avoidance objective functions then causes an especially close encounter.
Additionally, v2 incurs inefficiency with respect to its planned maneuver to port based
on standing on for both v1 and v4.
Figure 6-63 illustrates a similar three-vehicle encounter with the use mode intent.
In this case, the v2 inefficiency is eliminated through communication by v1 of intended
passage to port. However, by freeing up maneuvers to starboard with respect to v1,
219
(ufa rtci - Mtrnu CPAWohum(10 per1 ibaftfloPttert - Severe Minimasto CM MIomn (o6m3n per 1,000
Encoufibrae .hInent Tpe sta Cantedt Count Encounters vs. Intent 1WweenConts Count
1.67 wbli04
108
I040.8
081
~03
021
(a) repa < 10 m per 1,000 Encounters (b) ra< <6 m per 1,000 Encounters
Figure 6-61: Traffic pattern - ifea violation rate by intent type and contact count.
Violations in single contact encounters are due to the same issues exposed in the
canonical encounter experiments. Many of the violations in two-contact encounters,
under no intent use and both mode and non-kinematic trajectory are a result of not
accounting for ownship kinematics or dynamics in the establishment of utility in as-
sociated collision avoidance objective functions. Violations in two-contact encounters
under non-kinematic trajectory intent illustrate the need for trajectory planning and
negotiation that accounts for the collision avoidance maneuvers taken with respect to
all vehicles.
v2, which is in CPA mode with v4 and is not accounting for ownship kinematics or
dynamics, chooses a starboard maneuver that results in a collision with v4. This case
illustrates how unsafe scenarios are created with the application of mode intent that
Figure 6-64 illustrates another three-vehicle encounter with the use of non-kinematic
trajectory intent. In this encounter, v3 has agreed to avoid the communicated trajec-
tories of both v1 and v2; v1 has agreed to avoid the communicated trajectory of v2;
and v2 has agreed to avoid the communicated trajectory of v1 based on initial en-
with respect to v1, v1 is not applying constraints with respect to v3 and closes within
in-extremis range before maneuvering away from v1. This case illustrates the short-
coming of the proposed method for handling non-kinematic trajectory intent, where
vehicles do not account for changes in planned trajectories based on interactions with
220
more than one vehicle.
In addition to the inefficiencies exposed in the previous section and the improvements
over these inefficiencies offered by the application of mode or non-kinematic trajectory
intent, Figures 6-65 and 6-66 illustrate two more three-vehicle encounters in which
account for intent information provides for improved efficiency.
Figures 6-67 and 6-68 show that application of the updated mode intent objective
function formulation for the overtaking scenario, as presented in Section 5.5.3, per-
forms at least as well as the non-intent-aware system in two-contact encounters in
terms of both safety and efficiency. Figure 6-67 shows that statistically significant
differences in means were not achieved, though means under the use of updated mode
are improved. Additionally, Figure 6-68 shows that safety violation rates are im-
proved over the initial mode objective function formulation and are comparable to
Figures 6-69 and 6-70 show that application of kinematic trajectory intent performs at
safety and efficiency. Figure 6-69 shows that statistically significant improvements
Figure 6-70 shows that safety violation rates are improved over or comparable to the
ters. It is observed that the on-water encounters are generally consistent with the
221
(1)
(2)
(3)
(4)
Figure 6-62: Traffic pattern - three-vehicle encounter without intent use resulting in a
collision between vi and v4. Objective functions are the collective objective functions
for each vehicle. In (1), v1 chooses to overtake both v2 and v4 on the port side, despite
the fact that v1 has an upcoming maneuver to starboard and v4 has an upcoming
maneuver to port. Additionally, v4 chooses to overtake v2 on the starboard side
despite a similar conflict in intent. In (2), the overtaking vessels continue overtaking
maneuvers, while v2 maintains course and speed. In (3), v1 and v4 open range
to the point that they enter CPA mode, and waypoint utility causes each to choose
maneuvers towards one another. In (4), a failure to account for ownship kinematics or
dynamics in the collision avoidance objective functions by v1 and v4 causes selection
of maneuvers that result in an especially close encounter.
222
- A
(1)
(2)
(3)
Figure 6-63: Traffic pattern - three-vehicle encounter with mode intent use resulting in
a collision. Objective functions are the collective objective functions for each vehicle.
Similar to the encounter in Figure 6-62, in (1), v1 chooses to overtake both v2 and
v4 on the port side and v4 is overtaking v2 on the starboard side. In (2), range
was opening between v2 and v4 causing v2 to go into CPA mode with respect to
v4. Additionally, v1 has communicated to v2 that v1 is overtaking on the port side,
thereby allowing v2 to make its desired maneuver to starboard. In (3), a failure to
account for ownship kinematics or dynamics in collision avoidance objective functions
by v2 and v4 causes selection of maneuvers that result in an especially close encounter,
as in the 90 degree to starboard maneuver selected by v2 in (2).(c).
223
(1)
(2)
(3)
(4)
224
-I
(1)
(2)
(3)
Figure 6-65: Traffic pattern - three-vehicle encounter with improved efficiency due
to non-kinematic trajectory intent. Objective functions are the collective objective
functions for each vehicle. In (1), v3 maneuvers to port to avoid the trajectories of
v1 and v2 and simultaneously steer towards its next waypoint. In (2), v2 continues
towards its next waypoint while vi maneuvers to starboard around v2 to get ahead
of v2 prior to its upcoming turn to starboard. In (3), vi completes this maneuver
by cutting ahead of v2 to make way towards its next waypoint and v2 is also able
to turn toward its next waypoint. Also, v3, having avoided v1 and v2, is close to its
next waypoint.
225
(1)
P~
(2)
0* N
(3)
Figure 6-66: Traffic pattern - three-vehicle encounter with improved efficiency due to
non-kinematic trajectory intent use. Objective functions are the collective objective
functions for each vehicle. In (1), v1 chooses to overtake v2 on its port side based on
v2's intended maneuver to starboard. Also, v1 and v4 agree that no trajectory-based
risk of collision exists and therefore do not take crossing scenario actions. In (2), v2 is
able to maneuver to port based on having standon priority in the encounter with v1.
In (3), v2 completes its maneuver to port based on priority with respect to v1 and
not having a trajectory-based risk of collision with v4. Also, v1 is able to maneuver
towards its next waypoint based on freedom from constraints with respect to v4.
226
82.0
UDL
81.9
20.40
UDL 81.8
20.36 81.7
81.6
20.30
I Avg 20.2807 81.5
Avg = 81.585
I
=
20.28 81.4
81.3
20.20
81.2 LDL
INTENT 2 INTENT 2
a = 0.05 a = 0.05
Figure 6-67: Traffic pattern - ANOM comparison of means between none and
updated mode intent.
Mkfimum CPAMoteone (,c10Cm) per 1.06 Enoowiters vs, Intent Severe IMbimum CA Ml-tetlon (46.4n per 1,000 Srnommtere w.L
WEN 2A TENT2
so 953
*001
0-1 1 1 1 - 1- tA
(a) rcpa < 10 m per 1,000 Encounters (b) rpa < 6 m per 1,000 Encounters
Figure 6-68: Traffic pattern - f~pa violation rate by intent type with updated mode
intent objective functions.
227
21.4
21.2 85
I DL
. = 84.39
21.0
UDL 0384
Avg = 20.873
CPA 20.8 L=
.
83
20.6
82
20.4
81
-
INTENT 2 IMTENT 2
a = 0.05 a = 0.05
Figure 6-69: Traffic pattern - ANOM comparison of means between none and kine-
matic trajectory intent.
WThile Ple wA~bemetio Traectory Gmatln - inmu CPA Traffic Pattern wAUWac
-- Tajectory Generion - Severe
Vioaetions 1-c 10 m) per1,000 Encountern va. hIent l7pe Minimum CP0A V~oletie (,c 6mo per 1000 Encounters va. Intn
Type
*am
1ams
(4.< ,0 E ImeOn r
(a) rp~a < 10 m per 1,000 Encounters (b) <6 m per 1,000 Encounters
rpa<
Figure 6-70: Traffic pattern - repa violation rate by intent type with kinematic tra-
jectory intent. Safety violation rates are improved over or comparable to the non-
kinematic trajectory formulation and non-intent-aware baseline.
228
1.M.acPMrn On-WAr Rasmoi- 4sctaRy Emdyncy w. rC
Pkamlated Deb- Palio Inkakrund)
12
401.
ir vs. rep
Figure 6-71: Traffic pattern - on-water efficiency and rqa in comparison to simulated
encounters, where on-water data points are overlaid on the grayed-out simulation
data points.
"I" indicates improvement over the baseline non-intent-aware system , "D" indicates
degradation, and "-" indicates no significant change. Results for kinematic trajec-
tory intent generation (Section 5.2.2) are shown for all experiments. Results for the
updated mode intent objective formulation (Section 5.5.3) are shown only for the over-
taking and traffic pattern experiments, since this formulation applies only to these
experiments. In general it was found that the relationships summarized in Tables 6.12
and 6.13 were maintained with variation of other independent variables. Additionally,
229
Table 6.12: Summary of Canonical Encounter Results
fcpa I D I I - I
rpa <10 m Rate - D - I -
-
rqpa<6mRate - I - I - I
I I I I - I
9,I I I I - I
updated kinematic
Parameter mode trajectory
cpa
rea< 6 m Rate -
-
- I
The following observations are made with respect to the application of mode intent:
* Stand-on objective function formulations should not naively assume that all ma-
neuvering actions away from a side of passage will result in a safe encounter. The
230
distribution over trajectories or a more precise, deterministic description of the
portion of the state space that will occupied by the communicating vehicle; this
can trigger unsafe interactions that otherwise would not occur; such scenarios
are not necessarily a result of constraint relaxation but may also be a function
The following observations are made with respect to the application of trajectory
intent:
" The degradation of safety performance in the overtaking scenario reflects the
taking vehicle from continuously switching the desired side of passage until a
" Additional parameter values that effect the use of trajectory intent, specifically
that greater values of fp,,, will result in fewer safety violations based on a
" Generation and communication of trajectories that more accurately reflect fu-
" The absence of protocol constraints in cases where both vehicles have collision
precise reason for the existence of the COLREGS protocol, whereby, for in-
non-conflicting maneuvers.
231
* Lack of account for contact maneuvers in response to ownship and other contacts
collision avoidance methods should examine techniques for accounting for these
interactions, as discussed in [671.
The following observations are made with respect to use of intent information in
" The assumption that vehicle will follow communicated intent does result in
unsafe encounters in the case where vehicles are allowed to change their original
intent or do not precisely follow the original intent, as in the case of non-
This chapter has presented the experimental evaluation of the proposed intent-aware
ning methods would result in overall improvement. The bulk of our analysis has been
based on simulation but have also included on-water experiments. Even though the
small number of on-water experiments prevents statistical analysis of this field data,
general agreement of the data with the larger, more extensive simulation analysis is
encouraging.
232
Chapter 7
Conclusions
The overall conclusion of this thesis is that shared vehicle motion intent information
recognition that there are many contributing factors to the ability to apply intent
information towards improving safety and efficiency, including the need to account
for certain aspects of the problem that were either not considered or treated as as-
7.1 Contributions
This thesis evaluated the impact of allowing autonomous surface vehicles to share
certain types of intent information and to apply that intent information to the process
scenario communicates its intent to maneuver in a certain direction with respect the
stand-on vessel, and discrete trajectory intent in which vehicles communicate planned
positions and times associated with those positions and then agree on responsibility
ios this work finds that both COLREGS mode intent and discrete trajectory intent
233
can improve both safety and efficiency, however further investigation of parameter set-
tings for the proposed multi-objective optimization approach is required in order to
ensure an appropriate balance between safety and efficiency in all possible encounter
work finds that application of intent information in these encounters where vehicles
are closing one another at a high rate can be unsafe as vehicles may not have enough
time to process intent information and begin executing associated collision avoidance
maneuvers. In this case it may be more appropriate to simply execute the protocol
defined maneuver, which does not require communication or agreement between ve-
hicles; otherwise, the time horizon on which intent is considered should be extended.
more complex as the intent itself may evolve based on interactions between all of the
vehicles. For example, a vehicle may communicate intent to pass on one side, and
the receiving vehicle will maneuver based on this information, but interaction with a
third vehicle may cause the original intent to change such that the receiving vehicle's
maneuver is now unsafe. Accounting for such a scenario requires providing the vehicle
While the improvements in safety and efficiency achieved in this work were ac-
future work should consider application of other planning methods, especially if con-
sidering a strategic application of intent information in which vehicles adjust nominal
sidering encounters with more than one contact, methods should be explored that
account for change in vehicle trajectories due to interaction between all vehicles and
do not simply assume that collision avoidance will be achieved through assignment of
pair-wise collision avoidance responsibility, as in the case of the proposed trajectory
intent application.
9 The assumption that communicated intent will be followed precisely does not
generally hold and and can result in vehicles selecting unsafe maneuvering ac-
234
tions. This includes the communication of trajectory intent without adequate
consideration for vehicle kinematic or differential constraints. Significant im-
provements in safety were observed in simulation when applying an accurate
vehicle motion model to trajectory intent generation. Since such accurate ve-
hicle models are not always available, impact of deviation from communicated
intent should be further evaluated in addition to methods that identify when a
vehicle is deviating from intended trajectory.
* In the case where vehicles agree upon actions to avoid collision, if this agreement
involves both vehicles taking action then either protocol constraints must be
applied or some cooperative replanning algorithm must be applied such that
the vehicles do not choose conflicting actions.
Further research into the area of intent-aware collision avoidance in the marine
domain should leverage similar research in other domains, as many of the above
concerns and areas for further investigation are addressed in that research, as covered
in Chapter 2 of this thesis. The following section provides specific recommendations
for further research in this area.
235
vehicles counts (i.e. more vehicles in a given encounter), and parameter values
should be conducted to confirm and expand upon the results of this thesis.
236
case of human-robot interaction is of particular interest from the perspective of
nicate its intentions. Specifically, merging the concepts presented in this thesis
with the IMO and IALA e-Navigation efforts [50,53] could be highly beneficial
proves upon a vehicles ability to safely and efficiently execute collision avoidance
maneuvers over the ability to do so with an inferred intent, for which it may
into a world that has no explicit rules for unmanned vehicles begs the question
of what, if any, new rules should be applied to these vehicles. Such rules could
also account for the ability of unmanned vehicles to share intent, as in COL-
Rules [109].
Formal Marine Intent Languages: Like the development of AIDL in the aviation
domain, development of a similar formal intent language for the marine domain
of intent information into marine navigation systems, both for manned and
unmanned use. Such work could build upon existing efforts to share vessel
237
validates intent information before applying it to the motion planning process.
This work might be incorporated with intent inference, whereby levels of con-
probabilistic model.
array of challgenging, real-world conditions with both human and robotic agents.
Such work is critical to assessing the viability of employing intent-aware systems, and
238
Appendix A
The behavior and general intent-aware system introduced in Chapter 5 were developed
The concept of payload autonomy is illustrated in Figure A-1, where the processes
of mission planning, payload autonomy, and the vehicle are separated such that the
payload autonomy accepts inputs from mission planning and provides general com-
mands to the vehicle [20]. Using this system, a designer of the payload autonomy
system need only be concerned with the interface to the vehicle control system and
not with the particulars of the system itself. This may mean that low level control is
handled entirely by the vehicle computer while the payload computer only provides
high level actions such as desired heading and speed. Additionally, each system may
have its own representation of the environment in some world model, the contents
of which can be shared between systems. The payload autonomy system may rely
entirely on the vehicle for sensory data input, or may be connected directly to a set
of its own sensors.
system realization, the payload autonomy concept is employed such that the colli-
sion avoidance system is executed on a payload computer where the interface with
239
Mission <_Autonomy Vehicle
World Model World Model World Model
User
Input
__F100"_____
.
Paykoed sows] Vshidle Sensor
the vehicle computer allows updates to ownship state data from vehicle sensors and
motion execution through actuator commands to the vehicle computer. The mission
perform high-level control of mission execution, such as the starting and stopping of
and publish to output variables, all of which are stored in the MOOS database, or
240
A.2 Intent-Aware System Applications
The intent-aware system described in Chapter 5 was built using the MOOS mid-
dleware and the payload autonomy concept discussed in the previous section, where
Figure A-2. Each of the following software components were developed to perform
REGS collision avoidance behavior for use by pHelmIvP, which contains the
tion employs the interval programming method (IvP) introduced in [21] and is
in [19], which accepts updates to contact state and intent information and per-
forms any necessary intent inference or validation and outputs the processed
" Intent communication application: selects updated intent information and broad-
Figure A-2 and discussed in Section A.3 were used to realize the full vehicle autonomy
system.
241
* I.t.
ICoCutp
,
WBWOh COLREGS COLES
sahaar Behavwr BWWWWa
Figure A-2: The MOOS-IvP autonomy system used for intent-aware system evalu-
ation. Applications communicate through information published to the MOOSDB.
The IvP helm accepts objective functions from individual behaviors and solves for the
optimal maneuvering action. Maneuvering actions are executed by a PID controller,
the output of which is passed to the vehicle through a vehicle interface. High-level
vehicle control is executed through a shoreside interface.
" Value of repa,, above which trajectory-based risk of collision is considered not
" Distance at which the COLREGS encounter is considered complete and the
242
" The angle used to define a nearly reciprocal course, as defined in Section 4.1.2:
Oheadon
* The utility values for constructing COLREGS collision avoidance mode intent-
based objective functions as described in Section 5.5.3: 9mode, go,iow,intent, ge,min,intent,
triggered for a contact when contact range is within -,,t and is terminated when
contact range is beyond rcomplete, meaning that several instances of the behavior may
exist simultaneously, each producing an objective function for a different contact. The
parameter intentuse allows the selection of the type of intent information employed
avoidance mode intent only, and intentuse = trajectory results in use of trajectory
intent only.
lishes the current COLREGS collision avoidance mode for each contact as well as any
collision avoidance priority assessments for use by the intent communication applica-
tion.
The intent-aware collision avoidance behavior parameters contained in Table A.1 were
243
Table A. 1: Fixed Collision Avoidance Parameters
The intent processing application performs the intent processing functions described
in Section 5.1 in addition to processing and managing contact state data as described
in [19].
244
The application accepts contact "node reports" which are simulated AIS reports
that, amongst other data, may include all of the following information about the
contact [19]:
or acknowledgement communications.
Given the assumption that all received intent information is valid, the current
instantiation of this application simply publishes updated mode and trajectory infor-
mation as received through contact node reports for consumption by the intent-aware
COLREGS behavior. If any contact node reports contain collision avoidance priority
information addressed to ownship, i.e. priorityn and priority", this information
any data manipulation to express the intent as defined in Section 5.2, and publishes
245
Communicating COLREGS Collision Avoidance Mode Intent
" Contact identification: the contact for which the mode intent applies.
* Any assessment of collision avoidance priority, along with the associated contact
identification: priority's, priority'
246
application of vehicle model as described in Section 5.2.2. If the waypoint pattern will
" Heading and speed at the final point in the trajectory: an = (On, vn)
Configuring pIntentComms
" Trajectory horizon time: thorizon E [0, oc], where use of tho,izon = oc results
247
Using the above parameters, Iavoid and Itraj will be continuously updated and
published at a frequency of fcoms as long as communicate mode = true and
communicate-trajectory = true, respectively. Additionally, trajectories will only
be reported up theizo, seconds into the future or to latest time at which the trajectory
The following applications, as documented in [19], were employed to realize the com-
provided weights and solves for the globally optimal maneuvering action a* =
(Od, Vd).
" pMarinePID: basic PID controller that converts a maneuvering action a = (9, v)
into vehicle actuator commands in terms of desired thrust and desired rudder.
" pNodeReporter: aggregates ownship state and intent data to produce a single
node report that is effectively broadcast to all other vehicles as a simulated AIS
report.
Broker, uFldMessageHandler
e Helm iteration frequency, the frequency at which objective functions are pro-
248
" Heading decision domain, as a discrete subset of [0, 360)
" Speed decision domain, as a discrete subset of [0, Vmax] for some maximum speed
Vmax
" The ratio, as a percentage, of wo to w,, as discussed in Sections 4.3.1 and 4.3.3.
" Waypoint lead distance, lead damper, capture radius, and slip radius as de-
scribed in [191.
" The maximum priority weight of the waypoint objective function, i.e. the value
The pMarinePID application provides for separate PID control of ownship yaw and
* Gains for the yaw pid controller: kp,,,, kd,yaw, and ki,yaw
" Gains for the speed pid controller: kp,,pd, kd,spd, and ki,spd
* Frequency at which ownship node reports are created and communicated: freport
249
A.3.5 Fixed Parameters in Supporting Applications
250
Appendix B
Supplemental Figures
soao~ iMode
4M00
2000-
21j II j j I ___IiI
V- I.- _ _ _ _
25 75 50 25 75
100 150 200
AVQDPWT / CRS-SPD RATIO / IJT2N
251
Overtaking Scenario - Mean r-CPA vs. intent Tyrpe, Course-Speed Overakig Sceamlo - SWd Deviation of r.cPA vs. Inte'N pe.
Ratio, and Avoid Priority WalgM Cours"-Peed Rtio, and Avoid Prrtyf WeW~i
25
23. 42 24, M~a rm 72 5.9 69 IWENT
5.i2 6.9 67 5.66
65 6 628 6.48 G64 6.59m Nw
,mMM*
.
1.6 21.22191 21.7 6.12 6.21 Mms
5.55.73 =
15
I
U
tin-
Iij
4 -I125 9
25 7.5 w 25 --I 2II5 I
148
-
Figure B-2: Overtaking scenario - range at CPA (rca) mean and standard deviation
by experiment.
Overtaking Sceario - Meom Distace Efficiency vs. Itent 1IVe Overtaking Scenario, - Sid Deviation of Distance Efliolanc vs.
Course-Speed Ratio, and AvodM Prioit MblWt Intui Twa Course-Speed Rati, mat Avoid Priot Weght
65- 96 3.7 94.7 5.9 96
965 .9
IA 4A 91 eHEJ 16 14. "TENT
mou&
2. n~l
127 Me Mode
as- ibMode
11IMPuar
9.57
27
I
0.
76.6 60 25 7
I
IL&.I
111IT40r
I I
114151
liji1 1j1!j11
25 75 so 25 75
Ion 199 259
MrkPr/0WV61l/51U9
Figure B-3: Overtaking scenario - distance efficiency (rid) mean and standard devia-
tion by experiment.
252
Overtablagwa Scd- Mm Tim Elffloany vs& 1 -en 1 Tpe. Overtaking Scenario - S Devialon of Tw EMolency vs. bintnt
lam Course-Speed Rafto and Avoid Prorllyf t 1W, Cose-Seed Rat dAvomid ort Wet
80.4 ~ 942 9.2 Wf91 rmEW vow
BR
inM5
in99959 13.1a 13.2 MN"o
INNOM
12.3 11111TV01-1
I 7Me
I
I 4- tQ 4.22 4.11 4.9 .2 4.2 4M2
0
ijIiI1jiI jI iIf
25 76 51) 2 7S 2TS 75 so 76 26
-52 IN 2w 100 IS 2w
AWD PW
I CRtB-W M M
(a) ?t (b) o,
Figure B-4: Overtaking scenario - time efficiency (rq) mean and standard deviation
by experiment.
OvertaMng Scenario - nknmu CPA Violations (< 10 i per 1,000 Overtaidtg Scenario - Sever. hUn*man CPA Violations (Si<6n) per
Encountars vs. Intent 1Ipe, Counw-Sped Raio, and Avoid 1,000 Encomtirs vs. Itent 1Tipe, Course-Speed Ratio, end Avoid
52 Porty Wee _ Prioft VIN_
W8ThT 3.2 WENT
J
14
117
Mmwa.
MAabf
I
I
2-
MMX*52
Ij 1.31 1.32
1.56
212
j 011i
'l67Ow I
- -U-w
25 75 2 S (b 76 so pe 1 c
AWD PWT C1B4PO RAME M AD-PW IC I.= . ~I
(a) rpa < 10 m per 1,000 Encounters (b) rep, < 6 m per 1,000 Encounters
253
B.2 Crossing Scenario Simulations
30001
2000
-
1000
-
III I
ra
~ii II-
z
iii
25 75 50 25 75
100 150 200
AVD PWT / CRS-SPD RATIO / INTENT
254
Cmaang Scenari - Mean- r-CPAm vI Dt 1Tme Couaae-Spewd Cmaalng Scenaro - Sd Deviation of r-CPA vs. Intent 1Te.
Raft ndAvoid PriodtyWgi Courae-Sped Reaot en Avoid Priority Ub~q
27 6.01M 67651 DITEW
Z 4.9 25246 4 42 246
2216 *0 5 7.12 7 7.92 76
?A.
7.6 781
4179
7.15 7.065 ~ ~M141166
23-
21A I z75 56
I - 26
1_2"12 S3.
20.1. ~ Ij _ _
Os 6 7 i
Figure B-7: Crossing scenario - range at CPA (rca) mean and standard deviation by
experiment.
CroeahfgScenulo- Mean Dftance EffIiec Pvis. -~n 1iPe, Cmealnkg Scenari - did Devian at Dbtwne Efficenc vs. Intent
Couse4kwed PaUkhua" Avmid Priort VlMW Toop, Cousee-Speed Rat and Avoid Prioit W*VA9
100- ~ 96&4 62 6 964 NTE7E 7.0- 6.75 WOTIEi
2.6 1615
4.1.0
2575-6 2
in16 9
(a)s-
1906 1966 096
.
(b) an,,
Figure B-8: Crossing scenario - distance efficiency (rid) mean and standard deviation
by experiment.
255
Crossing Scenato - Mom Time EfMciency vs. Intent Type, Course- Crossing Scenario - Std Deviation of Time Efficiency vs. Intent
Speed Ratio, and Avoid Priority Weight 7 p Course-Speed Ratio- and Avoid Priority Weight
6M TEN
as&80.03
.m12 Nam
ad_~r SA&S/fTENT
i
26 25 10 36 25
(a) i (b) o,
Figure B-9: Crossing scenario - time efficiency (77t) mean and standard deviation by
experiment.
Crossing Scenario - Mimumn CPA Violations 1-c 10 m) per 1,000 Crossing Scenerlo - Severe Minimum CPA Violallons (< 6 m) per
Encounsters vs. Intem, Type, Course-Speed Ratio, and Avoid 1,000 Encounters vs. Intent Type, Course-Speed Relic, and Avoid
M3 Prio"iyWeight ___ Priorft1110lgt
211 "N 402 WENT7
2D218
13
1101
I1
70:
40-
36.5
so 20.0 356520
25 75 SO 36 75
(a2110m e ,0 E c u tr M5
AM pWT/Rs
ISO
AWIO101/2T
200
256
B.3 Head-on Scenario Simulations
Head-on
5000_T
Sceario - Total Number of Encounters by Experiment
INTENT
MMode
5000-
MM~aetory
4M00
3000-
2000-
1000-
I
n-
I I
iii25
100
75
I
50
j I
25
200
75
I
AVD PWr I CRS-SPO RATIO fINrENT
257
Head-on Soenerlo - Mean r _CPA vs. intent Tpe Course-Speed Head-on Scsaado - SWd Deviation of r._CPA vs. Intent Tyvpe,
Ratio. and Avoid Prioritv Weight 706 Course-Speed Ratio, and Avoid Priesty Weight
16.9 1.1 6.02 &B INTENT
1961 93 19 . 7 ITENT 65 &. 37
17.9
$0 6.7 .4 . Mode
55- 5&42 6.463m5.28
12- 0-
11.1
m1a78,8,y 5 25-
3-
27
.1 I I I I
1ClSOPD 200 -
~--urn
7 50 5 -
MO
75
Figure B-12: Head-on scenario - range at CPA (r a) mean and standard deviation
by experiment.
Heed-on Scenario - Mean Dhetnc Efficiency vs. Wetnt Type, Head-on Scenario - Id Detiation of Disiano. Efficiency vs. Intent
CourseSpeed Raedo, and Avoid Priodl VWh T*VeCourse-Speed Rati, and Avoid Piody Weight
INTENT 6,592
5 96 90.7 ebb.085.4 97 911.7 W,.8
milod
M' NI,
9.55M
LO.
75-
ta
70
Figure B-13: Head-on scenario - distance efficiency (rid) mean and standard deviation
by experiment.
258
Head-on Scenalo - Momn Tbne Efficiency ve. Intent 1Ipe Coame- Heed-on Scenaio - Sid Deviation of Thne Efficienc vs. intent
IG] Speed Ratio, W iAvod Prlof WelgM 514 pVM Coume-Speed Rat and Avoid Pvlolt Weigt
25-921 W.8 n48~ 9.49~ ~WNT2
18 9. 194
&9 .7
19~
4.72
4.88 2.13 4.9M 4TN
M90"1
4 4
mMN~
4.16, 4-27 mw299
36 3.75
&
3,6 3U4
so-
3j 3.3.0
14
Ii 5
2907 9
7
1
191
301
79904 t
100 200
me PWT I Cfl-4PD F11 WF EMT
(a) 7 t (b) o
Figure B-14: Head-on scenario - time efficiency (rqt) mean and standard deviation by
experiment.
Heed-on Scenario - nbnemow CPA Vklaimm (< 10 m)~ per 1,00 Heed-on Soanaio - Seve 1Gn lnp n CPA Violatona 1cSm) per
Encountera ys. 2.0mnt 1Tpe Couee-Speed Ratio,and Avoid 1000 Enootstr vs. Intent TVe Couree-Upeed Ratio, and Avoid
37.9
ISO- PdTEM~~s
J
147
140- 137M
w9 129 BX .b
In
j321
17.6
I
10
4:~
9.5 43 II. 825&Ii.
4343
III'
AOWIIII I-MIIIENT
(a) rp',a, <10 mn per 1,000 Encounters (b) ri>pa < 6 m per 1,000 Encounters
259
B.4 Traffic Pattern Simulations
5000
INTENT
m None
M Mode
inTrajectory
4000-
3000
2000-
1000-
25
100
75
I 50
150
AVD PWT / PAM1ENCE / INTENT
25
II
200
. 0
75
I
260
flalfc Patten - Mau r-.CPA v&. -ite- Tyme Coase-Spd Rati. ibifc Patter - Std Deviaton of r.-CM vs.krWt 1IWp, Come-
Speed Ratio, mid Avod Prirt We4h
272d AviWTEMtTWih - 24 8054
IN8W
6.46 ~MM"
54,91 4.38
22- 28. 19 m a2
W121 192.8 1072. 20262.2242. 20.
4.21 418~4 43
2.00
0.5 0.8
I I
is__ SO 25 26 78 so 28 75
2w 101) 1382 738
Figure B-17: Traffic pattern - range at CPA (rca) mean and standard deviation by
experiment.
1hiafo Patteon - Mm Tuectery Effleac- r v&'. bil 1IVPe, 1l'aIW Pater - Sat Deviatn of Waojectory Wflckenc vs. Inten
1Tpe Courae-Speed Rati, and Avoid Prioit VWh
380- ous-~ PAto, mid Avod Pdaft Weigh
174 14KE 181.5
9D- 37 7 87 M7j w87iNwfe
5. Mod 6 1 8. 1886 508I
as* 81.9 81. 81 512 85
75A M7 SSAo
I
S 28 78 58 28 7
1 027
go-lvsm l,484
0i3Wl/4038O6
(b) orl
Figure B-18: Traffic pattern - trajectory efficiency (r) mean and standard deviation
by experiment.
261
Terffic Pattern - Minimum CPA Violations (<10 n) per 1,000 Tluffic Pattern - Svrs Minimum CPA Violations (< 6 m) per 1,000
Encounters vs. Intent Type, Course-Speed Ratio, and Avoid Encounters vs. Inthit 1spe, Course-Speed Ratio, and Avoid
Prioly WMb 4 dm8WegAW
Prkxf
2-21.7 210 KMNT CS
TfENT
I
=am
-
t44
Iis-I 173
I i,-~18.6 I7N mode
42 1i
4A
m**
-
I
am-
ZA-
32-
-t
11111
so-
1012
.757 I 2-4-
Z2-
II ii
2A- 170
12-
7- 5.2
'-6
-
1A
Ili-
10
02-
-
Ion IO 200
AVD- PWr / CR"PD PJM / WTEKr 'srsWT/R'SSPDmno In1B.1
(a) r,a < 10 m per 1,000 Encounters (b) re,a < 6 m per 1,000 Encounters
262
Appendix C
The algorithms in this appendix are adapted from the mode check procedures pre-
sented by Benjamin in [17].
263
Algorithm 10 Checking COLREGS Crossing Give-way Mode Entry Criteria
1: procedure CHECKCROSSINGGIVEWAYENTRY()
2: evaluate a > Section 4.1
3: evaluate / > Section 4.1
4: if (a < 247.5) or (3 > 112.5) then
5: return false
6: end if
7: return true
8: end procedure
264
Appendix D
Notation
265
Symbol Section Defined Description
tion
weight
weight
weight
mode 4.4.1 COLREGS collision avoidance mode
submode 4.4.1 COLREGS collision avoidance submode
risk, 5.3.2 Boolean trajectory-based risk of collision
266
Symbol Section Defined Description
sion avoidance
priority 5.5.1 Trajectory-based collision avoidance maneu-
vering priority
priorityset 5.5.1 Boolean indicating priority has been set for
a particular contact
agree 5.5.1 Boolean agreement between vehicles for
267
268
Bibliography
[1] Single European Sky ATM Research Joint Undertaking (SESAR). http:/
www.sesarju.eu/.
[2] FAA General Operating and Flight Rules, 14 C.F.R., Chapter 1, Subchapter
F, Part 91, 2017.
[7] Georges Aoude, Joshua Joseph, Nicholas Roy, and Jonathan How. Mobile agent
trajectory prediction using Bayesian nonparametric reachability trees. Proc. of
AIAA Infotech@ Aerospace, pages 1587-1593, 2011.
[10] Tirthankar Bandyopadhyay, Chong Zhuang Jie, David Hsu, Marcelo H. Ang Jr,
Daniela Rus, and Emilio Frazzoli. Intention-aware pedestrian avoidance. In
Experimental Robotics, pages 963-977. Springer, 2013.
269
[11] Tirthankar Bandyopadhyay, Kok Sung Won, Emilio Frazzoli, David Hsu,
Wee Sun Lee, and Daniela Rus. Intention-aware motion planning. In Algo-
rithmic Foundations of Robotics X, pages 475-491. Springer, 2013.
[12] Richard Barhydt, Parimal Kopardekar, Vernol Battiste, Nathan Doble, Wal-
ter Johnson, Paul Lee, Thomas Prevot, and Nancy Smith. Joint NASA
Ames/Langley experimental evaluation of integrated air/ground operations for
en route free maneuvering. 2005.
[13] Richard Barhydt, Michael Palmer, and Todd Eischeid. Development and eval-
uation of an airborne separation assurance system for autonomous aircraft op-
erations. In 24th International Congress of Aeronautical Sciences, Yokohama,
Japan, 2004.
[18] Michael R. Benjamin, John J. Leonard, Joseph A. Curcio, and Paul M. New-
man. A method for protocol-based collision avoidance between autonomous
marine surface craft. Journal of Field Robotics, 23(5):333-346, May 2006.
[19] Michael R. Benjamin, Henrik Schmidt, and John J. Leonard. The MOOS-IvP
Open Source Autonomy Project. http: //www.moos-ivp. org.
[20] Michael R. Benjamin, Henrik Schmidt, Paul M. Newman, and John J. Leonard.
Nested autonomy for unmanned marine vehicles with MOOS-IvP. Journal of
Field Robotics, 27(6):834-875, November 2010.
270
[23] Mads Bentzen Billeso. ACCSEAS Project - Service Description: Tactical Ex-
change of Intended Routes. http://www.accseas. eu, 2015.
[24] K. Bimbraw. Autonomous cars: Past, present and future a review of the de-
velopments in the last century, the present scenario and the expected future
of autonomous vehicle technology. In 2015 12th International Conference on
Informatics in Control, Automation and Robotics (ICINCO), volume 01, pages
191-198, July 2015.
[25] Ole Bakman Borup. ACCSEAS Project - Service Description: Maritime Cloud.
http://www.accseas.eu, 2015.
[26] S. Campbell, W. Naeem, and G.W. Irwin. A review on improving the au-
tonomy of unmanned surface vehicles through intelligent collision avoidance
manoeuvres. Annual Reviews in Control, 36(2):267-283, December 2012.
[27] Clearpath Robotics. Heron Unmanned Surface Vessel
for Bathymetry. https://www.clearpathrobotics.com/
heron-bathymetry-unmanned-surface-vessel/.
[29] Joseph A. Curcio. Rules of the Road for Unmanned Marine Vehicles. In Man-
har R. Dhanak and Nikolaos I. Xiros, editors, Springer Handbook of Ocean
Engineering, pages 517-526. Springer International Publishing, 2016. DOI:
10.1007/978-3-319-16649-0_23.
[34] Gilles Dowek, Cesar Mu6z, Alfons Geser, and Institute for Computer Applica-
tions in Science and Engineering. Tactical conflict detection and resolution in
a 3-D airspace[microform!; [prepared ... under contract NAS1-97046. ICASE
report ; no. 2001-7, NASA contractor report ; NASA CR-210853. ICASE, Na-
tional Aeronautics and Space Administration, Langley Research Center ; Avail-
able from NASA Center for Aerospace Information, Hampton, VA : Hanover,
271
MD, 2001. Shipping list no.: 2002-0173-M. "April 2001." Distributed to depos-
itory libraries on microfiche.
[35] Les Elkins, Drew Sellers, and W. Reynolds Monach. The Autonomous Maritime
Navigation (AMN) project: Field tests, autonomous and cooperative behaviors,
data fusion, sensors, and vehicles. Journal of Field Robotics, 27(6):790-818,
November 2010.
[36] Ian Fairclough and David McKeever. PHARE Advanced Tools Trajectory Pre-
dictor. 2000.
[38] Paolo Fiorini and Zvi Shiller. Motion Planning in Dynamic Environments Using
Velocity Obstacles. The InternationalJournal of Robotics Research, 17(7):760-
772, July 1998.
[41] J. Giesbrecht. Global path planning for unmanned ground vehicles. Technical
report, DTIC Document, 2004.
[43] George E. Hagen and Ricky W. Butler. Towards a Formal Semantics of Flight
Plans and Trajectories. Technical report, December 2014.
272
[461 Cheng-Neng Hwang, Joe-Ming Yang, and Chung-Yen Chiang. The Design of
Fuzzy Collision-Avoidance Expert System Implemented by H,,-Autopilot. Jour-
nal of Marine Science and technology, 9(1):25-37, 2001.
[48] Yong K. Hwang and Narendra Ahuja. Gross Motion Planning - a Survey. A CM
Comput. Surv., 24(3):219-291, September 1992.
[49] Yong K. Hwang and Narendra Ahuja. Gross Motion PlanninghA'Ta Survey.
ACM Comput. Surv., 24(3):219-291, September 1992.
[571 Vasiliy Karasev, Alper Ayvaci, Bernd Heisele, and Stefano Soatto. Intent-aware
long-term prediction of pedestrian motion. In Proceedings of the International
Conference on Robotics and Automation (ICRA)(May 2016), 2016.
[58] Sven Koenig, Maxim Likhachev, and David Furcy. Lifelong Planning A*. Arti-
ficial Intelligence, 155(1):93-146, May 2004.
[59] Michael Konyak, Scott Doucett, Robab Safa-Bakhsh, Eduardo Gallo, and
Paul Parks. Improving Ground-Based Trajectory Prediction through Com-
munication of Aircraft Intent. In AIAA Guidance, Navigation, and Control
273
Conference. American Institute of Aeronautics and Astronautics, 2009. DOI:
10.2514/6.2009-6080.
[60] Michael Konyak, Dan Warburton, Javier Lopez-Leones, and Paul Parks. A
demonstration of an aircraft intent interchange specification for facilitating
trajectory-based operations in the national airspace system. In AIAA Guid-
ance, Navigation and Control Conference and Exhibit, page 7145, 2008.
[611 Voemir Kunchev, Lakhmi Jain, Vladimir Ivancevic, and Anthony Finn. Path
planning and obstacle avoidance for autonomous mobile robots: A review. In
InternationalConference on Knowledge-Based and Intelligent Information and
Engineering Systems, pages 537-544. Springer, 2006.
[62] Yoshiaki Kuwata, Michael T. Wolf, Dimitri Zarzhitsky, and Terrance L. Hunts-
berger. Safe maritime navigation with colregs using velocity obstacles. In Intel-
ligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference
on, pages 4728-4734. IEEE, 2011.
[63] Yoshiaki Kuwata, Michael T. Wolf, Dimitri Zarzhitsky, and Terrance L. Hunts-
berger. Safe Maritime Autonomous Navigation With COLREGS, Using Veloc-
ity Obstacles. IEEE Journal of Oceanic Engineering, 39(1):110-119, January
2014.
[66] Young-il Lee and Yong-Gi Kim. A Collision Avoidance System for Autonomous
Ship Using Fuzzy Relational Products and COLREGs. In Intelligent Data
Engineering and Automated Learning - IDEAL 2004, pages 247-252. Springer,
Berlin, Heidelberg, August 2004.
[67] St6phanie Lefevre, Dizan Vasquez, and Christian Laugier. A survey on motion
prediction and risk assessment for intelligent vehicles. ROBOMECH Journal,
1(1):1, July 2014.
[68] 0. Levander. Autonomous ships on the high seas. IEEE Spectrum, 54(2):26-31,
February 2017.
[69] Timothy A. Lewis, Nipa Phojanamongkolkij, and David J. Wing. The effects of
limited intent information availability on self-separation in mixed operations. In
Integrated Communications, Navigation and Surveillance Conference (ICNS),
2012, pages B6-1. IEEE, 2012.
274
[70] Mikael Lind, Mikael Hagg, Ulf Siwe, and Sandra Haraldson. Sea Traffic Manage-
ment - Beneficial for all Maritime Stakeholders. In Proceedings of 6th Transport
Research Arena, April 18-21, 2016 Warsaw, Poland, 2016.
[71] Yu-Hong Liu and Chao-Jian Shi. A fuzzy-neural inference network for ship
collision avoidance. In 2005 International Conference on Machine Learning
and Cybernetics, volume 8, pages 4754-4759 Vol. 8, August 2005.
[72] Oivind Aleksander G. Loe. Collision Avoidance for Unmanned Surface Vehicles.
141, 2008.
[73] J. Lopez, M. Vilaplana, I. Bayraktutar, J. Klooster, J. M. Asensio, G. McDon-
ald, and P. Kappertz. Towards an open test bed for the study of trajectory
synchronization in the future ATM system: The ASIS initiative. In 2009 Inte-
grated Communications, Navigation and Surveillance Conference, pages 1-14,
May 2009.
[75] Javier Lopez Leones. Definition of an aircraft intent description language for
air traffic management applications. PhD, University of Glasgow, 2008.
[78] St6phane Mondoloni. Flight Object - A Recommendation for Flight Script and
Trajectory Description, October 2006.
[80] Cengiz Pasaoglu, Nursel Akcam, Emre Koyuncu, A. Farabi Tarhan, and Gokhan
Inalhan. Collaborative Intent Exchange Based Flight Management System with
Airborne Collision Avoidance for UAS. Journalof Intelligent & Robotic Systems,
84(1-4):665-690, December 2016.
[81] Thomas Porathe. Transmitting intended and suggested routes in ship opera-
tions: cognitive off-loading by placing knowledge in the world. Work, 41(Sup-
plement 1):4873-4878, January 2012.
275
[82] Thomas Porathe and George Shaw. Working with the human element: Human
Factors and technical innovation from EfficienSea and on to ACCSEAS. In
Chalmers PublicationLibrary (CPL), 2012.
[83] Tom G. Reynolds and R. John Hansman. Analysis of separation minima using a
surveillance state vector approach. In 3rd USA/Europe Air Traffic Management
R&D Seminar, volume 10, 2000.
[84] C. Rihacek, T. Lutz, B. Weinert, A. Bolles, K. Nielsen, and J. K. Jensen. Effi-
cienSea2 - Conceptual Model of the Maritime Cloud. http: //ef f iciensea2.
org, 2015.
[87] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach.
Prentice Hall Press, Upper Saddle River, NJ, USA, 3rd edition, 2009.
[88] Anders Rydlinger. Mona Lisa 2.0 - STM Voyage Exchange Format and Archi-
tecture. http: //stmvalidation. eu, 2015.
[90] Brual C. Shah, Peter $vec, Ivan R. Bertaska, Wilhelm Klinger, Armando J. Sin-
isterra, Karl von Ellenrieder, Manhar Dhanak, and Suneet K. Gupta. Trajectory
planning with adaptive control primitives for autonomous surface vehicles op-
erating in congested civilian traffic. In Intelligent Robots and Systems (IROS
2014), 2014 IEEE/RSJ International Conference on, pages 2312-2318. IEEE,
2014.
[91] Brual C. Shah, Petr Svec, Ivan R. Bertaska, Armando J. Sinisterra, Wilhelm
Klinger, Karl von Ellenrieder, Manhar Dhanak, and Satyandra K. Gupta.
Resolution-adaptive risk-aware trajectory planning for surface vehicles oper-
ating in congested civilian traffic. Autonomous Robots, pages 1-25, December
2015.
276
[94] CheeKuang Tam, Richard Bucknall, and Alistair Greig. Review of Collision
Avoidance and Path Planning Methods for Ships in Close Range Encounters.
Journal of Navigation, 62(03):455, July 2009.
[98] U.S. Coast Guard Navigation Center. Vessel Traffic Services. https://www.
navcen.uscg.gov/?pageName=vtsMain.
[102] Haozhou Wang, Han Su, Kai Zheng, Shazia Sadiq, and Xiaofang Zhou. An effec-
tiveness study on trajectory similarity measures. In Proceedings of the Twenty-
Fourth Australasian Database Conference- Volume 137, pages 13-22. Australian
Computer Society, Inc., 2013.
[103] Anthony Warren. Trajectory prediction concepts for next generation air traffic
management. In 3rd USA/Europe ATM R&D Seminar, page 171, 2000.
[106] David Wing, Bryan Barmore, and Karthik Krishnamurthy. Use of Traffic Intent
Information by Autonomous Aircraft in Constrained Operations. American
Institute of Aeronautics and Astronautics, August 2002.
277
[1071 David Wing, Robert Vivona, and David Roscoe. Airborne Tactical Intent-Based
Conflict Resolution Capability. In 9th AIAA Aviation Technology, Integration,
and Operations Conference (ATIO). American Institute of Aeronautics and As-
tronautics, 2009. DOI: 10.2514/6.2009-7020.
[109] David J. Wing and William B. Cotton. Autonomous Flight Rules - A Concept
for Self-Separation in U.S. Domestic Airspace. Technical report, November
2011.
[112] Yiyuan Zhao, Christine Haissig, and Mary Jo Hoffman. Analysis of pilot intent
parameters in air traffic management. In American Control Conference, 1998.
Proceedings of the 1998, volume 3, pages 1789-1792. IEEE, 1998.
[113] Yiyuan Zhao and Qian Maggie Zheng. Modeling Uncertainties in Intents, Guid-
ance, and Pilot Actions for Reliable Trajectory Synthesis. American Institute
of Aeronautics and Astronautics, September 2012.
278