NAVAL

POSTGRADUATE
SCHOOL
MONTEREY, CALIFORNIA

THESIS

A MACHINE LEARNING APPROACH TO

ENABLE MISSION PLANNING OF TIME-OPTIMAL
ATTITUDE MANEUVERS

by

Reed Ransom Smith

September 2020

Co-Advisors: Mark Karpenko

Brian M. Wade
Approved for public release. Distribution is unlimited.
THIS PAGE INTENTIONALLY LEFT BLANK
 Form Approved OMB
REPORT DOCUMENTATION PAGE No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing
instruction, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of
information. Send comments regarding this burden estimate or any other aspect of this collection of information, including
suggestions for reducing this burden, to Washington headquarters Services, Directorate for Information Operations and Reports, 1215
Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302, and to the Office of Management and Budget, Paperwork Reduction
Project (0704-0188) Washington, DC 20503.
1. AGENCY USE ONLY 2. REPORT DATE 3. REPORT TYPE AND DATES COVERED
(Leave blank) September 2020 Master's thesis
4. TITLE AND SUBTITLE 5. FUNDING NUMBERS
A MACHINE LEARNING APPROACH TO ENABLE MISSION PLANNING
OF TIME-OPTIMAL ATTITUDE MANEUVERS
6. AUTHOR(S) Reed Ransom Smith

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING

Naval Postgraduate School ORGANIZATION REPORT
Monterey, CA 93943-5000 NUMBER
9. SPONSORING / MONITORING AGENCY NAME(S) AND 10. SPONSORING /
ADDRESS(ES) MONITORING AGENCY
N/A REPORT NUMBER
11. SUPPLEMENTARY NOTES The views expressed in this thesis are those of the author and do not reflect the
official policy or position of the Department of Defense or the U.S. Government.
12a. DISTRIBUTION / AVAILABILITY STATEMENT 12b. DISTRIBUTION CODE
Approved for public release. Distribution is unlimited. A
13. ABSTRACT (maximum 200 words)
Time-optimal spacecraft rotations have been developed and implemented on orbiting spacecraft,
highlighting opportunities for improving slew performance. Double-digit reductions in the time required to
slew from one attitude to another have been demonstrated. However, the ability to perform mission planning
to make use of minimum time slewing maneuvers is largely precluded by the need to compute a numerical
solution to find a single minimum time maneuver control trajectory. Machine learning approaches can
eliminate the need to generate problem solutions by approximating time-optimal maneuver times with
sufficient accuracy for planning using only the initial and final attitude requirements. The advantages of
time-optimal spacecraft maneuvers, a planning construct for evaluating legacy and machine learning
maneuver time generators, and the machine learning processes that enable this approach are outlined.
Compared to legacy planning techniques, time-optimal slew approximations yield target collection increases
of 3% to 24% for an example planning framework.

14. SUBJECT TERMS 15. NUMBER OF

optimal control, trajectory optimization, genetic algorithm, machine learning, supervised PAGES
learning, neural network, linear regression, path planning, trajectory planning, optimal 171
maneuvers, time-optimal rotations, spacecraft control, mission planning, remote sensing, 16. PRICE CODE
remote sensing planning
17. SECURITY 18. SECURITY 19. SECURITY 20. LIMITATION OF
CLASSIFICATION OF CLASSIFICATION OF THIS CLASSIFICATION OF ABSTRACT
REPORT PAGE ABSTRACT
Unclassified Unclassified Unclassified UU
NSN 7540-01-280-5500 Standard Form 298 (Rev. 2-89)
Prescribed by ANSI Std. 239-18

i
THIS PAGE INTENTIONALLY LEFT BLANK

ii
 Approved for public release. Distribution is unlimited.

A MACHINE LEARNING APPROACH TO ENABLE MISSION PLANNING OF

TIME-OPTIMAL ATTITUDE MANEUVERS

Reed Ransom Smith

Lieutenant Commander, United States Navy
BA, Duke University, 2008
M, Systems Analysis, Naval Postgraduate School, 2016

Submitted in partial fulfillment of the

requirements for the degree of

MASTER OF SCIENCE IN ASTRONAUTICAL ENGINEERING

from the

NAVAL POSTGRADUATE SCHOOL

September 2020

Approved by: Mark Karpenko

Co-Advisor

Brian M. Wade
Co-Advisor

Garth V. Hobson
Chair, Department of Mechanical and Aerospace Engineering

iii
THIS PAGE INTENTIONALLY LEFT BLANK

iv
 ABSTRACT

Time-optimal spacecraft rotations have been developed and implemented on

orbiting spacecraft, highlighting opportunities for improving slew performance.
Double-digit reductions in the time required to slew from one attitude to another have
been demonstrated. However, the ability to perform mission planning to make use of
minimum time slewing maneuvers is largely precluded by the need to compute a
numerical solution to find a single minimum time maneuver control trajectory. Machine
learning approaches can eliminate the need to generate problem solutions by
approximating time-optimal maneuver times with sufficient accuracy for planning using
only the initial and final attitude requirements. The advantages of time-optimal spacecraft
maneuvers, a planning construct for evaluating legacy and machine learning maneuver
time generators, and the machine learning processes that enable this approach are
outlined. Compared to legacy planning techniques, time-optimal slew approximations
yield target collection increases of 3% to 24% for an example planning framework.

v
THIS PAGE INTENTIONALLY LEFT BLANK

vi
 Table of Contents

1 Introduction 1
1.1 Remote Sensing Planning . . . . . . . . . . . . . . . . . . . . . 1
1.2 Current State of Practice . . . . . . . . . . . . . . . . . . . . . 4
1.3 Enabling Greater Sensing Efficiency. . . . . . . . . . . . . . . . . 5
1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Kinematics of Sensing from Space 7

2.1 Orbital Mechanics . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Rotational Kinematics . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Spacecraft Rotations . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Development of a Mission Path Planning Algorithm 29

3.1 Problem Classification . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Developing the Remote Sensing Iterative Planner . . . . . . . . . . . 37
3.3 Remote Sensing Iterative Planner Mission Models . . . . . . . . . . . 53
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 Time-Optimal Maneuvers 57
4.1 Parameters of Example Spacecraft . . . . . . . . . . . . . . . . . 58
4.2 Dynamic Modeling . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Trajectory Optimization Problem . . . . . . . . . . . . . . . . . . 62
4.4 Example Solution . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Approximating Optimal Slew Times Using Machine Learning 81

5.1 Supervised Learning for Maneuver Time Approximation. . . . . . . . . 83
5.2 Design of Experiments . . . . . . . . . . . . . . . . . . . . . . 84

vii
5.3 Minimum Time Rotation Surrogate Models . . . . . . . . . . . . . . 94
5.4 Mission Plan Analysis . . . . . . . . . . . . . . . . . . . . . . 109
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

6 Conclusion 121
6.1 Implications for Current and Future Missions . . . . . . . . . . . . . 121
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Appendix A Simulation Target Locations 123

Appendix B Optimal Control Nomenclature 125

Appendix C Spacecraft High Fidelity Model 127

Appendix D Supervised Learning Dataset 129

List of References 147

Initial Distribution List 151

viii
 List of Figures

Figure 1.1 Eigenaxis rotation, duration = 35.5 sec. . . . . . . . . . . . . . . 2

Figure 1.2 Time-optimal rotation, duration = 29.6 sec. . . . . . . . . . . . . 3

Figure 1.3 Over-subscription planning flow diagram. . . . . . . . . . . . . 6

Figure 2.1 Selected classical orbital elements. . . . . . . . . . . . . . . . . 9

Figure 2.2 Two reference frames, 𝐴 and 𝐵. . . . . . . . . . . . . . . . . . . 12

Figure 2.3 Principal axis rotation around 𝑎®3 . . . . . . . . . . . . . . . . . . 14

Figure 2.4 Body axis rotations, 3-2-1. . . . . . . . . . . . . . . . . . . . . . 15

Figure 2.5 Example eigenaxis rotation. . . . . . . . . . . . . . . . . . . . . 17

Figure 2.6 State space schematic for a remote sensing planner. . . . . . . . . 21

Figure 2.7 Acceleration-limited slew. . . . . . . . . . . . . . . . . . . . . . 22

Figure 2.8 Rate-limited slew. . . . . . . . . . . . . . . . . . . . . . . . . . 23

Figure 2.9 Taxicab rotation, duration = 46.8 sec. . . . . . . . . . . . . . . . 26

Figure 3.1 Over-subscription planning flow diagram (Mission Planner). . . 30

Figure 3.2 Best performing population member in the first generation. . . . 40

Figure 3.3 Labeled targets (cities in the traditional TSP discussion). . . . . . 44

Figure 3.4 Genetic algorithm performance, minimizing distance between cities. 45

Figure 3.5 Genetic algorithm performance, minimizing travel time. . . . . . 47

Figure 3.6 Fastest time and shortest distance solutions superimposed. . . . . 48

Figure 3.7 Best-time constrained collection plan, final generation by shortest

travel time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

Figure 3.8 Targets with collection value assigned. . . . . . . . . . . . . . . 51

ix
Figure 3.9 Time constrained, value maximizing genetic algorithm solution. . 52

Figure 3.10 Spacecraft mission plan using taxicab slews, value maximizing. . 54

Figure 3.11 Spacecraft mission plan using eigenaxis slews, value maximizing. 55

Figure 4.1 Over-subscription planning flow diagram (Engineering Model). . 58

Figure 4.2 Scaled quaternion states and corresponding costates. . . . . . . . 72

Figure 4.3 Scaled angular rate states and corresponding costates. . . . . . . 73

Figure 4.4 Control trajectories and control torques mapped to reaction wheels. 73

Figure 4.5 Time lapse of slewing maneuver. . . . . . . . . . . . . . . . . . 74

Figure 4.6 Verified state trajectories in engineering units. . . . . . . . . . . 75

Figure 4.7 Hamiltonian as a function of time. . . . . . . . . . . . . . . . . . 76

Figure 4.8 Reaction wheel torque trajectories and corresponding covectors. . 77

Figure 4.9 State variable path constraint plotted with its covector. . . . . . . 78

Figure 4.10 Scaled path constraints and constraint covectors. . . . . . . . . . 80

Figure 5.1 Over-subscription planning flow diagram (Slew Time Generator). 82

Figure 5.2 Example Latin Hypercube. . . . . . . . . . . . . . . . . . . . . . 87

Figure 5.3 Input data set eigenaxes. . . . . . . . . . . . . . . . . . . . . . . 88

Figure 5.4 Fit of final maneuver time by relative quaternion input data. . . . 90

Figure 5.5 Fit of final maneuver time by Euler parameter input data. . . . . . 91

Figure 5.6 Input data set histograms. . . . . . . . . . . . . . . . . . . . . . 92

Figure 5.7 Relative quaternion DoE histograms. . . . . . . . . . . . . . . . 93

Figure 5.8 Example neural network with two hidden layers. . . . . . . . . . 95

Figure 5.9 Hyperbolic tangent sigmoid activation function. . . . . . . . . . 96

Figure 5.10 Neural network validation set performance. . . . . . . . . . . . . 100

x
Figure 5.11 Neural network averaged predictions versus target outputs for test
data set, 𝑅 2 = 0.998. . . . . . . . . . . . . . . . . . . . . . . . . 100

Figure 5.12 Feed-forward neural network with two, 13-node hidden layers. . . 102

Figure 5.13 Test data predicted vs. target outputs. . . . . . . . . . . . . . . . 109

Figure 5.14 Mission plan using neural network optimal slew approximations,
value maximizing. . . . . . . . . . . . . . . . . . . . . . . . . . 110

Figure 5.15 Mission plan using stepwise regression optimal slew approximations,
value maximizing. . . . . . . . . . . . . . . . . . . . . . . . . . 111

Figure 5.16 Average number of targets collected, by planning method and space-
craft altitude. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

Figure 5.17 Average mission value collected, by planning method and spacecraft
altitude. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

Figure 5.18 Mission time residual error, by planning method. . . . . . . . . . 115

Figure 5.19 Test data predicted vs. target outputs over small angles. . . . . . . 116

Figure 5.20 Mission residuals from optimal execution, by method. . . . . . . 117

Figure A.1 Simulation targets. . . . . . . . . . . . . . . . . . . . . . . . . . 123

xi
THIS PAGE INTENTIONALLY LEFT BLANK

xii
 List of Tables

Table 2.1 Slew time comparison by maneuver for a representative rotation. . 25

Table 3.1 Typical genetic algorithm population parameters. . . . . . . . . . 42

Table 3.2 Comparison of shortest distance and fastest time methods. . . . . 46

Table 3.3 Average slew time by maneuver. . . . . . . . . . . . . . . . . . . 56

Table 5.1 Neural network predicted time error with relative quaternion inputs. 99

Table 5.2 Neural network predicted time error with Euler parameter inputs. . 101

Table 5.3 Stepwise multiple linear regression terms. . . . . . . . . . . . . . 106

Table 5.4 Linear regression terms. . . . . . . . . . . . . . . . . . . . . . . . 108

Table 5.5 Targeting gains resulting from optimal approximation techniques. 115

Table 5.6 Value gains resulting from optimal approximation techniques (VU). 116

Table 5.7 Summary statistics for small angle test set. . . . . . . . . . . . . . 117

Table 5.8 Floating-point operations required for maneuver time methods. . . 118

Table A.1 Simulation targets. . . . . . . . . . . . . . . . . . . . . . . . . . 124

Table D.1 Supervised learning labeled data, 𝑛 = 500. . . . . . . . . . . . . . 129

xiii
THIS PAGE INTENTIONALLY LEFT BLANK

xiv
 List of Acronyms and Abbreviations

DCM direction cosine matrix

DoE Design of Experiments
DU distance units
FLOP floating-point operation
GA genetic algorithm
GCI Geocentric Inertial
LVLH Local Vertical/Local Horizontal
RSIP Remote Sensing Iterative Planner
TRACE Transition Region and Coronal Explorer
TSP Traveling Salesman Problem
TU time units
VU value units

xv
THIS PAGE INTENTIONALLY LEFT BLANK

xvi
 Acknowledgments

I would like to express my gratitude for the considerable direction and support provided to
me by my advisors, Professor Mark Karpenko and Professor Brian Wade. Professor
Karpenko’s extensive experience and ingenuity in this area offered me the opportunity to
explore a critical and timely topic that can directly improve the efficiency of space
systems. Professor Wade has been exceedingly generous with his time and expertise,
which has introduced me and many of my colleagues to machine learning as a discipline,
from which I expect to benefit for many years.

I would also like to thank my colleague and classmate Lieutenant Benjamin Diehl for his
assistance and collaboration with the development of our example spacecraft’s kinematics
and graphics.

I have benefited immensely from the instruction and collaboration here at the Naval
Postgraduate School, and for that, I am grateful.

xvii
THIS PAGE INTENTIONALLY LEFT BLANK

xviii
 CHAPTER 1:
Introduction

Remote sensing from space provides a useful service to a growing number of customers,
however the barriers to entry remain high. Satellite, launch, and mission operations costs are
significant while the industry becomes increasingly crowded. To excel, satellite operators
are well-incentivized to pursue increases in efficiency in their operations. This could mean
more bandwidth, higher platform availability, or more sensing products. Increasing cost
efficiency could be the result of taking pictures more quickly or decreasing overhead time
where sensing is not occurring. This thesis focuses on reducing unproductive time by
incorporating minimum time maneuvers into operations.

1.1 Remote Sensing Planning

Slewing a sensor from one target to the next target with traditional optics systems requires a
sensor attitude change, assuming the field of view is fixed to a boresight axis. For spacecraft,
this process is frequently accomplished by fixing the boresight to an axis of the vehicle and
rotating the entire vehicle until the desired alignment is achieved.

An intuitive method to slew a spacecraft from one attitude to another as quickly as possible
is to rotate about the axis that is perpendicular to both the starting and final boresight
vectors. This effectively traces a direct arc with the sensor’s boresight across a flat target
surface; an example of this maneuver is depicted in Figure 1.1. However, such a maneuver
(an eigenaxis maneuver) is not the fastest way to accomplish the rotation [1].

Real spacecraft have non-uniform mass properties and can maneuver more quickly around
certain axes of rotation and less quickly around others. These characteristics are exploited
by dynamic optimization to produce maneuvers that are more time-efficient. This type of
maneuver was first demonstrated on the NASA Transition Region and Coronal Explorer
(TRACE) spacecraft on August 10, 2010 [2]. To illustrate the potential gain, the same
sample maneuver shown in Figure 1.1 was accomplished using a time-optimal control
trajectory, depicted in Figure 1.2. The time-optimal maneuver is accomplished in 29.6
seconds, which is 16.6% faster than the original eigenaxis maneuver.

1
 Figure 1.1. Eigenaxis rotation, duration = 35.5 sec.

While this seems like an obvious technique to exploit, developing and implementing a time-
optimal slewing maneuver is only half the battle. To be useful, time savings from optimal
maneuvers must be incorporated into planning algorithms, which presents a problem in
the current state of practice, as spacecraft slew plans can include hundreds or thousands
of attitudes and slews. Although computing a single time-optimal slew can be done quite
quickly (tens of seconds), planning algorithms must solve a combinatorial optimization to
obtain a high value plan. In such a system, the number of optimal slews that would need to
be determined to produce a single mission plan would be on the order of millions or more
because of the iterative nature of the planning algorithm. Thus, incorporating the solution
of time-optimal maneuvers as part of the planning process is impractical for an operational
system.

The TRACE demonstration produced maneuver time reductions ranging from 5% to 21%
[2]. However, the ability to develop a realistic mission plan based on these capabilities
remains out of reach because of the scale of the computations required to combine planning
algorithms and optimal slewing maneuver calculations.

2
 Figure 1.2. Time-optimal rotation, duration = 29.6 sec.

Machine learning can bridge this gap today. While the maneuvers discussed for the TRACE
mission required high fidelity, flight quality solutions, a planning algorithm need only
evaluate the cost of accomplishing a given maneuver. In the case of an imaging mission
planner, a fitness function is used to evaluate how well a given parameter accomplishes its
objective. For mission planning, the fitness of a given maneuver sequence will be evaluated
by summing the individual maneuver times; faster sequences that accomplish the same
purpose are favored.

To iterate a mission plan, the precise engineering solution for a time-optimal slew need not
be available. The algorithm must provide maneuver times with sufficient accuracy for the
fitness function to correctly distinguish the relative fitness of various candidate maneuvers.

Many mission planning routines use legacy maneuvers that are inefficient to fly but easy to
model. If the optimal slew time could be efficiently approximated by a machine learning
tool, then that approximated time may suffice to generate mission plans that incorporate
efficiencies gained from optimal slewing maneuvers.

Machine learning models are lightweight (i.e., fast running) approximations of computation-
ally intensive engineering models. These are called metamodels and may be the mechanism

3
required to sensibly combine iterative planners and optimal maneuvers.

Assuming such a machine learning model could be created, what kind of accuracy would
be required for a useful product? If TRACE demonstrated maneuver time reductions of
5% to 21%, it is reasonable to infer that accuracy to within the low single digits would be
sufficient to implement a mission plan that incorporates optimal slews and yields overall
greater efficiency for a remote sensing enterprise, even if slight inaccuracies are introduced
in the planning phase by the machine learning metamodel. Any remaining discrepancies
between the approximated mission plan and flight-fidelity mission plan would be trivial
compared to the efficiencies gained.

Once a mission plan is generated for time-optimal slews, the full engineering model provides
the accuracy required to generate a flight-fidelity solution, without running computationally
intensive calculations during the iterative planning process. Flight-quality solutions can be
generated by the engineering model after the fact, with the broad mission plan in hand. Even
with slight inaccuracies in the approximated slew times when compared to the full-fidelity
engineering model slew times, the overall mission plan time is likely to approach the true
optimal time as the number of included slews increases.

To determine whether or not machine learning models provide a sufficiently accurate optimal
slew time, a common framework for comparison is required. Once this framework is created,
mission plans can be generated using legacy techniques as well as approximated optimal
slew techniques. Those approximated mission plans can then be evaluated against flight-
fidelity mission plans for accuracy and potential time savings.

1.2 Current State of Practice

Dimensionality associated with these problems leads operators to simple, suboptimal plan-
ning routines augmented heuristically by analysts to improve performance or remove ob-
viously detrimental results. “Traditional planning and scheduling relies on humans to
pre-program a set of actions to accomplish a pre-planned task or series of tasks” [3]. These
human-in-the-loop systems would naturally tend to slow or delay the planning process, and
would preclude evaluating hundreds, thousands, or more potential options that are required
to find an optimal or near-optimal solution to the problem. Frequently, there are insufficient

4
resources to collect all potential value for a given problem. For remote sensing missions, this
simply means that there is not enough time and assets to collect all the imagery desired. This
type of problem is called an over-subscription problem. One approach to over-subscription
problems is the greedy approach, in which the planner simply schedules the highest priority
collection first, and proceeds down the list in priority order. “If the observation being con-
sidered is still possible, it is added to the schedule; otherwise it is discarded. This approach
can work reasonably well for problems in which the cost of achieving an objective does not
depend on the order in which the objectives are achieved” [4]. This, of course, is not the
case for an overhead collection vehicle attempting to find the most efficient path through a
list of possible objects of varying value.

1.3 Enabling Greater Sensing Efficiency

The task at hand then is to develop a planning method that incorporates optimal slewing
maneuvers into a mission plan while being executable with computing resources available
today. The flow for such a planner is depicted in Figure 1.3. Currently, over-subscription
planners are widely available [5] and engineering models that provide time-optimal slew
control trajectories have been demonstrated [2]. The “Slew Time” function block that
provides slew times to the “Path Planner” is currently being accomplished using legacy
techniques. Two of those legacy techniques are the taxicab method and the eigenaxis
method, both of which will be discussed further in Chapter 2.

Augmenting the Slew Time function block with the capability to provide optimal slew times
can lead to significant gains in mission efficiency. However, given the limits of modern
processors, instead of attempting to determine the precise optimal slew time for a candidate
maneuver (which would be required millions of times for a relatively simple mission plan),
a machine learning metamodel can be used to approximate the optimal slew time, which
allows the iterative planner to quickly determine whether that maneuver should be saved for
further consideration or discarded as relatively inefficient.

For early planning, control trajectories and maneuver details are irrelevant—only the ma-
neuver time is required. If this metamodel provides approximate maneuver times with
sufficient accuracy to generate feasible mission plans, then the demonstrated 5% to 21%
gains in efficiency for a single maneuver can be extended to an enterprise scale.

5
 Mission Parameters
Flight path
Sensor characteristics
Target locations

Mission Mission Sequence Mission time

Flight Model
Planner E.g., 1, 4, 8, 11, 16 Control trajectories

Vehicle location Legend

Slew time
Target locations
Iterative Operations
Maneuver Generator
High Fidelity Operation

Figure 1.3. Over-subscription planning flow diagram.

1.4 Thesis Outline

A method for incorporating time-optimal rotations into the mission planning process will be
developed in the following chapters. Chapter 2 will discuss the mechanics of remote sensing
from space, including the limits of legacy techniques. Chapter 3 will describe the use of
a genetic algorithm to solve the over-subscription problem for a remote sensing mission
plan with targets of varying values. Chapter 4 will address the specifics of producing
a mathematically optimal maneuver, which is used to generate labeled data for training
machine learning models and for generating high fidelity control trajectories. Finally,
Chapter 5 will describe the techniques used to generate fast-running machine learning
metamodels to approximate optimal slew times at speeds that are practical for inclusion in
an iterative mission planner.

6
 CHAPTER 2:
Kinematics of Sensing from Space

This chapter will introduce and discuss the orbital mechanics and attitude rotations required
to orient a spacecraft for remote sensing operations. First, the orbital mechanics required
to predict a spacecraft’s velocity relative to Earth as a function of orbital altitude are
introduced. Next, multiple methods for determining a spacecraft’s attitude will be discussed,
including direction cosine matrices, Euler angles, and quaternions. Methods to describe
kinematic rotations of a spacecraft are identified, including basic methods for determining
the kinematics required to achieve a given orientation relative to Earth. Finally, the state
space kinematics model used in the development of this planner will be defined for future
inclusion in a mission planning program.

2.1 Orbital Mechanics

Newton’s first law of motion states that a body will remain in its state of rest or motion
until it is acted upon by an outside force to change that state. With knowledge of planetary
motion and the existence of satellites in orbit around the Earth, it is apparent that some
external force is acting upon orbiting bodies.

Law of Gravity The magnitude of that force is described by Newton’s Law of Gravity

𝐺𝑚 1 𝑚 2
𝑓 = (2.1)
𝑟2

where, 𝐺 is the universal gravitational constant, 𝑚 1 and 𝑚 2 are the masses of the two objects
(in our case, the mass of Earth and the mass of the orbiting satellite), and 𝑟 is the radius
between the two points. The gravitational acceleration of an object with respect to the Earth
is then
𝐺 𝑀⊕
𝑔= (2.2)
𝑟2
where, 𝑀⊕ is the mass of the Earth. Equation (2.2) describes the acceleration acting
radially on an object in Earth’s orbit between that object and Earth. This acceleration varies
as a function of the radius from the point mass of Earth to the orbiting object. With this

7
information, it is apparent that the speed with which an object accelerates towards the body
that it is orbiting is a function only of its distance from that body, assuming that the mass
of the orbited body is fixed.

Classical Orbital Elements The radius of an orbit need not be constant, in fact constant
radius circular orbits are a special case of elliptical orbits. Very few orbits are truly circular,
however circular orbits are useful and benefit from some simplified dynamics. From
these basic principles, six classical orbital elements can be derived. The orbital elements
completely describe any orbit and are shown in Figure 2.1. Further information can be
found in Wie’s textbook [6]. The six elements are
𝑎 = semimajor axis
𝑒 = eccentricity
𝑡 𝑝 = time of perigee passage
Ω = right ascension longitude of the ascending node
𝑖 = inclination of the orbit plane
𝜔 = argument of the perigee.
These elements provide all the information required to determine the radius and velocity of
an Earth orbiting vehicle. Altitude above the Earth’s surface can easily be inferred from the
average radius to the center of the Earth.

2.1.1 Elliptical Orbits

The two cases of closed orbits are elliptical orbits and circular orbits, of which elliptical
orbits are more generic. Circular orbits are a subset of elliptical orbits with eccentricity
𝑒 = 0. Before considering circular orbits, consider the information necessary to obtain the
altitude and velocity of a spacecraft at a given point in its elliptical orbit.

One method to define the position of the spacecraft is its true anomaly, 𝜃, which is the
angle between the argument of perigee and the current position of the spacecraft. The true
anomaly vertex is coincident with the main focus of the ellipse (which is the position of the
planet being orbited). Assuming that a spacecraft’s true anomaly and its orbital elements
are known, the spacecraft’s radius at a given true anomaly can then be obtained by

𝑎(1 − 𝑒 2 )
𝑟= . (2.3)
1 + 𝑒 cos 𝜃

8
 𝑧

ℎˆ

𝜃
Ecliptic
Plane 𝜔

Ω 𝑦

Line of Nodes

𝑥
Figure 2.1. Selected classical orbital elements.

To determine the spacecraft’s altitude above the surface, 𝑧, simply subtract the radius of the
Earth.
𝑧 = 𝑟 − 𝑅⊕ (2.4)

The magnitude of the spacecraft’s velocity in an elliptical orbit is a function of the distance
between the two objects. Where 𝜇 = 𝐺 𝑀⊕ , the velocity of Earth orbiting satellites is
determined by s 
2 1

𝑣= 𝜇 − . (2.5)
𝑟 𝑎

With these relationships in hand, a known orbit’s altitude and velocity information can
be determined for a mission planning algorithm. However, accounting for the changing
altitudes and velocities associated with elliptical orbits adds unnecessary complications for
a concept demonstration. Circular orbits provide a simpler mission framework to exercise
potential planning algorithm improvements.

9
2.1.2 Circular Orbits
For a spacecraft in a circular orbit, altitude and velocity determination is somewhat sim-
plified. For a circular orbit, 𝑒 = 0. The formula for the radius at a given point then
becomes
𝑟 = 𝑎. (2.6)

The constant orbital radius is a helpful new feature of circular orbits, and assuming a
spherical Earth (which is not true, but is an acceptable simplification for this purpose), the
spacecraft’s altitude is now constant.

Now that 𝑟 = 𝑎 for circular orbits, orbital velocity becomes

r
𝜇
𝑣= . (2.7)
𝑎

The length of the semimajor axis and the standard gravitational parameter 𝜇 are constant
for a given orbit, therefore a constant velocity is also a feature of circular orbits.

Equations (2.6) and (2.7) highlight the advantages of circular orbits where they can be used.
Many remote sensing spacecraft are in circular or nearly circular orbits to take advantage
of fixed distance and velocity parameters for the sensor. This can be advantageous for
uniformity of perspective or consistency in the resolution of an imaging product.

2.1.3 Flat Earth Model

For circular orbits in which vehicle altitude and velocity is constant, spacecraft position
relative to Earth can be effectively modeled over sufficiently small areas using a flat Earth
perspective for low earth orbit. This assumption significantly simplifies the planning
problem by allowing the use of polar coordinates to be neglected. This approach was
chosen for this thesis because it meets all the required attributes for determining relative
geometry of a spacecraft and sensing targets on the surface.

2.2 Rotational Kinematics

In addition to a spacecraft’s position relative to a target, information about the spacecraft’s
attitude compared to a known reference frame is essential for modeling and performing

10
remote sensing operations. There are multiple reference frames available for considering
the attitude of orbiting vehicles. For Earth-centered applications, the most encompassing of
these frames is the inertial reference frame, which is Earth-centered. The inertial reference
frame is sometimes referred to as Geocentric Inertial (GCI). Its 𝑧-axis is defined as the
Earth’s pole, and the 𝑥-axis is the direction from Earth to the Sun on the first day of Spring,

also referred to as the vernal equinox, . While this is not truly an inertial frame thanks to
the precession of the Earth’s axis around the ecliptic pole, it is sufficient for most practical
purposes.

While the GCI provides a useful reference frame for orbit problems, the local horizontal or
Local Vertical/Local Horizontal (LVLH) reference frame is useful to consider a satellite’s
attitude relative to the Earth’s surface. This reference frame fixes the 𝑧-axis to the nadir, or
directly below the spacecraft. The 𝑥-axis is aligned in the direction of the velocity vector,
perpendicular to nadir. This reference frame is sometimes referred to in terms of roll, pitch,
and yaw for rotations around the 𝑥, 𝑦, and 𝑧 axes, respectively.

“Typically, we choose the Earth’s center as the origin for problems in orbit analysis or
geometry on the Earth’s surface and the spacecraft’s position for problems concerning the
apparent position and motion of objects as seen from the spacecraft or analysis of payload
observations” [7].

2.2.1 Transformation Matrices

With multiple reference frames ready for use and an agile spacecraft ready to maneuver, a
method of mathematically describing a vehicle’s rotation relative to one or all of these frames
is necessary. The mathematical mechanism that effects this change from one coordinate
system to another is a transformation matrix.

Generally, one right-hand orthogonal reference frame can be described in terms of another
right-hand orthogonal reference frame using the direction cosines, which are the cosines of
the angles between the vector and the coordinate axes used to describe it.

Figure 2.2 shows a reference frame 𝐵 with basis vectors { 𝑏® 1 , 𝑏® 2 , 𝑏® 3 } that can be described
in terms of a reference frame 𝐴 with basis vectors { 𝑎®1 , 𝑎®2 , 𝑎®3 } by using the direction cosines

11
 𝑎®3
𝑏®2
𝑏®3
𝐴

𝑎®2
𝑎®1 𝐵
𝑏®1
Figure 2.2. Two reference frames, 𝐴 and 𝐵.

𝐶𝑖 𝑗 = 𝑏®𝑖 · 𝑎® 𝑗 , taking advantage of the inner product to derive the following relationships.

𝑏® 1 = 𝐶11 𝑎®1 + 𝐶12 𝑎®2 + 𝐶13 𝑎®3 (2.8a)

𝑏® 2 = 𝐶21 𝑎®1 + 𝐶22 𝑎®2 + 𝐶23 𝑎®3 (2.8b)
𝑏® 3 = 𝐶31 𝑎®1 + 𝐶32 𝑎®2 + 𝐶33 𝑎®3 (2.8c)

The direction cosines can also be presented in matrix form.


 𝑏® 1   𝐶11 𝐶12 𝐶13
 


 𝑎®1 
 
 𝑎®1 
 
𝑏® 2  =  𝐶21 𝐶22 𝐶23  𝑎®2  = C𝐵/𝐴  𝑎®2  (2.9)
       
 
       

 𝑏® 3   𝐶31 𝐶32 𝐶33
 


 𝑎®3 
 
 𝑎®3 
 

When these operations are applied in sequence, the transformation matrix that results is
a direction cosine matrix (DCM), 𝐶𝑖 𝑗 . 𝐶𝑖 𝑗 is similar to the form C𝐵/𝐴 which more
   

specifically indicates the DCM for a transformation to 𝐵 from 𝐴.

 𝐶11 𝐶12 𝐶13 

 
(2.10)
 
C𝑖 𝑗 =  𝐶21 𝐶22 𝐶23
 

 
 𝐶31 𝐶32 𝐶33 
 

  −1   𝑇
Direction cosine matrices are orthonormal, meaning that 𝐶𝑖 𝑗 = 𝐶𝑖 𝑗 , which allows

12
for the following inverse operation.

 𝑎®1   𝐶11 𝐶21 𝐶31

  



 𝑏® 1 

 𝑎®2  =  𝐶12 𝐶22 𝐶32 𝑏® 2 (2.11)
     
  
     
 𝑎®3   𝐶13 𝐶23 𝐶33
  



 𝑏® 3 


Now, an arbitrary vector 𝐻® can be expressed in terms of reference frame 𝐴 and 𝐵.

𝐻® = 𝐻1 𝑎®1 + 𝐻2 𝑎®2 + 𝐻3 𝑎®3

(2.12)
= 𝐻10 𝑏® 1 + 𝐻20 𝑏® 2 + 𝐻30 𝑏® 3

Since 𝐻® is known and expressed in terms of 𝐴, Wie expresses the same vector using
reference frame 𝐵 as 𝐻® 0 [6].

𝐻10 ≡ 𝑏® 1 · 𝐻® = 𝑏® 1 · (𝐻1 𝑎®1 + 𝐻2 𝑎®2 + 𝐻3 𝑎®3 ) (2.13a)

𝐻20 ≡ 𝑏® 2 · 𝐻® = 𝑏® 2 · (𝐻1 𝑎®1 + 𝐻2 𝑎®2 + 𝐻3 𝑎®3 ) (2.13b)
𝐻30 ≡ 𝑏® 3 · 𝐻® = 𝑏® 3 · (𝐻1 𝑎®1 + 𝐻2 𝑎®2 + 𝐻3 𝑎®3 ) (2.13c)

Algebraic reorganization yields the familiar direction cosine matrix, C𝐵/𝐴 .

 𝐻®10   𝑏® 1 · 𝑎®1 𝑏® 1 · 𝑎®2 𝑏® 1 · 𝑎®3   𝐻®1   𝐻®1 

       
(2.14)
  ®
𝐻®20  =  𝑏 2 · 𝑎®1 𝑏® 2 · 𝑎®2 𝑏® 2 · 𝑎®3 𝐻®2  = C𝐵/𝐴 𝐻®2
     
    
       
 𝐻® 0   𝑏® 3 · 𝑎®1 𝑏® 3 · 𝑎®2 𝑏® 3 · 𝑎®3   𝐻®3   𝐻®3 
 3       

The same DCM that transforms orthogonal basis vectors of different reference frames also
transforms the representation of an arbitrary vector from one frame to another.

2.2.2 Euler Angles

Consider three special case DCMs that result from transforming a matrix by rotating it
about a single axis. Three simplified and useful DCM cases describe the rotations about

13
the first, second, and third axes of reference frame 𝐴.


 1 0 0 
C1 (𝜃 1 ) =  0 cos 𝜃 1 sin 𝜃 1  (2.15a)
 
 

 0 − sin 𝜃 1 cos 𝜃 1 

 cos 𝜃 2 0 − sin 𝜃 2 
C2 (𝜃 2 ) =  0 1 0 (2.15b)
 

 

 sin 𝜃 2 0 cos 𝜃 2 

 cos 𝜃 3 sin 𝜃 3 0 
C3 (𝜃 3 ) =  − sin 𝜃 3 cos 𝜃 3 0  (2.15c)
 
 

 0 0 1 

One method for orienting a body is to rotate it in succession about the axes of the body-fixed
reference frame; an example of such a rotation is provided in Figure 2.3. Wie calls this
a body-axis rotation and explains that the first rotation can be about any axis, the second
rotation is about either of the unrotated axes, and the third rotation is about either axis not
used for the second rotation.

Consider a rotation from reference frame 𝐴 to reference frame 𝐵 using three successive
rotations, first about the axis 𝑎®3 , then about 𝑎®2 , and finally about 𝑎®1 (shown in Figure 2.4).
Wie symbolically denotes this operation by C1 (𝜃 1 ) ← C2 (𝜃 2 ) ← C3 (𝜃 3 ).

𝑎®3
𝑏®3

𝑏®2
𝜓
𝐵 𝑎®2
𝜓
𝑎®1 𝑏®1
Figure 2.3. Principal axis rotation around 𝑎®3.

14
 𝑎®3
𝑎®003 𝑎®03
𝑎®3
𝑎®03 𝜃2
𝐴0

𝐴 𝐴
𝑎®02 𝑎®02
𝑎®1
𝑎®01 𝑎®2 𝑎®002
𝐴0 𝑎®2 𝐴00
𝜃3
𝑎®1 𝑎®01 𝑎®001

𝑎®3
𝑏®3 ®
𝑎300

𝜃1
𝐵 𝑏®2

𝐴 𝐴00
𝑎®1 𝑎®002
𝑎®2
𝑏®1
𝑎®001
Figure 2.4. Body axis rotations, 3-2-1.

The intermediate rotations described by this symbolic notation are

C3 (𝜃 3 ) : 𝐴0 ← 𝐴 (2.16a)
C2 (𝜃 2 ) : 𝐴00 ← 𝐴0 (2.16b)
C1 (𝜃 1 ) : 𝐵 ← 𝐴00 . (2.16c)

15
The elements of each of these rotations are:

 𝑎®01   cos 𝜃 3 sin 𝜃 3 0   𝑎®1   𝑎®1 

       
𝑎®02  =  − sin 𝜃 3 cos 𝜃 3 0 𝑎®2 = C (𝜃 ) 𝑎®2 (2.17a)
       
    3 3  
       

 𝑎®03  
  0 0 1 


 𝑎®3 


 𝑎®3 

 𝑎®001   cos 𝜃 2 0 − sin 𝜃 2   𝑎®01   0
𝑎®1 
      
𝑎®002 = 0 1 0 𝑎®02  = C2 (𝜃 2 )  𝑎®02  (2.17b)
       
  
       
 𝑎®003   sin 𝜃 2 0 cos 𝜃 2   𝑎®03   𝑎®03 
      
 𝑏® 1   1 0 0   𝑎®001   𝑎®001 
      
𝑏® 2  =  0 cos 𝜃 1 sin 𝜃 1 𝑎®002  = C1 (𝜃 1 )  𝑎®002  (2.17c)
       
  
       
 𝑏® 3   0 − sin 𝜃 1 cos 𝜃 1   𝑎®003   𝑎®003 
      

By algebraically combining these equations, the condensed mathematical transformation is

expressed:
 𝑏® 1   𝑎®1 
   
(2.18)
 ® 
 𝑏 2  = C1 (𝜃 1 )C2 (𝜃 2 )C3 (𝜃 3 )  𝑎®2 
 
   
 𝑏® 3   𝑎®3 
   
And therefore,
C𝐵/𝐴 = C1 (𝜃 1 )C2 (𝜃 2 )C3 (𝜃 3 ). (2.19)

The three defining angles for this process, 𝜃 1 , 𝜃 2 , 𝜃 3 , are called Euler angles. “In general,
Euler angles have an advantage over direction cosines in that three Euler angles determine a
unique orientation, although there is no unique set of Euler angles for a given orientation” [6].

2.2.3 Eigenaxis Rotation

Direction cosine matrices are particularly valuable for transforming a vector from one
reference frame to another; however, to describe the rotation between two vectors (both
within the same reference frame), Euler’s eigenaxis rotation is a useful tool. This method
makes use of an eigenaxis, 𝑒® or ê, which is a unit vector that is the axis of rotation, fixed both
in the body frame and in the inertial frame. The angle of the rotation about the eigenaxis,
𝜓, is the eigenangle. An example eigenaxis rotation is depicted in Figure 2.5.

To calculate the axis around which a body must rotate for a given maneuver without changing
reference frames, the cross product can be exploited to find the vector normal to the plane

16
 𝑎®3
𝑏®3
Eigenaxis
𝑏®2

Initial

𝑎®1

Final 𝑎®2
𝑏®1
Figure 2.5. Example eigenaxis rotation.

that contains the current vector (e.g., a) and the desired vector (e.g., b). This will yield the
eigenaxis (ê), rotating about which will provide a direct path for the maneuver.

ê = a × b (2.20)

With the axis of rotation obtained, the angle of rotation must also be calculated. This
eigenangle is obtained by exploiting the inner product of the two attitudes. The eigenangle
is indicated by 𝜓 and the magnitude of a vector is indicated by k k.

a · b = kak kbk cos 𝜓 (2.21a)

a·b
𝜓 = arccos (2.21b)
kak kbk

In this manner, both the eigenaxis and eigenangle can be obtained from knowledge of the
two original vectors.

17
2.2.4 Quaternions
Quaternions, invented by William Hamilton in the 19th century, are commonly used to
describe spacecraft rotations [6]. Quaternions are defined as


 𝑒 1 sin ( 𝜓/2) 

𝑒 2 sin ( 𝜓/2)
 
(2.22)
 
q =  .
 𝑒 3 sin ( 𝜓/2) 

cos ( 𝜓/2)
 
 
 

One parameterization of the direction cosine matrix in terms of quaternions is

 𝑞2 − 𝑞2 − 𝑞2 + 𝑞2 2(𝑞 1 𝑞 2 − 𝑞 3 𝑞 4 ) 2(𝑞 1 𝑞 3 + 𝑞 2 𝑞 4 ) 
 1 2 3 4 
C𝐵/𝐴 2 2 2
= C(q) =  2(𝑞 1 𝑞 2 + 𝑞 3 𝑞 4 ) −𝑞 1 + 𝑞 2 − 𝑞 3 + 𝑞 4
 2 2(𝑞 2 𝑞 3 − 𝑞 1 𝑞 4 )  . (2.23)


−𝑞 21 − 𝑞 22 + 𝑞 23 + 𝑞 24
 
 2(𝑞 1 𝑞 3 − 𝑞 2 𝑞 4 ) 2(𝑞 2 𝑞 3 + 𝑞 1 𝑞 4 ) 
 

Quaternions have some advantageous properties. The 2-norm of all quaternions is 1 (i.e.,
𝑞 21 + 𝑞 22 + 𝑞 23 + 𝑞 24 = 1). Additionally, they do not have geometric singularities as Euler
angles do, which makes them frequently preferable for aerospace applications. “Moreover,
quaternions are well suited for onboard real-time computation because only products and
no trigonometric relations exist in the quaternion kinematic differential equations. Thus,
spacecraft orientation is now commonly described in terms of quaternions” [6].

2.2.5 GCI Transformation

Wie describes the process of representing a vector 𝑟® with known GCI coordinates 𝑋 𝐼®+𝑌 𝐽® +
𝑍 𝐾® into a reference frame that is fixed to the orbit plane with its origin coincident with
the body being orbited. Such a reference frame is called the perifocal frame (𝑥, 𝑦, 𝑧) and
has basis vectors (®𝑖 , 𝑗®, 𝑘)
® [6]. To accomplish this, a series of three successive rotations are
used. This method uses a 3–1–3 method, C3 (𝜔) ← C1 (𝑖) ← C3 (Ω), taking advantage of
the orbital elements right ascension (Ω), inclination (𝑖), and argument of perigee (𝜔).

Wie defines the vector using intermediate rotations

𝑟® = 𝑋 0 𝐼®0 + 𝑌 0 𝐽®0 + 𝑍 0 𝐾® 0 = 𝑋 00 𝐼®00 + 𝑌 00 𝐽®00 + 𝑍 00 𝐾® 00 (2.24)

18
where, single prime (0) indicates components of the position vector in the first intermediate
reference frame and double prime (00) indicates the components in the second intermediate
reference frame.

Using the nomenclature from Figure 2.1, Wie computes the three successive rotations in
the following form.

 𝐼®0   cos Ω sin Ω 0   𝐼® 

    
 𝐽  =  − sin Ω cos Ω 0   𝐽®  (2.25a)
 ®0    
    
 𝐾® 0  
   0 0 1   𝐾® 
 
 𝐼®00   1 0 0   𝐼® 
  0 
  
 𝐽  =  0 cos 𝑖 sin 𝑖   𝐽®0  (2.25b)
 ®00    
    
 𝐾® 00   0 − sin 𝑖 cos 𝑖   𝐾® 0 
    
 ®𝑖   cos 𝜔 sin 𝜔 0   𝐼®00 
    
 𝑗  =  − sin 𝜔 cos 𝜔 0   𝐽®00  (2.25c)
 ®   
    
 𝑘®  
   0 0 1   𝐾® 00 
 

Here, the DCM is defined with the following elements.

𝐶11 = cos Ω cos 𝜔 − sin Ω sin 𝜔 cos 𝑖

𝐶12 = sin Ω cos 𝜔 + cos Ω sin 𝜔 cos 𝑖
𝐶13 = sin 𝜔 sin 𝑖
𝐶21 = − cos Ω sin 𝜔 − sin Ω cos 𝜔 cos 𝑖
𝐶22 = − sin Ω sin 𝜔 + cos Ω cos 𝜔 cos 𝑖
𝐶23 = cos 𝜔 sin 𝑖
𝐶31 = sin Ω sin 𝑖
𝐶32 = − cos Ω sin 𝑖
𝐶33 = cos 𝑖

19
And therefore the total transformation is described by

 𝑥   𝐶11 𝐶12 𝐶13   𝑋 

     
 𝑦  =  𝐶21 𝐶22 𝐶23 (2.26)
     
  𝑌 .
     
 𝑧   𝐶31 𝐶32 𝐶33   𝑍 
     

2.3 Spacecraft Rotations

Ross, Proulx, and Karpenko [8] have developed a state space model with an elementary
kinematics representation that is a suitable framework to model attitude changes for a remote
sensing mission planner operating over a flat Earth. Their state space model is reproduced
in Figure 2.6.

2.3.1 Kinematic State Space Model

As a basis for a three-dimensional spacecraft attitude model, tilt (𝜃) and pan (𝜙) angles
from the vehicle’s body-fixed Cartesian coordinate system must be defined. Tilt and pan
are defined as the angles from the body-fixed triad to the target (ℓ 1 , ℓ 2 , 0) along the 𝑏 1 and
𝑏 2 axes respectively. This assumes a staring-type sensor where the yaw of the boresight is
not relevant. Certain spacecraft sensors (e.g., pushbroom-type) must control yaw as well,
which is neglected for this purpose. The principle of superposition provides the following
basic relationship.
𝑥 ℓ1 ª
© ª®
­ 𝑦 ® + 𝑟 LOS = ­ ℓ 2 ®® (2.27)
©
­ ® ­ ®
« 𝑧 ¬ « 0 ¬

Applying kinematic principles, individual relationships are derived.

𝑥 + 𝑟 LOS sin 𝜃 = ℓ 1 (2.28a)

𝑦 + 𝑟 LOS sin 𝜙 cos 𝜃 = ℓ 2 (2.28b)
𝑧 − 𝑟 LOS cos 𝜙 cos 𝜃 = 0 (2.28c)

Algebraic rearrangement of these equations allows tilt and pan angles to be derived from

20
 (𝑥, 𝑦, 𝑧)
𝑛3

NPS
𝑏1

𝑏2
𝑏3

𝑛2 𝑟 LOS

(ℓ 1 , ℓ 2 , 0)

𝑛1
Figure 2.6. State space schematic for a remote sensing planner.
Adapted from [8].

known target and vehicle locations in the inertial frame.

tan 𝜃
𝑥+𝑧 = ℓ1 (2.29a)
cos 𝜙
𝑦 + 𝑧 tan 𝜙 = ℓ 2 (2.29b)

2.3.2 Kinematic Slew Time Derivation

An individual rotation is assumed to occur according to a known acceleration rate (𝛼) and
a maximum slew rate (𝜔max ), beginning from rest, 𝜔 = 0 rad/s. The vehicle rotates over an
angle, 𝜓.

There are two conditions that must be derived.

1. An acceleration-limited slew where coasting is not required.

2. A rate-limited slew where a coast phase at 𝜔max is required.

21
Acceleration-Limited Slew
For a rest to rest maneuver, half of the maneuver will be accelerating and half will be
decelerating. The maximum rotation for an acceleration limited maneuver occurs at the
critical angle, 𝜓crit , which corresponds to a peak angular velocity of 𝜔max , shown in Figure
2.7. The critical time, 𝑡crit , is the midpoint of any acceleration-limited maneuver.

𝜔
𝜔max

𝑡 crit 𝑡 𝑓 at 𝜓crit 𝑡

Figure 2.7. Acceleration-limited slew.

From rest, begin with the relationship

1
𝜓 = 𝛼𝑡 2 ,
2

then, the rotation time can be determined.

1 1 2
𝜓 = 𝛼𝑡 crit (2.30a)
2 2
r
𝜓
𝑡crit = (2.30b)
𝛼
r r
𝜓 4𝜓
𝑡𝑓 = 2 = (2.30c)
𝛼 𝛼

Rate-Limited Slew
In the event that 𝜓 > 𝜓crit , a coast period is required prior to deceleration (Figure 2.8).

22
 𝜔
coast
𝜔max

𝑡 crit 𝑡𝑓 𝑡

Figure 2.8. Rate-limited slew.

The total time required for the maneuver is the sum of the acceleration, coast, and deceler-
ation phases, Equation (2.31).

𝑡 𝑓 = 2𝑡crit + 𝑡coast (2.31)

𝜓 − 𝜓crit
𝑡coast = (2.32)
𝜔max

Apply the kinematic formulas relating angular acceleration and velocity to the critical
conditions.

𝜔max = 𝛼𝑡crit (2.33a)

𝜔max
𝑡crit = (2.33b)
𝛼

Using equations (2.30a) and (2.33b), the critical angle can be derived from a vehicle’s
kinematic parameters.

2
𝜓crit = 𝛼𝑡 crit (2.34a)
 𝜔 2
max
𝜓crit = 𝛼 (2.34b)
𝛼
2
𝜓crit = max
𝜔
(2.34c)
𝛼

From equations (2.31), (2.32), (2.33b), and (2.34c), the total maneuver time can be obtained.

23
 𝜔max 𝜓 − 𝜓crit
𝑡total = 2 + (2.35a)
𝛼 𝜔max
𝜔max 𝜓 𝜓crit
𝑡total =2 + − (2.35b)
𝛼 𝜔max 𝜔max
max𝜔2
𝜔max 𝜓
𝑡total =2 + − 𝛼 (2.35c)
𝛼 𝜔max 𝜔max
𝜔max 𝜓
𝑡total = + (2.35d)
𝛼 𝜔max

The resulting total time equation is a function of known elements and is used for computing
slew times that are rate-limited. Kinematically determining slew times based on vehicle
parameters will be used for two legacy planning techniques that serve as benchmarks for
current practices in this concept demonstration.

2.3.3 Legacy Rotation Methods

Two methods are commonly used to change a spacecraft’s attitude. The first method
discussed will be the taxicab method. It minimizes the 1-norm of the slew, similar to
driving through city blocks in a grid configuration. An example taxicab slew is provided
in Figure 2.9. Using this method, the spacecraft rotates only about its principal axes, in
sequence (i.e., moving in each axis direction in the surface plane sequentially). The second
method is the most common for modern spacecraft and is the eigenaxis method, which is
shown in Figure 1.1. An eigenaxis slew minimizes the 2-norm of the slew angle between
two points (i.e., the sensor boresight follows a direct arc from the current position to the
desired position).

For both of these methods, there is a nonlinear relationship between the moving vehicle and
the pan and tilt angles to a given target. Therefore, to simplify the calculation the following
process is used.

1. The time required to perform the slew is computed, assuming the slew is performed
from a static position at (𝑥, 𝑦, 𝑧).
2. Once the slew time is computed, the vehicle is advanced along the 𝑛1 axis according
to kinematics (𝑥 new = 𝑥 previous + 𝑣vehicle 𝑡slew ).

24
 3. Vehicle velocity is constant and directed along the 𝑛1 inertial axis, which is coincident
with 𝑏 1 body-fixed axis.
4. Slew time used for the cost function is recomputed from the attitude at the previous
vehicle position with the sensor boresight pointed at the previous target to the attitude
at the new vehicle position with the sensor boresight pointed at the new target.

Taxicab Method
Absent other circumstantial benefits, the taxicab method is an inefficient method of slewing
a sensor. This concept is extended to a spacecraft sensor by first maneuvering around one
spacecraft principal axis and then another in sequence, ultimately arriving at the desired
sensor orientation. In the event that there are power, design, or external limitations in a
system, this may be a desirable or necessary practice. From an efficiency perspective, it is
undesirable.

Eigenaxis Method
The eigenaxis slew is a simple method to improve slewing efficiency over taxicab slews by
maneuvering directly to the desired attitude. This rotation takes its name from the Euler
parameters that define a rotation [ 6]. While the eigenangle describes the minimum angular
distance to accomplish the desired attitude change, it does not infer any guarantees of time-
optimality [1]. Time-optimality of rotations will be explored in more detail in Chapter 4.
A brief comparison of the time required to accomplish an identical three-axis reorientation
by both legacy methods discussed here, as well as the time-optimal method introduced in
Chapter 1 is provided in Table 2.1. In context, the relative inefficiency of a taxicab rotation
and the potential advantages of a time-optimal reorientation are evident.

Table 2.1. Slew time comparison by maneuver for a representative rotation.

Method Time per Maneuver

Taxicab 46.8 s
Eigenaxis 35.5 s
Time-optimal 29.6 s

25
 Figure 2.9. Taxicab rotation, duration = 46.8 sec.

2.4 Summary
This section introduced orbital mechanics, rotational kinematics, as well as a model for
describing spacecraft rotations using multiple reference frames. These relationships are
the basis for computing slew times needed by a mission planning system. That planning
process will be developed in Chapter 3. Finally, this section also introduced two legacy
slewing techniques, as well as the mathematics required to compute slew times for each.

The eigenaxis technique assumes a kinematics model in which the vehicle can be accelerated
in any direction, up to a maximum angular velocity. This assumption is a simplification
and neglects significant realities including the physical properties of the spacecraft, which
time-optimal techniques will account for in Chapter 4. Taxicab rotations are further limited
to slews about a spacecraft’s principal body-axes, exacerbating the inefficiencies of the
eigenaxis method.

To their advantage, eigenaxis maneuvers are simple and direct. They are intuitive, provide
feasible results, and have been effectively used for decades for planning and on-orbit.
However, as the quality and reliability of spacecraft increase, enterprise efficiencies are
increasingly attainable and desirable.

26
Moving away from simple, suboptimal maneuvers is a necessary step for today’s respon-
sible stewards of satellite resources, given the availability of optimal control techniques.
The following chapters will propose a method for implementing these optimal control
approaches.

27
THIS PAGE INTENTIONALLY LEFT BLANK

28
 CHAPTER 3:
Development of a Mission Path Planning Algorithm

To take advantage of the rotational agility of imaging spacecraft, a process is needed to

select which targets to image and in what order. This planning process should be based
on the slewing maneuvers that would be required, with the goal of selecting a sequence of
operations that includes as much efficiency as possible. This would allow an organization to
take as many pictures as possible, given a finite amount of the time in which the spacecraft
may have access to an imaging target.

This chapter will classify the planning problem using combinatorial literature, describe the
challenges associated with combinatorial problems of this nature, and present an approach
to this planning problem that uses that uses genetic algorithm (GA) techniques. The
interactions of the planner relative to the larger problem are highlighted in Figure 3.1.

3.1 Problem Classification

Finding an optimal solution in a set of possible solutions is referred to as combinatorial
optimization. For mission planning problems, it is not feasible to simply try every combina-
tion because of the scale and computing limitations introduced in Chapter 1. Recall, for an
imaging target list of 31 potential targets, there are 31 factorial possible sequences through
the entire list without repetition. The formula for permutations quantifies this phenomenon

𝑛!
𝑃(𝑛, 𝑟) = (3.1)
(𝑛 − 𝑟)!

where there are 𝑛 distinct objects and 𝑟 is the number of objects selected [9]. Note that
this refers to a permutation where the order in which the objects are selected matters, as
opposed to a combination where the order of selection does not matter. For a mission
planning problem, the sensor will slew from one target to the next in a prescribed order; the
correct approach here is to analyze permutations.

29
 Mission Parameters
Flight path
Sensor characteristics
Target locations

Mission Mission Sequence Mission time

Flight Model
Planner E.g., 1, 4, 8, 11, 16 Control trajectories

Legend
Vehicle location
Slew time
Target locations Iterative Operations

Maneuver Generator High Fidelity Operation

Figure 3.1. Over-subscription planning flow diagram (Mission Planner).

3.1.1 Traveling Salesman Problem

To determine the total number of sequences through all 31 targets, simply calculate the
number of permutations of 31 objects while selecting 31:

31! 31!
𝑃(31, 31) = = = 8 × 1033 (3.2)
0! 1

Equation (3.2) refers to the solution space for a common combinatorial optimization problem
called the Traveling Salesman Problem (TSP). TSP describes a traveling salesman with
a list of cities to visit and who wishes to find the most efficient route through all of these
cities. This problem is classically addressed using the two-dimensional distance between
each city, with the solution being the shortest possible path that visits all cities.

Assume a situation in which a time or distance constraint prevents the traveling salesman
from visiting all cities. For example, assume only 11 cities may be visited. The number of

30
possible solutions is then
31!
𝑃(31, 11) = = 3 × 1015 . (3.3)
20!
While this number of solutions is significantly less than the full permutation result, is is still
a prohibitively large number of possibilities to solve naively; the author’s personal computer
has difficulty solving problems with permutations in excess of 11! = 4 × 107 .

3.1.2 Orienteering Problem

To further develop the problem at hand, we must consider the value attached to various
imaging targets. Assume some targets have greater value to the planner than others. With
this concept incorporated, the problem begins to resemble an orienteering problem, which
is a subset of the TSP. “In an orienteering problem, we are given a set of cities, a prize
for each city (possibly zero), and a network of roads between cities. The objective is for a
salesman to collect as much prize money as possible given a fixed amount of gas” [4]. So in
an orienteering problem, two concepts are introduced: constraints on the number of cities
that can be visited and non-uniform value associated with different cities. David Smith
proposes a methodology for selecting an optimal path in such an over-subscribed scenario,
which is a situation where the resources required to visit every city exceed the resources
available [4]. However, the number of nodes (cities) in his example is small, so an exact
solution can be computed.

In the spacecraft planning problem, the resource requirement for collecting against one
target is the same as any other. This may contrast to other problems, for example a Mars
rover planning problem where different science experiments may require different resource
allocation. The planning techniques introduced in this thesis have been refined into a
methodology the author refers to as the Remote Sensing Iterative Planner (RSIP).

Smith addresses the greedy approach to orienteering problems, in which the highest value
targets are collected first regardless of resource requirement. Greedy approaches repeat this
process until resources are exhausted, which is a simple and intuitive method. Unfortu-
nately, the greedy approach often does not yield the best results. For the remote sensing
applications, greedy planning would be an appealing approach, however the problem of
dimensionality precludes naively solving every possible solution for comparison. Smith’s
orienteering problems have only a few nodes, so all possible paths can be explored. That is

31
not possible for mission planning problems with the number of potential targets that must
be considered for remote sensing from space.

While the brute force approach will not provide insight, the principle of assigning values to
targets is relevant. RSIP attempts to achieve some of the advantages of the greedy approach
in a more advantageous way by incorporating the values of the targets in the style of an
orienteering problem using a fitness function, J . Fitness functions are the formulas for
calculating the “goodness” of a candidate solution, accomplished by defining a metric to
be minimized. Candidate solutions are evaluated against the fitness function throughout
an iterative solver. Fitness functions for a solver can be refined to improve performance,
resulting in an algorithm that converges on desirable solutions.

3.1.3 RSIP Fitness Functions

The spacecraft image planning problem has all the elements of an over-subscribed orienteer-
ing problem—a complex set of imaging targets with varying distances between them, time
constraints due to a fixed orbital velocity and field of view, and differing values assigned
to different imaging targets. Fitness functions are used to evaluate the goodness of each
potential solution during the course of the GA. Each version of the planner uses only one
fitness function; however, several versions were created to build to the capability of the final
version of RSIP.

In the first version, the fitness function (J ) minimized distance between all targets in each
target sequence
𝑛−1
Õ
J1 = 𝑑𝑖 (3.4)
𝑖=1

where 𝑛 is the number of cities and 𝑑 is the distance between two sequential cities in the
sequence.

Next, the approach was enhanced by changing the fitness function to minimize the time
required to travel between one city and the next. This allows for different vehicle speeds in
different directions.
𝑛−1
Õ
J2 = 𝑡𝑖 (3.5)
𝑖=1

32
where 𝑡 is the time required to travel between two sequential cities.

After accounting for travel time, the method of computing the travel time between cities was
changed from modeling two-dimensional travel to the time required to slew a spacecraft
sensor from one target to the next. This required an adjustment to the method of calculating
𝑡𝑖 , however the fitness function, Equation (3.5), was unchanged.

The fitness function was next adjusted to account only for the targets that could be collected
in the constrained environment, 𝑛𝑐 .

𝑛−1
Õ 𝑐 −1
𝑛Õ
J3 = 𝛼𝑡𝑖 + 𝛽𝑡 𝑗 (3.6)
𝑖=1 𝑗=1

This function is a multi-objective fitness function in contrast to the single objective fit-
ness functions in equations (3.4) and (3.5). To generate an efficient solution for a time-
constrained problem, generally efficient paths through the entire target set must be achieved
before time-constrained performance takes priority. Without this groundwork, the algo-
rithm consistently converges prematurely. This effect is achieved with the scalarization
method of multi-objective optimization [10]. The values of the scale factors 𝛼 and 𝛽 are
pre-determined and changing over the total number of generations; this will be discussed in
more detail in Section 3.2.5.

Finally, the multi-objective fitness function was adjusted to include value as a decision
variable in lieu of constrained travel time in the final version of RSIP. The formulation for
Equation (3.7) will be further developed in Section 3.2.6.

𝑛−1
Õ  𝑛𝑐
Õ 
(3.7)
 
J4 = 𝛼𝑡𝑖 + 𝛽 𝑣max − 𝑣 𝑗  ,
𝑖=1  𝑗=1 
 

where 𝑣max is the total value available in the target set, 𝑣 𝑗 is the value associated with a given
target and 𝛼 and 𝛽 are again scale factors.

This sequence of planning process development began with a simple Traveling Salesman
Problem construct that was solved with a genetic algorithm. From there, the use of travel time
as a decision variable was introduced, and finally orienteering problem characteristics were

33
included in the form of a multi-objective fitness function accounting for constrained time
and target collection value. The underlying genetic algorithm approach remained constant
through each of these versions. The implications of these functions will be discussed in
more detail in Section 3.2.

3.1.4 Genetic Algorithm Nomenclature

John Holland was one of the developers of the genetic algorithm in the 1960s and 1970s,
in an effort to model evolutionary behaviors outlined in Charles Darwin’s theory of natural
selection [5]. A GA attempts to leverage evolutionary traits found in biology and apply those
principles to other complex problems that seek to converge on an optimal solution. GAs
borrow much of their nomenclature from biology as well. A group of possible solutions
to a problem, or paths through a problem, is referred to as a population. The individual
solutions are called genomes and are encoded into chromosomes; therefore, a population is
made up of chromosomes.

A population evolves over generations, where each generation draws its population from
some combination of chromosomes from the preceding generation, and the results of genetic
modifications to the preceding generation’s makeup. Chromosomes are evaluated by some
fitness function, J , wherein the best performing chromosomes are selected to pass their
genetic makeup (their genomes) on to the next generation. The concept of passing high-
performing genomes into the next generation is referred to as elitism and depending on the
mechanisms chosen to build out future populations, elitism can be relatively weak or strong.

Some members of the next generation are generated through random changes to chromo-
somes. This occurs in two different ways: either crossover which combines portions of
different parent chromosomes, or mutation which randomly manipulates some portion of
a parent chromosome. The next generation is populated by way of selection, crossover,
and mutation, and the process repeats over successive generations (frequently hundreds
or thousands of generations) until an acceptable solution to the problem as been found,
typically by converging on a nearly optimal solution.

34
3.1.5 Properties of Genetic Algorithms
“Genetic algorithms have some distinct advantages for certain classes of problems, notably
their ability to handle complex problems and their use of parallelism—the ability to explore
the search space in many directions simultaneously” [5]. They are able to pursue multiple
solutions in parallel because individual chromosomes act independently from one another
and can explore the search space independently as well. Additionally, GAs can deal with
stationary fitness functions that remain static over time or with non-stationary fitness
functions that change with time.

Some problem attributes such as genome sequence and relative fitness are produced and
tracked by the algorithm. However, other elements of the algorithm must be specified by
the developer. For example, the fitness function, the structure of chromosomes, and the
rates of selection, crossover, and mutation must all be predetermined and coded into the
algorithm. As with any selection, there is the risk of selecting disadvantageous parameters.
For example, defining the correct fitness function is critical—a poor definition here may lead
to an excellent solution to a problem that the designer did not intend to solve. Similarly,
the exact mechanisms for crossover and mutation must be defined. The mechanism of
combining parent chromosomes for crossover must be defined, as well as the frequency and
extent of mutations. Additionally, the size of the population and the number of generations
over which the population will be evolved must be determined; too small and an optimal
solution may not be found, too large and excessive computing time may be required. Careful
design as well as trial and error are frequently required to create an effective GA. With this
in mind, it is an effective tool for efficiently generating solutions that converge on optimality.

3.1.6 Genetic Algorithm Design

In this section, the implications of genetic operators and population parameters will be
explored in more detail.

Genetic Operators
Genetic algorithms typically use three mechanisms to pass genes from one generation to
the next—selection, crossover, and mutation.

35
Selection New generations are populated with some portion of the highest performing
chromosomes from the previous generation. This selection is determined by the fitness
function where better performing chromosomes are preferred and it ensures that high fitness
solutions are included in subsequent generations. The concept of selection of the fittest is
referred to as elitism in the GA context. Algorithm designers decide to what degree selection
of high performing solutions are flowed into subsequent generations, however very strong
elitism could cause the algorithm to converge on a solution too soon, similar to finding
a local extremal solution instead of the global extremal solution. Conversely, excessively
weak elitism can delay converging on a solution. These concerns must be considered when
assigning weights and probabilities in the GA design process.

Crossover Crossover is typically the primary method of populating a new generation.

Crossover can be implemented in different ways. One option is to combine parent chro-
mosomes at a single point (i.e., the beginning portion belongs to one parent chromosome
and the ending portion belongs to the other). Another option is to use multiple swap
points, where the child chromosome is a collection of multiple segments from the parent
chromosomes [5].

Generally, crossover is the main method for making evolutionary progress. For a given
crossover chromosome, genetic segments are chosen from parent chromosomes that have
remained in the population due to their evolutionary progress. As the populations become
more efficient with later generations, it follows that the chromosomes within the population
as well as their progeny are also becoming more efficient. These actions enhance the chances
of finding increasingly efficient paths from desirable building blocks. This principle is the
basis for relying on the crossover component to aid convergence to a solution in a GA. More
precisely, the crossover operation provides “mixing of the solutions and convergence in a
subspace” [5].

Mutation While the goal of crossover is to aid convergence, Yang attributes increases in
the genetic diversity of a population to the mutation operation [5]. Probability of mutation
is typically much smaller than the crossover probability. Mutations commonly swap bits
or elements of a chromosome, however that could happen at only two sites, multiple sites,
or more complicated schemes could be developed. Mutation at multiple sites can add
evolutionary efficiency to a routine, however Yang cautions that excessive mutation can

36
make it difficult for a population to converge. Thanks to these injections of randomness into
genomes, mutation tends to find a solution outside of the subspace, and in that way expands
genetic diversity [5].

Population Parameters
Population size must be considered as a balance between having a sufficiently large popu-
lation to explore all of the relevant search space, and avoiding excessively large populations
that require extensive computing resources to evaluate each chromosome’s fitness. A pop-
ulation that is too small risks being caught in a premature solution if one chromosome
shows significantly better fitness than the others and then overwhelms all other chromo-
somes in a small population. Yang recommends a population size of 𝑛 = 40 to 200 for most
problems [5].

Consideration must also be given to the appropriate number of generations. One simple
method is to define a set number of generations and evaluate the performance over those
generations. If performance improvements diminish well before the end of the routine
(i.e., performance flattens out early), then fewer generations may be needed to arrive at an
acceptable solution. Conversely, if significant improvements continue to be found through
the final generation, then a convergence likely has not occurred and the routine may benefit
from additional generations.

In some cases, a stopping criterion could also be used. If performance improvements

drop below a certain threshold, or if a specified performance level has been reached, a
routine could be instructed to terminate. These approaches require sufficient knowledge of
the system to ensure that the stopping criteria is eventually reached. This may be feasible,
however for initial development of a GA, some amount of trial and error should be expected.

3.2 Developing the Remote Sensing Iterative Planner

Based on the discussion to this point, a TSP construct is a reasonable starting framework
for overhead remote sensing planning. For a set of targets or cities, the goal of this system
is to develop a path that travels through the cities as efficiently as possible. Two significant
differences from the classic TSP arise: there is no guarantee that an airborne collection plan

37
will be able to collect every potential target and there is no guarantee that all targets will be
equally valued. Regardless, a TSP approach is a promising starting point.

A naive, brute force method of solving a TSP is only feasible for very small target sets.
Generally, target sets of 10 to 12 targets can be brute forced with commonly available
computers; each additional node adds significant complexity. A TSP solution for an 11-
node mission plan where all possible paths are computed yields 4 × 107 possibilities. Add
one additional target and the total number possible permutations increases by an order of
magnitude to 5 × 108 . Quickly this becomes prohibitively complex.

When choosing an approach for solving a TSP-like problem for mission planning purposes,
we can take advantage of the fact that true optimality is not required. This opens the
decision space up to feasible solvers that do not guarantee optimality, but likely converge
on an optimal solution.

The technique employed should be well suited to the complexity resulting from the number
of permutations in a likely problem, and be able to search in multiple directions of the search
space simultaneously. A technique that begins with some subset of all possible paths, then
selects the more efficient solutions, and finally iterates those selected paths to find further
improvements is clearly advantageous. Ultimately, genetic algorithms excel at this type of
problem and that is the primary approach for this remote sensing mission planner, RSIP.

3.2.1 Features of the RSIP Genetic Algorithm

Several features of the GAs are well suited to this planning problem. This section will
explore these features and their specific implementation in RSIP.

Problem Complexity
The target list used for this thesis contains 31 elements. For context, the number of
targets that can be collected by this model from low earth orbit is typically in the range
5 to 15 for a given mission. To begin unscrambling a randomly initialized permutation,
the algorithm must consider the full permutation. From Equation (3.2), that search space
contains nearly 1034 possibilities. Even if reduced to 11 targets out of the 31 total candidates,
the number remains unreasonable at around 1015 possibilities, Equation (3.3). Assuming

38
this complexity precludes a brute force approach, a genetic algorithm continues to be an
appealing approach to the problem.

Search Space Exploration and Mutation

To determine whether a randomly generated permutation might be broadly efficient by
chance, the best performing genome in a randomly initialized population was plotted. Ex-
amining Figure 3.2, there is no apparent efficiency in this randomly generated permutation.
Heuristically, this was the case consistently. From this, it was inferred that convergence
to an efficient solution does not lie in the original randomly generated permutations. To
the contrary, mutation is required to expand the search subspace outside of the genetic
makeup of initial randomly generated genomes. It should be advantageous to rely heavily
on mutation to expand the search space, particularly given a population size of 100 and
reasonable solution spaces on the order of 1015 possibilities.

Crossover
Of the three genetic operators, crossover operations are problematic. Crossover relies on
combining elements of two parent chromosomes. In binary or combination operations, this
usually does not present a problem. However, for a image collection planning problem, each
chromosome is a permutation, and therefore repetition or skipped elements are unacceptable.
All chromosomes must be complete permutations of the range of cities. Therefore, crossover
is particularly problematic in this case and was not used. This leaves only selection and
mutation as the feasible genetic operators. Based on Yang’s analysis, we should be wary of
the algorithm failing to converge on a solution due to the lack of crossover. However, given
the heavy reliance on mutation to expand the search space beyond the initial population,
the infeasibility of crossover for this application may not be problematic. That is to say,
convergence is not the biggest challenge here; expanding the search space is the more
pressing concern.

Selection and Elitism

To balance elitism that preserves high performing chromosomes without stifling the benefits
of parallelism, a strict 25% selection parameter was used, but not over the entire population.
This was a design characteristic of Kirk’s approach that was adopted here [11]. Early trials

39
 Figure 3.2. Best performing population member in the first generation.

demonstrated that selection from the best performing subset of the entire population (i.e.,
the top 25% of 100 genomes) quickly eliminated almost all diversity in the population;
mutation was unable to provide sufficient balance.

In Kirk’s approach, the population is broken into groups of four chromosomes and the highest
performing chromosome within each group is selected. The other three are discarded, and
their places are taken by mutations of the selected chromosome. In this way, the diversity
associated with parallelism is enforced while still maintaining a strong sense of elitism to
ensure that well-performing members are maintained and only replaced if a better version
of that chromosome is created through mutation.

Parallelism
Structuring genetic competition for a place in subsequent generations into bouts of four
chromosomes is a key design choice. In early iterations of the GA, the entire population
consisted of very similar genomes within a few hundred generations. This early convergence
consisted of genomes that outperformed the genomes formed randomly at initialization, but
did not offer a path to a globally efficient solution and resulted in inefficient solutions
overall. Enforcing parallelism by instituting bouts of four for selection eliminated this early
convergence problem and allowed the algorithm to consistently progress toward a globally
optimal solution.

40
Stationary and Non-stationary Fitness Functions
The flexibility of GAs to handle different types of fitness functions was a key enabler
to developing a planner that applied GA concepts to a large scale orienteering problem.
Equations (3.4) and (3.5) are stationary fitness functions in that the metric used to evaluate
a chromosome’s relative fitness does not change during the course of the algorithm (i.e.,
from one generation to the next).

Equations (3.6) and (3.7) are examples of multi-objective non-stationary fitness functions.
Equation (3.7) is also the fitness function used in the final version of the mission planner.

𝑛−1
Õ  𝑛𝑐
Õ 
(3.7)
 
J4 = 𝛼𝑡𝑖 + 𝛽 𝑣max − 𝑣 𝑗 
𝑖=1  𝑗=1 
 

Here, the scalarization method is implemented with scale factors 𝛼 and 𝛽, which adjust over
time to emphasize the total path time objective of the fitness function early in the process and
the constrained value objective toward the end of the evolutionary process. To accomplish
this effect, the total path time scale factor, 𝛼, decrements linearly from one to zero over the
course of the GA, while the constrained value scale factor, 𝛽, increments linearly from zero
to one over the same period.

This fitness function results from the need to generate modestly efficient choices early in
the evolutionary process. Without generally efficient solutions in the population by the time
that constrained choices take precedence, the algorithm will not be able to converge on a
low time or high value solution. These relationships cause efficient paths through the entire
target list to be favored early on and high value solutions that satisfy algorithm constraints
to be favored in the later portion of the solver.

3.2.2 RSIP Design Choices

The first iteration of the Remote Sensing Iterative Planner relied on several simplifying
assumptions. This simplified planner was used to develop the general technique and validate
the genetic algorithm parameters chosen by the designer. The initial assumptions were

1. All targets are visited; time is not a constraint.

2. Travel from one target to the next occurs at the same speed in all directions.

41
 3. Travel from one target to the next occurs in a straight, two-dimensional line.

None of these assumptions will carry through to the final model, but all are assumed in
the first iteration of the GA. Several techniques for GA implementation were considered,
however all shared a few attributes fundamental to genetic algorithms:

1. The first generation of the algorithm is initialized with a population of 𝑛 samples of

randomly generated, unique permutations that represent possible collection plans.
2. A certain portion of the best performing members (or genomes) of the population are
selected for inclusion in the next generation.
3. Remaining members of the next generation are created through some combination of
parts of well-performing genomes, selected in the previous step. This is accomplished
by way of mutation of parent genomes.
4. Iterative selection of the best performers resulting from random genetic adjustments of
preceding generations is repeated until a satisfactory solution to the planning problem
is achieved.

This technique is flexible over many parameters chosen by the designer; however, typical
population parameters used in this planner are listed in Table 3.1.

Table 3.1. Typical genetic algorithm population parameters.

Parameter Value
Population Size 100
Iterations 1000
Imaging Targets 31

To guarantee the optimal solution using a brute force method, a 31-city mission plan would
require evaluating 31 factorial (8 × 1033 ) permutations, which is too many for any practical
system. By contrast, a feasible result is achieved by evaluating 105 (100 · 1000) different
permutations using the genetic algorithm. Here the advantages of the GA become apparent.

After considering various implementations of a GA, the essential elements of the RSIP GA
are:

42
 1. Population genomes are broken into groups of four.
2. The best performing genome of each group of four is selected for inclusion in the
subsequent generation, and the remaining three genomes are discarded.
3. The best performing genome is replicated and mutated to generate three child se-
quences to fill the remaining three spots in each group of four.
(a) One child sequence has a randomly selected portion of the gene reversed.
(b) One child sequence has two randomly selected elements swapped.
(c) One child sequence has a randomly selected portion shifted by one position,
keeping the elements’ relative order intact.
4. This iterative process is repeated for the desired number of generations.

3.2.3 Results of Genetic Algorithm Minimizing Distance

Using total path length as the fitness function (i.e., minimizing the two dimensional distance
over the permutation path through the set of targets presented in Figure 3.3), typical GA
results are presented in Figure 3.4. Decaying exponential progress is indicated, implying
that a near-optimal solution may have been obtained. This method does not guarantee
optimality, nor does it preclude the routine from becoming trapped at a local minimum,
however this result is feasible in that valid permutations are produced and those permutations
become more advantageous as the routine progresses.

This first iteration of the algorithm does not address movement in different directions at
different rates nor being time constrained and therefore unable to visit every city. These
features are added in the next versions of the GA.

43
Figure 3.3. Labeled targets (cities in the traditional TSP discussion).

44
 (a) Shortest path between all cities in the first generation.

(b) Shortest path between all cities in the 1000th generation.

(c) Best performing population member by generation.

Figure 3.4. Genetic algorithm performance, minimizing distance between
cities.

45
3.2.4 Results of Genetic Algorithm Minimizing Time
Finding the shortest two-dimensional path through a set of targets is certainly advantageous,
however in a spacecraft application a vehicle’s sensor will not necessarily move at equal
speeds in all directions. The GA can be modified to accommodate this.

To demonstrate the concept, the scenario from Figure 3.4 is repeated, but instead of mini-
mizing the total distance traveled, total time for the route is minimized. The case presented
in Figure 3.5 assumes that the traveling salesman can move along the vertical axis at twice
the speed that the salesman can move along the horizontal axis.

Compared to Figure 3.4, the time-minimizing parameters in Figure 3.5 show similar trends
in performance improvement, as well as similar heuristic differences from the best randomly-
generated genome in the first generation when compared to the best performer in the final
generation. The first generation is of course unimpressive, as it is a randomly generated
permutation. It is depicted in Figure 3.5a.

The total time minimizing solution is presented in Figure 3.5b. The fastest total time solution
superimposed on the shortest distance solution is presented in Figure 3.6 for comparison.
This figure highlights the tendency of the algorithm to favor travel in the vertical direction,
as there is a speed advantage to be exploited. Table 3.2 lists the time and distances for
the travel paths using generic distance units (DU) and time units (TU). The fastest time
approach finds its best solution over a much longer ground path than the shortest distance
approach.

Table 3.2. Comparison of shortest distance and fastest time methods.

Optimization Method Distance [DU] Time [TU]

Shortest Path 896.3 696.5
Fastest Time 1009.7 633.0

46
 (a) Shortest travel time between all cities in the first generation.

(b) Shortest travel time between all cities in the 1000th generation.

(c) Best performing population member by generation.

Figure 3.5. Genetic algorithm performance, minimizing travel time.

47
Figure 3.6. Fastest time and shortest distance solutions superimposed.

48
3.2.5 Results of Genetic Algorithm Minimizing Time with Time Con-
straints
Now that the genetic algorithm is configured to account for different rates along different
axes, overall time constraints must be addressed. Spacecraft are constrained by orbital
mechanics, therefore the altitude and access window for a remote sensing vehicle may not
permit the vehicle to collect all the desired targets. In fact, this is likely to be the case.
The mission planner must be able to account for time constrained scenarios in which only
a portion of the desired targets can be collected.

Practically, the planning algorithm must be able to generate a mission plan for an 𝑛 choose 𝑟
permutation where the number of targets collected (𝑟) is not strictly defined, rather enforced
by a maximum allowable mission time. Later, the mission time will be enforced as a
product of the spacecraft’s orbital elements. This means that the final version of RSIP will
determine maximum mission time by calculating the window in which the spacecraft falls
within an acceptable field of regard based on look angle constraints. Before building up to
that model, time is constrained as an operator input.

Figure 3.7 presents the results of a time constrained solution. In this case, the maximum
allowable mission time is 300 TU, chosen by the designer. The GA for this scenario operates
in largely the same manner as previously described, with an updated fitness function,
Equation (3.6).
𝑛−1
Õ 𝑐 −1
𝑛Õ
J3 = 𝛼𝑡𝑖 + 𝛽𝑡 𝑗 (3.6)
𝑖=1 𝑗=1

In this scenario, 17 out of 31 targets are collected within the 300 TU constraint.

The different paths and travel times when minimizing time versus distance inform a key
takeaway for mission planning programs. Adopting a simplifying assumption that travel
distance is a reasonable proxy for image collection planning comes with significant costs.
In Table 3.2, the time calculations assume identical velocity capabilities for the shortest
distance and fastest time approaches, however the shortest distance approach is 10% less
efficient than the approach that takes vehicle speed into account.

As spacecraft generally do not have uniform agility, extending the shortest distance assump-
tion to spacecraft maneuvers would have similarly detrimental effects. This result reinforces

49
 Figure 3.7. Best-time constrained collection plan, final generation by shortest
travel time.

the need for path planning programs that account for the non-uniform maneuvering capa-
bilities associated with spacecraft.

3.2.6 Results of Genetic Algorithm Maximizing Value with Time Con-

straints
Values of 1, 2, or 3 value units (VU) have been assigned to the targets, as depicted in Figure
3.8.

The value-based multi-objective fitness function was generated in a similar fashion as the
time-constrained fitness function discussed in Section 3.2.5. An adjustment was made to
the fitness function so that uncollected value is minimized in the later part of the algorithm.

𝑛−1
Õ  𝑛𝑐
Õ 
(3.7)
 
J4 = 𝛼𝑡𝑖 + 𝛽 𝑣max − 𝑣 𝑗 
𝑖=1  𝑗=1 
 

Using the value-based approach, the value maximizing solution depicted in Figure 3.9a
was generated. In this solution, the path tends to favor the right-hand side of the target

50
 Figure 3.8. Targets with collection value assigned.

field where more high-value targets are located. Convergence on a solution occurs in the
value-based version using the same tools and techniques as the time-minimizing approach.

Figure 3.9b illustrates progress in both time and value improvement over the evolution of
this scenario, as a result of the multi-objective value-based fitness function. Some spikes
over the final value collected are the result of scalarization causing some time-efficient paths
to slightly out-perform value-efficient paths due to the scalar weights for those particular
generations. Generally, this performance is indicative of the algorithm converging on an
optimal solution.

This latest addition to RSIP is potentially open to criticism because of the manner in which
the collected portion of the permutation is selected. In this planner, only the first portion of
any given permutation is collected; therefore, if a more efficient or higher value path through
the middle or end of the permutation was available, it would not be considered. Ultimately,
the algorithm is designed to select for the most valuable sequences as they appear at the
beginning of each sequence, and therefore that design is relied upon to ensure that the most
advantageous paths evolve in that position.

51
 (a) Time-constrained collection plan, final generation by most value.

(b) Population performance over entire path, by generation.

Figure 3.9. Time constrained, value maximizing genetic algorithm solution.

52
3.3 Remote Sensing Iterative Planner Mission Models
Section 3.2 has discussed several techniques for path finding for remote sensing purposes,
culminating in the value-maximizing, time-constrained model. This model accounts for
maneuvering abilities that are not uniform in all directions, includes the ability to assign
different values to different targets, and selects for efficient plans that meet external time
constraints that preserve maximum collection value in the constrained environment.

The next logical step is to extend these principles from a two-dimensional model to a
remote sensing spacecraft model that incorporates orbital mechanics to derive the vehicle’s
position at a given time and vehicle kinematics to determine path times to reorient from one
target to the subsequent target. Two legacy methods of accomplishing these calculations
are the taxicab method which rotates about one principal axis of the vehicle, and then
around a second, and the eigenaxis rotation, which slews the vehicle in a direct arc from
one orientation to the next. Both of these methods were introduced in Chapter 2 and neither
are the most efficient method of accomplishing an attitude change [1]. However, these
approaches are commonly used today and are therefore useful to analyze.

Taxicab Method
A mission plan for an orbital altitude of 675 km over a target field described in Appendix A
using the taxicab maneuvering approach is presented in Figure 3.10. This mission plan and
all subsequent mission plans are value-maximizing and time-constrained, using Equation
(3.7) as their fitness function.

In this example, five targets are collected for a value of 13 VU out of 62 available VU.

Eigenaxis Method
For the same scenario that was shown in Figure 3.10, changing only the slewing method
from taxicab slews to eigenaxis slews yields value gains with no change in hardware. The
taxicab and eigenaxis scenarios, along with all future spacecraft scenarios discussed, use
identical vehicle and target parameters and can be directly compared. Figure 3.11 depicts
such a mission plan, which when compared to the taxicab approach, collects two additional
targets and achieves an additional five value u nits. Table 3 .3 l ists t he average t ime per
slewing maneuver from the same vehicle collecting on the same target field, a lbeit not

53
 Figure 3.10. Spacecraft mission plan using taxicab slews, value maximizing.

identical targets. Based on this comparison, there is a clear advantage to eigenaxis slews
over taxicab slews when possible. A simple Pythagorean inspection of the different paths
also supports this conclusion.

3.4 Summary
This section introduced the path planning problem for spacecraft applications, including
classifying the problem according to current combinatorial optimization literature. A
genetic algorithm was offered as a candidate solution to this path planning problem, and
such an algorithm was developed for remote sensing planning. Finally, two mission plans
were presented based on this mission planner and using two different legacy spacecraft
maneuvering techniques. The comparison of these two plans highlights the ability to gain
value and path efficiency by adjusting the slewing method used. Chapter 4 will introduce the

54
 Figure 3.11. Spacecraft mission plan using eigenaxis slews, value maximizing.

concept of mathematically optimal attitude rotations that can achieve further efficiencies for
remote sensing vehicle maneuvers, if such a maneuver can be incorporated into the planning
process.

55
Table 3.3. Average slew time by maneuver.

Method Time per Maneuver

Taxicab 20.73 s
Eigenaxis 13.23 s

56
 CHAPTER 4:
Time-Optimal Maneuvers

Two path planning techniques for spacecraft reorientation that are simple to calculate and
execute were introduced in Chapter 2, however these techniques are inefficient in terms of
time. The least efficient technique addressed in this thesis is the taxicab technique, which
rotates a vehicle around one principal axis and then another principal axis to achieve the
final desired attitude. The eigenaxis technique is more efficient than the taxicab technique in
that it accomplishes the reorientation in a single rotation along a direct arc, but it is also not
time-optimal. Bilimoria and Wie showed decisively that even the direct arc reorientation
of the vehicle will not be the optimal, minimum time reorientation in almost all cases [1].
The theory behind time-optimal reorientation maneuvers has been developed in literature
for decades. Furthermore, flight i mplementation o f t ruly t ime-optimal m aneuvers has
since been demonstrated on operational spacecraft, such as NASA’s Transition Region and
Coronal Explorer (TRACE) spacecraft [2].

This chapter develops the engineering model that produces a mathematically optimal min-
imum time reorientation maneuver of an example spacecraft. The relationship between the
engineering model and the larger planning problem is illustrated in Figure 4.1. The ap-
proach relies on Pontryagin’s principle to develop the necessary conditions for optimality by
way of a cost function, dynamic modeling, endpoint constraints and path constraints. This
implementation and problem solution is accomplished using the DIDO software package in
MATLAB. Further details about DIDO and Pontryagin’s principle for optimal control are
available in two publications by Ross [12], [13].

This chapter will present the framework for minimum time reorientations of the example
spacecraft, including scaling and balancing the problem, and verification and validation of
the results for an example maneuver. Pontryagin’s principle is briefly introduced, followed
by the analytic and numerical implementation for this specific p roblem. The engineering
model developed herein will produce a set of sample maneuvers that will be used to train
machine learning models to approximate maneuver times based on the required slew, thereby
enabling mission planning that incorporates minimum time maneuvers.

57
 Mission Parameters
Flight path
Sensor characteristics
Target locations

Mission Mission Sequence Engineering Mission time

Planner E.g., 1, 4, 8, 11, 16 Model Control trajectories

Vehicle location Legend

Slew time
Target locations
Iterative Operations
Maneuver Generator
High Fidelity Operation

Figure 4.1. Over-subscription planning flow diagram (Engineering Model).

4.1 Parameters of Example Spacecraft

To investigate the viability of applying machine learning to a planning problem, an example
reaction wheel spacecraft model was developed. The mass and agility parameters for this
spacecraft are provided here. A summary of the notation used for optimal control problems
is available in Appendix B.

The inertia matrix I and the reaction wheel configuration matrix Z for the spacecraft are
given here.
 222.17 2.58 −9.10 

I =  2.58 264.51 2.42  kg · m2 (4.1)
 
 
 −9.10
 2.42 171.91  

58
  cos(𝛼) cos(𝜂) sin(𝛼) cos(𝜂) − cos(𝛼) cos(𝜂) − sin(𝛼) cos(𝜂) 
 
Z =  − sin(𝛼) cos(𝜂) cos(𝛼) cos(𝜂) sin(𝛼) cos(𝜂) − cos(𝛼) cos(𝜂) (4.2)
 

 

 sin(𝜂) sin(𝜂) sin(𝜂) sin(𝜂) 

𝛼 = 45 deg 𝜂 = 30 deg

The reaction wheel configuration matrix maps the reaction wheel cluster to the body prin-
cipal axes. Torque exerted by an individual reaction wheel is represented by 𝑟𝑤 𝜏 and
spacecraft angular velocity is given by 𝜔. The maximum torque that can be exerted by a
reaction wheel and the maximum angular velocity about any principal axis are given as

𝑟𝑤
𝜏max = 0.6 N · m
𝜔max = 1.00 deg/sec .

4.2 Dynamic Modeling

Previous sections have addressed spacecraft rotations in terms of kinematics only; this
chapter introduces spacecraft dynamic modeling. This model is derived from a high fidelity
example developed by Lippman, Kaufmann, and Karpenko [14], which is reproduced in
Appendix C.

4.2.1 Spacecraft Dynamics Model

The state vector for the spacecraft dynamics model is a four element quaternion represen-
tation for attitude and a three element vector for angular rates of the body. The control is
a three element vector of the torque on the rigid body, 𝝉 𝑏 , subject to torque constraints on
each of four reaction wheels. The configuration matrix Z maps the contributions of the
individual reaction wheels to the spacecraft body.

Spacecraft attitude kinematics are parameterized in terms of rotational velocity using quater-

59
nion representation, where Q(𝜔) is defined as [6]:

 0 𝜔3 −𝜔2 𝜔1 

0
 
 −𝜔3 𝜔 1 𝜔 2
(4.3)

Q(𝜔) =  .
 𝜔2 −𝜔1 0 𝜔3 
 −𝜔1 −𝜔2 −𝜔3 0 
 
 

The transformation matrix Q(𝜔) enables modeling of quaternion dynamics of the spacecraft
using the relationship q¤ = 12 Q(𝜔)q. In this model, spacecraft body torque is controlled
directly, subject to agility and reaction wheel limitations. Spacecraft angular acceleration
can be obtained from Euler’s equations of motion, where external torques are assumed to
be zero [6].
¤ = 𝝉 𝑏 − 𝝎 × I𝝎
I𝝎 (4.4)

Torque on the body of the spacecraft is represented as the opposite of internal torque, which
is generated by reaction wheel momentum exchange devices for the example spacecraft:

𝝉 𝑏 = − (Z𝝉𝑟𝑤 + 𝝎 × Z𝒉𝑟𝑤 ) (4.5)

where 𝒉𝑟𝑤 represents the angular momentum of the reaction wheels. Substituting for 𝝉 𝑏 ,
Equation (4.4) becomes
I𝝎¤ = −Z𝝉𝑟𝑤 − 𝝎 × (Z𝒉𝑟𝑤 + I𝝎) (4.6)

Angular momentum is conserved, therefore for a zero net-bias system, Z𝒉𝑟𝑤 + I𝝎 = 0. This
simplifies the dynamics function to

¤ = I−1 [𝝉 𝑏 ] .
𝝎 (4.7)

In this form, angular acceleration on the spacecraft, 𝝎,

¤ is defined as a function of the control
variable, which is torque on the spacecraft body.

The inertia matrix, I, has been simplified from the true inertia tensor describing the non-
aligned principal axes of inertia and the body-fixed axes. The eigenvalues are used to
represent spacecraft inertia parameters in a more convenient manner.

60
  223.66 0 0 
 
I=  0 264.71 0  kg · m2 (4.8)
 
 

 0 0 170.23 

4.2.2 Minimum Time Reorientation Problem

The goal of this effort is to improve performance over legacy rotation techniques by im-
plementing a solution that minimizes time. Problem P is a minimum time optimal control
problem that produces the control trajectories required to implement such a solution. The
cost function, dynamics, boundary conditions, and path constraints, collected here for ease
of reference, are

Minimize: 𝐽 [x(·), u(·), 𝑡 𝑓 ] = 𝑡 𝑓






Subject to: q¤ = 21 Q(𝜔)q







¤ = I−1 [𝝉 𝑏 ]



 𝝎


𝝉𝑟𝑤 = Z+ [𝝉 𝑏 ]






𝑡0 = 0





P: q0 = (𝑞 10 , 𝑞 20 , 𝑞 30 , 𝑞 40 )𝑇



q 𝑓 = (𝑞 1𝑓 , 𝑞 2𝑓 , 𝑞 3𝑓 , 𝑞 4𝑓 )𝑇







𝝎0 = (0, 0, 0)𝑇






𝝎 𝑓 = (0, 0, 0)𝑇






h1 (x(𝑡), u(𝑡)) , |𝑟𝑤 𝜏𝑖 | ≤ 𝜏𝑚𝑎𝑥 , 𝑖 = 1, 2, 3, 4






h2 (x(𝑡), u(𝑡)) , |𝜔 𝑗 | ≤ 𝜔max , 𝑗 = 1, 2, 3





where q ∈ R4 , 𝝎 ∈ R3 , and u := {𝝉 𝑏 ∈ R3 }. 𝜏𝑟𝑤 represents the torque exerted by

each of four reaction wheel momentum exchange devices and Z+ is the Moore-Penrose
pseudoinverse of the reaction wheel configuration matrix, Z. All external torques are
assumed to be zero.

In this model, torque on the spacecraft body is the control variable, however torque limits are
imposed upon the reaction wheels and not the body axes. This is enforced by path constraint

61
h1 , which is the maximum allowable torque magnitude that can be applied by each reaction
wheel (|𝜏𝑟𝑤 | ≤ 𝑟𝑤 𝜏max). The second path constraint, h2, restricts the rotational rate around
each principal axis of the vehicle to less than or equal to 𝜔max, which is 𝜔max = 1 deg/sec in
this problem.

The potential for saturating the angular momentum of a reaction wheel is implicitly ad-
dressed by the h2 path constraint. By controlling the maximum angular velocity of the
spacecraft, reaction wheel saturation is also limited due to the conserved momentum rela-
tionship I𝝎 = −Z𝒉𝑟𝑤 .

4.3 Trajectory Optimization Problem

Before any attempt to solve this problem using a numerical tool can be made, the full
problem formulation must be developed. The first step of this process was to define the
system dynamics ( 𝒇 (𝑡) = x¤ (𝑡)), system state space (x ∈ R7 ), system control space (u ∈ R3 ),
and constraints (h(x(𝑡), u(𝑡))), which are collected in Problem P. The problem formulation
techniques and principles discussed in this chapter are available in more detail in Ross [13].

4.3.1 Optimal Control Cost Functional

A cost functional, 𝐽 [x(·), u(·), 𝑡 𝑓 ], must also be identified.1 This general form of a func-
tional is a function of the entire state and control trajectories for the problem as well as
time, indicated by x(·), u(·), and 𝑡 𝑓 ; it is not a function of any individual value at a point in
time, which would be annotated by the more common notation x(𝑡) and u(𝑡).

A cost functional comprises an endpoint cost (𝐸 (x 𝑓 , 𝑡 𝑓 )) and a running cost (𝐹 (x(𝑡), u(𝑡), 𝑡))
and will define the metric that the solver will minimize; Equation (4.9) is the standard cost
functional for an optimal control problem [13].
∫ 𝑡𝑓
𝐽 [x(·), u(·), 𝑡 𝑓 ] := 𝐸 (x 𝑓 , 𝑡 𝑓 ) + 𝐹 (x(𝑡), u(𝑡), 𝑡) 𝑑𝑡 (4.9)
𝑡𝑜

It is crucial that the cost functional be correctly defined, otherwise an analyst may find no
solution, or a solution to an unintended problem. This problem finds the fastest maneuver
1This cost functional is specific to the optimal control problem and should not be confused with the fitness
function for a genetic algorithm, J , discussed in Chapter 3.

62
to accomplish a given rotation without regard to other parameters that might limit control
effort, outside of h1 and h2 . Accordingly, the only factor impacting the cost functional is
the maneuver final time, which is represented as an endpoint cost.

𝐸 (x 𝑓 , 𝑡 𝑓 ) = 𝑡 𝑓 (4.10)

As there are no other considerations, running cost for this problem is zero.

𝐹 (x(𝑡), u(𝑡), 𝑡) = 0 (4.11)

The cost functional for Problem P is then:

𝐽 [x(·), u(·), 𝑡 𝑓 ] = 𝑡 𝑓 (4.12)

4.3.2 Necessary Conditions for Optimality

This section develops the necessary conditions for optimality, which are the set of conditions
that must be satisfied for a solution to be optimal. Satisfying these conditions cannot
guarantee global optimality, however such a guarantee is not needed to achieve a valuable
result. The minimum time maneuvers produced by this solver will be shown to be feasible,
and with that result, a feasible minimum time rotation is a valuable improvement over legacy
techniques, even without a guarantee of global optimality.

Problem Variable Summary

x = [𝑞 1 , 𝑞 2 , 𝑞 3 , 𝑞 4 , 𝜔1 , 𝜔2 , 𝜔3 ] 𝑇
u = [ 𝑏 𝜏1 , 𝑏 𝜏2 , 𝑏 𝜏3 ] 𝑇
𝐹 (x(𝑡), u(𝑡)) = 0
𝐸 (x 𝑓 ) = 𝑡 𝑓

Pontryagin Hamiltonian
A full account of the significance of the Hamiltonian is not practical here, but some relevant
concepts will be discussed. “For the control function 𝒖(·) to be optimal, Pontryagin’s
Principle requires that u globally minimize the Hamiltonian for every 𝑡 ∈ [𝑡0 , 𝑡 𝑓 ]” [13].
The general form of the Pontryagin Hamiltonian is provided in Equation (4.13). The

63
following steps implement the mathematical processes needed to define the Hamiltonian
function for this problem and accomplish the control minimization.

For a minimum time optimal control problem, the Hamiltonian units are time-units divided
by time-units (i.e., unitless). The covector 𝝀 is referred to as a costate for the state variable
covector specifically, or alternatively the adjoint covector. The units of the adjoint covector
allow the evaluation of the algebraic functions in Equation (4.13) to be mathematically
legal. The costate also varies as needed for the optimization process. Equation (4.14) is the
Hamiltonian for this problem.

𝐻 (x, 𝜆, u, 𝑡) = 𝐹 (x(𝑡), u(𝑡)) + 𝝀 · f (𝑥, 𝑢) (4.13)

1
𝐻 (x, 𝝀, u, 𝑡) = [𝜆 𝑞1 (𝑞 2 𝜔3 − 𝑞 3 𝜔2 + 𝑞 4 𝜔1 )
2
+ 𝜆 𝑞2 (−𝑞 1 𝜔3 + 𝑞 3 𝜔1 + 𝑞 4 𝜔2 )
+ 𝜆 𝑞3 (𝑞 1 𝜔2 − 𝑞 2 𝜔1 + 𝑞 4 𝜔3 )
− 𝜆 𝑞4 (𝑞 1 𝜔1 + 𝑞 2 𝜔2 + 𝑞 3 𝜔3 )] (4.14)
−1
+ 𝐼 (1,1) 𝜆 𝜔1 𝑏 𝜏1
−1
+ 𝐼 (2,2) 𝜆 𝜔2 𝑏 𝜏2
−1
+ 𝐼 (3,3) 𝜆 𝜔3 𝑏 𝜏3

Lagrangian of the Hamiltonian

The Lagrangian of the Hamiltonian introduces path constraints to the problem. This is
accomplished by adding the dot product of the vector containing constraints on problem
variables and a covector for measuring the path constraints, 𝝁, such that

𝐻¯ (x, 𝝀, 𝝁, u, 𝑡) = 𝐻 (x, 𝝀, u, 𝑡) + 𝝁 · h(u) (4.15)

64
 1
𝐻¯ (x, 𝝀, 𝝁, u, 𝑡) = [𝜆 𝑞1 (𝑞 2 𝜔3 − 𝑞 3 𝜔2 + 𝑞 4 𝜔1 )
2
+ 𝜆 𝑞2 (−𝑞 1 𝜔3 + 𝑞 3 𝜔1 + 𝑞 4 𝜔2 )
+ 𝜆 𝑞3 (𝑞 1 𝜔2 − 𝑞 2 𝜔1 + 𝑞 4 𝜔3 )
− 𝜆 𝑞4 (𝑞 1 𝜔1 + 𝑞 2 𝜔2 + 𝑞 3 𝜔3 )]
−1
+ 𝐼 (1,1) 𝜆 𝜔1 𝑏 𝜏1 , (4.16)
−1
+ 𝐼 (2,2) 𝜆 𝜔2 𝑏 𝜏2
−1
+ 𝐼 (3,3) 𝜆 𝜔3 𝑏 𝜏3
+ 𝜇1𝑟𝑤 𝜏 1 + 𝜇2𝑟𝑤 𝜏 2 + 𝜇3𝑟𝑤 𝜏 3 + 𝜇4𝑟𝑤 𝜏 4
+ 𝜇5 𝜔1 + 𝜇6 𝜔2 + 𝜇7 𝜔3
where reaction wheel torques are defined as

𝑟𝑤 𝜏
1
 



  𝑏 𝜏1 
𝑟𝑤 𝜏
2
 
 = Z+  𝑏 𝜏2  . (4.17)
 

𝑟𝑤 𝜏
3
   



  𝑏 𝜏3 
𝑟𝑤 𝜏
4
   
 

Adjoint Equations
The partial derivative of the Hamiltonian with respect to the costate vector provides a
powerful result: the system dynamics.

𝜕𝐻
= 𝒇 (x, u, 𝑡) = x¤
𝜕𝝀

Further analysis indicates that this result is related to the adjoint covector, Equation (4.18).
This set of equations defines important relationships in an optimal control problem and
enables the minimization that leads to an extremal result.

𝜕 𝐻¯
𝜆¤ = − (4.18)
𝜕x

65
 𝜕 𝐻¯ 1
𝜆¤ 𝑞1 = − (4.19a)

= 𝜆 𝑞 2 𝜔3 − 𝜆 𝑞 3 𝜔2 + 𝜆 𝑞 4 𝜔1
𝜕𝑞 1 2
𝜕 𝐻¯ 1
𝜆¤ 𝑞2 = − (4.19b)

= −𝜆 𝑞1 𝜔3 + 𝜆 𝑞3 𝜔1 + 𝜆 𝑞4 𝜔2
𝜕𝑞 2 2
𝜕 𝐻¯ 1
𝜆¤ 𝑞3 = − (4.19c)

= 𝜆 𝑞 1 𝜔2 − 𝜆 𝑞 2 𝜔1 + 𝜆 𝑞 4 𝜔3
𝜕𝑞 3 2
𝜕 𝐻¯ 1
𝜆¤ 𝑞4 = − (4.19d)

= − 𝜆 𝑞 1 𝜔1 + 𝜆 𝑞 2 𝜔2 + 𝜆 𝑞 3 𝜔3
𝜕𝑞 4 2
𝜕𝐻 ¯ 1
𝜆¤ 𝜔1 = − −𝜆 𝑞1 𝑞 4 − 𝜆 𝑞2 𝑞 3 + 𝜆 𝑞3 𝑞 2 + 𝜆 𝑞4 𝑞 1 − 2𝜇5 𝜔1 (4.19e)

=
𝜕𝜔1 2
𝜕 𝐻¯ 1
𝜆¤ 𝜔2 = − 𝜆 𝑞1 𝑞 3 − 𝜆 𝑞2 𝑞 4 − 𝜆 𝑞3 𝑞 1 + 𝜆 𝑞4 𝑞 2 − 2𝜇6 𝜔2 (4.19f)

=
𝜕𝜔2 2
𝜕 𝐻¯ 1
𝜆¤ 𝜔3 = − −𝜆 𝑞1 𝑞 2 + 𝜆 𝑞2 𝑞 1 − 𝜆 𝑞3 𝑞 4 + 𝜆 𝑞4 𝑞 3 − 2𝜇7 𝜔3 (4.19g)

=
𝜕𝜔3 2

Hamiltonian Minimization As introduced previously, an optimal solution requires the

minimization of the Hamiltonian with respect to the control vector. This is the Hamiltonian
Minimization Condition, Equation (4.20).

𝜕 𝐻¯
=0 (4.20)
𝜕u

𝜕 𝐻¯ −1
= 𝐼 (1,1) +
𝜆 𝜔1 + 𝜇1 𝑍 (1,1) +
+ 𝜇2 𝑍 (2,1) +
+ 𝜇3 𝑍 (3,1) +
+ 𝜇4 𝑍 (4,1) =0 (4.21a)
𝜕 𝑏 𝜏1
𝜕 𝐻¯
𝑏
−1
= 𝐼 (2,2) +
𝜆 𝜔2 + 𝜇1 𝑍 (1,2) +
+ 𝜇2 𝑍 (2,2) +
+ 𝜇3 𝑍 (3,2) +
+ 𝜇4 𝑍 (4,2) =0 (4.21b)
𝜕 𝜏2
𝜕 𝐻¯
𝑏
−1
= 𝐼 (3,3) +
𝜆 𝜔3 + 𝜇1 𝑍 (1,3) +
+ 𝜇2 𝑍 (2,3) +
+ 𝜇3 𝑍 (3,3) +
+ 𝜇4 𝑍 (4,3) =0 (4.21c)
𝜕 𝜏3

Hamiltonian Value Condition In this problem, the Endpoint Lagrangian is a function of

the final time and therefore the value of the Hamiltonian at 𝑡 𝑓 is not zero.

𝐸¯ (x 𝑓 , 𝜈, 𝑡 𝑓 ) = 𝐸 (x 𝑓 ) + 𝜈 · e(x 𝑓 )

66
 𝐸¯ (x 𝑓 , 𝜈, 𝑡 𝑓 ) = 𝑡 𝑓 + 𝜈𝑞1 (𝑞 1 − 𝑞 1 ) + 𝜈𝑞2 (𝑞 2 − 𝑞 2 ) + 𝜈𝑞3 (𝑞 3 − 𝑞 3 ) + 𝜈𝑞4 (𝑞 4 − 𝑞 4 )
𝑓 𝑓 𝑓 𝑓
(4.22)
+ 𝜈𝜔1 (𝜔1 ) + 𝜈𝜔2 (𝜔2 ) + 𝜈𝜔3 (𝜔3 )

𝜕 𝐸¯
𝐻 [@𝑡 𝑓 ] = − = −1 (4.23)
𝜕𝑡 𝑓
This result is expected for a minimum time problem.

Hamiltonian Evolution Equation For a properly minimized Hamiltonian and an optimal

trajectory, the Hamiltonian Evolution Equation provides the following relationship [13].

𝑑H 𝜕𝐻
=
𝑑𝑡 𝜕𝑡

Given a value for the Hamiltonian at 𝑡 𝑓 obtained from the Hamiltonian Value Condition,
and information about the derivative of the Hamiltonian with respect to time, a plot of the
Hamiltonian can be used to verify optimality. In this problem, the Hamiltonian is not a
function of time.
𝜕𝐻 𝑑H
= 0, ∴ =0 (4.24)
𝜕𝑡 𝑑𝑡

It is therefore expected to be constant as a function of time.

Transversality Equations Transversality allows the discovery of boundary conditions that

may not be provided in the problem formulation by determining the final costate value as the
partial derivative of the Endpoint Lagrangian with respect to the final state whose condition
is not known, Equation (4.25).

Using the endpoint conditions and the Endpoint Lagrangian, where

 𝑞 𝑓  
  1 
  𝑓 
𝑞  
q 𝑓 −  2𝑓  

e(x 𝑓 ) :=  𝑞 3   = [0]
  𝑓 
𝑞  

  4 
 𝝎𝑓 
 

67
and

𝐸¯ (x 𝑓 , 𝜈, 𝑡 𝑓 ) = 𝑡 𝑓 + 𝜈𝑞1 (𝑞 1 − 𝑞 1 ) + 𝜈𝑞2 (𝑞 2 − 𝑞 2 ) + 𝜈𝑞3 (𝑞 3 − 𝑞 3 ) + 𝜈𝑞4 (𝑞 4 − 𝑞 4 )

𝑓 𝑓 𝑓 𝑓
, (4.22)
+ 𝜈𝜔1 (𝜔1 ) + 𝜈𝜔2 (𝜔2 ) + 𝜈𝜔3 (𝜔3 )

the costate values 𝜆𝑖 at 𝑡 𝑓 are defined. In this problem, all required boundary conditions are
provided, so transversality principles are not expected to provide new information. This is
indicated by the final costate value being equal to the endpoint covector element for each
state, which is also unknown.

𝜕 𝐸¯
𝝀𝑡 𝑓 = (4.25)
𝜕x 𝑓

𝜕 𝐸¯
𝜆 𝑞1 𝑓 = = 𝜈𝑞1 (4.26a)
𝜕𝑞 1 𝑓
𝜕 𝐸¯
𝜆 𝑞2 𝑓 = = 𝜈𝑞2 (4.26b)
𝜕𝑞 2 𝑓
𝜕 𝐸¯
𝜆 𝑞3 𝑓 = = 𝜈𝑞3 (4.26c)
𝜕𝑞 3 𝑓
𝜕 𝐸¯
𝜆 𝑞4 𝑓 = = 𝜈𝑞4 (4.26d)
𝜕𝑞 4 𝑓
𝜕 𝐸¯
𝜆 𝜔1 𝑓 = = 𝜈𝜔1 (4.26e)
𝜕𝜔1 𝑓
𝜕 𝐸¯
𝜆 𝜔2 𝑓 = = 𝜈𝜔2 (4.26f)
𝜕𝜔2 𝑓
𝜕 𝐸¯
𝜆 𝜔3 𝑓 = = 𝜈𝜔3 (4.26g)
𝜕𝜔3 𝑓

4.3.3 Scaling and Balancing

Problems that deal in orbital magnitudes have the potential to develop significant mismatches
of magnitude between the states and controls and their corresponding covectors. Therefore,
scaling and balancing techniques were employed to make the models more robust and better

68
conditioned for numerical analysis. Scaling and balancing techniques for optimal control
problems are discussed at length by Ross et al. [15].

Canonical Scaling
First, canonical scaling was considered using common units for mass, length, time, and
angle. To reduce run times and improve the ability of DIDO to solve otherwise poorly con-
ditioned problems, the dynamics were considered according to the following relationships.

The variables must be defined in terms of their scaled states; a given variable’s scaled state is
indicated by an overbar and the selected scale factors are indicated by a capitalized constant
(e.g., 𝑥 = 𝑥¯ 𝑋).

𝑞 = 𝑞¯ (4.27a)
𝜃 𝜃¯ 𝐴 𝐴
𝜔= = = 𝜔¯ (4.27b)
𝑡 𝑡¯ 𝑇 𝑇
¯ 𝑀𝐿 𝑀 𝐿2
𝜏 = Force · radius = 𝐹 · 𝑟 = 𝐹 2 · 𝑟¯ 𝐿 = 𝜏¯ (4.27c)
𝑇 𝑇2
I = Ī 𝑀 𝐿 2 (4.27d)

The goal of canonical scaling is to ensure that states fall within a desirable range (e.g.,
x ∈ [−1, 1]). Quaternions fall into this range by definition, and are therefore not adjusted
by a scale factor in this problem.

Using the scaled variable relationships, scaled dynamics equations are derived.

1
 
𝑑q 𝑑 q̄ 𝐴
= ¯
= Q( 𝜔) q̄
𝑑𝑡 𝑑 𝑡¯ 𝑇 2 𝑇
𝐴
q¤̄ = Q( 𝝎) ¯ q̄ (4.28)
2

Rotation rate dynamics are a function of torque.

𝑑𝝎 𝑑 𝝎 ¯ 𝐴 
2
 −1 𝑀 𝐿2
= = Ī 𝑀 𝐿 [ ¯
𝝉 ]
𝑑 𝑡¯ 𝑇 2 𝑇2
𝑏
𝑑𝑡
1
¤̄ = Ī−1 [ 𝝉¯ 𝑏 ]
𝝎 (4.29)
𝐴

69
For this model, the only canonical unit that remains for the given dynamics is angle, 𝐴.
While the goal of canonical scaling is to ensure that states fall within a desirable range, it
cannot also guarantee that the states and their covectors are properly balanced. Properly
balanced variables have vectors and covectors that fall within similar orders of magnitude,
which is desirable for numerical computation. With only one canonical unit to adjust and the
inability to simultaneously ensure effective balancing, canonical scaling is not a particularly
useful technique for this application.

Designer Units
Instead of canonical scaling, scaling with designer units was used because it allows for each
state-costate, control-covector, or path-covector pair to be scaled individually, which also
allows for balancing [15]. Using the selected scale factors, well-scaled and balanced states,
constraints, and covectors were achieved, ensuring that well-conditioned parameters were
provided to the optimization solver. Equation (4.30) lists the scale factors used.

𝐾𝑞1 = 0.2 𝐾 ℎ1,1 = 3

𝐾𝑞2 = 0.2 𝐾 ℎ1,2 = 1
𝐾𝑞3 = 0.2 𝐾 ℎ1,3 = 6
𝐾𝑞4 = 0.1 𝐾 ℎ1,4 = 1 (4.30)
𝐾𝜔1 = 0.01 𝐾 ℎ2,1 = 0.02
𝐾𝜔2 = 0.01 𝐾 ℎ2,2 = 0.02
𝐾𝜔3 = 0.05 𝐾 ℎ2,3 = 0.02

4.4 Example Solution

An example trajectory optimization problem was solved using DIDO in MATLAB. This
maneuver is a rest-to-rest maneuver (i.e., zero rotational velocity at the beginning and
end of the maneuver) with the quaternion boundary conditions q0 = (0, 0, 0, 1)𝑇 and
q 𝑓 = (0.2660, 0.4234, 0.0472, 0.8647)𝑇 .

70
4.4.1 Results and Analysis
Figure 4.2 shows the scaled quaternion state trajectories and their respective scaled costates;
Figure 4.3 shows the scaled rotational rate state trajectories and their respective scaled
costates. The control trajectories are depicted in Figure 4.4. The minimum time required
for this slewing maneuver is 49.04 seconds with this problem formulation. Figure 4.5 is
a three-dimensional rendering of the rotation of this spacecraft and the boresight ground
trace.

In Figure 4.3, the effects of the state path constraint can be seen early in the maneuver, when
the angular rates plateau. This behavior is examined in more detail in Section 4.4.2.

The fluctuations in the control trajectories (Figure 4.4) that start before the ten second mark
and end after the 40 second mark is a byproduct of the active velocity state constraint, which
is preventing the steady application of torque so as not to exceed the path constraint defined
by h2 . Before and after this time frame, the path constraint is not active and traditional
switching function behavior is observed. This solution is consistent with optimal control
theory and the required conditions described by Karush, Kuhn, and Tucker [13], which will
be addressed in more detail in the next section.

4.4.2 Verification and Validation

This problem formulation will be verified and validated along two lines of inquiry: investi-
gating the feasibility of the solution and checking for adherence to the necessary conditions
for optimality.

Model Feasibility Conditions

The feasibility of this solution will be addressed in two main ways: whether or not the model
adheres to the problem dynamics while satisfying path constraints and whether or not the
desired boundary conditions have been achieved. Visually inspecting Figure 4.4, it appears
to be a switching function type controller as expected. This is evident from the control
trajectories that are alternately at the upper bound or the lower bound, with the exception
of fluctuating inputs in the control trajectories while the path constraint h2 is active.

71
 Figure 4.2. Scaled quaternion states and corresponding costates.

The control trajectory solution was propagated through the plant dynamics, and the resulting
state trajectories are overlaid with the DIDO solver trajectories and displayed in Figure 4.6.
From this figure, it is apparent that the plant dynamics have been satisfied and that the initial
and final boundary conditions have also been met. Furthermore, this result supports the
conclusion that the fluctuations in the control trajectories due to the active path constraint
are not degrading the overall solution.

Model Optimality Conditions

While adhering to the necessary conditions for optimality developed for this model is a
requirement for an optimal solution, that is distinct from satisfying the sufficient conditions
for optimality which would guarantee an optimal solution [13]. Here, evidence is presented
that supports the conclusion that the solution may be optimal, or even is likely to be optimal.
However, this cannot guarantee that the solution is globally optimal. This process also seeks
to uncover any evidence that the necessary conditions have not been satisfied. Under such
conditions, the solution cannot be optimal.

72
 Figure 4.3. Scaled angular rate states and corresponding costates.

Figure 4.4. Control trajectories and control torques mapped to reaction wheels.

Costate Behavior From Equations (4.19a) through (4.19g), it can be inferred that the
adjoint equations will likely be non-constant, but little else can be guaranteed. Figures 4.2
and 4.3 do indeed show non-constant costate behavior.

Hamiltonian Behavior According to the Hamiltonian Evolution Equation, Equation

(4.24), the ordinary derivative of a properly minimized Pontryagin Hamiltonian with respect
to time will be equal to the partial derivative of the Pontryagin Hamiltonian with respect to
time, which in this case is zero, as the Hamiltonian is not an explicit function of time.

73
 (a) Initial orientation. (b) One-third of maneuver complete.

(c) Two-thirds of maneuver complete. (d) Final orientation.

Figure 4.5. Time lapse of slewing maneuver.

It is further shown from the Hamiltonian Value Condition, Equation (4.23), that the Hamil-
tonian at the final time is equal to −1. As the derivative with respect to time is zero, the
Hamiltonian should be constant with respect to time. This will allow the Hamiltonian
resulting from the numerical solution to be inspected for constancy. While a constant
Hamiltonian does not guarantee optimality, it is a requirement for optimality [13].

The Pontryagin Hamiltonian is plotted against time in Figure 4.7. Given the scaling of
the vertical axis and the discretization of the problem by the DIDO solver, this behavior is
consistent with a properly minimized Hamiltonian and is evidence that an optimal solution
has been found.

74
 Figure 4.6. Verified state trajectories in engineering units.

Control Constraint Behavior Karush, Kuhn, and Tucker developed two conditions that
describe control variable behavior: the stationarity condition and the complementarity
condition [13]. These conditions rely upon the Lagrangian of the Hamiltonian developed
for this problem, Equation (4.16). The Lagrangian of the Hamiltonian in its general form
makes use of the Lagrange multiplier 𝜇.

𝐻¯ (x, 𝝀, 𝝁, u, 𝑡) = 𝐻 (x, 𝝀, u, 𝑡) + 𝝁 · h(u)

75
 Figure 4.7. Hamiltonian as a function of time.

𝜕 𝐻¯
The 𝜇 multiplier is allowed to vary as necessary to ensure that = 0 is always possible.
𝜕u
This is the definition of the stationarity condition, and was used in the development of the
necessary conditions for this problem [13].

The complementarity condition describes the behavior of the control covector, 𝝁, which
varies in response to the control constraint, h1 (u). The complementarity condition is
defined [13] as


 ≤ 0 if 𝑢𝑖 = 𝑢 lower
𝜇𝑖 = 0 if 𝑢 lower < 𝑢𝑖 < 𝑢 upper .



 ≥ 0 if 𝑢𝑖 = 𝑢 upper




To check for adherence to the stationarity and complementarity conditions, the control tra-
jectories are plotted over their corresponding control covector, 𝜇𝑖 in Figure 4.8. While there
are slight deviations from strict adherence to the values of 𝜇 relative to the zero crossings of
the corresponding control trajectory, these are artifacts of the numerical solution process.
The effects of the binding path constraint h1 are again evident through the fluctuations in the
middle portion of the reaction wheel torque trajectories. During this condition the values
for 𝜇 are zero or approach zero because extremal control effort is not being applied.

76
 Figure 4.8. Reaction wheel torque trajectories and corresponding covectors.

Adherence to the complementarity principles is observed: 𝜇 values are negative when the
constrained parameter is at its lower bound and the values are positive when it is at its upper
bound.

Path Constraint Covector Behavior The second set of constraints, h2, is defined by the
three angular rate state variables and ensures that the rotational rate about each principal
axis of the vehicle is kept below the maximum allowable value. The explicit value of 𝜇 𝑗 over

77
 Figure 4.9. State variable path constraint plotted with its covector.

time is not readily verifiable; however, the behavior can be verified graphically. Figure 4.3
depicted each scaled state’s behavior over time, and Figure 4.9 depicts the path constraint
h2 using unscaled inputs, along with the control covector 𝜇 𝑗 and the maximum allowable
vehicle rotation rate, 𝜔max .

This plot demonstrates that the h2 path constraint becomes active for each rotational axis
during the middle portion of the maneuver, where 𝜇 𝑗 is non-zero and |𝜔 𝑗 | = 𝜔max .

To check for effective scaling of the model constraints, the scaled constraints and their
covectors are presented in Figure 4.10; here, it is evident that the constraints are properly
balanced for efficient computation.

4.5 Summary
Performing the same example rotation described in this chapter using the legacy eigenaxis
technique requires 68.0 seconds, compared to 49.0 seconds for the minimum time maneuver
derived here. That is a 27.9% improvement over the status quo eigenaxis maneuver for a
roughly 1 radian slew. A performance improvement of this magnitude will likely appeal
to any spacecraft operations manager. As these maneuvers have been demonstrated on
orbit, the last remaining hurdle to wide implementation of these techniques is the ability to
incorporate them into mission planning processes. This optimal maneuver model will be
the basis for fast-running machine learning models that can approximate optimal slew times

78
with enough accuracy to enable enterprise-wide implementation of minimum time satellite
rotations.

This chapter has established the process needed to develop minimum time slewing ma-
neuvers, which will enable significant gains if incorporated into mission planning. The
development of the necessary conditions and extensive verification and validation make this
model a reliable framework from which a library of slewing maneuvers and their associated
time-optimal slew times can be developed. In Chapter 5, such a library will be generated,
and that data will be used to train supervised machine learning models. These, in turn, will
be used to approximate optimal slew times for path planning purposes.

79
Figure 4.10. Scaled path constraints and constraint covectors.

80
 CHAPTER 5:
Approximating Optimal Slew Times Using Machine
Learning

The task at hand now is to incorporate minimum time maneuvers into a mission planner.
Figure 5.1 highlights the function under consideration in this chapter as it relates to the larger
system. As discussed previously, the computational resources required to produce a single
minimum time maneuver make it impractical to incorporate optimal maneuver solvers
directly into an iterative path planner such as Remote Sensing Iterative Planner (RSIP).
To date, simple-to-calculate but physically inefficient solutions have been used to generate
feasible mission plans (e.g., the taxicab and eigenaxis maneuvers). To enable improvements
to the planning process, supervised machine learning concepts can be used to incorporate
approximations of certain aspects of optimal spacecraft slews into mission planners without
otherwise changing or modifying the algorithm.

Fortunately, the complete solution for an optimal maneuver it not required to generate a
mission plan sequence. Principally, only those concepts required by the fitness function,
Equation (3.7), need be produced. Externally, the fitness function requires maneuver time
and the value of the target in question. The value of the target in question is defined by the
planner, so only the maneuver slew time must be approximated.

𝑛−1
Õ  𝑛𝑐
Õ 
(3.7)
 
J4 = 𝛼𝑡𝑖 + 𝛽 𝑣max − 𝑣 𝑗 
𝑖=1  𝑗=1 
 

A slew time can be derived from a starting orientation and a final orientation in the orbital
reference frame, discussed in Chapter 2. The mission planner uses these attitude pairs to
determine slew times for legacy techniques, just as the optimal solver uses these attitude
inputs for the same purpose, as addressed in Chapter 4.

As RSIP already provides initial and final attitudes for a maneuver, a supervised learning
model needs only to approximate the optimal maneuver time, 𝑡𝑖 , from these attitudes and

81
 Mission Parameters
Flight path
Sensor characteristics
Target locations

Mission Mission Sequence Engineering Mission time

Planner E.g., 1, 4, 8, 11, 16 Model Control trajectories

Vehicle location Legend

Slew time
Target locations
Iterative Operations
Slew Time
Generator High Fidelity Operation
Available Methods
Taxicab
Eigenaxis
Neural Network (Optimal approx.)
Stepwise Regression (Optimal approx.)
Figure 5.1. Over-subscription planning flow diagram (Slew Time Generator).

provide that input to the planner fitness function for evaluation, in the same fashion as the
legacy approach.

This chapter will introduce supervised learning, the methods for generating a set of a
training data that enables the supervised learning process, and the resulting surrogate
models that approximate minimum time maneuvers. Two supervised learning methods will
be discussed: neural networks and stepwise multiple linear regression. Mission planner
results using both methods will be presented and the relative strengths of each will be
discussed.

82
5.1 Supervised Learning for Maneuver Time Approxima-
tion
Supervised learning is a machine learning method that relies on a labeled data set of known
inputs and outputs. This labeled data is used to initially fit a model by training model
parameters to minimize prediction error based on the known true outputs, given a set of
corresponding inputs. Supervised learning is the appropriate approach here because of
the availability of a labeled dataset for training (i.e., a dataset of optimal maneuvers and
their corresponding optimal slew times). This contrasts with the unsupervised learning
approach, in which input data is not associated with a response variable, and is instead used
to find patterns or to group inputs. This approach would be appropriate in cases where the
goal of the model is to understand relationships between the input variables by grouping
them into output classes. In this thesis, the labeled dataset of minimum time maneuvers for
supervised learning was created using the engineering model discussed in Chapter 4.

5.1.1 Supervised Learning Approaches

Within supervised learning, models are typically either regression models, associated with
continuous data, or classification models, which are appropriate in cases of discrete data.
For maneuver time approximation, the input data is a vector of real data that represents a
spacecraft orientation or rotation. The associated output data is the time required to perform
the rotation using a minimum time maneuver. These data are continuous, and therefore
regression approaches are employed here.

Two common frameworks for supervised learning problems with continuous data are neu-
ral networks and linear regression. These two techniques were explored for this problem,
leading to two models for optimal maneuver time approximation, one based on each frame-
work. These “models of models” are referred to as surrogate models, or metamodels of
the system they are intended to approximate [16]. In this case, they are surrogate models
of the engineering model from Chapter 4. Broadly speaking, metamodels are light weight,
fast-running approximations of other models, which in this case is a high fidelity optimiza-
tion model that requires significant computational resources to produce a minimum time
maneuver solution. Of course, without the high fidelity model, accuracy is sacrificed; by
definition, surrogate models are simply approximations of the high fidelity solution.

83
5.1.2 Surrogate Model Concept
A machine learning surrogate model is developed in three main parts. First, the training
data to be used must be developed or obtained. The process of intelligently sampling the
input space is called Design of Experiments (DoE) [17], and will be discussed further in the
next section. Once the desired set of model inputs is generated, the corresponding model
outputs must be generated. This is accomplished through a wrapper function, which is the
application of the full fidelity engineering model to the input DoE to obtain the labeled
target outputs. Once the inputs and outputs are associated, they are referred to as labeled
data and then used for model training. Before a supervised learning model can be trained
on the labeled data, it must be created in a software package, which necessitates decisions
about how the model will operate. For this remote sensing application, some of these
decisions have already been made (i.e., the decision to develop a neural network and a
linear regression model). Additional decisions remain, however.

For the neural network, the type of network must be decided (e.g., a feed-forward network, a
convolutional network, a recurrent network), the number of hidden layers must be assigned,
the number of nodes in each layer must be decided, and the activation function used
in each node must be chosen. These choices are made by the designer and are called
hyperparameters, which are not adjusted by the model during training [18]. These options
will be explored in detail in subsequent sections. For the linear regression model, choices
must be made about the error function used, method for determining regression terms, and
statistical confidence thresholds.

With these three broad steps completed—selection of input DoE, generation of labeled data
with the engineering model, and selection of hyperparameters for the surrogate model—
model training and evaluation can be accomplished. This thesis will evaluate multiple
approaches and compare them for effectiveness.

5.2 Design of Experiments

An effective supervised learning model depends on its training data. A model with poorly
designed or incomplete training data is likely to have decreased performance, therefore
DoE theory for designing the data collection process governs this discipline. DoE aims
to intelligently sample the input space so that maximum information can be gleaned from

84
the fewest number of trials. An effective DoE will also ensure that the entire relevant
input space is sampled, allowing a model to interpolate future inputs, instead of having to
extrapolate conclusions that are not included in the training data.

This model uses training data that is analytically generated and noiseless. The only concerns
are developing the correct data and committing to the processing time to produce it. This
is distinct from any physical experiments which may require limited resources, such as
reagents, equipment, personnel support, etc. Furthermore, analytic results do not have to
account for noise in the data, which would increase output error and could potentially mask
important relationships between the input data and maneuver time. It remains important,
however, to ensure than the entire sample space is covered and that the model is designed
with enough flexibility to capture the relevant features that predict maneuver time.

Six significant principles that apply to DoE are assembled here for discussion [17]. These
principles guided the development of the DoE used for RSIP.

1. Interactions. While individual factors can play a large role in the output of a process,
the interactions between these factors can also have significant impacts.
2. Scarcity of Effects. Most outcome variability tends to be the result of a subset of
design variables, rather than the full set of variables.
3. Hierarchy of Factors. Variability tends to result mostly from the main (first order)
effects, followed by second order effects and third order effects, and so on.
4. Heredity of Factors. If a higher order effect is deemed important (e.g., a second order
effect), then the lower order effects that make up the higher order effect (e.g., its
composite first order effects) are also important and should be included in the model.
5. Aliasing. Also referred to as confounding, the source of some variability may not be
identifiable due to a DoE that does not include every possible combination of input
factors.
6. Orthogonality. Trials within a DoE should contain as little overlap in information
as possible. This can be complicated by inputs which are not independent of one
another.

Three of these principles are inherently respected by the structure of the models chosen for
this thesis, as well as by the training algorithms used to determine final weights, biases,
and coefficients. Scarcity of effects, hierarchy of factors, and heredity of factors principles

85
are all implemented by the algorithms that determine how to weight, include, or exclude
different factors. The role of interactions, aliasing, and orthogonality are addressed in the
creation of the DoE itself, which will be discussed further in the next section. Broadly
speaking, the models employed here account for interactions and the DoE minimizes the
potential for aliasing while maximizing orthogonality.

5.2.1 Space Filling Design

Maneuver time approximation requires a DoE that is appropriate for continuous data on the
interior of a sample space. That is to say, there is no need for the model to extrapolate inputs
that are outside the domain of the training set. The domain that covers rotations is limited
by geometry (i.e., there is no need to model rotations in excess 𝜋 radians for this planner).
The class of approaches that satisfy these requirements is Space Filling Design [17], an
example of which is a Latin Hypercube shown in Figure 5.2.

Latin Hypercube Sampling

A Latin Hypercube maximizes “the minimum distance between design points but requires
even spacing of the levels of each factor. This method produces designs that mimic the
uniform distribution” [17]. This occurs by dividing each input variable domain into 𝑛
uniform subdivisions (where 𝑛 is the number of data points) and placing a single data point
within each subdivision in such a way that information overlap is minimized by maximizing
orthogonality between the data points.

Recall that the inputs for a rotation can be expressed in multiple ways: as a pair of quaternions
including an initial and a final quaternion (q𝑖 and q 𝑓 ), as a relative change in quaternion
units (Δq), or as the Euler parameters that define the change in terms of an axis and angle
of rotation (ê and 𝜓).

The domain of values for quaternion inputs is 𝑞𝑖 ∈ [−1, 1] subject to 𝑖 𝑞𝑖2 = 1 and for
Í

Euler parameters the domain of inputs is 𝑒𝑖 ∈ [−1, 1] subject to 𝑖 𝑒𝑖2 = 1 for the eigenaxis
Í

elements and 𝜓 ∈ [−𝜋, 𝜋] for the rotation angle. The input domains include continuous
data and the domain is bounded without the risk that future inputs may exceed the training
domain, because all rotations can be represented by elements within these bounds.

In summary, a Latin Hypercube uses random sampling techniques that cover the interior

86
 𝑧

𝑦
𝑥

𝑦 𝑧 𝑧

𝑥 𝑥 𝑦
Figure 5.2. Example Latin Hypercube.

of the design space, and maximize information while minimizing overlap between samples
over 𝑝 variables. The 𝑛 values of data points are distributed over equal sized intervals along
each variable.

RSIP Latin Hypercube

Input sampling for RSIP was accomplished using a MATLAB generated Latin Hypercube
with four variables and 2000 samples, however initial training data included only a subset
of randomly selected vectors. These vectors represent sensor boresight orientations, Figure
5.3.

87
 Figure 5.3. Input data set eigenaxes.

5.2.2 Input Parameterization

The hypercube generated for optimal maneuver time approximation consisted of four vari-
ables (𝑒 1 , 𝑒 2 , 𝑒 3 , 𝜓) and 500 samples (𝑛 = 500), chosen because Euler parameters can be
used to define a rotation independently, or transformed into quaternion representation for
the same purpose. Euler parameters are used as the starting point because the eigenaxis
(ê) is independent of the eigenangle (𝜓); if quaternion samples had been chosen directly,
random sampling would need to ensure that the quaternion definition is respected. Recall
that a quaternion is
 𝑒 sin ( 𝜓/2) 
 1 
 𝑒 2 sin ( 𝜓/2) 
 
q =   (2.22)
 3
𝑒 sin ( 𝜓/2) 

 cos ( 𝜓/2) 
 
 
and therefore 𝑞 1,2,3 are not independent of 𝑞 4 due to the eigenangle, and any analysis would
need to consider these relationships. A simpler approach is to simply generate the rotations
using Euler parameters and convert to quaternion when convenient.

Three configurations were considered for input parameterization:

1. 8 × 1 vector of initial quaternion attitude (q0 ) and final quaternion attitude (q 𝑓 ) in the

88
 orbital reference frame.
2. 4×1 vector of relative quaternion rotation (Δq) between the initial and final quaternion
attitudes.
3. 4 × 1 vector of Euler parameters (i.e., 3 × 1 eigenaxis of rotation (ê) and the angle of
rotation about that axis, the eigenangle [𝜓]).

The first option consisting of the starting and stopping orientation of a rotation was discarded
because of the superfluous information it injects into the model. For a spacecraft, the
important information that affects the reorientation of the vehicle is the change of the body
attitude in the body-fixed reference frame. The starting orientation of the spacecraft is
relevant to the orientation of the vehicle in the inertial and orbital frames, but not relevant
to the spacecraft itself, which only experiences rotations relative to the body frame. While
orientation relative to the inertial or orbital frame is important to the planning problem, it
is not required to compute a specific reorientation maneuver, as its information can easily
be condensed into one of the other two forms.

The two remaining candidates are both 4 × 1 vectors: Δq and [ê, 𝜓] 𝑇 . Given the risk of
introducing undesired aliasing into the DoE, the Euler parameters are the more desirable
choice as they are the more direct mathematical representation of a rotation. Quaternion
representation risks aliasing the effects of the ê and 𝜓 elements for 𝑞 1,2,3 because 𝑞 1,2,3 are a
function of ê and 𝜓 while 𝑞 4 is a function of 𝜓 alone. Therefore in quaternion representation,
there are no inputs that are functions of the eigenaxis exclusively. Figures 5.4 and 5.5 depict
the fit of each factor by the target maneuver time; patterns resulting from these distinctions
are apparent.

Relative Quaternion
The relative quaternion in the body frame (Δq) can be computed from initial and final
orientations in the orbital frame using the orthonormal quaternion transmuted matrix Q [6].


 𝑞 40 −𝑞 30 𝑞 20 𝑞 10 𝑇


 𝑞 1𝑓 
𝑞 30 𝑞 40 −𝑞 10 𝑞 20 𝑞 2𝑓 
   
(5.1)
𝑇
  
Δq = Q q 𝑓 =   
3

 −𝑞 20 𝑞 10 𝑞 40 𝑞 30 


 𝑞 𝑓 
−𝑞 10 −𝑞 20 −𝑞 30 𝑞 40 𝑞 4𝑓 
   
  
  

89
 Fit Group
Bivariate Fit of tf By dq_1 Bivariate Fit of tf By dq_2
200 200
180 180

160 160
140 140
120 120

tf 100 tf 100
80 80

60 60
40 40

20 20

0 0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
dq_1 dq_2

Bivariate Fit of tf By dq_3 Bivariate Fit of tf By dq_4

200 200
180
160
150
140
120

tf 100 tf 100
80
60
50
40

20
0 0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
dq_3 dq_4

Figure 5.4. Fit of final maneuver time by relative quaternion input data.

An analysis of Δq inputs versus target time was performed by fitting target time (𝑡 𝑓 ) by
the quaternion elements independently, shown in Figure 5.4. These plots illustrate the fit
of target maneuver time by relative quaternion elements for the labeled data set. There is
a strong correlation between maneuver time and 𝑞 4 . Meaningful correlations or patterns
are not visually evident between maneuver time and the other three quaternion elements,
however, the shape of these three distributions is strongly influenced by the 𝑒𝑖 · sin 𝜓/2
operation, which indicates potential masking and may make relationships harder to identify
during the model training process. There is also a strong curvature indicated as 𝑞 4 nears 1.

90
 Fit Group
Bivariate Fit of tf By eig_1 Bivariate Fit of tf By eig_2
140 140

120 120

100 100

80 80
tf tf
60 60

40 40

20 20

0 0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
eig_1 eig_2

Bivariate Fit of tf By eig_3 Bivariate Fit of tf By eig_4

140 140

120 120

100 100

80 80
tf tf
60 60

40 40

20 20

0 0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
eig_3 eig_4

Figure 5.5. Fit of final maneuver time by Euler parameter input data.

Euler Parameters
The fit of the Euler parameter elements against maneuver time (Figure 5.5) shows similar
strong correlation to maneuver time in the fourth element (the eigenangle). The other
three elements lack the distinctive shaping from the Δq approach and resemble a much
more visually random pattern. This implies that the triangularly shaped distributions of
𝑞 1,2,3 against 𝑡 𝑓 from Figure 5.4 are products of the mathematical mapping that creates
quaternions, and not inherent in the input data itself. The curvature at small slew times
is also less apparent in the Euler parameter representation than the relative quaternion
representation, indicating that it may also be exaggerated by the quaternion transformation.

91
Building of DoE

Figure 5.6. Input data set histograms.

DoE using both Euler parameters and the relative quaternion construction were generated
from the same underlying data; the relative quaternion form is simply an extra transformation
on the Euler parameters. A Latin Hypercube of four variables was created to represent the
three elements of an eigenaxis and a fourth element for the eigenangle. This ensures that
each set of Euler parameters represents a deliberate input point. Most Latin Hypercube
generating programs create their elements on a normalized or uniform distribution of either
[0, 1] or [−1, 1], so for this problem the elements of the eigenaxis were generated over
uniformly distributed subdivisions from [−1, 1], then normalized to a unit vector, and the

92
eigenangle was mapped to a domain of [−𝜋, 𝜋]. For transformation into a relative quaternion
vector, the quaternion definition from Equation (2.22) is applied. Distributions of the DoE
elements are provided in Figure 5.6.

The DoE distributions show the effects of starting with a uniformly distributed DoE, nor-
malizing the three eigenaxis elements to a unit vector, and then applying the quaternion
definition. In Figure 5.6, smaller slews appear to be favored, indicated by the concentration
of 𝑞 1,2,3 elements around zero and 𝑞 4 elements around one, however, this is somewhat of
a product of applying trigonometric functions of half the eigenangle, per the quaternion
definition. The cos 𝜓/2 term yields more inputs closer to one over the domain 𝜓 ∈ [−𝜋, 𝜋],
while the 𝑒𝑖 · sin 𝜓/2 terms do the opposite, yielding more terms close to zero.

To ensure that the DoE contained adequate representation over the whole input space,
random subsets of starting and stopping quaternions were selected to form rotations in the
orbital frame, and then transformed into relative quaternion maneuvers in the body-frame.
This operation is largely redundant; however, it was done to validate the effectiveness of
reducing orbital inputs into relative quaternion inputs as a simplification technique. Also,
the engineering model discussed in Chapter 4 requires an 8 × 1 input vector consisting of
a starting and stopping quaternion, so it is convenient to have a rotation expressed in all
three manners. The distribution of relative quaternion elements is depicted in Figure 5.7;
this operation had the effect of applying a more negative kurtosis to the 𝑞 1,2,3 elements and
reducing the severity of the left skew in the 𝑞 4 element.

Figure 5.7. Relative quaternion DoE histograms.

93
Both of these methods were used to train candidate neural networks, where their comparative
performance was evaluated. All input candidates had strong correlation with output times,
however the Euler parameters had slightly better performance. As the basis of a quater-
nion rotation, but without the mathematical operations required to generate a quaternion,
this result is intuitive, as it is the most direct expression of the physical rotation desired.
Accordingly, Euler parameters were selected as the input form for this model. Specific
configuration performances are addressed in more detail in the next section.

5.3 Minimum Time Rotation Surrogate Models

With the DoE available in both relative quaternion and Euler parameter form, a neural
network was the first surrogate model built and trained.

5.3.1 Neural Networks

Broadly, neural networks are nonlinear models that rely on many linear modeling processes
as well as some non-linear processes. “A neural network is a two-stage regression or
classification model, typically represented by a network diagram.... This network applies
both to regression [and] classification. For regression, typically...there is only one output
unit 𝑦1 ” [19]. For this thesis, the neural network is a regression model that has a single
output. Figure 5.8 is an example of a small neural network of this type.

Feed-Forward Neural Networks

A common type of neural network is a feed-forward network, in which layers are positioned
in sequence and inputs flow from the input layer sequentially through the output layer.
Figure 5.8 is a fully-connected feed-forward network in that information flows in one
direction through the network, and also each node in a layer is connected to each node in
the adjacent layers.

In this example, the input layer consists of two nodes, which is analogous to two input
variables. There are two hidden layers, each consisting of two nodes or neurons. Finally,
the output layer is a single neuron representing one output, which is a typical configuration
for a regression, according to Hastie [19].

94
Input Layer Hidden Layer 1 Hidden Layer 2

+1
12
𝑤11 +1

𝑤0111 Output Layer

𝑝1 −1 −1
𝑤 23
𝑤 12 11
𝑤 01 12
12
𝑎 11 = 𝑓 [ 𝑤𝑝] + 𝑏 11 𝑎 21 = 𝑓 𝑤𝑎 1 + 𝑏 21
Í
Í 

01 1 ŷ
𝑤2 12
𝑤 21 23
𝑤 21
𝑤2021 12
𝑝2 𝑎 31 = 𝑤𝑎 2 + 𝑏 31
Í 
+1 𝑤22 +1

−1 −1

𝑎 12 = 𝑓 [ 𝑤𝑝] + 𝑏 12 𝑎 22 = 𝑓 𝑤𝑎 1 + 𝑏 22
Í
Í 

Figure 5.8. Example neural network with two hidden layers.

The input layer performs no operations on the network inputs, it merely passes them along
to the next layer. Weights are assigned to all node edges. All neurons except the input layer
receive all weighted inputs, sum them, and add a bias term for that neuron. The resulting
scalar value is simply a linear combination of its inputs and is referred to as the logit. For
example, the logit for the first node in the first hidden layer in the example network from
Figure 5.8 would be 𝑧11 = 𝑤11
01 𝑝 + 𝑤01 𝑝 + 𝑏 1 . After calculating the logit, the non-linearity
1 21 2 1
of the neuron is introduced by an activation function.

For each neuron, the non-linear activation function is applied to the logit, 𝑧, in the form
𝑎 = 𝑓 (𝑧) where 𝑎 is the output of the neuron. Many different activation functions can be
used; among those, sigmoid functions are common. In this model, the hyperbolic tangent
sigmoid function is the activation function applied to all hidden and output layer neurons.
Figure 5.9 plots this function, showing its domain of all real numbers, which maps to an
output range between one and negative one, 𝑎 ∈ (−1, 1). It is differentiable and returns
values close to zero when the input argument is close to zero; both of those traits aid
in training network. The formula for a hyperbolic tangent sigmoid function is given in
Equation (5.2).

95
 +1

−1
Figure 5.9. Hyperbolic tangent sigmoid activation function.

𝑒 𝑧 − 𝑒 −𝑧
𝑓 (𝑧) = (5.2)
𝑒 𝑧 + 𝑒 −𝑧

The same process occurs in all neurons downstream of the input nodes until the output
predictions are produced by the output layer. These predictions are typically referred to by
the variable 𝑦.
ˆ

Figure 5.8 labels the relevant mathematical operations with their positions in the network,
highlighting the relationships between the layers and the forward flow of information through
the network used to make predictions. The scripting convention for network parameters is
𝑗 𝑚 , where 𝑘 is the previous layer, 𝑛 is the next layer, 𝑗 is the previous neuron, and 𝑚 is the
𝑤 𝑘𝑛
next neuron. Similarly for activation functions and biases, 𝑎 𝑛𝑚 represents the 𝑚th neuron in
layer 𝑛. All mathematical operations are linear with the exception of the activation function,
Equation (5.2); superscripts represent layers and not exponential operations [18].

Neural Network Hyperparameters In the previous section, several features of neural

networks were described, some aspects of which are adjusted by the model itself and some
of which are set in advance by the designer. The aspects of the network that are set in
advance are called hyperparameters. Hyperparameters are typically aspects of the model’s
structure. The number of hidden layers, input nodes, output nodes, and nodes per hidden
layer are all hyperparameters, as is the activation function applied in neurons. The remaining
parameters are adjusted by the model during the training process. Examples of parameters
adjusted by the model are weights and biases.

Training Neural Networks After a supervised learning model is constructed, it must be

trained to the labeled data. During this process the model makes adjustments to weight and
bias parameters.

96
Neural networks are trained by randomly initializing model weights and biases, and then
propagating training inputs through the model. The predicted values are evaluated against
the target data set and adjustments are made to the weights and biases of each neuron based on
the output error through a process called backpropagation. “The backpropagation algorithm
leverages the chain rule of differential calculus, which computes the error gradients in
terms of summations of local-gradient products over the various paths from a node to the
output” [20]. Updates to the weights and biases are then computed based on these error
gradients. The Levenberg-Marquardt backpropagation technique was used in this thesis; it
is the recommended training method in MATLAB’s Deep Learning Toolbox for supervised
learning problems. Together, one forward and backward sweep through the network is an
epoch. Training continues through multiple epochs until a stopping criterion is reached.
Additional details and the mathematical derivations for backpropagation are available from
Hastie [19] or Aggarwal [20].

RSIP Neural Network

When designing the neural network for the Remote Sensing Iterative Planner, the hyperpa-
rameters that have the greatest effect on model performance were found to be the number
of hidden layers and the number of nodes in each hidden layer. “Generally speaking it is
better to have too many hidden units than too few. With too few hidden units, the model
might not have enough flexibility to capture the nonlinearities in the data; with too many
hidden units, the extra weights can be shrunk toward zero” [19].

More than four nodes in the hidden layers are desired to ensure that features from each of the
four element input vector can be captured. Several different layer and node combinations are
presented in tables 5.1 and 5.2 for performance comparison. For a relatively simple problem,
a shallow network should suffice. Accordingly, possible neural network configurations
investigated here consisted of one or two hidden layers.

As the initial weights and biases are initiated with some degree of randomness, training
the same model with different initializations or random number seeds typically results in a
slightly different final model. In some cases, training multiple models and averaging their
predictions yields a better result than relying on an individual model. This approach was
explored here as well. Multiple configurations were investigated for the DoE in two forms:
relative quaternion elements and Euler parameters.

97
Relative Quaternion Input Evaluation The relative quaternion results are provided in
Table 5.1, which reports the predicted final time (𝑡 𝑓 ) mean absolute error (MAE) and
standard deviation (𝜎) for different hidden layer configurations. This table presents results
for relative quaternion input (Δq) performance of the average of five network outputs and
the expected performance of an individual network. Here, the performance achieved by
averaging the outputs of five individually trained neural networks exceeds the performance
the five networks evaluated independently. This chart implies that for a given model,
better performance should be expected from the average of five different predictions when
compared to the prediction of any one of those networks alone. This is the result of averaging
the effects of random error from the individual neural networks that is the product of their
random initializations during the training process.

A representative training performance plot is presented in Figure 5.10. MATLAB was con-
figured to discontinue neural network training after ten consecutive increases in validation
error, and then selected the epoch with the least validation error as the final model. For
this network, training was stopped at the 138th epoch, and the 128th epoch had the best
validation performance at 1.7804 s2 mean squared error. The validation set was 15% of the
labeled data while 85% of the data was used for training. An independent 𝑛 = 100 test set
was used evaluate performance with new data, depicted in Figure 5.11.

Searching for the overall best performer using a relative quaternion input, the networks that
appear to perform best both contain two hidden layers, one of 13 nodes each and one of 20
nodes each. These two networks have similar performance, with each excelling in either
mean absolute error or standard deviation.

Euler Parameter Input Evaluation The same analysis was conducted with Euler parameter
input vectors. These results are presented in Table 5.2. Here, we see the same conclusion
that averaging the prediction of five neural networks leads to a better overall result than
should be expected from any individual network. Furthermore, the predictions using Euler
parameter inputs tend to outperform relative quaternion inputs in both mean absolute error
and standard deviation. This is an intuitive conclusion, given that both inputs provide
the same information, however relative quaternion elements are masked behind an extra
mathematical transformation which could confound variation due to the eigenaxis and
variation due to the eigenangle.

98
 Table 5.1. Neural network predicted time error with relative quaternion inputs.

Average of
Hidden Layer Using 5 Networks Individual Networks
Configuration MAE (s) 𝜎 (s) MAE (s) 𝜎 (s)
[50] 0.888 1.558 1.414 2.482
[13] 1.036 1.616 1.310 1.982
[120] 1.170 1.748 2.067 3.722
[13, 13] 0.892 1.467 1.415 2.453
[50, 50] 1.242 1.889 2.336 4.117
[50, 13] 1.072 1.727 1.562 2.736
[120, 50] 1.503 2.753 3.127 6.123
[120, 13] 1.104 1.865 1.961 3.346
[7, 7] 1.157 1.837 1.742 2.670
[20, 20] 0.817 1.536 1.373 2.803
[13, 7] 0.917 1.489 1.365 2.233
[20, 7] 1.007 1.624 1.482 2.326
[20, 13] 1.037 1.909 1.443 2.600

Notably, for both sets of inputs, the two best performing models both consisted of two hidden
layers comprised of 13 or 20 nodes. In general, Euler parameter inputs tended to outperform
relative quaternion inputs. Due to its performance and the reduced computational burden
associated with 13-node hidden layers over 20-node hidden layers, the final neural network
architecture comprises two hidden layers of 13 nodes each with Euler parameter inputs.
The structure of the final neural network is graphically depicted in Figure 5.12.

Using the final neural network architecture, a 100 sample point test set was used to evaluate
model performance with a very high coefficient of determination, adjusted for sample size
(adjusted 𝑅 2 = 0.998, Figure 5.11).

99
 Figure 5.10. Neural network validation set performance.

Figure 5.11. Neural network averaged predictions versus target outputs for
test data set, 𝑅 2 = 0.998.

100
Table 5.2. Neural network predicted time error with Euler parameter inputs.

Average of
Hidden Layer Using 5 Networks Individual Networks
Configuration MAE (s) 𝜎 (s) MAE (s) 𝜎 (s)
[50] 0.800 1.454 1.126 1.938
[13] 0.782 1.316 1.187 1.814
[120] 1.050 1.580 1.840 2.824
[13, 13] 0.716 1.270 1.022 1.845
[50, 50] 1.270 2.309 2.295 4.272
[50, 13] 0.872 1.623 1.535 2.642
[7, 7] 1.079 1.628 1.606 2.364
[20, 20] 0.756 1.184 1.182 2.057
[13, 7] 0.785 1.353 1.157 1.996
[20, 7] 0.855 1.345 1.225 2.084
[20, 13] 0.769 1.296 1.114 1.948

101
 +1 +1

−1 −1

+1 +1

−1 −1

+1 +1

−1 −1

+1 +1

𝑒1 −1 −1

+1 +1

−1 −1

+1 +1

𝑒2 −1 −1

+1 +1

−1 −1
𝑡𝑓

𝑒3 +1

−1
+1

−1

+1 +1

−1 −1

+1 +1

𝜓 −1 −1

+1 +1

−1 −1

+1 +1

−1 −1

+1 +1

−1 −1

Figure 5.12. Feed-forward neural network with two, 13-node hidden layers.

102
5.3.2 Stepwise Multiple Linear Regression
Neural networks are an exciting and robust approach to non-linear approximations, and
while they are significantly less computationally demanding than the optimal maneuver
solver addressed in Chapter 4, they still require a significant number of mathematical
operations to execute, depending on the number of nodes involved. The “linear model has
distinct advantages in terms of inference and, on real-world problems, is often surprisingly
competitive in relation to non-linear methods” [21]. In this case, there are multiple inputs
that map to a single output, which is well suited to the linear regression approach. For those
reasons, a linear regression technique was built using the same DoE as the neural network
to compare the relative advantage of each model.

One potential advantage of linear regression models over neural networks is model inter-
pretability. “It is often the case that some or many of the variables used in a multiple
regression model are in fact not associated with the response. Including such irrelevant
variables leads to unnecessary complexity in the resulting model” [21]. Analyzing a re-
gression model after fitting for significant variables by their 𝑝-value and coefficient can
lead to insights in system performance. A “𝑝-value is the probability that, if the null hy-
pothesis is true, the results from another randomly selected sample will be as extreme or
more extreme as the results obtained from the given sample” [22]. Thus for low 𝑝-values,
typically 𝑝 < 0.05, an analyst can generally assume that the observed effect is statistically
significant and not due to random chance. This kind of peak-under-the-hood is not available
with neural networks, due to the complex nature of propagating neural network inputs.

The specific technique chosen for RSIP is stepwise multiple linear regression or stepwise
regression, which is a variable selection technique in that it uses a statistical method to
determine which terms in a linear regression are significant to include. “Variable selection
is used when the analyst is faced with a series of potential predictors but does not have (or use)
the necessary subject matter knowledge to enable her to prespecify the ‘important’ variables
to include in the model” [23]. This allows the regression to select what combination of Euler
parameter main effects, interactions, and second-order effects provide significant influence
over the system output.

The basic form for a general linear regression is provided in Equation (5.3), where 𝛽𝑖 are
coefficients for input variables, 𝑥𝑖 are the input variables, and 𝜖 is random error. Our model

103
is analytical, so while error between predictions and target values will remain, this is a
noiseless system. Also, the input terms 𝑥𝑖 apply to all combinations of inputs, and not
simply the main effects (𝑒 1 , 𝑒 2 , 𝑒 3 , 𝜓); an 𝑥𝑖 term also accounts for the interaction 𝑒 1 · 𝑒 2
and another for the quadratic term 𝜓 2 , etc.

y = 𝛽0 + 𝛽1 𝑥 1 + . . . + 𝛽 𝑝 𝑥 𝑝 + 𝜖 (5.3)

Note that linear regression refers to the linear relationship between the model coefficients,
𝛽𝑖 and the model variables, 𝑥𝑖 [21]. A term such as 𝜓 2 is treated as an independent input
variable for the regression and maintains a linear relationship with its coefficient 𝛽; it does
not imply a non-linear action.

Stepwise Multiple Linear Regression Model Fitting

This model was fitted using MATLAB. In general, stepwise multiple linear regression
regression models use forward and backward selection to add or remove terms to the model
one at a time. The addition or removal of a term is based on the statistical significance of
its effect on the variance of the outcome, based on an 𝐹-test.

“Forward stepwise selection begins with a model containing no predictors, and then adds
predictors to the model, one-at-a-time, until all of the predictors are in the model. In
particular, at each step the variable that gives the greatest additional improvement to the fit
is added to the model.... [Backward stepwise selection] begins with the full least squares
model containing all... predictors, and then iteratively removes the least useful predictor,
one-at-a-time” [21].

F-test This model began with all quadratic terms, which includes the constant (𝛽0 ), main
effect (e.g., 𝑝 1 ), interaction (e.g., 𝑝 1 · 𝑝 2 ), and second-order effect (e.g., 𝑝 21 ) terms. The
algorithm then fits the regression to the training data by determining the coefficients (𝛽𝑖 )
and terms that minimize prediction error using a least squares method.

104
The F-statistic is determined by the relationship between the variance of the model before
the addition or removal of a term and afterwards [22]:

(𝑆𝑆total − 𝑆𝑆𝐸 between )(𝑘 − 1)

𝐹= (5.4)
𝑆𝑆𝐸 between (𝑛 − 𝑘)

The training algorithm uses this test to determine if the difference in variance between the
model with the term included and with the term removed was significant. The criterion
used for this model was the sum of squared errors (SSE). The threshold for adding terms is a
𝑝-value of 0.05 or less and 0.10 or greater for removing them. Thus, if the 𝐹-test determines
that there is a greater than 0.10 probability that the true value for the term’s coefficient is
zero, the term is removed. This method is a type of linear regression and therefore requires
linear regression assumptions: errors are normally distributed, the variance of errors is
constant, and the errors are uncorrelated [22]. As the training and test data for this model
come from analytical sources and not nature, the impacts of noise are negligible, though
prediction error will not be eliminated.

If several terms are available to be added or removed from the model, MATLAB will only
act on one term at a time, choosing the term with the lowest 𝑝-value. A higher order or
interaction term will not be added unless both of the main effects are included as well.
Similarly, a main effect term will not be considered for removal if significant higher order
or interaction terms are present. These rules ensure adherence to DoE principles including
scarcity of effects, hierarchy of factors, and heredity of factors.

The stepwise regression model fitted for this problem using the Euler parameter data set is
presented in Table 5.3. The final stepwise linear regression is also presented in Equation
(5.5).

ŷ = 𝛽0 + 𝛽1 𝑒 1 + 𝛽2 𝑒 2 + 𝛽3 𝜓 + 𝛽4 𝑒 1 𝑒 2 + 𝛽5 𝑒 2 𝜓 + 𝛽6 𝑒 22 + 𝛽7 𝜓 2 (5.5)

The main effects that were found to be statistically significant based on the DoE are 𝑒 1 , 𝑒 2
and 𝜓. Recall from the DoE analysis that the eigenangle of rotation has a strong impact on
maneuver time. Its inclusion with an extremely low 𝑝-value and a relatively large coefficient
is intuitive. For context, with all other terms held constant, an increase in the eigenangle

105
 Table 5.3. Stepwise multiple linear regression terms.

Term (𝑥𝑖 ) Coefficient (𝛽𝑖 ) 𝑝-value

(Intercept) 4.1513 7.28 × 10−14
𝑒1 0.40122 0.17733
𝑒2 −1.6608 0.00466
𝜓 40.376 3.82 × 10−285
𝑒1 · 𝑒2 1.416 0.017756
𝑒2 · 𝜓 1.8547 2.09 × 10−10
𝑒 22 5.0066 1.22 × 10−17
𝜓2 −0.52705 0.012203

of 1 rad would lead to a 40.376 second increase in maneuver time. Of course, given the
interaction and higher order terms that also include the eigenangle, this thought experiment
is not directly realizable, but it is a helpful way to gain intuition about the relative effects of
the input variables based on their coefficients.

Recall that the inertia matrix for the spacecraft in question is

 223.66 0 0 
 
I=  0 264.71 0  kg · m2 . (4.8)
 
 

 0 0 170.23 

The inclusion of the 𝑒 1 and 𝑒 2 terms of the eigenaxis implies that those terms are more
significant than the 𝑒 3 term, which is not included. This is perhaps the case because they map
to the higher moments of inertia along those axes, shown in Equation (4.8). Rotations that
require movement around principal axes with higher moments of inertia should be
expected to take longer, given a minimum time bang-bang control solution. In this
particular model, Equation (5.5) indicates that the time prediction for any rotation about
the 𝑒3 axis, for example, will be exclusively a function of 𝜓 from the 𝛽3 and 𝛽7 terms, as
well as the intercept, 𝛽0.

Ultimately, the model defined in Table 5.3 and used here is a Response Surface Model be-
cause its input variables include main effects, interactions, and second-order (or “quadratic”)

106
terms [16]. The terms give curvature and twist to the 𝑛-dimensional hyperplane that results
from the regression. Here, the model represents a four-dimensional hyperplane after drop-
ping the 𝑒 3 variable, leaving three independent input variables plus the prediction output.
These characteristics allow the model to capture some non-linear features while still using
linear regression to determine model coefficients.

Full Linear Regression Comparison

Stepwise regression can be a valuable tool for models that contain a large number of
independent input variables. For example, in some cases it might be undesirable or infeasible
to evaluate a complete quadratic model with 100 independent input variables and forward
stepwise selection may be an appealing framework to find the interactions and quadratic
terms that have a significant effect on the output’s variance. However, some of the available
literature is skeptical about stepwise regression in principal because the 𝑝-values used for
selection of terms are based on the estimates of the term coefficients, as well as several
assumptions that are explored in more detail in Harrell [23].

“The problems of 𝑝-value-based variable selection are exacerbated when the analyst (as she
so often does) interprets the final model as if it were p re-specified. Copas and Long stated
one of the most serious problems with stepwise modeling eloquently when they said, ‘The
choice of the variables to be included depends on estimated regression coefficients rather
than their true values, and so 𝑋 𝑗 is more likely to be included if its regression coefficient is
over-estimated than if its regression coefficient is underestimated”’ [23].

The importance of the final i nput t erms i s r elatively e vident i n t his s tepwise m odel, as
they are directly mapped to a physical system. Longer rotations will take longer, as should
rotations that require large reorientations around axes with larger moments of inertia. Only
the smallest principal axis moment of inertia component (𝑒3) was deemed insignificant by
the stepwise model.

However, Harrell concluded that “[u]nless most predictors are either very significant or
clearly unimportant, the full model usually outperforms the reduced model. Full model fits
have the advantage of providing meaningful confidence intervals using standard formulas”
[23].

107
 Table 5.4. Linear regression terms.

Term (𝑥𝑖 ) Coefficient (𝛽𝑖 ) 𝑝-value

(Intercept) 4.0401 2.63 × 10−9
𝑒1 0.29854 0.62412
𝑒2 -1.7053 0.003986
𝑒3 0.73392 0.2257
𝜓 40.367 7.5 × 10−281
𝑒1 · 𝑒2 1.3959 0.020321
𝑒1 · 𝑒3 0.55279 0.37967
𝑒1 · 𝜓 0.042562 0.89433
𝑒2 · 𝑒3 -0.87833 0.1482
𝑒2 · 𝜓 1.9047 1.21 × 10−10
𝑒3 · 𝜓 -0.34041 0.29706
𝑒1 2 0.17883 0.81871
𝑒22 5.1237 7.69 × 10−13
𝑒 23 0 N/A
𝜓 2 -0.5163 0.01500

To explore the impact of this concern in this model, a full linear regression with no selection
or removal of terms was fitted to compare performance with the stepwise model. This linear
regression included all independent input terms as main effects, their interactions, and their
quadratic terms. The resulting model is defined in Table 5.4. A graphical comparison of
test set performance is presented in Figure 5.13.

The test sets show a model and fit that is very similar between the stepwise and full linear
approaches for this particular case. Accordingly, there is no apparent detriment to removing
some less significant terms from the regression for optimal maneuver time approximation.
It remains for future work to further explore various linear regression implementations to
find a desirable balance between computational efficiency and model accuracy.

108
 (a) Stepwise regression, model in Table 5.3. (b) Linear Regression, model in Table 5.4.
Figure 5.13. Test data predicted vs. target outputs.

5.4 Mission Plan Analysis

The advantage of the two supervised models developed in this chapter over legacy techniques
is their ability to output maneuver times that are close to the optimal time without the need to
explicitly solve the optimal control problem. Chapter 2 discusses the taxicab and eigenaxis
legacy maneuvers, including their suboptimality. Chapter 3 posited that the full maneuver
profile is not needed to support a path planner; only the time required to execute the
maneuver is needed. Chapter 4 provided the engineering model needed to produce time-
optimal slewing maneuvers. Now that two supervised learning techniques are available,
path planning programs should be able to take advantage of the approximated maneuver
times to generate more efficient mission plans.

5.4.1 Minimum Time Approximation Mission Plans

RSIP was augmented to run with approximated optimal slew times using these new tech-
niques; the results of this effort are presented in this section.

Neural Network Mission Model

A sample mission model using a neural network to approximate minimum time maneuver
slew times is shown in Figure 5.14. This particular mission was flown at 675 km altitude
and collected eight different targets for a total collection value of 20 VU.

109
 Figure 5.14. Mission plan using neural network optimal slew approximations,
value maximizing.

Stepwise Regression Mission Model

A sample mission model using a stepwise regression function to approximate minimum
time maneuver slew times is shown in Figure 5.15. This mission was flown at 675 km
altitude and collected nine different targets for a total collection value of 19 VU.

5.4.2 Legacy Mission Plan Comparisons

The sample mission plans provided in Figures 5.14 and 5.15 are enticing results, but
require further inspection. The feasibility of the mission plans must be investigated prior
to being implemented on an actual system, however these plans were created in the exact
same fashion as the plans discussed in Chapter 3 that used legacy slewing maneuvers
(figures 3.10 and 3.11). Only the code used to generate the maneuver time, 𝑡 𝑓 , required
adjustment. In the case of the legacy models, that code provided a simple analytical

110
 Figure 5.15. Mission plan using stepwise regression optimal slew approxi-
mations, value maximizing.

solution. However, the neural network calculates maneuver times through a sequence
of algebraic scaling operations, matrix multiplications, closed-form combinations, and
algebraic unscaling operations. The stepwise model produces a maneuver time through
simple linear operations on input variables. With these distinctions, the path planner
operates with the exact same mechanisms.

To facilitate comparisons, nine mission plans were generated for each of the four maneuver
time calculation techniques previously discussed, and the actual optimal mission time
required to execute each plan was also produced, using the engineering model from Chapter
4. For each method, mission plans were created from three different orbital altitudes:
675 km, 875 km, and 1200 km. For each altitude, three plans were created using different
random number seeds for neural network initialization and genome permutations in the
genetic algorithm. This led to distinct solutions of similar value and efficiency.

111
 Average Targets Collected 20.0

15.0

10.0

5.0

0.0
675 km 875 km 1200 km
Planning Method, by Altitude

Taxicab Eigenangle Neural Net Stepwise Regression

Figure 5.16. Average number of targets collected, by planning method and

spacecraft altitude.

Figure 5.16 compares the average number of targets collected, broken down by altitude
and maneuver method. Consistently, the optimal approximation techniques outperform
legacy techniques in number of targets collected. Figure 5.17 makes the same comparison,
but measures the value of the mission plan instead of the number of targets collected.
In this form, the relative performance of the methods remains the same. For different
orbital altitudes over the same target field, it is intuitive that a higher altitude would enable
additional collections due to the slower orbital speed as well as a slightly larger access
window thanks to additional slant range for a given look angle constraint, which in this
model is 30 degrees.

Feasibility and Improvement

By design, the planning program is a low fidelity mission sequence planner, in that it provides
merely a sequence of targets to collect and does not issue a high fidelity solution without
the use of the full engineering model. RSIP attempts to achieve that function by integrating
the engineering model into the final phase of the planner, where the promulgated mission
sequence is input into the high fidelity engineering model to compute actual maneuver
times, in lieu of approximations.

112
 Average Value Collected 35.0
30.0
25.0
20.0
15.0
10.0
5.0
0.0
675 km 875 km 1200 km
Planning Method, by Altitude

Taxicab Eigenangle Neural Net Stepwise Regression

Figure 5.17. Average mission value collected, by planning method and space-
craft altitude.

A successful implementation of the planning principles proposed in this thesis does not
require an exact approximation or a precise high fidelity mission plan. Rather, success is
defined in two ways:

1. Does this mission planner produce a feasible mission sequence that assumes optimal
maneuvers?
2. Does the mission plan allow more value to be obtained than with legacy techniques?

If the answers to both of these questions are “yes,” then the model is successful in that it
represents an improvement over the status quo of mission planning that can be achieved
simply by incorporating proven techniques to increase mission efficiency.

While the individual maneuvers produced by an engineering model like the one presented in
Chapter 4 are proven to be mathematically optimal, convergence on an efficient solution for a
genetic algorithm-driven path planner like RSIP does not come with the same assurances of
optimality. The possibility of suboptimality in the mission planner should not be considered
a risk, however. For similar planning effort, more mission accomplishment can be obtained
and that advance alone is valuable.

113
Slew Time Prediction Error
An important metric for the supervised learning methods is their accuracy in approximating
slews. Much discussion has been made of accuracy in approximating individual slews, as
well as test data and training datasets. A useful macro metric is the overall residual error
from a mission plan generated using approximations of optimal slews with that same mission
plan executed with the actual optimal slews. Those residual comparisons are presented in
Figure 5.18.

The two optimal approximation techniques are considerably closer to the optimal slew,
noting the logarithmic scale on the vertical axis. For this model, the stepwise regression
technique yields smaller residuals than the neural network, though both are significant
improvements over legacy techniques. This chart encompasses data from the same mission
plans discussed herein, where the slews are relatively small. It is reasonable to infer that
the gains from this method would be more advantageous for a plan that required generally
longer slews.

Relative Value of this Approach

The amount of efficiency gained through this technique must be determined for each indi-
vidual system to which it is applied, as it will be a function of the physical properties of the
spacecraft, the planning techniques currently in use, and typical slews for the system. For
example, systems with longer slews may see more significant efficiency gains than systems
that generally make small slews. For the six mission sequences presented here for direct
neural network–stepwise comparison, the median slew is 5.4 degrees and the maximum
slew is 24.6 degrees. Whether or not the potential gains are worth the effort of reworking
planning systems and incorporating optimal maneuvers will depend on individual satellite
operators, however with demonstrated gains of 5% to 21% for individual maneuvers as
previously discussed, the potential could be significant and should be explored [2].

The direct comparisons presented in this section make a compelling case for taking advan-
tage of these techniques. Using the the results in Figures 5.16 and 5.17 and the eigenaxis
technique as a baseline, tables 5.5 and 5.6 quantify these advantages. For example, using
a neural network enabled optimal slew approximation technique for a system at 675 km,
operators could expect an increase in targets imaged on the order of 14%. Using a stepwise
regression enabled technique for the same mission could yield an increase on the order of

114
 100.0 73.8 89.4
51.8
Mission Time Residuals (s)
35.1
22.3 27.5 22.7
9.9 12.0
10.0 6.7

1.5 1.7

1.0
675 km 875 km 1200 km
Planning Method, by Altitude

Taxicab Eigenangle Neural Net Stepwise Regression

Figure 5.18. Mission time residual error, by planning method.

Table 5.5. Targeting gains resulting from optimal approximation techniques.

Taxicab Eigenaxis (baseline) Neural Network Stepwise Regression

675 km 4.7 (-33%) 7.0 8.0 (14%) 8.7 (24%)
875 km 7.0 (-28%) 9.7 10.0 (3%) 11.0 (14%)
1200 km 10.7 (-18%) 13.0 13.7 (5%) 15.7 (21%)

24% for a system similar to this example. For this concept demonstration, double digit
efficiency gains are consistently achieved.

5.4.3 Small Angle Performance

Surprisingly, in this model the stepwise regression technique outperforms the neural net-
work. An intuitive assumption may be that the neural network would have more opportunity
to capture relationships and non-linearities in the data given the complex interactions asso-
ciated with feed-forward networks and the number of hidden layer nodes. This expectation
is also supported by coefficients of determination for test set performance. The adjusted 𝑅 2
for the neural network was 0.998, whereas for stepwise regression it was 0.974.

115
 Table 5.6. Value gains resulting from optimal approximation techniques (VU).

Taxicab Eigenaxis (baseline) Neural Network Stepwise Regression

675 km 12.0 (-31%) 17.3 18.7 (8%) 19.7 (13%)
875 km 17.3 (-15%) 20.3 22.7 (11%) 24.3 (20%)
1200 km 22.7 (-18%) 27.7 29.3 (6%) 31.3 (13%)

(a) Neural network over small angles, (b) Stepwise regression over small angles,
architecture in Figure 5.12. model in Table 5.3.
Figure 5.19. Test data predicted vs. target outputs over small angles.

A possible explanation for this phenomenon is that while the neural network performs better
over the full domain of possible slews (𝜓 ∈ [−𝜋, 𝜋]), the vast majority of slews required
for the missions modeled here are small slews, generally less than 25 degrees. While the
neural network should be expected to perform better over the full input space, over small
slews the neural network is biased slightly higher than the stepwise model, leading to a more
consistent over-estimation of the true optimal slew time. A comparison of the performance
of the trained models over small angle slews is presented in Figure 5.19. For this test set with
eigenangles from zero to 0.26 rad (15°), bias in both models is visually evident. Table 5.7
provides summary statistics for small angle mission residuals from optimal performance.
While both optimal approximation techniques have similar residual variances, the mean
residual is considerably higher for the neural network. Finally, Figure 5.20 depicts this
phenomenon graphically; while the variance for both techniques is similar, the mean error
for the neural network is biased higher.

116
 Figure 5.20. Mission residuals from optimal execution, by method.

Table 5.7. Summary statistics for small angle test set.

Neural Network Stepwise Regression

Mean (s) 1.44 0.19
Variance (s2 ) 3.36 3.21

With a relatively small number of samples, one should not draw the conclusion that one
technique is better than the other, rather that both yield feasible results, and for these
two specific models with this training dataset, the stepwise regression model does tend to
perform slightly better.

Subsequent investigations of these techniques and real-world implementation should not

discount either method; however, further effort and feature engineering is needed to develop
a DoE that achieves best model performance over the types of slews most likely to be
required. For RSIP, the model does not achieve its best performance over small slews,
which make up nearly all of its mission maneuvers. Subsequent versions of this planner
will take that into account and design a more appropriately tailored DoE.

117
5.4.4 Computational Comparisons
Separate from performance comparisons, the simple nature of regressions may make them an
appealing option when additional computational efficiency is desired. While both a neural
network and a regression are significantly s impler a lgorithms t han t he f ull engineering
model, stepwise regression will typically be simpler still than a neural network. In this
case, the number of floating-point o perations (FLOPs) t o c ompute a m aneuver t ime is
orders of magnitude different for the different methods. A FLOP is a single arithmetic
operation of one value on another. In computer memory, these values are stored as floating-
point numbers, which is simply a protocol for storing non-binary numbers in a computer’s
memory. Determining the number of floating-point operations required to complete a set
of code can be a useful proxy for the complexity of the code and the computing resources it
requires. Table 5.8 compares the maneuver time methods used in RSIP to compute a single
maneuver time.

Taxicab and eigenaxis methods are both very efficient computationally (and of course
inefficient physically). The taxicab computation is simply two eigenaxis maneuvers
around different principal axes, followed by their summation. The stepwise regression
prediction is also relatively efficient. With seven terms and four interactions, the stepwise
method requires 17 FLOPs to compute a maneuver time. For the neural network, two
hidden layers of 13 nodes is a far more complex prediction scheme. Furthermore, to lower
the prediction variance to the desired level, the average prediction of five networks is used
in RSIP, which requires five times the floating-point operations as a single neural network.

Table 5.8. Floating-point operations required for maneuver time methods.

Method FLOPs
Taxicab 7
Eigenaxis 3
Stepwise Regression 17
Neural Network 756
5 Neural Networks 3785

118
One conclusion that should be considered, is that even if a neural network is outperforming a
regression technique, it may still be advantageous to consider a regression if its performance
is adequate and the decreased computational requirements are significant. In this particular
case, the the regression performs slightly better and runs considerably faster, but it is likely
that many applications will have better performance with a neural network. For the current
version of RSIP, stepwise regression for calculating maneuver times is clearly the most
desirable approach.

5.5 Summary
The goal of this thesis is to explore the possibility of integrating time-optimal maneuvering
as part of a mission planning process, which would enable widespread adoption of these
techniques in space enterprises. This chapter addresses the final discipline required to
achieve such a mission planning process using supervised machine learning.

Design of Experiments provided a useful sampling of the the input space for this problem,
which was developed into labeled data using the optimal solver discussed in Chapter 4. With
this labeled data, multiple formats for the input data were evaluated, as well as multiple
approaches for supervised learning models. For this system, a neural network model and a
stepwise multiple linear regression model were selected based on their performance, both
using Euler parameter inputs.

When augmenting the mission planner with these metamodels, both minimum time approx-
imation techniques performed significantly better than legacy techniques. This outcome
reinforces the conclusion that with simple planning changes, space system operators can
realize double-digit increases in mission efficiency using techniques already demonstrated
on-orbit.

119
THIS PAGE INTENTIONALLY LEFT BLANK

120
 CHAPTER 6:
Conclusion

This thesis addresses the absence of time-optimal attitude maneuvers in current space
operations. Time-optimal rotations are significantly faster than legacy slewing techniques
that are commonly used today. These minimum time maneuvers have been demonstrated
on-orbit, first in 2010; the lack of wide-spread adoption of this maneuvering technique is
remarkable. The inability to incorporate time-optimal slews into existing mission planning
frameworks is the likely source of this arrested development, and this thesis offers an
approach for minimum time maneuver implementation that leverages supervised learning.

Two viable supervised learning approaches that approximate the minimum maneuver time
were developed: a neural network and a stepwise multiple linear regression. Both of these
models were trained on a dataset of time-optimal slews, generated from the optimal control
problem presented in Chapter 4. Using these metamodels in existing planners can improve
mission efficiency without the need to solve optimal control (or hybrid optimal control)
problems as part of the planning process. Both of these approaches yield feasible mission
sequences that are significant improvements over the legacy mission planning approach.
This thesis thus demonstrates that existing mission planners can be augmented to account
for time-optimal spacecraft maneuvers by using supervised machine learning to approximate
optimal maneuver times. To incorporate these ideas into a mission planner, all that is needed
is to update the model used for slew time prediction.

6.1 Implications for Current and Future Missions

The potential added value from an approach such as this is significant and should be con-
sidered by any spacecraft operator with an over-subscription problem for mission planning.
Any operator with more targets to collect than time to collect them could directly and imme-
diately benefit from the incorporation of time-optimal maneuvers into their satellite slews,
enabled by a supervised learning augmented planning process. Today, double-digit per-
centage increases in imaging target collection is an attainable goal with the implementation
of the mission planning approach presented herein.

121
6.2 Future Work
The feasibility of the supervised learning approach has been demonstrated, however future
work remains to iterate upon this approach and refine its efficacy. Some potential areas for
further investigation follow.

Feature Engineering The relative superior performance of linear regression over neural
networks in this example system is a reminder of the efficiency and utility of regressions;
however, it also indicates that neural network performance could very likely be improved.
Feature engineering should be undertaken to more precisely identify the drivers of neural
network performance. Future investigation should identify whether training separate super-
vised learning models for acceleration-limited slews and rate-limited slews reduces overall
variance in the minimum time prediction.

Constrained Mission Plans Real spacecraft frequently have maneuver constraints based
on hardware limitations or bright bodies that may damage sensitive sensors inside certain
look angles. The incorporation of look-angle constraints and keep-out zones has been
demonstrated for optimal control problems previously [14], [24]. However, methods to
introduce the constraints into the combinatorial mission planning problem remain to be
demonstrated.

122
 APPENDIX A:
Simulation Target Locations

Simulation targets used for the RSIP mission models are presented here.

Figure A.1. Simulation targets.

123
 Table A.1. Simulation targets.

Target ℓ 1 (km) ℓ 2 (km) Value

1 0 0 1
2 88.44 -138.273 2
3 107.185 126.426 2
4 107.465 -19.578 3
5 123.24 191.754 2
6 130.535 -292.635 1
7 163.97 35.901 1
8 175.835 -86.796 3
9 223.145 -177.849 2
10 290.595 24.819 2
11 307.3 -110.184 3
12 308.345 -222.396 1
13 362.25 159.099 1
14 378.94 255.582 3
15 384.745 -266.841 3
16 411.665 -113.43 3
17 456.33 45.921 2
18 509.18 -61.731 3
19 528.025 285.438 2
20 561 -199.086 1
21 588.535 -117.477 1
22 592.365 111.837 2
23 604.555 46.176 1
24 663.385 -293.616 2
25 663.775 252.444 3
26 691.29 -162.771 2
27 772.55 260.529 2
28 784.98 -255.42 3
29 787.81 74.547 2
30 788.855 151.11 1
31 789.325 -86.208 2

124
 APPENDIX B:
Optimal Control Nomenclature

The following nomenclature is used to develop the minimum time optimal control model
used for this project. Ross’ primer on optimal control theory is a useful guide for further
information on this subject [13].

Trajectory Optimization Nomenclature

x = State Vector
u = Control Vector
𝐹 (x(𝑡), u(𝑡)) = Running Cost Function
𝐸 (x 𝑓 , 𝑡 𝑓 ) = Endpoint Cost Function
e(𝑥 𝑓 , 𝑡 𝑓 ) = Endpoint Condition Vector
h(x(𝑡), u(𝑡)) = Constraint Function
𝝀 = State Covector
𝝂 = Endpoint Condition Covector
𝝁 = Constraint Covector
𝐻 (𝑥, 𝜆, 𝑢, 𝑡) = Hamiltonian Functional
𝐻¯ (𝑥, 𝜆, 𝜇, 𝑢, 𝑡) = Hamiltonian Lagrangian
𝐸¯ (x 𝑓 , 𝝂, 𝑡 𝑓 ) = Endpoint Lagrangian
q = Quaternion Attitude Vector
𝜔 = Spacecraft rotational velocity [rad/s]
𝛼 = Spacecraft rotational acceleration [ rad/s2 ]
I = Inertia matrix of the spacecraft [kg · m2 ]
Z = Reaction wheel configuration matrix
𝜏 = torque [N · m]
ℎ = Reaction wheel angular momentum [N · m · s]
Ω𝑤 = Reaction wheel angular rate [ rad/s]
𝐼𝑤 = Reaction wheel moment of inertia [kg · m2 ]

125
Scaling Nomenclature

¯· = An overbar indicates that the variable or parameter has been scaled

𝑀 = Mass Units
𝐿 = Length Units
𝑇 = Time Units
𝐴 = Angle Units

126
 APPENDIX C:
Spacecraft High Fidelity Model

The high fidelity model from which the optimization and eigenaxis models were derived
is presented here. It can be found in Reference [14]. The model was modified to this
spacecraft’s characteristics and a path constraint that is not considered in this problem was
removed.

Minimize: 𝐽 [x(·), u(·), 𝑡 𝑓 ] = 𝑡 𝑓






Subject to: q¤ = 12 Q(𝜔)q







h¤ 𝑅𝑊 = 𝝉 𝑅𝑊






𝝎 = u1






𝝉 𝑅𝑊 = u2





High Fidelity: 𝑡0 = 0



x0 = q0 , h0𝑅𝑊

  𝑇



 h i𝑇


 x = q 𝑓 , h𝑓

 𝑓 𝑅𝑊






 −𝜏max ≤ 𝜏RW ≤ 𝜏max





 −ℎmax ≤ ℎRW ≤ ℎmax

ℎ(𝑡) , I𝝎 + Zh 𝑅𝑊 = 0





This high fidelity model takes into account the capabilities of the reaction wheels, as well
as torque and angular momentum saturation. It ensures that momentum is conserved by
using a constraint that defines the sum of the momentum of the spacecraft and the angular
momentum of the reaction wheel assembly in the body-fixed frame to be zero. The model
represents the spacecraft’s attitude with quaternion representation and includes the actual
spacecraft moment of inertia.

127
THIS PAGE INTENTIONALLY LEFT BLANK

128
 APPENDIX D:
Supervised Learning Dataset

The labeled data for time-optimal rotations that was used to train the neural network and
regression models is presented in this section.

Table D.1. Supervised learning labeled data, 𝑛 = 500.

Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)

1 -0.6265 -0.4512 0.6355 0.0801 9.2476
2 0.1120 0.5636 0.8184 0.1489 11.8064
3 -0.5548 0.5979 0.5785 1.9122 71.5721
4 -0.9404 0.0629 -0.3343 2.4987 108.8106
5 0.4730 -0.7356 0.4850 1.3772 57.1958
6 0.0099 -0.9999 0.0027 2.7790 130.7753
7 0.2485 -0.9683 -0.0263 0.9796 52.2292
8 -0.2857 -0.7171 0.6358 0.6209 29.7438
9 0.6615 0.6082 0.4388 0.2231 15.5634
10 -0.2498 -0.7766 -0.5784 1.5555 66.8205
11 -0.6249 0.0981 -0.7745 1.6805 72.7039
12 -0.0708 0.5646 -0.8224 2.4904 105.9342
13 -0.5373 -0.8434 0.0078 0.9187 43.7938
14 0.5759 0.2701 -0.7716 2.6069 106.0559
15 0.8071 0.3403 0.4824 0.7062 33.6040
16 -0.7252 0.3355 0.6013 2.9610 115.9445
17 0.9728 0.2243 0.0586 0.5883 33.7513
18 0.4298 -0.4633 -0.7750 0.3784 20.0895
19 -0.9457 0.3175 0.0701 0.3043 18.8720
20 0.1549 0.8135 0.5606 0.3210 18.5293
21 0.8709 0.4852 -0.0774 0.5550 29.0070
Continued on next page
129
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
22 -0.4988 0.5694 0.6534 2.3820 88.7381
23 0.8341 0.1452 0.5321 0.5381 27.4708
24 -0.6127 0.2728 -0.7417 0.9365 41.5137
25 -0.7924 -0.1405 -0.5936 2.9641 123.0128
26 0.5562 -0.7660 0.3223 1.9937 82.3595
27 -0.3391 0.5949 -0.7288 2.6207 104.0127
28 -0.1014 0.9922 0.0722 2.4193 113.5224
29 0.5818 -0.5006 -0.6410 1.9406 73.1156
30 -0.4348 -0.7578 -0.4865 0.1987 14.6915
31 -0.1549 0.9553 0.2520 2.1067 96.7004
32 -0.6548 0.7557 0.0152 0.0479 6.7436
33 -0.1211 0.9565 0.2654 0.9048 48.0866
34 0.9190 0.3942 0.0120 2.1301 93.3973
35 -0.6544 -0.1229 0.7461 1.4387 62.9082
36 0.3736 0.4033 0.8353 2.9938 120.4141
37 -0.7884 0.2556 0.5595 1.7552 74.1041
38 -0.0922 -0.9800 0.1763 2.0951 98.9804
39 0.3615 -0.9319 0.0295 2.6110 113.4042
40 0.7919 -0.0270 -0.6101 2.4079 102.7230
41 -0.6932 -0.5089 -0.5105 0.7458 32.5431
42 -0.3292 -0.3637 0.8714 2.2851 96.6590
43 -0.6957 0.7180 0.0213 2.2365 96.3992
44 0.5300 0.7566 0.3829 3.1200 121.0292
45 -0.0159 -0.0443 0.9989 1.7722 88.9949
46 0.3593 -0.9173 -0.1717 1.8523 83.5005
47 0.7167 -0.2566 -0.6484 2.5235 102.5456
48 0.8941 0.3229 -0.3102 0.9483 46.6506
49 0.5806 0.5832 0.5682 0.2152 15.3435
50 0.0277 -0.8147 0.5793 0.7326 35.5979
51 -0.7134 -0.1758 0.6783 0.8827 40.1768
Continued on next page

130
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
52 -0.4107 -0.5585 0.7207 2.9219 111.9494
53 0.3425 0.9374 0.0631 0.5569 31.4791
54 -0.8958 0.0194 0.4441 2.6299 112.3392
55 0.4803 -0.5898 0.6492 2.6614 98.9055
56 -0.9144 -0.3849 0.1254 3.1273 132.3074
57 -0.4655 -0.5977 -0.6527 1.4386 57.0277
58 0.9265 0.3263 0.1875 2.7395 117.0523
59 0.6064 -0.3884 -0.6939 2.5118 97.6669
60 0.4393 0.8126 -0.3831 2.1609 89.2272
61 -0.7688 0.3500 0.5351 2.5503 101.7839
62 -0.6675 0.6306 -0.3960 2.3638 92.3673
63 -0.5756 -0.7900 0.2113 0.6885 33.2132
64 -0.9038 -0.3459 -0.2520 0.4903 27.0176
65 -0.2834 -0.6351 -0.7186 0.1540 12.6189
66 -0.2640 -0.8322 0.4877 0.0079 2.8657
67 0.4076 -0.0117 0.9131 0.1820 12.0439
68 -0.0499 -0.2111 0.9762 1.0117 53.4115
69 0.4296 0.6218 0.6548 0.2859 17.5679
70 0.8171 -0.5282 0.2311 1.8606 79.1884
71 0.8587 0.2397 -0.4530 0.7561 37.2574
72 0.9559 -0.0513 -0.2893 0.6702 37.0181
73 -0.3604 0.2892 0.8868 0.0163 3.9041
74 -0.4120 -0.4872 0.7700 2.3607 93.8414
75 0.9084 0.4113 -0.0755 1.8816 83.0699
76 -0.7092 0.6003 0.3698 1.9349 78.1048
77 -0.0970 0.7192 -0.6880 1.1003 49.9492
78 -0.3998 0.8648 -0.3039 2.1513 91.7879
79 0.0799 -0.9557 0.2833 2.0368 93.5405
80 0.5720 0.7484 -0.3358 1.1219 48.8914
81 -0.6189 -0.5737 0.5365 1.4229 55.3501
Continued on next page

131
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
82 0.8625 0.1422 -0.4856 2.0795 89.1389
83 0.0036 -0.9876 0.1572 1.7058 84.8581
84 -0.3133 -0.8334 0.4552 1.4579 64.2187
85 0.9731 0.2302 -0.0034 2.3879 107.7040
86 0.5268 -0.6627 -0.5323 2.4115 89.9254
87 -0.8473 -0.3694 0.3815 2.0237 85.7438
88 0.1734 -0.4704 -0.8652 1.8366 79.4191
89 0.7844 -0.3856 0.4859 0.2687 16.6017
90 0.9643 -0.0924 0.2484 2.3762 106.6353
91 0.6242 0.3298 -0.7083 2.7487 108.4174
92 -0.5230 0.8250 -0.2142 2.0695 88.1255
93 -0.2575 -0.7448 -0.6156 1.2889 55.9517
94 0.5850 -0.7158 0.3813 0.0898 9.8856
95 0.4754 -0.1566 0.8657 0.0541 6.9618
96 -0.9125 -0.1528 -0.3794 2.5676 109.8956
97 -0.4214 0.6667 -0.6148 1.7908 70.8645
98 0.7260 -0.6256 -0.2855 2.3089 94.3566
99 0.4009 0.8137 0.4209 0.3726 20.3155
100 -0.7285 -0.3233 -0.6040 1.7224 70.9472
101 -0.1397 0.9206 -0.3647 1.6461 75.9606
102 -0.5098 -0.0445 0.8591 1.3971 62.2975
103 0.5449 0.5490 0.6338 2.2552 82.8640
104 0.5235 0.4654 -0.7137 2.0177 78.6454
105 -0.0659 0.6823 -0.7281 0.8125 38.3064
106 -0.0633 -0.8257 -0.5605 3.0811 129.9339
107 -0.1445 0.8650 -0.4806 2.0511 88.6411
108 -0.4287 0.5967 0.6783 1.3283 54.0628
109 0.6805 -0.7121 -0.1726 0.4376 23.1772
110 0.4095 -0.6258 -0.6638 1.5939 63.9836
111 -0.1279 0.9614 0.2436 1.5539 75.6464
Continued on next page

132
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
112 0.3214 -0.9466 0.0256 3.0487 131.3356
113 -0.4184 0.8433 -0.3373 1.8161 78.2769
114 -0.9338 -0.2314 -0.2731 1.1233 55.8146
115 0.6347 -0.7437 -0.2100 1.5158 65.5296
116 -0.6777 0.6230 -0.3908 1.2550 52.2343
117 -0.2830 -0.3052 0.9093 1.6779 76.2182
118 -0.8008 0.4107 0.4359 0.0309 5.6504
119 -0.7734 -0.6308 0.0623 2.4794 105.7400
120 -0.4372 -0.7260 -0.5308 2.2298 87.4622
121 0.4398 -0.6042 0.6645 1.8419 72.1772
122 0.5027 0.4617 -0.7308 1.2710 52.3657
123 0.7712 0.4717 -0.4274 1.3034 55.3120
124 0.9152 -0.3596 -0.1816 2.2217 96.4844
125 -0.0167 -0.2042 -0.9788 2.1398 99.7532
126 0.8446 -0.0252 0.5349 1.2604 56.6161
127 -0.6383 -0.7691 -0.0331 1.7569 77.0035
128 0.4271 -0.3982 -0.8118 1.3770 59.6409
129 0.2762 0.6252 -0.7300 1.6436 69.2628
130 -0.4058 -0.7487 -0.5242 1.0413 45.2674
131 -0.8719 -0.4838 0.0754 1.5857 70.2619
132 0.3520 -0.5264 -0.7739 0.8335 37.5633
133 0.6548 0.3119 -0.6884 2.4570 98.2453
134 0.6550 0.5879 0.4747 0.2379 16.0910
135 0.6498 -0.6264 0.4306 2.0222 78.8245
136 -0.6154 -0.6007 0.5104 2.8486 103.5240
137 0.7996 -0.2656 0.5386 1.7302 73.1580
138 0.7952 0.5657 -0.2182 0.5195 25.9696
139 0.3430 0.3194 0.8834 0.5070 27.1615
140 0.5487 0.7124 -0.4376 0.0827 9.5313
141 0.4836 0.4240 -0.7658 0.5425 25.8022
Continued on next page

133
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
142 -0.0673 -0.8937 0.4436 2.0467 89.4600
143 -0.9767 0.1880 0.1037 0.0308 4.9245
144 -0.6699 -0.1587 0.7253 0.0671 7.9076
145 -0.5872 0.3475 0.7310 1.9410 78.6907
146 -0.3565 0.6780 -0.6429 0.3487 19.5182
147 0.7815 0.5041 0.3677 2.0278 82.8166
148 0.3983 -0.9126 0.0927 2.3480 102.1257
149 -0.7969 -0.5295 0.2908 0.7808 36.3170
150 -0.6111 0.5347 -0.5837 1.8231 68.3126
151 0.6462 -0.6856 -0.3352 0.3386 19.2497
152 0.7432 0.1158 0.6589 1.3947 61.1877
153 -0.6524 -0.7466 0.1300 0.8410 39.5431
154 0.4912 -0.8705 0.0314 1.2410 57.3527
155 -0.6460 0.6206 0.4443 1.5033 59.7719
156 -0.4780 -0.8662 0.1457 0.9074 44.0205
157 -0.3495 -0.0365 0.9362 2.6746 115.3636
158 0.6527 0.0882 0.7525 1.2710 56.3158
159 0.6017 -0.6921 0.3986 1.1715 49.3155
160 -0.6978 -0.6802 -0.2243 2.7690 113.1750
161 0.8242 0.4734 -0.3109 1.3576 59.4262
162 -0.8919 -0.2943 -0.3434 2.7310 114.4796
163 -0.3962 -0.2302 0.8888 2.3051 98.0166
164 -0.8182 0.0246 0.5743 2.2969 98.2537
165 0.5007 0.1255 0.8565 0.3193 18.0664
166 -0.0754 -0.0639 -0.9951 1.9564 95.2973
167 0.2259 0.7554 -0.6150 2.3965 99.3658
168 -0.8627 -0.4197 0.2822 1.8482 79.7669
169 0.8248 -0.3933 0.4063 2.8914 116.1525
170 -0.2033 0.4825 -0.8520 1.0281 47.6554
171 0.7920 0.3885 -0.4709 1.2591 54.4961
Continued on next page

134
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
172 -0.6806 0.5216 0.5145 1.4055 55.3970
173 0.3298 -0.3471 -0.8779 0.1222 10.7685
174 -0.7882 0.3920 0.4744 1.0579 46.7556
175 -0.2289 -0.4153 0.8804 2.5882 108.7767
176 -0.4545 0.5309 -0.7152 1.8715 73.7423
177 0.6393 -0.2593 0.7239 0.0751 8.6054
178 -0.6424 0.4182 0.6423 2.4581 94.3174
179 -0.6811 -0.3548 -0.6405 2.6549 103.9268
180 0.6498 -0.6847 -0.3300 0.8612 38.4553
181 0.7219 -0.5941 0.3549 1.0911 46.8568
182 0.7340 -0.0788 -0.6746 2.9806 124.6285
183 -0.2061 -0.5568 -0.8046 2.8900 119.2259
184 -0.2884 -0.6112 -0.7371 0.6096 28.9599
185 -0.8589 -0.4455 0.2526 1.9593 83.6064
186 -0.2418 -0.9510 -0.1929 0.6629 36.9803
187 0.4215 -0.6897 0.5888 2.2647 88.1292
188 -0.0428 -0.0708 0.9966 0.4604 28.0862
189 -0.0114 0.5016 -0.8650 0.6539 32.9341
190 0.4143 -0.6518 -0.6353 2.8732 109.2286
191 0.7650 0.0474 0.6423 1.5692 68.6832
192 -0.7926 -0.0254 -0.6092 1.1151 50.2273
193 -0.4479 -0.4923 0.7463 1.3373 55.3709
194 0.1679 0.9852 0.0359 2.5571 116.8595
195 0.5043 -0.7421 0.4416 1.7928 72.5008
196 -0.3084 -0.6339 0.7093 2.9391 116.3417
197 0.8569 -0.3486 -0.3799 2.0833 88.4566
198 0.3796 -0.7023 -0.6022 1.7408 70.5744
199 -0.9773 -0.1691 -0.1274 0.8438 45.9937
200 0.7990 0.5935 0.0961 0.2578 15.9284
201 0.3600 -0.0013 -0.9329 1.3309 63.8315
Continued on next page

135
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
202 -0.4213 0.3473 -0.8378 1.2260 55.0332
203 -0.3461 -0.5484 -0.7612 1.0513 45.6728
204 0.3192 0.6741 0.6661 0.9241 40.7992
205 0.5655 0.3988 0.7219 1.9086 76.2377
206 -0.8244 0.4410 -0.3548 2.9449 118.4392
207 -0.6086 -0.5069 0.6105 2.8991 105.1095
208 -0.2901 0.5771 0.7634 2.3248 95.2194
209 -0.2360 -0.2722 0.9329 2.2892 101.1594
210 -0.6842 0.5201 0.5112 0.7681 33.3545
211 0.5127 0.6658 -0.5420 1.0397 42.8502
212 -0.3698 -0.1428 -0.9181 2.9522 125.1680
213 -0.5239 0.4016 0.7512 0.8149 36.1441
214 -0.2026 0.8120 -0.5474 2.5392 106.2665
215 0.8262 -0.0790 0.5578 2.2228 95.0289
216 -0.0483 0.8018 -0.5956 2.5802 110.1526
217 0.8215 -0.1807 -0.5408 2.7211 113.3682
218 -0.7145 0.6224 0.3196 1.3356 56.7144
219 0.2621 -0.6940 0.6705 3.0339 121.4384
220 -0.7090 -0.5947 -0.3791 2.5577 100.1420
221 -0.4979 0.5680 -0.6553 2.5755 95.4535
222 0.8187 -0.5219 0.2395 1.3886 60.5773
223 0.8113 0.5826 0.0498 2.0883 89.8867
224 0.5185 -0.5920 -0.6170 2.5351 92.6308
225 -0.1380 -0.6448 -0.7518 0.3624 20.0052
226 -0.6449 -0.2827 0.7100 2.5913 104.2599
227 -0.7670 0.2985 -0.5679 2.9781 118.5607
228 0.4269 0.8450 0.3221 2.4288 101.0851
229 0.9931 0.0480 0.1070 2.8839 130.2935
230 0.4744 -0.1169 -0.8725 0.5446 28.4960
231 -0.1082 -0.9244 -0.3657 1.1656 57.3082
Continued on next page

136
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
232 0.6161 -0.5934 0.5179 0.4738 23.9351
233 0.3209 0.9442 0.0746 2.5077 109.8867
234 -0.1344 0.6480 0.7497 1.3549 59.6170
235 -0.0585 -0.7567 0.6511 0.3667 20.5626
236 -0.5829 -0.7873 0.2009 1.7616 75.7300
237 -0.1994 0.5793 0.7904 1.2126 53.4102
238 0.3972 0.5075 0.7646 2.4233 96.2314
239 0.8038 -0.2852 -0.5221 2.8599 115.6364
240 0.4197 -0.8191 0.3912 2.0620 86.1689
241 0.7096 -0.3073 -0.6341 3.1075 122.0555
242 -0.1631 -0.8760 -0.4539 1.2019 56.1692
243 0.6226 0.5384 0.5679 1.1017 44.4825
244 -0.4576 0.5659 -0.6858 2.0575 79.5192
245 -0.5653 0.6202 0.5439 2.1401 79.3096
246 -0.3865 0.1673 0.9070 2.1775 94.1351
247 -0.6978 -0.6933 0.1801 2.4772 103.3773
248 0.1155 0.8703 0.4787 1.6598 73.4760
249 0.4934 0.6000 -0.6298 1.0045 41.3965
250 -0.6694 0.4428 0.5965 0.7041 30.9433
251 -0.5892 0.6039 -0.5368 2.1693 80.1939
252 -0.4221 0.5329 0.7333 1.3166 54.3860
253 0.5322 -0.7567 -0.3798 1.8993 77.6134
254 0.7383 0.5449 -0.3974 2.6853 104.7808
255 0.2658 0.5627 0.7827 0.1759 13.3191
256 0.6438 -0.4586 -0.6125 2.8078 104.6939
257 -0.4865 0.8167 -0.3104 1.7834 75.7543
258 -0.1444 0.9721 -0.1847 0.7233 40.6073
259 0.6197 -0.7593 -0.1984 1.6292 70.2822
260 0.7190 -0.0572 -0.6927 0.6222 30.2583
261 -0.7968 0.3662 -0.4805 1.9215 79.1961
Continued on next page

137
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
262 0.4219 0.9019 -0.0932 2.1207 92.7021
263 0.0470 0.2080 -0.9770 1.8773 89.2491
264 0.5779 -0.7367 -0.3512 0.9946 43.6996
265 -0.0348 -0.6822 0.7303 1.5254 67.2032
266 -0.3898 0.8575 0.3357 0.9710 46.5226
267 0.5815 0.4805 0.6565 3.0499 112.0223
268 0.9321 0.1709 -0.3192 0.6407 34.7803
269 -0.1479 -0.7571 0.6363 0.4745 24.7355
270 -0.2723 0.6317 0.7258 0.3963 20.8801
271 0.1698 0.5505 0.8174 0.5924 29.2569
272 -0.3933 0.3048 0.8674 0.1963 13.6574
273 -0.1208 -0.6092 -0.7837 1.5400 67.1866
274 -0.8202 -0.0192 -0.5717 0.1323 10.4784
275 0.1340 -0.9294 0.3438 2.6961 116.3694
276 -0.0547 0.4295 0.9014 2.6976 115.1974
277 0.8454 0.4267 0.3213 2.9307 119.3375
278 -0.5963 0.3242 0.7344 0.7521 33.8596
279 0.7258 -0.2697 0.6328 1.3124 56.0900
280 0.2035 0.6990 0.6855 2.6305 108.4877
281 -0.8263 0.3829 0.4131 1.6128 69.3864
282 -0.2449 0.7734 -0.5847 2.6164 107.4934
283 -0.6312 0.7482 -0.2044 0.2914 17.4238
284 0.5594 0.8203 -0.1192 0.2599 16.2197
285 0.4596 0.8576 0.2308 1.6678 72.9885
286 -0.8054 -0.3332 -0.4902 1.3637 58.8237
287 0.8543 0.0729 0.5147 3.1095 130.9123
288 -0.7265 -0.0667 0.6839 3.0350 127.0233
289 0.9765 -0.1963 -0.0894 0.1616 11.6864
290 -0.6090 0.7575 -0.2350 2.3456 97.3626
291 -0.6971 0.3350 -0.6340 1.6571 67.8851
Continued on next page

138
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
292 0.3509 0.3951 0.8490 0.1162 10.6659
293 -0.9272 0.2699 0.2595 2.4481 106.8677
294 0.8693 0.2804 -0.4069 0.8228 40.4574
295 -0.5466 0.0442 -0.8362 0.4069 21.7447
296 0.6967 0.6815 0.2239 2.0000 83.8313
297 -0.3587 -0.9312 0.0645 0.6372 34.9868
298 -0.5886 -0.5627 -0.5804 0.1468 12.6679
299 -0.7191 -0.4993 -0.4833 0.2929 17.6925
300 0.3385 0.9081 -0.2465 1.0683 52.7094
301 -0.5200 -0.6153 -0.5925 2.1699 80.3768
302 0.4204 0.5551 -0.7177 2.2525 88.1652
303 -0.3193 0.2649 0.9099 1.0081 49.6737
304 -0.1416 0.9826 -0.1206 2.8280 127.0512
305 -0.8010 -0.1548 0.5783 1.1689 51.8777
306 0.5880 -0.5726 -0.5712 2.6664 96.4102
307 0.8126 0.4741 0.3390 2.1262 87.7401
308 -0.4341 -0.7118 -0.5521 2.8242 107.7893
309 -0.9074 -0.1290 -0.4000 2.2437 97.0697
310 -0.5776 -0.8075 -0.1199 0.0450 6.7107
311 0.1644 0.6587 -0.7342 3.0611 125.9021
312 0.0843 0.8536 -0.5140 1.5712 69.5848
313 0.3071 -0.9202 0.2426 0.8856 45.5908
314 0.6541 -0.3091 0.6904 1.9777 80.4332
315 -0.1380 0.9865 0.0886 0.4224 26.3884
316 0.2816 -0.0036 0.9595 0.5326 30.4882
317 -0.9668 -0.2549 0.0173 0.7335 40.4067
318 -0.1410 0.4637 0.8747 1.2353 56.5574
319 0.9423 0.3277 -0.0684 1.1348 56.4824
320 0.5406 -0.6349 -0.5520 2.7598 100.1047
321 0.3260 0.8714 -0.3665 1.9885 86.3280
Continued on next page

139
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
322 0.5911 0.6818 0.4310 1.5267 61.4416
323 -0.7430 0.4776 0.4689 1.2267 51.3711
324 0.3186 0.5480 0.7734 2.1941 89.7797
325 -0.6605 0.0614 0.7483 0.7690 36.1813
326 0.3276 0.2384 -0.9142 1.6901 77.0111
327 -0.8426 0.3704 0.3909 1.8288 78.2656
328 0.7989 -0.4010 0.4483 3.1398 123.3288
329 0.5368 0.7089 -0.4575 2.9087 110.1251
330 0.7423 -0.4256 0.5175 1.6145 65.5544
331 -0.6072 -0.2307 -0.7603 0.4278 21.8651
332 0.5148 -0.5461 -0.6609 1.9718 74.5908
333 0.4674 -0.6458 -0.6037 0.6700 30.3981
334 -0.4625 0.7499 0.4730 2.4504 96.4543
335 0.1826 0.9582 -0.2201 2.5306 113.2183
336 -0.6720 0.1118 0.7320 1.9824 84.6699
337 -0.0823 -0.9685 0.2349 0.4489 27.2397
338 -0.0855 0.8384 -0.5383 1.8028 78.6974
339 -0.1864 0.8019 -0.5676 1.4957 65.4493
340 0.7380 0.0718 0.6710 3.0181 126.2793
341 0.4738 0.8248 -0.3086 0.3800 20.9217
342 0.4148 0.7977 0.4378 1.6921 71.5067
343 -0.1035 -0.6887 -0.7177 1.9683 84.5390
344 0.3305 -0.5834 -0.7419 1.0536 45.6153
345 -0.8179 0.1301 -0.5605 2.2022 93.7164
346 -0.4430 0.4828 0.7554 1.4661 60.4734
347 0.5214 0.6609 0.5397 1.4858 58.0635
348 -0.4170 -0.7616 0.4960 3.0070 116.7194
349 0.0842 0.8763 0.4744 2.9195 124.0089
350 0.7487 -0.6240 0.2239 0.9867 44.3659
351 -0.7970 0.4481 -0.4049 0.8742 39.9825
Continued on next page

140
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
352 0.1406 0.3106 0.9401 2.8500 122.2074
353 0.5891 -0.5634 0.5793 2.8720 103.1727
354 -0.5613 -0.7816 0.2722 0.3026 17.9096
355 -0.0017 0.5976 -0.8018 0.7648 36.1410
356 0.9565 0.0940 -0.2763 1.6073 76.8111
357 0.5289 -0.6783 -0.5100 0.6500 29.8896
358 -0.5697 -0.7851 -0.2430 0.1012 10.2962
359 -0.1665 -0.9835 -0.0713 0.0117 3.1660
360 0.5438 0.3361 0.7689 1.0612 46.1803
361 -0.0200 0.7160 0.6979 0.7793 37.1952
362 -0.8300 -0.1841 -0.5266 0.5763 28.9494
363 -0.4110 -0.5973 -0.6887 1.1920 49.4398
364 -0.7404 0.2450 0.6259 2.6470 107.7090
365 -0.5830 -0.3320 -0.7415 0.7938 35.4174
366 0.7555 -0.5735 -0.3167 3.0121 119.1876
367 0.1906 0.7906 0.5819 0.6803 32.8837
368 0.6245 -0.2545 -0.7384 2.7981 113.0233
369 0.7391 -0.1669 0.6526 1.2876 56.4041
370 0.9644 -0.2023 -0.1704 1.8212 86.0065
371 -0.7947 -0.5168 0.3184 1.2455 54.0199
372 -0.6144 -0.7889 -0.0119 1.0718 49.3265
373 0.1841 -0.9775 -0.1028 1.1606 60.6229
374 0.2002 -0.7327 0.6504 0.8703 40.0829
375 0.6358 0.0083 -0.7718 0.2048 13.3986
376 -0.7999 0.4553 -0.3909 0.9015 41.1363
377 0.3579 0.5080 0.7835 2.7299 108.8463
378 0.8760 -0.4568 -0.1551 0.3568 20.2620
379 -0.6992 0.6492 -0.2994 2.5141 101.3692
380 -0.3094 -0.8448 0.4367 2.1024 88.7675
381 -0.4813 -0.3546 -0.8017 2.9706 117.9924
Continued on next page

141
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
382 -0.7311 -0.5129 0.4499 1.7070 68.3442
383 0.7147 -0.6914 0.1054 2.2650 96.5689
384 -0.6722 0.4700 0.5721 1.2185 49.1196
385 -0.4823 -0.5629 0.6712 1.5194 59.7510
386 0.6408 0.7272 0.2460 1.3029 56.6352
387 -0.3743 0.8630 0.3391 0.6941 35.3739
388 0.2239 0.4803 0.8480 2.9059 120.2507
389 -0.8764 0.4331 -0.2105 1.0254 48.8056
390 0.2400 0.6841 0.6887 0.4793 24.3743
391 0.4405 0.7883 -0.4297 0.1011 10.4487
392 -0.7614 0.3723 -0.5308 2.4647 97.9495
393 0.6366 -0.3874 -0.6668 0.3318 18.6699
394 -0.1804 0.6058 0.7749 0.4534 23.3297
395 -0.3887 0.6654 0.6373 0.2419 16.1222
396 0.4360 0.6900 -0.5777 2.7917 105.9288
397 0.2054 0.9250 -0.3198 0.4263 25.0478
398 0.6866 0.5625 0.4606 0.6511 29.6090
399 0.5886 0.4867 -0.6455 0.2402 16.0363
400 0.5933 -0.6469 0.4791 3.0615 112.6652
401 0.3066 -0.3097 -0.9001 1.5984 72.6162
402 -0.2523 0.4996 -0.8287 2.7120 111.8933
403 -0.5400 0.6284 -0.5599 0.7185 32.0745
404 0.5308 -0.8367 -0.1351 1.4741 65.2720
405 -0.8702 -0.0874 0.4848 1.8933 82.0759
406 0.5557 0.5691 0.6061 0.9424 39.2280
407 -0.8030 0.0917 -0.5889 1.7641 76.2107
408 -0.5919 -0.7977 -0.1149 3.1382 130.8816
409 0.3204 -0.4176 0.8503 1.9236 81.8379
410 0.7018 -0.3322 0.6301 0.2774 16.7892
411 -0.4405 -0.3282 -0.8356 0.9609 44.6187
Continued on next page

142
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
412 0.5983 -0.6731 0.4347 0.1682 13.5714
413 -0.8360 -0.4454 0.3206 2.3672 97.8113
414 0.9571 0.2514 0.1440 1.7094 80.8254
415 -0.6432 -0.5771 0.5032 2.7861 102.1828
416 -0.6527 -0.6969 -0.2972 2.4297 98.2710
417 0.6396 0.7098 -0.2950 2.1835 89.1791
418 -0.8323 -0.3940 0.3899 0.9832 45.5487
419 0.8574 -0.4696 0.2106 0.0237 4.8182
420 0.2405 0.8229 0.5148 0.6114 30.6809
421 0.6224 0.5077 -0.5957 1.8966 71.1564
422 0.5730 -0.6943 0.4353 3.0983 116.6390
423 -0.5787 0.0562 -0.8136 0.0576 7.0135
424 0.6719 0.2919 -0.6807 0.6940 31.8353
425 0.4512 -0.3846 -0.8053 2.8334 112.9630
426 -0.3433 0.7167 0.6070 1.1498 49.3957
427 0.8208 -0.5576 -0.1242 1.5632 68.3503
428 -0.3875 -0.1824 -0.9036 1.1902 56.6474
429 0.7791 0.5706 0.2596 2.4698 101.7948
430 -0.3497 0.6324 -0.6912 1.0872 46.5827
431 0.5978 -0.1474 0.7880 2.1177 89.8641
432 -0.7268 -0.2425 0.6426 0.8975 40.2780
433 0.5959 -0.5605 -0.5750 1.5757 60.1681
434 0.3469 0.0402 0.9370 3.0275 129.3556
435 0.2195 -0.1230 0.9678 1.4536 71.5702
436 -0.2530 0.0832 -0.9639 1.4163 69.6939
437 0.0918 -0.8168 0.5695 0.4519 24.0139
438 -0.0869 0.8229 0.5615 2.7770 117.6537
439 -0.0349 0.8024 0.5957 2.7171 115.7542
440 -0.0770 -0.7520 -0.6546 3.0766 128.8488
441 -0.0169 0.6885 -0.7251 0.5934 29.5817
Continued on next page

143
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
442 -0.7612 0.5903 -0.2686 1.8008 75.8068
443 -0.6793 -0.7180 -0.1520 1.4217 62.3973
444 0.0555 -0.0074 0.9984 1.1068 60.0214
445 -0.0789 0.1990 0.9768 0.5749 33.0936
446 0.8801 -0.0295 0.4739 1.5387 68.7483
447 0.0732 -0.0994 -0.9923 0.0937 8.0210
448 -0.4919 -0.4071 -0.7696 1.6478 67.7155
449 0.0360 -0.6863 0.7265 1.4831 65.4921
450 0.8721 0.4569 0.1750 0.7992 39.3048
451 -0.9337 0.3353 -0.1259 0.3969 23.2480
452 -0.5059 -0.4141 -0.7567 3.0677 118.1862
453 -0.4677 -0.8646 0.1835 2.3525 100.4609
454 -0.7252 -0.1256 -0.6769 1.3507 59.2927
455 0.4069 -0.8755 -0.2607 1.7411 77.0860
456 -0.4154 -0.3292 0.8480 2.8103 114.7175
457 0.2920 0.5910 0.7520 2.6808 108.1053
458 0.8492 0.5162 0.1117 1.4744 65.1589
459 -0.5360 -0.6128 -0.5807 2.1884 80.8584
460 0.5917 -0.6349 -0.4968 0.8637 36.9615
461 -0.0375 0.9909 0.1294 0.9504 52.4169
462 0.6306 -0.6514 0.4219 2.5963 99.4997
463 0.2818 -0.8608 0.4237 2.2870 96.6534
464 0.2359 -0.1069 0.9659 0.1777 12.3261
465 -0.5491 -0.5650 0.6158 0.1261 11.7323
466 0.8354 0.3332 0.4372 2.7611 112.4692
467 -0.8595 0.5106 -0.0240 2.2483 96.6996
468 0.0782 0.2636 -0.9615 2.9874 129.1738
469 0.7620 0.5007 -0.4107 1.4137 58.9080
470 0.2831 -0.7003 0.6554 1.4497 61.4418
471 -0.3254 -0.2197 0.9197 2.7082 115.2481
Continued on next page

144
 Table D.1 – Continued from previous page
Data Point 𝑒1 𝑒2 𝑒3 𝜓 (radians) 𝑡 𝑓 (seconds)
472 0.9078 0.3943 0.1429 3.0839 130.1808
473 -0.3324 0.9305 0.1540 2.8183 121.1262
474 -0.5104 0.8590 0.0405 0.8434 41.2135
475 0.1600 0.6724 -0.7227 0.2651 16.2527
476 -0.3288 -0.9055 0.2683 1.6268 74.7352
477 -0.1203 0.0893 -0.9887 2.3321 108.9069
478 0.9437 0.3293 -0.0321 0.2080 13.8725
479 0.8518 -0.0515 -0.5213 0.5179 27.0215
480 -0.4050 0.3493 0.8450 1.1443 52.2048
481 0.3201 -0.7452 0.5849 2.3181 93.9448
482 -0.5794 0.6039 -0.5474 2.6418 95.7617
483 0.8907 -0.3825 0.2459 2.8563 119.4223
484 0.7941 0.0125 -0.6076 3.0954 130.5968
485 -0.9320 0.2301 -0.2800 1.7188 79.4973
486 0.1958 0.6300 0.7515 0.8225 37.9692
487 0.8113 -0.1491 0.5654 1.1407 50.8706
488 -0.5456 -0.6927 -0.4716 3.0026 112.2904
489 0.1943 0.7599 -0.6203 0.4823 24.8988
490 0.0190 0.9867 0.1616 0.9643 52.8040
491 0.7291 -0.1966 0.6556 0.3498 19.0036
492 0.2530 -0.6050 0.7550 2.2066 91.5131
493 0.8629 -0.5049 -0.0234 2.0030 86.7206
494 0.5714 -0.6312 -0.5245 0.4069 21.7857
495 -0.3269 -0.6891 -0.6467 0.5112 25.1393
496 -0.3133 0.6211 0.7183 1.1940 51.2675
497 0.7138 -0.1978 0.6718 0.2355 14.9402
498 -0.6484 -0.5777 -0.4958 0.5787 27.3095
499 -0.2192 -0.6217 0.7519 2.3217 96.6485
500 -0.1720 0.6955 0.6976 0.4993 25.4664

145
146
 List of References

[1] K. D. Bilimoria and B. Wie, “Time-optimal three-axis reorientation of a rigid space-

craft,” Journal of Guidance, Control, and Dynamics, vol. 16, no. 3, pp. 446–452,
1993. Available: http://arc.aiaa.org/doi/full/10.2514/3.21030
[2] M. Karpenko, S. Bhatt, N. Bedrossian, and I. M. Ross, “Flight implementation of
shortest-time maneuvers for imaging satellites,” Journal of Guidance, Control, and
Dynamics, vol. 37, no. 4, pp. 1069–1079, 2014. Available: https://doi.org/10.2514/1.
62867
[3] K. Bollino, L. R. Lewis, P. Sekhavat, and I. M. Ross, “Pseudospectral optimal
control: A clear road for autonomous intelligent path planning,” in AIAA In-
fotech@Aerospace 2007 Conference and Exhibit. American Institute of Aeronautics
and Astronautics, 2007. Available: https://arc.aiaa.org/doi/abs/10.2514/6.2007-2831
[4] D. E. Smith, “Choosing objectives in over-subscription planning,” in Proceedings of
the Fourteenth International Conference on International Conference on Automated
Planning and Scheduling (ICAPS’04). Whistler, British Columbia: AAAI Press,
2004, pp. 393–401.
[5] X. Yang, “Genetic algorithms,” in Nature-Inspired Optimization Algorithms, 1st ed.
London: Elsevier, 2014, pp. 77–87.
[6] B. Wie, Space Vehicle Dynamics and Control, 2nd ed. Reston, VA: American Insti-
tute of Aeronautics and Astronautics, 2008. Available: http://arc.aiaa.org/doi/book/
10.2514/4.860119
[7] J. R. Wertz, “Space mission geometry,” in Space Mission Engineering: the New
SMAD (Space Technology Library). Hawthorne, CA: Microcosm Press, 2011, pp.
149–196.
[8] I. M. Ross, R. J. Proulx, and M. Karpenko, “Autonomous UAV sensor planning,
scheduling and maneuvering: An obstacle engagement technique,” in 2019 Ameri-
can Control Conference (ACC). Philadelphia, PA, USA: IEEE, July 2019, pp. 65–70.
Available: https://ieeexplore.ieee.org/document/8814474/
[9] J. P. Abramson, College Algebra (OpenStax). Houston, TX: Rice University, 2017.
Available: https://openstax.org/details/books/college-algebra
[10] N. Gunantara, “A review of multi-objective optimization: Methods and its appli-
cations,” Cogent Engineering, vol. 5, no. 1, July 2018. Available: https://www.
cogentoa.com/article/10.1080/23311916.2018.1502242

147
[11] J. Kirk, “Traveling Salesman Problem—Genetic algorithm,” May 2014. Available:
https://www.mathworks.com/matlabcentral/fileexchange/13680

[12] I. M. Ross, “Enhancements to the DIDO Optimal Control Toolbox,”

arXiv:2004.13112 [cs, math], Apr. 2020, arXiv: 2004.13112. Available: http:
//arxiv.org/abs/2004.13112

[13] I. M. Ross, A Primer on Pontryagin’s Principle in Optimal Control, 2nd ed. San
Francisco: Collegiate Publishers, Mar. 2015.

[14] T. Lippman, J. M. Kaufman, and M. Karpenko, “Autonomous planning of con-

strained spacecraft reorientation maneuvers” (Advances in the Astronautical Sci-
ences), Stevenson, WA, 2017.

[15] I. M. Ross, Q. Gong, M. Karpenko, and R. J. Proulx, “Scaling and balancing for
high-performance computation of optimal controls,” Journal of Guidance, Control,
and Dynamics, vol. 41, no. 10, pp. 2086–2097, 2018. Available: https://doi.org/10.
2514/1.G003382

[16] D. C. Montgomery, Design and Analysis of Experiments, 8th ed. Hoboken, NJ: John
Wiley & Sons Inc, 2013.

[17] SAS Institute Inc. 2017, JMP® 13 Design of Experiments Guide, 2nd ed. Cary, NC:
SAS Institute Inc.

[18] S. Skansi, Introduction to Deep Learning: From Logical Calculus to Artificial Intel-
ligence (Undergraduate Topics in Computer Science). Cham: Springer International
Publishing, 2018. Available: http://link.springer.com/10.1007/978-3-319-73004-2

[19] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning

(Springer Series in Statistics). New York, NY: Springer New York, 2009. Available:
http://link.springer.com/10.1007/978-0-387-84858-7

[20] C. C. Aggarwal, Neural Networks and Deep Learning: A Textbook. Cham: Springer
International Publishing, 2018. Available: http://link.springer.com/10.1007/978-3-
319-94463-0

[21] G. James, D. Witten, T. Hastie, and R. Tibshirani, An Introduction to Statistical

Learning (Springer Texts in Statistics). New York, NY: Springer New York, 2013,
vol. 103. Available: http://link.springer.com/10.1007/978-1-4614-7138-7

[22] B. Illowsky and S. Dean, Introductory Statistics (OpenStax). Houston, TX: Rice
University, 2018. Available: https://openstax.org/details/books/introductory-
statistics

148
[23] F. E. Harrell, Jr., Regression Modeling Strategies (Springer Series in Statistics).
Cham: Springer International Publishing, 2015. Available: http://link.springer.com/
10.1007/978-3-319-19425-7

[24] A. Fleming, P. Sekhavat, and I. M. Ross, “Constrained, minimum-time maneuvers

for CMG actuated spacecraft,” Advances in the Astronautical Sciences, vol. 135, pp.
1009–1028, 2009.

149
THIS PAGE INTENTIONALLY LEFT BLANK

150
 Initial Distribution List

1. Defense Technical Information Center

Ft. Belvoir, Virginia

2. Dudley Knox Library

Naval Postgraduate School
Monterey, California

151