Copyright

by

Patrick Foil Beeson

2008
 The Dissertation Committee for Patrick Foil Beeson
certifies that this is the approved version of the following dissertation:

Creating and Utilizing Symbolic Representations

of Spatial Knowledge using Mobile Robots

Committee:

Benjamin J. Kuipers, Supervisor

Dana Ballard

Gregory Dudek

Raymond Mooney

Brian Stankiewicz

Peter Stone
Creating and Utilizing Symbolic Representations
of Spatial Knowledge using Mobile Robots

by

Patrick Foil Beeson, B.S.

Dissertation

Presented to the Faculty of the Graduate School of

The University of Texas at Austin

in Partial Fulfillment

of the Requirements

for the Degree of

Doctor of Philosophy

The University of Texas at Austin

August 2008
 To all UTCS graduates that have held the Golden Turtle.
“The early bird may get the worm, but it’s the second mouse that gets the cheese.” –Steven Wright
Acknowledgments

I wish to thank the multitudes of people who helped me navigate this portion of my life.
Having Ben Kuipers as an advisor could not have worked out better for me. His
boundless faith in the scientific process and constant optimism in the face of adversity was
exactly what a pessimistic, sarcastic student needed to persevere. Rather than forcing me
into a research area that is fashionable, he allowed me the time and space to explore the
topics I felt were important. As a result, I believe that this thesis is an important contribution
towards finding a robust solution to the robot mapping problem.
My thesis committee is comprised of some of the smartest and nicest people I have
ever met. Peter Stone, Brian Stankiewicz, and Ray Mooney have been a great source of
intellectual stimulation, and have each helped to shape my education. In my few discussions
with Greg Dudek, he has certainly lived up to his reputation as one of the most forward-
thinking robotics researchers in the world. Dana Ballard came to the rescue by agreeing to
attend my dissertation defense in order to make a quorum. I am truly lucky to have such a
great group of people on my committee.
My colleagues in the AI lab deserve loads of praise, as they were the ones that
had to put up with me on a daily basis. Joseph Modayil, Jeff Provost, Matt MacMahon,
and Aniket Murarka have logged countless hours discussing robot and human navigation,
reading paper drafts, and humoring me by laughing at my constant jokes. I cannot forget
to mention Francesco Savelli, (Ram) Subramanian Ramamoorthy, Harold Chaput, Misha
Bilenko, Tal Tversky, Nick Jong, and Michael Quinlan, all of whom have been cohorts and

v
comrades at UT-Austin.
Additional thanks go to my group of non-CS friends here in Austin. Without their
alcohol-fueled taunts and insults about still being a student in my thirties, I may have stayed
in school forever. I especially cannot forget to mention my beautiful and inspiring love,
Denise. Her support keeps me going when things seem at their worst. She is my best
friend, and I cannot imagine life without her.
Finally, I would like to thank my family, who all wonder what exactly I have been
doing for the last 9 years, for their patience. My mother did not live to see me finish my
undergraduate studies, but she was aware that I had decided to pursue a doctorate degree. I
know she would be so proud to see this goal finally accomplished.

PATRICK B EESON

The University of Texas at Austin

August 2008

vi
 Creating and Utilizing Symbolic Representations
of Spatial Knowledge using Mobile Robots

Publication No.

Patrick Foil Beeson, Ph.D.

The University of Texas at Austin, 2008

Supervisor: Benjamin J. Kuipers

A map is a description of an environment allowing an agent—a human, or in our

case a mobile robot—to plan and perform effective actions. From a single location, an
agent’s sensors can not observe the whole structure of a complex, large environment. For
this reason, the agent must build a map from observations gathered over time and space. We
distinguish between large-scale space, with spatial structure larger than the agent’s sensory
horizon, and small-scale space, with structure within the sensory horizon.
We propose a factored approach to mobile robot map-building that handles qualita-
tively different types of uncertainty by combining the strengths of topological and metrical

vii
approaches. Our framework is based on a computational model of the human cognitive
map; thus it allows robust navigation and communication within several different spatial
ontologies. Our approach factors the mapping problem into natural sub-goals: building
a metrical representation for local small-scale spaces; finding a topological map that rep-
resents the qualitative structure of large-scale space; and (when necessary) constructing a
metrical representation for large-scale space using the skeleton provided by the topological
map.
The core contributions of this thesis are a formal description of the Hybrid Spatial
Semantic Hierarchy (HSSH), a framework for both small-scale and large-scale representa-
tions of space, and an implementation of the HSSH that allows a robot to ground the large-
scale concepts of place and path in a metrical model of the local surround. Given metrical
models of the robot’s local surround, we argue that places at decision points in the world
can be grounded by the use of a primitive called a gateway. Gateways separate different re-
gions in space and have a natural description at intersections and in doorways. We provide
an algorithmic definition of gateways, a theory of how they contribute to the description of
paths and places, and practical uses of gateways in spatial mapping and learning.

viii
Contents

Acknowledgments v

Abstract vii

Contents ix

List of Algorithms xiv

List of Tables xv

List of Figures xvi

Chapter 1 Introduction 1
1.1 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Major Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Chapter 2 Background and Related Work 11

2.1 Cognitive Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Human Spatial Representations . . . . . . . . . . . . . . . . . . . 12
2.1.2 Structural versus Object Landmarks . . . . . . . . . . . . . . . . . 12
2.1.3 Relevance of Cognitive Maps to this Thesis . . . . . . . . . . . . . 13
2.2 Robot Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

ix
 2.2.1 Metrical Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 Relevance of Metrical SLAM to this Thesis . . . . . . . . . . . . . 20
2.2.3 Topological Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.4 Spatial Semantic Hierarchy Theses . . . . . . . . . . . . . . . . . 28
2.2.5 Relevance of Topological Maps to this Thesis . . . . . . . . . . . . 30
2.2.6 Hybrid Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.7 Relevance of Hybrid Mapping to this Thesis . . . . . . . . . . . . . 32
2.3 Gateways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 The Foundations of Gateways . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Gateway implementations . . . . . . . . . . . . . . . . . . . . . . 36

Chapter 3 The Spatial Semantic Hierarchy 44

3.1 Hybrid SSH Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 The Basic SSH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 The SSH Control Level . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2.2 The SSH Causal Level . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.3 The SSH Topological Level . . . . . . . . . . . . . . . . . . . . . 54
3.2.4 The SSH Metrical Level . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Extending the SSH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Chapter 4 Modeling Obstacles and Trajectory:

The HSSH Local Metrical Level 62
4.1 Local Perceptual Map (LPM) . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 LPM Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3 Building an Accurate Localization Algorithm . . . . . . . . . . . . . . . . 66
4.3.1 Action Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.2 Adaptive Particle Filter Localization . . . . . . . . . . . . . . . . . 72

x
Chapter 5 Symbol Grounding:
The HSSH Local Topology Level 91
5.1 Formal Description of Local Topology . . . . . . . . . . . . . . . . . . . . 91
5.2 Grounding Local Topology in the Local Perceptual Map . . . . . . . . . . 94
5.2.1 Gateways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2.2 Local Topology Creation . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Detecting Places . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.3.1 Place Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3.2 Places from Local Topology . . . . . . . . . . . . . . . . . . . . . 100
5.3.3 Gateways versus Voronoi Junctions . . . . . . . . . . . . . . . . . 102
5.4 Selecting Local Motion Targets . . . . . . . . . . . . . . . . . . . . . . . . 104

Chapter 6 Implementing Gateways 107

6.1 Algorithm Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.1.1 The Voronoi Skeleton of an LPM . . . . . . . . . . . . . . . . . . 110
6.1.2 Defining the LPM Core . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2 Constriction-based Gateways . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2.2 Problems with the Constriction Algorithm . . . . . . . . . . . . . . 125
6.3 Current Gateway Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.3.1 The Extended Voronoi Graph (EVG) . . . . . . . . . . . . . . . . . 130
6.3.2 Anchor-based Gateways . . . . . . . . . . . . . . . . . . . . . . . 133
6.3.3 Evaluating Anchor-based Gateways . . . . . . . . . . . . . . . . . 145

Chapter 7 Human-Robot Interaction:

Evaluating the HSSH Local Levels 168
7.1 HSSH Interface for a Mobility Assistant . . . . . . . . . . . . . . . . . . . 169
7.1.1 Local Metrical Control . . . . . . . . . . . . . . . . . . . . . . . . 169

xi
 7.1.2 Local Topological Control . . . . . . . . . . . . . . . . . . . . . . 172
7.1.3 Global Topological and Metrical Control . . . . . . . . . . . . . . 173
7.2 Empirical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.2.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Chapter 8 Closing Loops:

The HSSH Global Topology Level 182
8.1 From Small-Scale to Large-Scale Star . . . . . . . . . . . . . . . . . . . . 183
8.2 The Tree of Possible Topological Maps . . . . . . . . . . . . . . . . . . . 185
8.3 Topological Mapping Example . . . . . . . . . . . . . . . . . . . . . . . . 188
8.4 Levels of Spatial and Temporal Granularity . . . . . . . . . . . . . . . . . 191

Chapter 9 Drawing Maps:

The HSSH Global Metrical Level 193
9.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.2 The Theory of the Global Metrical Map . . . . . . . . . . . . . . . . . . . 195
9.3 Global Mapping Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
9.3.1 Estimating F(λ ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
9.3.2 Estimating G(χ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.3.3 Estimating H(x) . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.3.4 Creating a map m∗ . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Chapter 10 Conclusion 204

10.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
10.1.1 The Local Metrical and Topology Levels . . . . . . . . . . . . . . 205
10.1.2 The Global Topology and Metrical Levels . . . . . . . . . . . . . . 208
10.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

xii
Bibliography 213

Vita 230

xiii
List of Algorithms

3.1 Building the tree of topological maps in the SSH . . . . . . . . . . . . . . . 58

6.1 Finding the core of an LPM from a pruned free space skeleton . . . . . . . . 118
6.2 Finding constrictions of an LPM. . . . . . . . . . . . . . . . . . . . . . . . 120
6.3 Finding constriction-based gateways . . . . . . . . . . . . . . . . . . . . . . 121
6.4 Finding anchors of an LPM . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.5 Finding anchor-based gateways . . . . . . . . . . . . . . . . . . . . . . . . 139

8.1 Building the tree of topological maps in the HSSH . . . . . . . . . . . . . . 187

xiv
List of Tables

2.1 Comparison of gateway implementations . . . . . . . . . . . . . . . . . . . 43

4.1 Action model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

4.2 Localization accuracy of 5 different algorithms . . . . . . . . . . . . . . . 89

7.1 User interface to the HSSH . . . . . . . . . . . . . . . . . . . . . . . . . . 171

xv
List of Figures

2.1 Closing large loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2 Incorrectly closing loops using odometry and scan matching . . . . . . . . 18
2.3 Voronoi graph of a global metrical map . . . . . . . . . . . . . . . . . . . 26
2.4 Detecting places with an offset path . . . . . . . . . . . . . . . . . . . . . 29

3.1 The Spatial Semantic Hierarchy . . . . . . . . . . . . . . . . . . . . . . . 48

3.2 SSH Control/Causal Level abstraction . . . . . . . . . . . . . . . . . . . . 51
3.3 Topological abduction example . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4 HSSH description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.1 Markov localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.2 Generalized action model . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Proposal distribution example: uncertainty in motion . . . . . . . . . . . . 73
4.4 Likelihood weighting example: uncertainty in observations . . . . . . . . . 74
4.5 Localization failure example: wheel slippage . . . . . . . . . . . . . . . . 75
4.6 Poor localization due to an overestimating proposal . . . . . . . . . . . . . 78
4.7 Improved localization using the shrink algorithm . . . . . . . . . . . . . . 79
4.8 Poor localization due to an underestimating proposal . . . . . . . . . . . . 81
4.9 Improved localization using the grow algorithm . . . . . . . . . . . . . . . 82
4.10 Poor localization due to an offset proposal . . . . . . . . . . . . . . . . . . 83

xvi
4.11 Improved localization using the shrink-and-grow algorithm . . . . . . . . . 84

5.1 Identifying gateways and local topology in the LPM . . . . . . . . . . . . . 97

5.2 Gateway examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3 The reduced generalized Voronoi graph (RGVG). . . . . . . . . . . . . . . 103
5.4 Grounding control using gateways . . . . . . . . . . . . . . . . . . . . . . 105

6.1 Overview of gateway detection process . . . . . . . . . . . . . . . . . . . . 109

6.2 Thinned skeleton of a global metrical map . . . . . . . . . . . . . . . . . . 111
6.3 Finding the skeleton of LPM free space . . . . . . . . . . . . . . . . . . . 113
6.4 Pruning a Voronoi graph using the LPM boundary . . . . . . . . . . . . . . 117
6.5 Examples of the base points and the core of different local environments . . 119
6.6 Defining the search space for constriction-based gateways . . . . . . . . . . 122
6.7 Defining gateways by connecting nearby obstacles . . . . . . . . . . . . . 124
6.8 Examples of constriction-based gateways in a variety of environments . . . 126
6.9 Unstable gateways near L intersections . . . . . . . . . . . . . . . . . . . . 127
6.10 Unstable gateways affect control . . . . . . . . . . . . . . . . . . . . . . . 128
6.11 Unstable gateways in coastal navigation . . . . . . . . . . . . . . . . . . . 129
6.12 Entering a room that is larger than the LPM . . . . . . . . . . . . . . . . . 132
6.13 The extended Voronoi graph (EVG) of a global metrical map . . . . . . . . 133
6.14 The thinning approximation of the EVG . . . . . . . . . . . . . . . . . . . 134
6.15 Intuition behind skeletal anchors . . . . . . . . . . . . . . . . . . . . . . . 135
6.16 Searching for anchors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.17 Anchors and gateways in coastal navigation scenarios . . . . . . . . . . . . 137
6.18 Finding gateway location hypotheses from anchors . . . . . . . . . . . . . 140
6.19 Examples of anchor-based gateways in a variety of environments . . . . . . 141
6.20 Stable anchor-based gateways near L intersections . . . . . . . . . . . . . . 142
6.21 Stable anchor-based gateways at coastal L intersections . . . . . . . . . . . 142

xvii
6.22 Handling environments without good anchors . . . . . . . . . . . . . . . . 144
6.23 Gateway (and place) detection in a university hallway . . . . . . . . . . . . 146
6.24 Traversing paths in a university hallway . . . . . . . . . . . . . . . . . . . 147
6.25 Detecting doorways in a university hallway . . . . . . . . . . . . . . . . . 148
6.26 Dealing with furniture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.27 Gateway (and place) detection using a simulated robot . . . . . . . . . . . 151
6.28 Gateway (and place) detection using a simulated robot . . . . . . . . . . . 152
6.29 Gateway evaluation conditions . . . . . . . . . . . . . . . . . . . . . . . . 154
6.30 Impact of rotations on gateways and local topology . . . . . . . . . . . . . 157
6.31 Variance of local topology classification across rotations . . . . . . . . . . 158
6.32 Time performance of local topology abstraction . . . . . . . . . . . . . . . 159
6.33 Stability of gateway orientations . . . . . . . . . . . . . . . . . . . . . . . 160
6.34 Impact of LPM resolution on gateways and local topology . . . . . . . . . 163
6.35 Impact of noise on gateways and local topology . . . . . . . . . . . . . . . 165
6.36 Determining the bounds of successful local topology abstraction . . . . . . 167

7.1 HSSH interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

7.2 Stimuli, map, and interface for the HRI experiment . . . . . . . . . . . . . 176
7.3 Example experimental runs . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.4 HRI experiment results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

8.1 Binding local-paths to distinctive states . . . . . . . . . . . . . . . . . . . 184

8.2 Finding the topological map an environment with multiple nested loops . . 189

9.1 Defining local frames of reference . . . . . . . . . . . . . . . . . . . . . . 195

9.2 Finding the global place layout . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3 Creating a global metrical map . . . . . . . . . . . . . . . . . . . . . . . . 200

xviii
Chapter 1

Introduction

Congratulations!
Today is your day.
You’re off to Great Places!
You’re off and away!
Dr. Seuss, Oh, the Places You’ll Go! (1990)

Intelligent machines have already begun to change our society. They have replaced human
workers in dangerous and/or repetitive tasks such as manufacturing or mining operations.
They have allowed exploration and scientific discovery of previously unseen worlds like
Mars and the ocean floor. They have even started to become consumer products, such as
toys, vacuums, and lawn mowers. In these domains, however, either the environmental
complexity or the autonomous control of the machines is still quite limited.
As the trend of automation continues, many scientists and engineers envision a fu-
ture where intelligent robots take on more autonomy, interacting with humans in our every-
day lives by taking on duties like health care for the elderly or chauffeuring people across
town. Similarly, there are initiatives for human-robot teams to continue exploration of outer
space and maintain long-term off-world settlements. There is still a large gap between
the kind of intelligence needed on a factory floor and the intelligence we expect to have
when deploying autonomous robots in homes, hospitals, or space stations. Specifically,

1
autonomous robots will need to handle communication with humans and to interact with
environments that may change over time.
One important substrate of human intelligence is spatial reasoning. We rely on our
understanding of space to move in the world, to make plans, to communicate with others,
and to bootstrap higher-level concepts that seem to make us intelligent beings [Lakoff and
Johnson, 1980]. Although psychologists do not agree exactly how human cognitive maps
are built, maintained, and used, we know that cognitive maps are robust to uncertainty, often
qualitative in nature, and inherently hierarchical [Kitchin, 1994]. Human spatial reasoning
is also fundamentally related to memorization, which is necessary for building intellect
from personal experiences.1
Mobile robots also need to understand space, but most robotics research has not
been motivated by robust, qualitative human spatial reasoning. Instead most research on
mobile robot navigation has been motivated by the characteristics of modern-day sensors.
These solutions focus on building and using accurate metrical maps, similar to blueprints,
which are monolithic and non-symbolic in representation. Though these maps may look
nice, and may be useful for real-time planning over short distances, they do not lend them-
selves to symbolic inference, such as verbal communication or well-studied AI knowledge
representation paradigms.
Can we overcome the large differences between robot and human perceptual capa-
bilities in order to bind symbolic spatial concepts, similar to the ones humans seem to have,
to the “pixel-level” data that robots experience? That is, can we produce robots that have
cognitive maps?
1 Expositions on rhetoric dating back to 85 BCE discuss mnemonic devices where large speeches are re-
membered via association of talking points to physical locations in the world. To deliver a speech, a person
simply has to take a tour in their mind through a well known place, like a market or town hall. Each memory
of a place helps recall a stored portion of the speech. Cicero credits Simonides of Ceos (c. 556 BCE-468 BCE)
with discovering this artificial memory technique [Yates, 1966]. This “method of loci” technique is still used
today by world champion memorizers (often termed the “journey method”). These individuals associate long
strings of random numbers or shuffled cards with detailed tours through their houses or neighborhoods.

2
1.1 Statement of the Problem

The animal kingdom has conquered the problem of reasoning about space by employing a
variety of techniques. Many animals utilize very precise path integration (or dead reckon-
ing) techniques, where they know the distance and direction they have traveled from home.
This is often improved by using the sun as a compass, calibrating flight speed from known
static landmarks, detecting the rotation direction of the stars, or using the earth’s magnetic
field. These navigation methods are remarkably accurate, but such animals can be easily
fooled by mischievous scientists [Gallistel, 1990].
On the other hand, we know that humans (without the use of tools) tend to have
fairly poor path integration skills over large-scale environments [Golledge, 1999]. Instead
we tend to impose a symbolic structure on the world, utilizing unique landmarks or im-
posing an artificial structure on the environment, to facilitate graph-like planning between
places in the world connected by paths. Thus in normal, everyday navigation, we rely
much more heavily on internal models of space, rather than external oracles, like the sun
or stars. Remarkably even animals thought to posses unparalleled dead reckoning abili-
ties sometimes prefer following longer, human-like routes, either along man-made highway
networks [Guilford et al., 2004] or along routes they define themselves with artificial land-
marks [Stopka and Macdonald, 2003], though scientists do not yet fully understand the
reasons behind these behaviors.
The human ability to abstract the 3D world into a symbolic representation has sev-
eral benefits. It allows humans to remember, plan in, and move across relatively large
regions with respect to our size. It also facilitates hierarchical representations that allow us
to think about space at a variety of levels, each with different types and amounts of infor-
mation. From an AI perspective, symbolic abstraction facilitates learning, inference, and
knowledge acquisition. Maybe most important, a symbolic representation of space allows
communication of directions and description of objects in the world, either through verbal
means or via drawing [MacMahon, 2007].

3
 A key challenge from a robotics point of view is how to encode human-style spatial
reasoning into a machine. This is especially difficult given that current robotic sensors are
so dissimilar to human perception. On the other hand, implementing robust path integration
on mobile robots has become quite trivial—especially given modern hardware that tracks
global position changes, e.g., inertial measurements units and differential GPS. This precise
localization, along with devices that provide precise distance measurements to nearby ob-
stacles, like laser radar (lidar), allows robots to generate metrical models of the immediate
surround using a variety of well-understood techniques [Thrun et al., 2005].
The problem we have set out to investigate in this dissertation is how to abstract
symbolic spatial concepts from these metrical models in a way that facilitates robust nav-
igational behavior while supporting symbolic topological reasoning (e.g., abductive and
non-monotonic reasoning) and effective communication with human users (e.g., following
symbolic route instructions). Can we bridge the gap between modern-day metrical map-
building and human-inspired, symbolic concepts of space? We argue that this is not only
possible, but that we can ground our symbolic spatial abstraction via a geometrically de-
rived primitive, called a gateway.

1.2 Thesis Overview

For decades, AI research has moved forward on the assumption that symbolic primitives can
be used for learning, inference, and planning. These symbols are often assumed to relate to
some real-world knowledge, but the hard problem of symbol grounding [Harnad, 1990], is
often ignored or left as future work. Robotics research in particular cannot overlook symbol
grounding, as the act of deploying a physical agent in the real world requires some form
of grounding discrete symbols to low-level perceptual inputs. There are two predominant
paradigms in robot navigation research that attack the symbol grounding problem from
different perspectives.

4
 For many years, robotics researchers took a classic GOFAI approach.2 These sci-
entists believed that graph-like representations of space, such as topological networks of
places and paths, could provide a low-level symbolic abstraction that would facilitate plan-
ning and may one day be useful for learning, communication, and hierarchical knowledge
acquisition. This early work suffered from a lack of reliable sensors; thus, the abstraction of
symbols from low-level inputs was often done using ad hoc techniques that were robot and
environment dependent. As a result, much of this work concentrated on theoretical issues
dealing with graph exploration, rather than on how the graph is actually abstracted from the
continuous environment.
Over the years, more precise measuring devices became affordable and portable
enough to integrate onto research robots. The improvement of global odometry tracking,
along with precise distance measurements to nearby obstacles, yielded research that fo-
cused on building very precise metrical models of the location of obstacles in the world.
These methods were so successful in building impressive looking maps, that much of the
current-day robotics research is focused around extending the same techniques to larger
and larger environments; this is despite the fact that these methods do not scale as well and
have little useful semantic information. Another problem with these methods is that the
robot’s knowledge of the world is quite separated from the human user. Although humans
can examine these maps for correctness, the robot itself has only a very primitive spatial
understanding of the world: what locations in the world contain obstacles.
This thesis examines the use of human-like symbolic representations in mobile
robotics. We use psychological evidence to motivate a theory of hybrid map-building that
uses the sensor-inspired advances along with symbolic spatial reasoning. This thesis shows
the first attempt to fully close the loop between local metrical sensations, symbolic topo-
logical inference, and global metrical map production. In particular, we use well-known
metrical mapping techniques to ground symbolic, topological representations of space. In
2 Haugeland [1985] is credited with coining the term “Good Old-Fashioned Artificial Intelligence”.

5
doing so we model the 3D world with a 2D metrical representation. Next, we abstract
the local metrical model into a symbolic model of local places and paths. Local symbolic
models are combined into a graph-like network of connected decision points. A global
metrical map can be built efficiently on top of the symbolic topological model if desired,
but global maps are unnecessary for robot navigation once a correct topological model has
been determined.
The core contribution of this thesis is to ground the concepts of place and path in
a straightforward metrical model of the local surround. This metrical model is currently
obtained by range-sensing, but we believe (1) that such models can soon be obtained by
other techniques, such as using machine vision [Sim and Little, 2006; Murarka et al., 2006]
and (2) the hybrid, hierarchical framework is independent of the sensors used to create the
local metrical models. Given metrical models of the robot’s local surround, we argue that
places at decision points in the world can be grounded by the use of a primitive called
a gateway. Gateways separate qualitatively different regions of space and have a natural
description at intersections and in doorways. We provide analysis of multiple gateway
algorithms, a theory of how they contribute to the description of paths and places, and
examples of the benefits of gateways in spatial mapping and learning.
Additionally, we introduce a formal description of our hybrid approach to spatial
reasoning, which grounds gateways (and consequently places and paths) from models of
small-scale space. This formal description is crafted in such a way as to allows us to con-
tinue using the logical topological map-building axioms of Remolina and Kuipers [2004],
while improving the computational complexity of map-building and the aesthetics of small-
scale and large-scale motion through the environment. We illustrate our initial implemen-
tation of hybrid map-building and show successful human-robot interaction in a pilot ex-
periment. We then discuss benefits of this hybrid approach to the global metrical mapping
community, and propose how global metrical maps should in turn benefit the topological
mapping community in the near future. Finally we discuss some of the new scientific prob-

6
lems that arise from our research.
Chapter 2 overviews background and related work. We discuss work on the human
cognitive map that has motivated the kinds of knowledge we want our robots to obtain.
We then give a summary of the recent history of mobile robotics research, as a contrast to
how humans represent space. We then discuss related work on gateways, explaining their
theoretical origins and detailing previous implementations. The various drawbacks of pre-
viously published gateway implementations motivates our current gateway implementation.
In Chapter 3, we describe the Spatial Semantic Hierarchy (SSH), a framework for
hierarchical abstraction of spatial knowledge from experience. The SSH is motivated by
the human cognitive map, but remains fairly agnostic about the type of sensations the agent
may posses. This motivates the need for the Hybrid Spatial Semantic Hierarchy (HSSH),
which directly addresses modern robotic sensors, by explicitly addressing the interactions
between small-scale space and large-scale space. Here, we introduce the HSSH framework,
including the concepts that are detailed in Chapters 4 through 9.
Chapter 4 discusses local mapping and localization for a particular class of robots
that can use vision or range-sensing devices to create metrical descriptions of the local
surrounding region. Local perceptual maps (LPMs), which model local, small-scale space
are useful both for robot control and for grounding symbols in the recent experience of the
robot. We discuss improvements to the traditional incremental localization schemes in order
to ensure high-quality local metrical models [Beeson et al., 2006]. Reducing localization
failures helps to ensure high-quality local metrical models, which in turn ensure correct
place/path detection, efficient place recognition, and accurate global layouts in the higher
levels of the HSSH.
Chapter 5 delves into the symbol grounding problem. We begin a formal description
of the HSSH, from local metrical models through local topological hypotheses, that extends
previous SSH formal descriptions [Remolina, 2001; Remolina and Kuipers, 2004; Savelli,
2005]. We focus on gateways and local topology representations of space, which are used

7
as primitives for the detection, description, and navigation of topological places and paths.
We also discuss using gateways as local motion targets to ground large-scale travel actions
in the LPM.
Chapter 6 discusses particular implementation details of gateways. We detail an
initial method for defining robust, stable gateways from a Voronoi skeleton that works well
in corridor-type environments [Beeson et al., 2005]. We illustrate scenarios where this ap-
proach becomes unreliable and introduce a new algorithm that overcomes these issues while
extending gateways to non-corridor environments. Gateways need to be fairly stable in or-
der to provide a robust local topology of places in the world, so we evaluate the reliability
of the current gateway algorithm across different types of noise.
In Chapter 7, we validate the HSSH approach with a human-robot interaction ex-
periment. We introduce an interface for the HSSH in a mobility-assistance domain, and we
test this interface at the HSSH Local Metrical and Local Topology Levels [Beeson et al.,
2007]. We validate the HSSH implementation and the initial human-robot interface with
an experiment that simulates visually impaired drivers of an intelligent personal transport.
We compare navigation performance with and without the use of local metrical and symbol
models of small-scale space.
Chapter 8 returns to the HSSH formal description; this time detailing global, sym-
bolic models of space. We discuss closing loops in topological map-building, and detail
the computational improvement of the HSSH approach over the basic SSH. Specifically,
we show that gateways and the resulting local topology provides a deterministic description
of decision points in the world, which can eliminate a large number of invalid topological
hypotheses [Beeson et al., 2003; Kuipers et al., 2004].
In Chapter 9, we discuss how to utilize the local metrical data experienced during
travel between places to synthesize a metrically accurate global map. This process uses the
topological map as a skeleton. The topological places can be relaxed in a global metrical
frame of reference prior to the full metrical map being hypothesized [Modayil, Beeson, and

8
Kuipers, 2004]. This approach scales efficiently with environment size and complexity.
Chapter 10 summarizes the contributions of this dissertation and proposes future
avenues of work. Particularly, we discuss how the global metrical layout, which is com-
putationally efficient when using a topological skeleton, can feed back into the search for
the correct topological map. Issues of active exploration, visual LPMs, and learning place
recognition are also discussed.

1.3 Major Contributions

There are several key contributions of this dissertation that span multiple disciplines. In the
area of cognitive science, we present a formal description of a spatial reasoning architecture
inspired by psychological evidence. This framework uses both small-scale and large-scale
frames of reference and both metrical and symbolic representations of space.
We improve localization techniques from the robotic metrical mapping (SLAM)
domain to ensure high-quality local metrical models of the nearby surround. We contribute
an efficient, online method for global metrical map-building, useful in surveying, mining
and other applications, by utilizing the learned symbolic topological map as a skeleton for
the global layout.
We introduce a novel algorithm for determining gateways, which separate qualita-
tively different portions of environment based on local structure. We tackle a specific AI
symbol grounding problem in the robotics domain by implementing place detection and
place description that improves on the state-of-the-art. Experimental validation of the hy-
brid metrical/symbolic navigation in a wheelchair domain is relevant to the human-robot
interaction (HRI) community.
This dissertation provides useful innovation in all the areas above. We attempt to
innovate across the breadth of the hierarchical framework of spatial knowledge, rather than
focusing one specific aspect; however, we believe the HSSH formal description, which is
consistent with (yet improves upon) the basic SSH logical theory, along with the current

9
implementation for determining gateways, which provides a variety of benefits at all levels
of the SSH framework, are the core contributions of this thesis.

10
Chapter 2

Background and Related Work

Knowledge is of two kinds. We know a subject ourselves, or we know where we can find
information on it.
Samuel Johnson, James Boswell’s Life of Johnson (1791)

This chapter introduces the reader to the background information needed to understand both
the motivations and the techniques used in our research. We begin with an overview of hu-
man cognitive map research. We then discuss recent history in the field of mobile robotics,
detailing research concentrating on building extremely accurate and precise metrical rep-
resentations of environments along with ideas about topological robot navigation as graph
exploration. We finish with a discussion of gateways, which have been examined by several
robotics researchers, and which we use as the anchoring primitive for grounding topological
concepts into well-understood metrical models of the world.

2.1 Cognitive Maps

Tolman [1948] coined the term “cognitive map” to represent the mental spatial representa-
tion of an animal. Since this time, there has been much study on the way animals, espe-
cially humans, solve navigation problems in the world. Most important to this thesis is the
research on human representations of large-scale space. Researchers from developmental

11
psychology all the way to urban designers were all pioneers in the early work of human
navigation [Piaget, 1954; Lynch, 1960; Appleyard, 1970].

2.1.1 Human Spatial Representations

Probably the most influential work on human acquisition and use of the cognitive map
is by Siegel and White [1975]. This research notes that adult humans go through three
stages when forming a cognitive map. First, landmarks define important locations in the
world. Next, routes between landmarks are remembered. Finally, configurations unify route
knowledge together into a more globally coherent representation. The authors remain fairly
agnostic as to what exactly constitutes a landmark, other than a discussion of distinctiveness
or novelty as a characteristic of landmarks. More recent work supports the notion that
stability can also be an important characteristic of landmarks [Biegler and Morris, 1996].
This idea of a few distinctive, stable places, connected together by remembered
routes, eventually leading to a cohesive connected network is the basis for several modern
theories that support topological maps as reasonable models of the human cognitive map,
including the framework we utilize in this dissertation. On the other hand, global, metrically
precise maps have long been recognized as limited in their cognitive plausibility [Lynch,
1960; Downs and Stea, 1973; Kuipers, 1982]. In fact, Lynch [1960] found that errors in
cognitive maps are rarely topological and frequently metrical. This is not to imply that
topological maps cannot use metrical information, but that symbolic abstraction is a vital
part of human spatial reasoning.

2.1.2 Structural versus Object Landmarks

Many researchers take Siegel and White’s concept of “landmarks” to refer to object land-
marks that facilitate homing-style navigation, which leads the human between beacons in
an otherwise perceptually uniform environment. While this may be true in certain environ-
ments, most man-made environments, such as office complexes, campus walkways, street

12
networks, and even hiking trails have a graph-like structure imposed on them. In these
environments, there exist object landmarks, though there will be structural landmarks in
addition to object beacons.
Specifically, decision points in environments are important, landmark locations
when building a cognitive map, as they are both very stable and quite distinctive from a
control point of view (new control opportunities arise at decision points).1 Lynch [1960]
was perhaps the first to present evidence that city dwellers abstract the world, at the finest
granularity, into streets and their intersections. Urbanites use these graph-like networks to
perform reliable navigation through the city. The idea of decision points (intersections)
in man-made environments as “first-level” spatial concepts has been reiterated by other
researchers [Golledge et al., 1985].

2.1.3 Relevance of Cognitive Maps to this Thesis

In our work, we are really concerned with detecting and describing places at decision points
in the world. That is, we constrain our domains to structured, man-made environments that
have natural graph-like connectivity. Environments like deserts, dense jungles, or oceans
can be learned and navigated successfully by humans, but normally only after acquiring
specific knowledge from experts, using external navigation tools, or utilizing detailed maps.
We are much more interested in having robots build up the types of knowledge the majority
of humans experience on a daily basis. Accordingly, we are currently not concerned with
places that do not occur at decision points.
We do not deny important places exist along paths where no decision really exists.
Places such as towers along roadways, water fountains in buildings, etc. are useful meeting
points or checkpoints along routes. However we argue that this object landmark knowledge
is actually secondary to knowledge of the structural connectivity of large-scale environ-
ments. Allen et al. [1979] performed an experiment where subjects were asked to pick the
1 Later we will show that decision points are actually distinctive from a structural point of view as well, as

gateway configurations change at decision points.

13
9 snapshots that would be best to help them remember the route captured by an ordered
collection of 52 snapshots. Adults chose snapshots taken at path intersections 94% of the
time. Appleyard [1969] found that mere visibility of a building is not sufficient for its recall,
but proximity to an intersection virtually guarantees it. Gale et al. [1990] and later Aginsky
et al. [1997] found that more visual information was remembered around intersections than
between intersections along paths. Heft [1979] followed by Stankiewicz and Kalia [2007]
showed that humans remember the structure of decision points better than object landmarks
when exploring unfamiliar environments.
This thesis discusses integrating local metrical models built from robot experience,
with large-scale symbolic models of space. We use these local models to detect and classify
decision points in the world, but we also utilize the small-scale representations to improve
control, exploration, and map-building efficiency.

If the organism carries a “small-scale model” of external reality and of its own
possible actions within its head, it is able to try out various alternatives, con-
clude which is the best of them, react to future situations before they arise,
utilize the knowledge of past events in dealing with the present and future, and
in every way to react in a much fuller, safer, and more competent manner to the
emergencies which face it. [Craik, 1943, p. 61]

As a result, we are combining useful dead-reckoning and metrical environment sensing in

local regions with symbolic reasoning about large-scale spatial knowledge.

2.2 Robot Maps

In modern day robotics, there are two main paradigms for modeling the spatial structure of
the environment. One, topological mapping, attempts to discretize the continuous world as
a graph-like representation of places connected by paths. The other, metrical mapping, tries
to create a CAD-like layout of the location of all obstacles in the world. We first discuss

14
metrical mapping, as it allows us to introduce common terminology and issues that arise in
mobile robot navigation and map-building. We then discuss topological mapping, followed
by a discussion of hybrid implementations that combine elements of both topological and
metrical maps.

2.2.1 Metrical Maps

Topological maps (discussed below) were studied early on in robotics. They were seen as
a symbolic way of describing of the environment that could provide a basis for higher-level
reasoning; however, due to the impoverished sensors of the time, robust topological imple-
mentations were not really obtainable. As more accurate, precise, and informative sensors
became available, researchers quickly realized that they could build much more detailed
maps than ever before. This was a real breakthrough for modeling indoor environments.
As a result, the idea of creating models of the world similar to human models of space was
largely abandoned by the robotics community in favor of creating more precise and visually
impressive CAD-like models of environments.2

Metrical Simultaneous Localization and Mapping (SLAM)

Powerful probabilistic methods have been developed for range-sensing mobile robots to
perform simultaneous localization and mapping (SLAM) within a single frame of reference
[Thrun et al., 2005]. These methods are accurate and reliable for online incremental lo-
calization within local neighborhoods. Sensing with sufficiently high frequency relative to
local motion guarantees large overlap between successive sensory images. Current sensory
information can be compared to the current map in order to improve localization. By anal-
ogy with radar signal interpretation, finding the correct match between observations and
model is called the data association problem. After improved localization occurs, the sen-
2 In our opinion some metrical mapping research focuses too much on creating precise maps—focusing their

robotics research on improving the visual quality of a map produced by offline computing, while forgetting that
the main purpose for a map is to facilitate robust, efficient navigation by a mobile agent.

15
sory information is used to update the map for the next SLAM iteration. In local regions,
many data association problems, such as the closing of large loops, can be excluded. The
absence of large loops means that the problem of large-scale structural ambiguity does not
arise in the local metrical map.

Scaling Metrical SLAM Techniques

While metrical SLAM methods work in small spaces, they do not extend well to larger
environments. Global metrical maps become more expensive to update and access without
clever storage schemes. More important, is the difficulty that arises when closing large
loops (Figure 2.1). Even with local SLAM methods that use perception to improve the ac-
curacy of localization, odometry error accumulates in the relation between the map’s global
frame of reference and the ground-truth reference frame of the real-world environment. This
global error becomes even more pronounced in environments with long paths that have few
distinguishing features. Without exact data association along paths, localization often drifts
from the ground-truth, both in the robot’s distance along the path and in the robot’s heading,
causing straight paths to compress, stretch, or curve in the map.
There are ad hoc methods for hypothesizing loop closures when the global odom-
etry error is small. When a large loop is closed, accumulated error will often result in the
robot’s current observations clashing with older portions of the map. Methods exist that
search for a nearby pose in the older portions of the map where perceptions match the pre-
diction (propagating detected global error backwards through the exploration trace) [Lu and
Milios, 1997; Hähnel et al., 2003a]; however, these solutions can fail in sufficiently large or
complex environments. For example, Cummins and Newman [2008] discuss closing loops
over kilometers of travel, where small rotational errors lead to large positional errors, and
the correct loop closure may never be considered by odometry-based solutions. Addition-
ally, if the environment is subject to perceptual aliasing (different locations look the same),
then the matching process may close the loop incorrectly, distorting the map as a whole

16
 (a) (b)

(c) (d)

Figure 2.1: Closing large-scale loops. Robots often need to explore, map, and navigate in large
environments with multiple, nested loops. (a) This environment and the robot’s trajectory through
it are used as an example throughout. (b) The data comes from a Magellan Pro research robot
with differential-drive odometry and a S ICK-brand lidar device for precise, planar range-sensing.
(c) This robot-made map of the environment in image (a) shows the effect of accumulated raw
odometry error. (d) This map shows the improvement in pose estimation over image (c) by using
metrical SLAM methods, but it also shows that significant errors still accumulate with respect to the
real environment.

17
Figure 2.2: Incorrectly closing loops using odometry and scan matching. Metrical methods that
attempt to match current observations to older portions of the map have been proposed. Sometimes
these work well. Sometimes they do not work at all. As illustrated here, sometimes they seem to
succeed as observations match older portions of the map, but the loop is incorrectly closed. (Adapted
with permission from Eliazar and Parr [2003].)

(Figure 2.2). Depending on the amount of symmetry in the environment, a single incorrect
match can lead the learner down an arbitrarily long “garden path” before the error is discov-
ered. It is still unclear how probabilistic methods can properly discover an incorrect map
and how they might efficiently backtrack to hypothesize a different loop closure [Hähnel
et al., 2003b].
Some research on map-building avoids loop-closing issues by explicitly assuming
that the correct data association is known [Leonard and Newman, 2003; Paskin, 2003]. In
some cases, even without an explicit assumption about data association, impressive feats
of large-scale map-making depend on locations in the environment being sufficiently dis-
tinguishable based on local cues [Montemerlo et al., 2002; Konolige, 2004]. Others accept
false negative matches in order to avoid false positives, sometimes improperly hypothe-
sizing that a previously seen location is new [Bosse et al., 2003]. This can eliminate the
possibility of closing a loop correctly and finding the correct map, which leads to poor
planning and navigation performance. In a rich environment with noise due to dynamic

18
changes, it could be that every location is in principle distinguishable, but it is difficult or
impossible to know which features identify the place, and which are noise. Methods cre-
ated to distinguish between perceptually aliased states can get confused under scenarios of
perceptual variability (the same place looks different on separate occasions) [Kuipers and
Beeson, 2002] causing a single physical location to be represented multiple times in the
same map.
Early approaches to probabilistic localization and mapping used particles to repre-
sent a distribution over robot poses for localization, but there was a single shared map that
was updated from the maximum-likelihood pose hypothesis [Thrun et al., 2000b]. This
could produce an incoherent map due to an incorrect and premature commitment to a
maximum-likelihood pose hypothesis that turned out to be incorrect. A more principled
approach uses Rao-Blackwellized particle filters to explicitly represent the distribution of
trajectories and maps by maintaining multiple metrical map hypotheses [Montemerlo et al.,
2003; Eliazar and Parr, 2003; Hähnel et al., 2003a]. These methods are run offline (due to
computational demands) after exploration is completed, forgoing useful active exploration
techniques capable of eliminating some loop-closing hypotheses. Additionally, in large,
symmetric environments, intractably large numbers of particles may be required to avoid
particle depletion when closing large loops. Particle depletion is a failure to have a particle
in the distribution that adequately models the correct map.
In addition to the problems that arise in building large metrical maps, there is the as-
sociated problem of navigating in large metrical models. While search is relatively straight-
forward in small, local metrical models or in compact topological representations, planning
over large distances in metrical models (represented as point clouds or large 2D or 3D grids)
is quite computationally intensive.3 Despite this fact, many robotic researchers continue to
focus on improving the accuracy and preciseness of large metrical models, many times
3 Probabilistic roadmaps [Kavraki et al., 1996], rapidly-exploring random trees [Kuffner and LaValle, 2000],

or other modern planning techniques may improve the speed of planning tasks; however, having hierarchical
representations facilitates efficient local replanning without having to replan the entire global route.

19
even driving the robot themselves to gather observations. This defeats the original purpose
of building a model, which is to facilitate robust exploration and navigation.

2.2.2 Relevance of Metrical SLAM to this Thesis

The work in this thesis depends on a small-scale model that we call a local perceptual map
(LPM). This local map is a highly accurate metrical map of the immediate local surround.
It has no long-term persistent state as it scrolls with the robot. It is useful for abstracting the
local, continuous environment into a stable model of small-scale space. It supports local
planning and motion control. Symbolic abstraction can then be performed using the LPM
representation.
The current implementation of this LPM is an occupancy grid map [Moravec, 1988;
Elfes, 1989] that uses planar lidar (laser range-sensing) devices to detect locations in the
world occupied by obstacles. The occupancy grid algorithm discretizes the world into cells,
each of which holds a probability of a cell being occupied by an obstacle—cells are initial-
ized to 0.5 probability for unmapped regions. Based on occupancy, cells in the grid can be
classified as occupied, free, or unknown. Our LPM is built using the SLAM particle filter-
ing techniques described above to create a single maximum-likelihood map hypothesis. As
a result of using range sensing, we are currently focusing our efforts on environments with
vertical walls.
We fully expect current progress in Visual-SLAM [Sim and Little, 2006], vision-
based road/sidewalk detection, and vision-based obstacle detection (along with 3D lidar
sensing as it becomes more available) to eventually give us LPMs of outdoor environments
with dynamic obstacles. Other related research is currently looking at ways of producing
semantically labeled LPMs to facilitate campus-wide navigation [Murarka et al., 2006].
Given LPMs as a base description of the local surround, our theory (and to a large extent
our implementation) of hybrid map-building and navigation can remain unchanged.

20
2.2.3 Topological Maps

Cognitive map research supports the creation of topological maps of large, complex en-
vironments [Lynch, 1960; Siegel and White, 1975; Kuipers, 1978; Chown et al., 1995;
Kuipers, 2000]. A topological map, in its most basic form, represents an environment as a
graph where nodes represent places and edges represent connections between places. Sev-
eral groups of robotics researchers have presented distinct topological implementations that
differ in their semantics for the graphs—how they define/describe places and actions be-
tween them [Kuipers and Byun, 1991; Mataric, 1992; Shatkay and Kaelbling, 1997; Duckett
and Nehmzow, 1999; Choset and Nagatani, 2001; Morris et al., 2005]. Some implemen-
tations build topological maps autonomously, some are given topological maps a priori
[Koenig and Simmons, 1996], and some let the robot explore autonomously while the re-
searcher provides place names to overcome perceptual aliasing issues [Kortenkamp and
Weymouth, 1994].
Topological navigation is a behavior that is used by a variety of different animal
species, including humans [Lynch, 1960; Gould, 1990; Stopka and Macdonald, 2003]. This
helps facilitate large-scale spatial reasoning, mainly due to the compactness of the repre-
sentation. The symbolic nature of topological representations allows higher-level reasoning
(such as containment, order, connectivity, regions) which can enhance large-scale spatial
knowledge. Because the environment is discretized into a graph, movement errors do not
accumulate globally. Additionally, the abstraction of continuous space into symbols facil-
itates communication about space between agents. Possibly the most important difference
for future robotics research is that topological maps allow hierarchical models that support
efficient navigation, planning, and communication about very large regions of space: e.g.,
a building may consist of many topological places, but may be a place in a higher-level
topological network of a campus.
The major hurdle for topological map-building has been the reliable abstraction
of useful symbols from continuous, noisy perceptions of the environment: i.e., how to

21
reliably detect and recognize places and paths. This is an instance of the more general
symbol grounding problem [Harnad, 1990] that has troubled the AI community for many
years. Probabilistic approaches are good at overcoming the kinds of local uncertainty and
systematic noise that can hinder reliable symbol extraction. Incorporating probabilistic
data association techniques into the topological map-building paradigm has sparked interest
in hybrid map-building, including the HSSH approach presented in this dissertation. We
discuss hybrid mapping approaches below; however first we detail research on exploring
and navigating graph-like models on environments.

Graph Exploration

Early work on topological mapping ignored the embodied agent altogether, bypassing dif-
ficult symbol grounding issues, and focused on the theoretical results of exploring graphs.

Deterministic Finite Automatons. Rivest and Schapire [1989] produced one of the ini-
tial pieces of work on formalizing topological-map building. They describe topological
map-building as the problem of determining the most simple deterministic finite automaton
(DFA) that explains symbolic input/output experience. A topological map can be thought
of as a six tuple, M = (S, A,V, ζ , s0 , γ), where:

• S is a finite nonempty set of states

• A is a finite nonempty set of inputs or basic actions

• V is a finite set of outputs or percepts

• ζ is the transition function, ζ : S × A → S

• s0 is the initial state

• γ is the output function, γ : S → V .

22
 No perceptual aliasing exists if ∀s1 , s2 ∈ S [γ(s1 ) = γ(s2 ) → s1 = s2 ]. In this case
determining the correct DFA is trivial. When ∃s1 , s2 ∈ S [γ(s1 ) = γ(s2 ) ∧ s1 6= s2 ], then
perceptual aliasing exists, making the general problem of inferring the smallest DFA NP-
complete [Gold, 1978; Angluin, 1978].
Rivest and Schapire [1989] show that there is a polynomial-time algorithm for in-
ferring the smallest DFA given that a distinguishing sequence can be learned in polynomial
time. A distinguishing sequence, d ∈ A∗ , results when ∀s1 , s2 ∈ S, s1 hdi = s2 hdi → s1 = s2 ,
where shdi denotes the sequence of outputs resulting from executing the sequence of ac-
tions d starting in s. Unfortunately, later it was shown that determining whether a DFA
has a distinguishing sequence is PSPACE-complete; however, an adaptive distinguishing
sequence can be determined in polynomial time [Yannakakis and Lee, 1991]. The output
symbols of the DFA determine the branches in a decision tree, which defines an adaptive
distinguishing sequence.
Dudek et al. [1991] investigate using multiple unique markers that can be dropped
and picked up by the robot. In a sense, this lets the robot create distinguishing sequences
on the fly—dropping markers at proposed loop closures before retracing its exploration to
look for the marker at some known state from the past. Although using markers still results
in polynomial time exploration, experimental evidence shows a reduction in the number of
actions necessary to properly model an arbitrary environment. This reduction levels off as
the number of unique markers approaches the number of aliased states in the environment.

DFAs and Uncertainty. Thinking of topological maps as deterministic finite automata

allows us to keep the map-building problem tractable—if not theoretically, at least in prac-
tice. However, in many mobile robot implementations, actions or observations may have
uncertainty, which is difficult to model using DFAs. Some researchers have attempted to
learn stochastic finite automata under various conditions of uncertainty [Dean et al., 1995;
Basye et al., 1997; Shatkay and Kaelbling, 1997].
In Rivest and Schapire [1989], the observation of a state was deterministic ϕ(s) = γ,

23
though perceptual aliasing could still exist. Dean et al. [1995] consider the case where per-
ceptual variability also exists (the observation function is non-deterministic), yet actions are
still deterministic. Here there exists some observation function, such that it is possible that
upon visiting state s on two occasions i and j, ϕ(si ) = γ and ϕ(s j ) 6= γ. They conclude that
it is possible to find a probably approximately correct (PAC) deterministic finite automaton
[Valiant, 2000]. This can be done in polynomial time given that the environment has a dis-
tinguishing sequence and such a distinguishing sequence is known prior to map-building
(or can itself be learned in polynomial time).
Shatkay and Kaelbling [1997] attempt to build a model of an environment in the
face of perceptual aliasing, perceptual variability, and uncertain actions. Given probabilis-
tic actions, an environment can be modeled as a Hidden Markov Model (HMM). They
utilize the Baum-Welch algorithm to converge to a model that maximizes the expectation
of the observed data. Specifically, they utilize odometry information between states in or-
der to find a model, M, that explains the experience, E = {v0 , a0 , ..., an−1 , vn }. This finds a
likely local maximum, which may be far from optimal: arg max p(E|M). In experiments,
the models generated were consistent with all available metrical information, but far from
topologically correct.
One major problem with these PAC learning approaches is that a large amount of
experience is needed in order to come up with a useful model. The amount of experience
needed is often far more than what is reasonably acceptable for a usable robotic imple-
mentation. Additionally, though models may be probabilistically optimal, they may not
accurately reflect the ground-truth of the environment. As a result, we prefer to utilize the
high-precision observations from modern sensors in order to assert that both observations
and actions are deterministic. Future work should look into the impact where rare non-
deterministic events occur: detecting and understanding these unusual events should allow
us to still assume deterministic actions (100 − ε)% of the time.

24
Voronoi Graph Exploration

Building DFA or HMM models of environments depends on some way of detecting and
categorizing places in the environment. Early work on place detection was often quite ad
hoc [Mataric, 1992; Kortenkamp and Weymouth, 1994], as most research robots used noisy
sensors. Using Voronoi graphs [Voronoi, 1907; Fortune, 1992] to model paths and places
became a well-known approach to detect and classify places in a more principled way.

Voronoi Junctions as Places. For some time, researchers have proposed utilizing vari-
ants of the Voronoi tessellation as tools for detecting places and proposing paths through
enclosed environments [Miller, 1985; Rao, 1995]; however, in recent years Choset and
his colleagues have been the main proponents for the use of Voronoi graphs in topologi-
cal map-building and navigation. Choset and Burdick [2000] define a generalized Voronoi
graph (GVG) as a one-dimensional set of points equidistant to the n closest obstacles in n
dimensions, where a preset threshold determines whether observations belong to the same
“obstacle”. When using planar range sensors (like common laser and sonar devices) in
walled environments, this results in a set of points equidistant from two or more obstacles
(Figure 2.3).
Researchers utilize the GVG both as a roadmap for robot control and as a way to
discretize the continuous environment into a finite set of places. Whenever the robot arrives
at a junction point in the Voronoi graph (a point equidistant from n + 1 or more closest
obstacles), it decides that it is at a topological place. In corridor environments, Voronoi
junctions occur at corners and intersections.
In order to describe a place for future recognition, they count the number of edges
emanating from a junction point—the degree of the graph node. In doing so, they eliminate
perceptual variability completely (assuming the environment, thus the GVG, is static). This
gives them an observation of a place equivalent to the one used by Dudek et al. [1993] in
work on graph exploration. However, simply using the degree of a graph vertex causes a

25
Figure 2.3: The generalized Voronoi graph (GVG) of a global metrical map. Using an occupancy
grid, the Voronoi graph of free space can be drawn by using any occupied cells as obstacles. Junction
points include intersections in corridor environments. For visual clarity, portions of the graph that
are close to obstacles are removed.

large amount of perceptual aliasing.

To reduce aliasing, Choset and Nagatani [2001] also use the Voronoi radius (dis-
tance to the nearest obstacle) of the junction node to help classify places. This still does not
help to eliminate aliasing between qualitatively different intersections, such as a + inter-
section which has two intersecting paths and a K intersection which has three intersecting
paths. Both of these intersections have four Voronoi edges and could have similar Voronoi
radii given the architecture of the building.
Choset and Nagatani [2001] also admit that the Voronoi graph can find many spuri-
ous topological places due to range-sensing noise or concave corners. They propose using
a reduced generalized Voronoi graph (RGVG) that removes many “spurs” in the graph. The
RGVG removes branches that terminate with only two nearest obstacles. This improves on
prior Voronoi navigation approaches, but the RGVG can still leave some spurs that create
false positive/unstable places. Additionally, some places may have no Voronoi junctions in
the RGVG, while other place neighborhoods may encapsulate multiple nearby junctions.
This is discussed in more detail in Section 5.3.3.

26
Voronoi-style Exploration. There are several implementations for place detection that
are similar to Voronoi junction detection. Kuipers and Byun [1991] detect “distinctive
places” (a precursor of the distinctive states discussed in Chapters 3 and 8) via wall-following
controls that end in hill-climbing to maximize a “distinctiveness measure”. The specific dis-
tinctiveness measure they use attempts to maximize the distance from the nearest distinct
obstacles, which essentially moves the robot to the Voronoi junction. This requires the robot
to physically move to a point place, rather than modeling the place as a region encapsulated
by the radius of the Voronoi junction or by gateways. A similar control-based approach is
used by Lee [1996], Menegatti [2002], and Victorino et al. [2003].
Silver et al. [2004] use local models of nearby obstacles to detect Delaunay trian-
gles. The Delaunay triangulation [Delaunay, 1934; Fortune, 1992] of a set of planar points
is a triangulation such that no point is inside the circumscribing circle of any triangle, with
the constraint that the triangulation maximizes the total edge length of all the triangles. The
Delaunay triangulation of a set of points is the dual graph of the Voronoi tessellation of the
points.
Delaunay triangles exist throughout the environment, so the algorithm of Silver
et al. [2004] ignores triangles that do not have all three sides longer than a predetermined
threshold. This eliminates many false positive place detections that junction-based methods
may find. At intersections, large Delaunay triangles exist; thus intersections are detected
well.
Intersections in this implementation are characterized by the size of the associated
triangles, but this is not a very distinctive measure in symmetric environments. Addition-
ally, at Voronoi junctions, Delaunay triangles connect the n + 1 nearby obstacles in the
n-dimensional model. This means that certain simple intersections, e.g., a noise-free + in-
tersection, must be divided into multiple places when using Delaunay triangles. Finally, this
method is only valid in corridor-style environments, where paths have two sides to anchor
the triangle vertices.

27
 Choi et al. [2002] also use skeletal graphs to detect places in local metrical mod-
els. Instead of the GVG, they utilize a thinning algorithm [Zhang and Suen, 1984] on the
discrete, “pixelized” occupancy grid. A thinned skeleton approximates the Voronoi skele-
ton, except that many spurs that appear in the GVG do not appear in the thinned skeleton,
due to the specifics of the thinning algorithm. This pre-pruned skeleton reduces false posi-
tive place detections, but like the reduced GVG (RGVG), it fails to detect important places
where junctions may not occur.
Thinned skeletons, along with discretized, pixel-based Voronoi approximations, can
be computed in linear time; thus they are useful, pruned approximations to the continuous
Voronoi skeleton. Thinned skeletons actually “look” nicer to the eye than Voronoi graphs,
as they tend to yield skeletons with straighter lines. However, these skeletons do not always
end up being equidistant from multiple nearby obstacles, which can be detrimental to any
algorithms that depend heavily on this Voronoi property. Section 6.1 discusses how this
approximation technique is used to help find gateways.

2.2.4 Spatial Semantic Hierarchy Theses

This dissertation extends the Spatial Semantic Hierarchy [Kuipers, 2000], a framework for
modeling large-scale space at a variety of related abstractions. There have been several dis-
sertations regarding the Spatial Semantic Hierarchy throughout the years [Kuipers, 2008].
Byun [1990] introduced the Spatial Semantic Hierarchy, focusing on the low-level
trajectory and hill-climbing controls needed to perform reliable exploration with a sim-
ulated robot (see Chapter 3). Lee [1996] improved upon this work by creating an SSH
implementation that worked for a real robot using sonar sensors. He specifically decided
against using local models as they were difficult to generate with noisy sonar devices, in-
stead relying on hand-crafted, behavior-based hill-climbing controls to detect places and
paths. Recently, Menegatti [2002] proposed using his omnidirectional vision hardware to
detect vertical edges in the surround in order to hill-climb to distinctive states in indoor

28
Figure 2.4: Detecting places with an offset path. Many place detection implementations, like those
that use Voronoi junctions or hill-climbing controls, cannot detect single places with paths that “jog”
slightly at the intersection. Instead they model these intersections as multiple nearby places. The
gateway approach we discuss in Section 6.2 detects this intersection as a single place with two paths.
(Image adapted from Lee [1996].)

environments.
All of these works provide distinctive state detection and recognition directly from
sensor data, which made the implementations highly dependent on rectilinear walls. Cer-
tain scenarios, like paths that become slightly offset at intersections, can cause single places
to be modeled as multiple places. This is because hill-climbing controls find multiple dis-
tinctive peaks when paths “jog” slightly at intersections (Figure 2.4). This is similar to the
problems faced by Voronoi junction approaches.
Remolina [2001] focused on a logical formalism that allowed a robot to build cohe-
sive topological networks from the causal experience between distinctive states. He focused
on generating and comparing map-hypotheses using nested abnormality theory [Lifschitz,
1995]. This required using a circumscription policy to compare different logical models.
Savelli [2005] investigated adding quantitative data to the circumscription policy in order
to improve the comparison between logical models. This work also investigated improve-

29
ments to the topological map-building theory via planarity constraints.

2.2.5 Relevance of Topological Maps to this Thesis

Due to the growing number of potential maps and the large amount of exploration needed
to handle uncertainty in place detection (non-deterministic actions) and place recognition
(uncertainty in observations), we prefer to assume a deterministic model of the world; how-
ever, this can be difficult without reliable place detection and place descriptions. As a result
of the place detection algorithm discussed in Section 5.3, we demonstrate that deterministic
actions and place descriptions are achievable in certain environments, and we are confident
that our methods will scale nicely to near-deterministic models in more complex domains
that may require visual sensing in highly dynamic environments. As a result, we model
environments as deterministic finite automata (DFAs), and should be able to utilize adapt-
ing distinguishing features or other theoretically sound policies for improved exploration
strategies in the future.
As we have seen, Voronoi graphs have been used to produce certain place detec-
tion/classification implementations, but these methods suffer from several shortcomings
due to relying on branch points in the Voronoi graph. It is our belief that with descriptive
metrical models of the local surround, we can improve upon this work, and detect/describe
places more robustly. As detailed in Section 5.3.3, our gateway approach (which also relies
heavily on Voronoi approximations), overcomes many of the problems of the junction-based
approach to place detection and classification. Additionally, we expect our gateway method
to work on true Voronoi skeletons or on the thinning-based approximations, which allows
us to utilize any state-of-the-art, efficient implementation of a Voronoi-style skeleton of free
space.
This dissertation improves on previous SSH work. Our implementation handles
non-rectilinear intersections, eliminates the need for awkward hill-climbing motions, and
improves the efficiency of generating and comparing logical topological models. Also, the

30
LPM representation is currently built by lidar sensors, but could also project 3D vertical
edges from omnidirectional vision into the 2D local metrical representation.

2.2.6 Hybrid Maps

One useful aspect of topological maps is that they discretize large, continuous environments
into a manageable set of connected places. Metrical maps are useful as they are a natural
model to produce with a robot using precise lidar sensors and high-frequency odometry
measurements. Additionally, they give a visually compelling picture about the model the
robot has built of an environment (detecting a loop-closing error in a symbolic representa-
tion is more difficult that simply seeing a misaligned hallway in a metrical map). Recently,
robotics researchers have begun to look at hybrid topological/metrical representations in or-
der to try to leverage the benefits of both approaches. There are too many hybrid approaches
to mention here, many with only very subtle differences. We refer the reader to the survey
of such approaches by Buschka [2005], as we refer to well-known or prototypical examples
in this discussion.
Basically, publications about hybrid metric/topological representations fall into two
categories, both of which are addressed by the work in this thesis. Many robot implementa-
tions use local metrical models as local observers that help filter out sensor noise, aggregate
observations over time, and create plans that avoid nearby obstacles. There are researchers
specifically interested in using metrical models to try and determine qualitatively distinct
or interesting places as the robot explores a new environment [Yeap and Jefferies, 1999;
Lankenau et al., 2002; Tomatis et al., 2002; Ko et al., 2004]. This is related to our work on
grounding places and paths in local metrical models, though our assumption of determinis-
tic places (detection and categorization) is more rigid than most approaches.
The second hybrid approach focuses on using “places” in order to reduce the num-
ber of locations in the world that must be considered when hypothesizing metrical loop
closures. That is, the goal of most hybrid mapping techniques is still to achieve a global

31
metrical map; however, they use some “topological” (i.e., graph) constraints to make the
closing of loops more efficient. Many of these implementations record places arbitrarily
[Duckett and Saffiotti, 2000; Zimmer, 2000; Blanco et al., 2008], e.g., every 5 meters trav-
eled, in order to reduce the number of locations in the world where loop closures can occur.
Others use a feature buffer, so the robot creates a new place at every n corners or wall seg-
ments [Bosse et al., 2003]. Some approaches simply have the researchers press a button to
define places in the world [Thrun et al., 1998]. Our research is related to these approaches as
well, in that given autonomous place detection at qualitatively distinct and metrically distant
places, we can provide a compact graph to make these approaches extremely efficient.
Finally, a third category of hybrid mapping has only recently been investigated.
There has been recent research looking at modeling the full Bayesian distribution over
topological hypotheses [Ranganathan et al., 2006; Blanco et al., 2008]. These methods
are still strongly grounded in using odometry knowledge (which can be unreliable over
large distances [Cummins and Newman, 2007]) and/or aligning raw lidar measurements.
As mentioned above, the “places” they utilize are determined by ad hoc means: by the
researcher via button presses or by using distance thresholds or finite-sized feature buffers.
We believe our compact topological representation is useful here as well. Section 10.1.2
addresses how this new area of hybrid research meshes with the HSSH.

2.2.7 Relevance of Hybrid Mapping to this Thesis

Labeling a region as a place using an arbitrary rule can improve the efficiency of global
relaxation by having fewer constraints. Likewise, it can improve Bayesian filtering of pos-
sible metrical loop closures as there are fewer low-level matches to consider. However,
these approaches to creating “places” do little to improve the spatial knowledge of the robot
itself.
In our experience, metrical and topological representations for space are very differ-
ent in character, or more precisely, in ontology. The topological map describes the structure

32
of large-scale space. It abstracts away the specific nature of sensory input and the specific
methods used for matching sensory images when the topological map is created. Metrical
mapping techniques that rely on local overlap of successive sensations, on the other hand,
precisely capture the structure within the local sensory horizon: small-scale space.
The Hybrid Spatial Semantic Hierarchy (HSSH) we present in this thesis is the
first framework and implementation to close the loop on the process of going from met-
rical sensations to both metrical and symbolic models of both small-scale and large-scale
space—moving from metrical models of small-scale space to symbolic representations of
small-scale space, inferring large-scale structure via symbolic inference, before producing a
consistent global metrical model from the symbolic structure. Our approach autonomously
detects and describes qualitatively distinct places, creating far fewer places than other “hy-
brid” approaches. Additionally, these places have meaning as they are inspired by the hu-
man cognitive map.
It is important to note that our model of large-scale space is topological; thus, the
problem of closing loops is one of topological place detection and recognition. As a result,
we must bridge the gap between local metrical models of small-scale space and global
symbolic models of large-scale space. In Chapter 5, we show that the concepts of places
and paths can be grounded in local maps via the gateway construct, which is discussed
below.
Although the rest of this dissertation focuses heavily on an initial HSSH imple-
mentation, our contributions of autonomous place detection and place description should
improve most other hybrid map-building implementations. As an example, Thrun et al.
[1998] illustrate a way of integrating topological maps and metrical maps to create large,
global metrical maps. The processes includes several passes of Expectation-Maximization
(EM) using both an alpha (Markov localization) step and a beta (future observations and
actions) step to converge to a consistent metrical map. Unique topological “places” are
defined by the researcher and the robot uses its odometry to figure out the best possible

33
map (large grained grid) of the world. Computation is done offline and assumes several
things that our work directly addresses: (1) places can be reliably detected, (2) places can
be recognized on revisits, (3) exploration of the environment is complete. Our topological
map-building addresses these issues to various degrees, as we will see in Chapters 5 and 8.

2.3 Gateways

In looking for a way to parse local environmental structure in order to abstract places and
paths, we came up with the idea for gateways. Gateways are intended to delineate the area
belonging to a place from the paths or doorways that lead away from the place. Upon
looking at relevant published work, we realized that the idea of a gateway has been “dis-
covered” (in some form or another) multiple times [Yeap, 1988; Chown et al., 1995; Thrun
and Bücken, 1996]. Below we discuss the history of gateways, detail previous attempts at
gateway implementations, and discuss their pros and cons.
As part of this dissertation, we provide our own implementation for grounding gate-
ways in the kinds of metrical models easily obtained by modern robotic sensors. Our ap-
proach improves on the previously published gateway approaches in that our algorithm is
meant to be environment independent, resistant to noise, resilient to scaling parameters,
and effective in both corridor-following and wall-following situations. This is detailed in
Chapter 6.

2.3.1 The Foundations of Gateways

Chown et al. [1995] propose a cognitive model of large-scale space that tried to unify the
network of object landmarks proposed by Siegel and White [1975] with visual scenes. Ad-
ditionally, they argue for using a neurally plausible associative network platform to asso-
ciate visual scenes with locations in the world. In this work, they recognize that the object-
landmark-based cognitive models that assume open spaces, dotted with beacon landmarks,
cannot handle certain common situations: doorways, entrances to caves, exiting forests,

34
and other scenarios where large visual changes occur. They propose that gateways marked
these types of places, and they note that gateways provide a structural-based definition of a
landmark.

In buildings, these [gateways] are typically doorways. . . . Therefore, a gateway

occurs where there is at least a partial visual separation between two neigh-
boring areas and the gateway itself is a visual opening to a previously obscured
area. At such a [location], one has the option of entering the new area or staying
in the previous area. [Chown et al., 1995, p. 32]

In outdoor environments gateways might include mountain passes, bridges, or

openings in a forest. Gateways are characteristically [locations] where people
stop to pause and look around. This may be because they occur at choice points,
such as intersections. . . . In their own way gateways are landmarks defined by
structural properties. [Chown, 2000, p. 2]

Chown [1999] further compares the traditional object-landmark-based models to ones that
utilize local maps (e.g., the local perceptual map of the HSSH).

Because the local map is invested in stronger assumptions about environmental

structure than the landmark structure, it is less likely to strongly apply to [an
arbitrary] environment, but more likely to yield significant payoff when it does.
The local map structure relies on some type of natural gateways. . . . Gateways
are useful because they mark transitions from one [subset of an environment]
to another.. . . Cognitively this allows the world to be broken up into smaller
pieces.. . . Another important attribute of an effective gateway [is] that it is both
an exit and an entrance. [Chown, 1999, p. 28]

35
2.3.2 Gateway implementations

Despite several publications that discuss the concept of gateways, Chown and colleagues
never explicitly discuss an algorithm or give a precise formal description of gateways. This
has led to several implementations over the years, including ours, which we discuss in
Chapter 6. Figure 2.1 at the end of this chapter, compares the abilities of each of these
approaches.

Sensor-based Gateways

A simple way to detect a gateway is to look for [locations] where the perceived
background distance suddenly changes. For example, when a vista that had
been occluded suddenly opens up. [Chown, 2000, p. 2]

The above quote is the closest description of a gateway algorithm given by Chown and col-
leagues. This approach, of looking for disparities in the range estimates to nearby obstacles
is utilized by several researchers investigating indoor hallway navigation [Kortenkamp and
Weymouth, 1994; Schröter et al., 2004].
A disparity approach relies on the measured distances from the robot’s current lo-
cation. Consequently, the approach can yield different results depending on the robot’s
location with respect to occlusions in the environment. This is an undesirable property, as
we would like stable place descriptions, independent of the robot’s location in the place.
Additionally, hallways of different sizes, or maps with different parameters (e.g., cell size),
demand different distance metrics, which affects the robustness of such implementations.
The approach we take in Section 6.3.2 is an attempt at a scale and location invariant algo-
rithm that overcomes these issues.
A disparity-based implementation for a mobile robot was proposed by Kortenkamp
and Weymouth [1994]. Their robot used sonar sensing to perform wall-following in a rec-
tilinear environment. The gateway implementation is quite simple in that the robot simply
looks for walls on the left or right to “disappear” or a wall to “appear” in front of it. Left,

36
right, and front are defined with respect to the orientation of the wall the robot is following.
Once a wall disappears or appears, the robot determines it is at a gateway. In this particular
implementation, a gateway is basically equivalent to a distinctive state of the Spatial Seman-
tic Hierarchy, as it corresponds to being at a particular place and facing along a particular
path. This implementation has the drawbacks mentioned above, along with the additional
limitation of only handling rectangular environments.
The most recent work on gateways, outside of our own, has been by Schröter et al.
[2004]. The authors discuss a lidar-based gateway implementation that uses what they call
virtual line models (VLMs). Like the approach above, VLMs are “based on the assumptions
that environments are rectangular and [all] hallways have the same width” [Schröter et al.,
2004]. The idea is quite straightforward, in that only L, T, and + intersections, open door-
ways, and dead-ends need to be modeled. The algorithm has a model for the left wall, right
wall, forward wall (orthogonal to the left and right walls), and a hidden wall (parallel to the
forward wall, but often unseen due to forward-facing lidar sensors). It fits laser data to this
model using Expectation-Maximization (EM) to hypothesize lines and preserve right an-
gles. It then looks for hallway-sized gaps in the line models where free space exists. These
gaps are the gateways. Given the size and configuration of gateways, the robot determines
the type of place.
For example, when traveling down a long corridor, the laser data fits the left wall
and right wall data well, but no data fits the forward wall or hidden wall data. Thus the
robot can hypothesize it is not near a place. As it approaches a + intersection, it begins to
see the far wall of the orthogonal hallway. This data fits the forward wall line model, and
the range disparity between the end of the left and right walls and the forward wall explains
the hidden wall. Thus the robot hypothesizes 4 gateways that form the square where the
two hallways meet. Smaller gaps are used for doorways. This method falls short of the
generalized local topology abstraction presented in Chapter 5, as it assumes specific sizes
and angles of hallways and can degrade in performance when obstacles clutter the range

37
measurements.
In his thesis, Schröter [2006] extends this gateway implementation by using a col-
lection of recent lidar returns. This is an attempt to overcome inaccurate gateway detections
due to clutter and other occlusions that affect fitting data from a single robot pose to the vir-
tual line models. This method still does not always detect gateways correctly, even with the
assumption of right angle intersections and uniformly sized hallways. More importantly,
his thesis examines visual detection of closed doors to hypothesize gateways unseen by
lidar sensing. He also investigates visual furniture detection to ignore non-architecture in
spatial reasoning. Adding in such visual object detection can only improve the gateway
methods discussed here, as well as the lidar generated LPMs that facilitate our gateway
implementation.

Model-based Gateways

In an attempt to overcome the problems of robot-centered, sensor-based approaches. There

has been some investigation into finding gateways in metrical models built from range sen-
sors. Youngblood et al. [2000] outline a gateway implementation that looks for openings
of a certain size in an occupancy grid. They define gateways as line segments connecting
occupied cells that are separated by free cells by more than a minimum distance but less
than a predetermined maximum distance. This finds gateways at doorways in their domain.
Gateways at intersections cannot be found using this method as hallways generally have a
reasonably uniform width throughout, including when they begin (or end) at an intersec-
tion. Like the implementations discussed above, this approach also suffers from needing
environment dependent thresholds.
Thrun and Bücken [1996] give an implementation that also looks at distances be-
tween occupied cells in a grid; however, their approach is a bit more general in that they
look for local minima in distances instead of looking for distances of a predetermined size.
This implementation takes a global metrical map of an environment, generates a Voronoi

38
skeleton, and traverses the skeleton looking for “critical points”, where the distance from a
point on the skeleton to the nearest obstacle is smaller than the corresponding distances of
the neighbors on the Voronoi skeleton.
At critical points, the algorithm connects the skeletal point to the nearest obstacle
on each side of the skeletal branch. The authors call these “critical lines”, though in general
a Voronoi point and the two closest obstacles that define it are not necessarily collinear—
making a line segment between the two obstacles seems sufficient. This approach works
fairly well in indoor, rectangular environments where local minima occur in doorways and
at the edges of intersections. Complex intersections often contain multiple Voronoi junc-
tions; thus, there can be many local minima along the Voronoi skeletons of these regions.
In these scenarios this method can create strange critical points and critical lines in complex
intersections.
This approach motivates our constriction-based implementation discussed in Sec-
tion 6.2, and works well in indoor, straight-walled environments, especially after noise is
smoothed out of either the grid or the Voronoi graph distances to the closest obstacles.
Our approach handles complex intersections by only looking for critical points (or con-
strictions) near the edges of place neighborhoods. The extent of place neighborhoods are
approximated by our concept of a core of a local region (Section 6.1.2).

Occlusion-based Gateways

Chown [1999] cites the earlier work of Yeap [1988] as partial inspiration for their concept of
gateways. Yeap [1988] proposes a cognitive mapping framework that relies on local maps
connected together in a topological network. The local maps, are defined as groups of line
segments that depict walls (they explicitly state that they ignore the issue of obstacles such
as furniture), with openings between walls marked as exits. These exits are very similar to
gateways. The approach here is similar to the sensor-based approaches of Kortenkamp and
Weymouth [1994] and Schröter et al. [2004], but no explicit assumptions about rectangular

39
intersections or fixed-width hallways are needed.
Yeap [1988] calls the local model of the robot’s surround an absolute space rep-
resentation (ASR). ASRs are generated by first extracting line segments, representing ver-
tical surfaces in the environment, from collections of experience in a local region. Given
line segments of a local region, the algorithm then hypothesizes the true boundary of en-
closed environments, like rooms, by attempting to connect radially neighboring endpoints
of the same type: “occluding” versus “occluded” endpoints. By connecting “neighboring”
endpoints (preferably of the same type), ASRs determine the boundary of a room by the
collection of walls and connections between endpoints. Connections that extend across free
space (or across disparities caused by occlusions) are related to our concept of gateways:
e.g., connected “occluding” endpoints almost always correspond to doorways in rooms.
There are several problems with this method, some discussed explicitly by Yeap
[1988]. Among these are possible non-symmetric ASRs in symmetric environments due to
the radial clustering technique utilized in looking for neighbors. Second no notion of what
determines a “neighbor” is given, but presumably this is some angular threshold. Yeap and
Jefferies [1999] overcome these problems by introducing a new set of rules for connecting
endpoints.
Instead of performing a radial sweep to look for neighboring endpoints of the same
type, Yeap and Jefferies [1999] explicitly look to connect each “occluding” endpoint to
another appropriate endpoint to create an exit. Most often, the closest other endpoint (in
Euclidean distance) is connected to an occluding endpoint. The new rules they use for
connecting endpoints fix the non-symmetry problems and eliminate “neighbor” thresholds.
This implementation defines exits at doorways and at intersections quite nicely, as doorways
create occlusions, and usually, the two door jambs are the closest obstacles to each other
(though counterexamples could be created that break this algorithm). As ASRs are meant
to represent local, non-overlapping space, only the minimal set of exits and walls are kept,
throwing out exits and walls that can only be reached by traveling through another exit.

40
 This new ASR approach by Yeap and Jefferies [1999] still has some disadvantages
that are inherited from the ASR approach of Yeap [1988]. First, there must be a metric
for deciding what constitutes an “occlusion”. This is the same problem mentioned above,
where disparity metrics may be environment dependent. Again we see that the ASR shape
is highly dependent on the robot’s location in the local region, as occlusions are defined by
a “virtual” 360◦ range sensor from the robot’s position.
More importantly, because a new ASR is only computed when the robot moves
through an exit (gateway) of a previous ASR, different starting points in a large-scale en-
vironment, and different exploration paths through the environment, can lead to drastically
different collections of ASRs that describe the same environment. Having non-deterministic
place abstractions is not a very good starting point for a topological framework—this presents
a problem for recognizing places, labeling places, etc.
Finally, there are specific, important places that may not be modeled at all. Suppose
the robot starts building an ASR in the middle of a +, L, or T intersection, where there
are no occlusions. In this situation, no endpoint connections (thus no exits/gateways) exist
near the intersection. This would mean two long, orthogonal paths can be grouped into one
large, complex place neighborhood. Such a place is contrary to the human defined places
used in English directions for hallway environments [MacMahon, 2007]. If the robot is just
outside the actual intersection when it begins creating the ASR, there would be an occlusion
(due to the “hidden” wall described by Schröter et al. [2004]). Here a gateway would be
defined across the path segment at the intersection (i.e., connecting the left wall and right
wall). Again, this is an example of having two drastically different place abstractions, due
to slightly different starting locations when exploring the same environment. This non-
deterministic discretization of an environment is not ideal.
Another slight problem discussed by Yeap [1988] is that two occluded endpoints
can form a connection (gateway) behind an occluding obstacle. This “gateway” may not
actually exist in the real world, as a wall could be hidden behind the occlusion. This problem

41
with occlusions is one encountered in any gateway algorithm, including ours. The simplest
solution we have found is to have the robot inspect the regions near proposed gateways
before finally accepting their existence. This is discussed in Section 5.3.2.

Summary

Table 2.1 summarizes the abilities of previous gateway implementations, along with our im-
plementations detailed in Chapter 6. The table compares various gateway implementations
by their properties: gateways are independent of robot pose; gateways are independent of
the size of paths; gateways have stable locations at places; gateways can be defined for arbi-
trarily shaped intersections; gateways are defined at intersections and doorways; gateways
apply in wall-following, coastal navigation, scenarios.
As the table indicates, the methods of Yeap [1988] and Yeap and Jefferies [1999]
work poorly if the robot starts looking for gateways when it is in the middle of intersec-
tion. The algorithm of Thrun and Bücken [1996] does not handle complex places where
multiple Voronoi junctions (thus many “critical points”) exist; thus, multiple gateways can
be placed inside complex shaped intersections or rooms. The method of Schröter et al.
[2004] can potentially handle coastal navigation in rectangular environments, though this is
never discussed by the authors. Section 6.2.2 shows cases where our initial implementation
for detecting gateways can produce somewhat unstable gateways. Our final anchor-based
solution to detecting gateways fulfills all the desired properties mentioned here.

42
 Pose Scale Stable Non- Doorways Coastal
rectangular
independent independent locations intersections and
intersections navigation
Yeap no no no yes sometimes no
[1988]
Kortenkamp and
Weymouth no no yes no yes yes
[1994]
Thrun and
Bücken yes yes yes no yes no
[1996]
Yeap and
Jefferies no no no yes sometimes no
[1999]
Youngblood
et al. yes no yes no no no
[2000]
Schröter
et al. no no yes no yes possibly
[2004]
Beeson:
Section mostly yes mostly yes yes no
6.2
Beeson:
Section yes yes yes yes yes yes
6.3

Table 2.1: Comparison of gateway implementations. Pose independent: gateways at places are not
determined by robot’s location (uses a local model of the surround); Scale independent: method
attempts to overcome high numbers of user-defined thresholds and metrics about door width, hall-
way widths, etc.; Stable locations: theoretically (ignoring noise), gateways are located in the same
locations upon different explorations of the same environment; Non-rectangular environments:
reliably handles both non-rectangular (e.g., 5-way or K) intersections and non-rectangular rooms;
Doorways and intersections: gateways are (theoretically) found near both hallway intersections
and doorways; Coastal navigation: gateways can be determined along non-corridor paths and
places, e.g., following walls, the outside of buildings, or the edge of a forest.

43
Chapter 3

The Spatial Semantic Hierarchy

Maps encourage boldness. They’re like cryptic love letters. They make anything seem possible.
Mark Jenkins, To Timbuktu: A Journey Down the Niger (1997)

This chapter describes the Spatial Semantic Hierarchy (SSH), a framework for representing
large-scale space at multiple levels of abstraction, and introduces its extension, the Hybrid
Spatial Semantic Hierarchy (HSSH) [Beeson et al., 2003; Kuipers et al., 2004]. We give a
formal description of the multiple levels of knowledge of the SSH as they pertain to map-
building (the most difficult navigation task). This formal description is an abridged version
of the logical formalism given by Remolina [2001], with a few slight changes that allow
us to make logical connections to the spatial concepts that are introduced via the HSSH.
We then discuss some of the limitations of the basic SSH approach, which motivates the
need to integrate small-scale space models that will benefit all levels of the SSH framework
(Chapters 4 through 9).

3.1 Hybrid SSH Overview

Most metrical approaches to mobile robot map-building define a single, global frame of
reference in which to create the map. Range measurements are used to perform proba-

44
bilistic inference about the location of features or about the occupancy of discretized cells
in the map. Existing SLAM (simultaneous localization and mapping) methods are highly
effective for building local metrical models of small-scale space and for providing reliable
localization in the frame of reference of the local map; however, maintaining global consis-
tency over large-scale environments is difficult, particularly when closing large loops in the
environment. A popular approach is to use particle filters, where each particle represents a
hypothesized exploration trajectory. The researcher must hope that with enough particles
the distribution will include one that closes the loop correctly [Thrun et al., 2005]. Since
the space of trajectories can be enormous, this hope is often optimistic.
The fundamental problem is representational: loop-closing hypotheses are alterna-
tive topological structures for the map, not alternative metrical structures. To be able to
solve complex, multi-hypothesis loop-closing problems in a tractable manner, the robot
must reason with symbolic topological maps. The space of metrical maps in a single frame
of reference does not appropriately represent the states of incomplete knowledge that arise
during exploration and map-building in complex, large-scale environments.
Our factored mapping framework is based on the Spatial Semantic Hierarchy (SSH)
[Kuipers, 2000, 2008], which uses multiple coordinated representations for knowledge of
large-scale space. The Hybrid SSH (HSSH) extends the basic SSH by including representa-
tions for small-scale space and defining the relationship between large-scale and small-scale
spatial representations. Symbolic topological mapping methods such as the SSH provide a
concise representation for the structural alternatives that arise in investigating loop closures.
Topological maps provide the ability to store and access multiple local maps with separate
frames of reference and topological connections annotated with weak metrical constraints.
By separating small-scale from large-scale space, we postpone the problem of coordinating
the local frames of reference until the global structure of the topological map has been iden-
tified. At that point, the global metrical map can be constructed, efficiently and accurately.
Therefore, our approach factors the mapping problem into four natural sub-goals:

45
(1) building a metrical representation for local small-scale spaces; (2) detecting places and
determining their symbolic descriptions; (3) finding a topological map representing the
qualitative structure of large-scale space; and (4) constructing a metrical representation for
large-scale space in a single global frame of reference, building on the skeleton provided
by the topological map. While the global metrical map is useful for some purposes, it is
worth noting that many autonomous planning and navigation goals can be achieved effec-
tively using only the global topological map and/or the local metrical maps. Therefore, this
approach to hybrid mapping is more robust than one that extracts topological relations from
a global metrical map that must be built first [Thrun and Bücken, 1996].
The multiple representations of the HSSH are described independently, while their
semantic dependencies imply that they build on each other. However, this does not imply
a simple serial processing pipeline. In fact, processing of sensory input to build represen-
tations of the different kinds is interleaved, providing various sorts of synergies. Two are
particularly important. First, the local metrical map of small-scale space is a useful “ob-
server” both for detecting and describing places and for low-level control with obstacle
avoidance. And second, metrical knowledge of relative displacement and then layout of
places may be useful for ordering candidate topological maps. Nonetheless, to clarify the
distinct representational ontologies, we often describe them in this dissertation as though
they operate independently.
The Hybrid SSH improves mobile robot capabilities in a variety of ways—efficient
and robust map-building and navigation, “natural” human-robot interaction due to the mul-
tiple representations of space (Chapter 7), and hierarchical control. This thesis describes the
HSSH theory and demonstrates key points of HSSH map-building using a particular imple-
mentation that focuses on perception using range sensors, though other sensory modalities
can also be utilized in the HSSH framework [Murarka et al., 2006]. The hybrid, hierarchical
framework is largely independent of the sensors used to create the local metrical model of
small-scale space.

46
3.2 The Basic SSH

The Spatial Semantic Hierarchy (SSH) [Kuipers, 2000, 2008] represents knowledge of
large-scale space with four distinct representations. Figure 3.1 illustrates the framework.
At the SSH Control Level, control laws provide reliable motion among distinctive states
(dstates) qi . At the SSH Causal Level, state-action-state schemas hq, a, q0 i explain how
the distinctive states are linked by turn and travel actions, and relations o(q) = v between
a state and its observable view describe the potential experiences of the robot. Thus the
Causal Level abstracts the continuous world to a deterministic finite automaton [Rivest and
Schapire, 1989; Dean et al., 1995], related to the way humans utilize route instructions in
navigation. At the SSH Topological Level, a map consisting of places, paths, and regions,
describes the connectivity, order, containment, and boundary relations of large-scale en-
vironments. At the SSH Metrical Level, local metrical information about the location of
obstacles, the magnitudes of actions, the lengths of path segments, and the directions of
paths at place neighborhoods are incorporated into local and global metrical maps. One
contribution of the Hybrid SSH is to clarify the relation between the metrical information
and the symbolic abstractions of the basic SSH levels.
The Spatial Semantic Hierarchy factors spatial uncertainty into distinct compo-
nents, controlled in distinct ways. Movement uncertainty is controlled by the behavior
of feedback-driven motion control laws. Pose uncertainty is controlled in the basic SSH by
hill-climbing to dstates (and in the Hybrid SSH by incremental localization within a local
metrical map). Structural ambiguity about the large-scale topology of the environment is
controlled by search in a space of alternative topological maps. Global metrical uncertainty
is controlled by relaxing metrical information from separate local frames of reference into
a single global frame of reference, guided by the topological map.

47
 Qualitative Quantitative
Continuous Analog
Attributes Model

Sensory names Sensor values

Control laws Local 2−D

Control HC and TF geometry
Distinctive states

Views Turn angle

Causal Actions Travel distance
Causal schemas

Places Local headings

Topological Paths 1−D distances
Connectivity
Order

Metrical Global 2−D

geometry

Figure 3.1: The Spatial Semantic Hierarchy. Closed-headed arrows represent dependencies; open-
headed arrows represent potential information flow without dependency. (From Kuipers [2000].)

48
3.2.1 The SSH Control Level

The SSH Control Level describes the system consisting of the agent and its environment as
a piecewise continuous dynamical system. The agent’s experience is represented at a fine-
grained sequence of time-steps 0 ≤ t ≤ N. At any time t, the agent-environment system is
described by the state vector xt (the agent’s pose in a static world), the agent’s sense vector
zt , and its motor vector ut . We assume that both the environment and the agent’s sensory
system are very rich, so the sense vector zt is very high-dimensional.
The dynamical system is described by the following equations, in which the func-
tions F and G represent the physics of the agent’s body in the environment and its sensori-
motor system respectively. These two functions are not explicitly known or available to the
agent. The control law Hi , on the other hand, can be selected by the agent.

xt+1 = F(xt , ut )

zt = G(xt )

ut = Hi (zt )

The agent acts by selecting a control law Hi to determine its motor output signals
as a function of its sensor input. In the basic SSH [Kuipers, 2000], motion is controlled
by alternating between two types of controllers. Trajectory-following control laws take the
robot from one distinctive state (dstate) to the neighborhood of another. A hill-climbing
control law guides the robot to the destination dstate x̄ from anywhere in its surrounding
neighborhood.
Hill-climbing localizes the agent by moving it reliably to a distinctive state within
the local neighborhood, preventing the accumulation of position error, and paving the way
for a discrete abstraction of the continuous space. Furthermore, hill-climbing control makes
very weak assumptions about the properties of the sensors and the agent’s knowledge of
those properties; however, hill-climbing control laws can be difficult to define and may

49
vary across domains. An agent often does have stronger knowledge of the properties of its
sensorimotor system, and physical motion to distinctive states seems awkward and unnec-
essary in light of that knowledge. A key insight behind the Hybrid SSH is that accurate
localization in the small-scale space model of a place neighborhood can substitute for the
physical motion of hill-climbing to a particular distinctive state in that neighborhood. In
Chapter 4, we discuss how the Hybrid SSH exploits metrical knowledge of small-scale
space to build local perceptual maps of place neighborhoods, within which localization is
reliable and effective.

3.2.2 The SSH Causal Level

Given pairs of trajectory-following (TF) and hill-climbing (HC) controls that represent mo-
tion between neighboring dstates at the Control Level, we begin to represent the robot’s
experiences as a set of symbolic abstractions (Figure 3.2). First, we define an action a ∈ A,
to represent a pair of trajectory-following and hill-climbing controls that connect dstates.
Since the sensory image at a dstate z̄ = G(x̄) is a point in a very high-dimensional space,
it will, in general, never be experienced twice. We will therefore assume that each dis-
tinctive state x̄ has an associated view, o(x̄) = v ∈ V , which is an abstracted description of
the sensory image z̄. Kuipers and Beeson [2002] describe a bootstrap-learning method for
obtaining a view representation suitable for high-performance place recognition. For this
thesis, we will not require that the observation function o be learned autonomously, but only
that it exists.
The SSH Causal Level describes the agent’s experience as a deterministic finite
automaton (DFA) [Rivest and Schapire, 1989; Dean et al., 1995]. The Causal Level deter-
ministic finite automaton
MC = hQ, A,V, R, oi

consists of sets of states Q, actions A, observable views V , a transition function R : Q × A →

Q, and an observation function o : Q → V . As the robot travels from one distinctive state x̄

50
Figure 3.2: SSH Control/Causal Level abstraction. In the SSH, dstates are defined by pairs of
trajectory-following and hill-climbing control laws. These sequences are abstracted into actions,
and the observations at dstates are abstracted into views. (Adapted from Kuipers [2000].)

to the next, its experience is an alternating sequence of views and actions. Some actions are
turns, while others are travels.

v0 a1 v1 a2 v2 · · · vn−1 an vn

At the SSH Control Level, a view vi is experienced only when the agent is at a distinctive
state x̄i , so the view vi is an observable manifestation of the distinctive state: vi = o(x̄i ).

x̄0 a1 x̄1 a2 x̄2 · · · x̄n−1 an x̄n

| | | | |
v0 v1 v2 · · · vn−1 vn

At the Causal Level, each state q ∈ Q represents an equivalence class of distinctive states
x̄ in the physical world.1 Two distinctive states x̄i and x̄ j are equivalent if they represent
different experiences of the same distinctive state q ∈ Q. (We use the notation [x̄i ] = [x̄ j ] = q
for this.) The set Q of distinctive states thus represents a specific hypothesis about which
1 The term distinctive state, abbreviated dstate, is thus overloaded. It refers both to the state x̄ resulting from

a hill-climbing control law at the SSH Control Level, and to the state q = [x̄] at the SSH Causal Level which is
part of the discrete abstraction of the continuous environment. It is this abstraction from continuous to symbol
that facilitates causal/topological mapping in the basic SSH.

51
experiences x̄i represent repeated encounters with the same state q in the environment; that
is, Q specifies data association for loop closures.
All distinctive states in the same equivalent class q must have the same view.2

[x̄] = [x̄0 ] → o(x̄) = o(x̄0 )

Thus, o(q) is well-defined, and we can write

q = q0 → o(q) = o(q0 )

The full sensory input from high bandwidth sensors in a realistically complex environment
is so rich that sensory images will never match exactly. Views must be defined in terms of
some observation function that allows the same dstate to be reliably detected on separate
occasions. Thus, an experience with repeated states such as

q0 a1 q1 a2 q2 · · · q0 an q1
| | | | |
v0 v1 v2 · · · vn−1 vn

can only be consistent if vn−1 = v0 and vn = v1 . However, abstracted views are subject
to perceptual aliasing (different places look the same), leading to ambiguities about the
topological structure of the map: o(q) = o(q0 ) 6→ q = q0 .
The transition function R : Q × A → Q is represented as a set of schemas r =
hq, a, q0 i, where context(r) = q, action(r) = a, and result(r) = q0 . As new observations
are added to the robot’s experience, new schemas h[x̄n ], an+1 , [x̄n+1 ]i are learned by the tran-
sition function R. The causal map is constructed by searching for an appropriate set Q of
2 The axioms provided here describe the nature of the spatial knowledge represented at each SSH level, but

we omit auxiliary axioms required for logical completeness (e.g., unique names axioms, etc). A complete set of
axioms is provided by Remolina and Kuipers [2004]. For clarity and conciseness, we use a typed logic in which
variable names encode their types, and we assume that all free variables in axioms are universally quantified.

52
states (i.e., equivalence classes of distinctive state observations), such that MC has a de-
terministic transition function R, predicted and observed views are consistent, and MC is
consistent with the axioms for topological maps [Remolina and Kuipers, 2004].
For the purpose of building the SSH Causal Level from exploration experience,
building and using a DFA is far more tractable than building and using a probabilistic state
model, such as a hidden Markov model (HMM) (as discussed in Section 2.2.3). The key
benefit of a DFA over HMMs (or stochastic finite automata in general) are that both the
transition function and the observation function are deterministic. The deterministic tran-
sition function follows from the nature of the abstraction that results from moving reliably
between dstates via TF and HC control laws. The deterministic observation function fol-
lows from the abstraction that defines the observation function o. One improvement of the
HSSH over the SSH is that local small-scale space models make place detection and ob-
servational classification of states deterministic without the need for hill-climbing (Section
5.1).
The effect of the SSH hill-climbing (and HSSH place detection and localization) is
that the Causal Level representation can assume that actions are deterministic. The deter-
minism of the observation function rests on the abstraction from sensory images to views
being sufficiently aggressive to eliminate perceptual variability. Although observations
are deterministic, they are not necessarily unique since there may still be perceptual alias-
ing. This ambiguity is handled by creating multiple hypotheses of topological (thus causal)
models, as explained in Section 3.2.3. In general, it is not possible for a robot to recover
the complete spatial structure of any arbitrary environment [Dudek et al., 1991]; therefore,
keeping around the tree of possible maps allows the robot to continue navigation when the
best hypothesis is refuted by an experienced counter example.3
3 Long-term experience with the HSSH has yielded deterministic actions and views in research settings, but

we can envision rare scenarios, e.g., an intersection crowded with people, that could lead to an undetected or
misclassified place. Detecting and understanding these unusual events should allow us to still assume deter-
ministic actions (100 − ε)% of the time.

53
3.2.3 The SSH Topological Level

In the SSH, a topological map is an instantiated model for two sets of axioms: one that de-
scribes topological maps in general and another that describes the exploration experience of
the agent in a particular environment. We identify the global topological map by generating
potential models of these axioms, discarding those that violate the axioms, and applying an
ordering on the remaining ones so that a single best model can be selected. If there is no
single best model, then a few closely competing models can be identified and can be used
to make an exploration plan to help discriminate between models.
The SSH Topological map M T describes the environment in terms of dstates, places,
paths, regions, and the qualitative relations among them such as connectivity, order, and
containment. A dstate q ∈ Q represents a distinctive state or pose of the agent in the envi-
ronment, a place p ∈ P represents a zero-dimensional location, a path π ∈ Π represents a
one-dimensional structure, and a region r ∈ R represents a two-dimensional subset of the
environment. In this dissertation, we will not discuss regions or their relations, which are
described by Remolina [2001].
We formalize a topological map as

M T = MC ∪ Ob jects ∪ Relations.

Here Ob jects = hP, Π, Ri, where P is a set of places, Π is a set of paths, and R is a set of
regions. M T thus includes (via MC ) the sets of states, actions, and views. Relations encodes
the relations over this set of objects. These relations, including at, along, on, order (and
the “star” relations that describe the local topology of places), allow for a richer description
of the connectivity of places and paths, and are introduced below as needed.
At the SSH Causal Level, the experience is represented as an alternating sequence
of states (qi ∈ Q) and actions (a j ∈ A).

q0 a1 q1 a2 q2 · · · qn−1 an qn

54
At the Topological Level, each distinctive state q ∈ Q corresponds to being at a place, and
facing along a path in some direction. Since a path is one-dimensional, it has two directions
d ∈ {+, −}, for which opp(+) = − and opp(−) = +. We define a directed path, π d , to
represent facing along a path in a particular direction.

∀q ∈ Q ∃p ∈ P, π ∈ Π, d ∈ {+, −} [at(q, p) ∧ along(q, π, d)] (3.1)

along(q, π d ) ≡ along(q, π, d)

Additionally, there are two kinds of basic actions, turns and travels, and there is a TurnAround
action.
A = Turns ∪ Travels TurnAround ∈ Turns

A place p ∈ P corresponds to a set of states linked by turn actions.

hq, a, q0 i ∈ S ∧ a ∈ Turns ∧ at(q, p) → at(q0 , p) (3.2)

Similarly, a path π ∈ Π corresponds to a set of states linked by travel actions, or by a

TurnAround.

hq, a, q0 i ∈ S ∧ a ∈ Travels ∧ along(q, π, d) → along(q0 , π, d) (3.3)

hq, a, q0 i ∈ S ∧ a = TurnAround ∧ along(q, π, d) → along(q0 , π, opp(d)) (3.4)

The relation on(π, p) means that the place p ∈ P is on the path π ∈ Π.

at(q, p) ∧ along(q, π, d) → on(π, p) (3.5)

A path defines an order relation over the places on it:

hq, a, q0 i ∈ S ∧ a ∈ Travels ∧ at(q, p) ∧ at(q0 , p0 ) ∧ along(q, π, d) → order(π, d, p, p0 ) (3.6)

55
 Causal schema Equation # Topological relation

hq1 ,travel, q2 i ∈ S (3.1) → at(q1 , p1 ), along(q1 , πa+ )

(3.5) → on(πa , p1 )
(3.3) → along(q1 , πa+ ), along(q2 , πa+ )
(3.1) → at(q2 , p2 )
(3.5) → on(πa , p2 )
(3.6) → order(πa+ , p1 , p2 )
(3.8) → p1 6= p2

hq2 ,turn, q3 i ∈ S (3.2) → at(q3 , p2 )

(3.1) → along(q3 , πb− )
(3.5) → on(πb , p2 )

hq3 ,travel, q4 i ∈ S (3.3) → along(q4 , πb− )

(3.1) → at(q4 , p3 )
(3.5) → on(πb , p3 )
(3.6) → order(πb− , p2 , p3 )
(3.8) → p2 6= p3

Figure 3.3: Topological abduction example. Here we illustrate the abduction process, using the
topological axioms to model exploration. Starting at dstate q1 , the agent reaches dstate q2 at a place
p2 having traveled along directed path πa+ . It then turns to dstate q3 , still at place p2 , and is ready to
travel along another path, say πb− , from q3 to dstate q4 at some other place.

order(π, d, a, b) → on(π, a) ∧ on(π, b) (3.7)

¬order(π, d, p, p) (3.8)

order(π, d, a, b) ⇐⇒ order(π, opp(d), b, a) (3.9)

order(π, d, a, b) ∧ order(π, d, b, c) → order(π, d, a, c) (3.10)

In order to create a Topological Level map from a Causal Level experience, such as hq1 ,travel, q2 i,
hq2 ,turn, q3 i, hq3 ,travel, q4 i, the agent uses abduction to hypothesize the existence of sev-
eral places and paths at which these distinctive states occur. Figure 3.3 shows an example
of the abduction process.
Remolina and Kuipers [2004] provide a non-monotonic axiomatization of the SSH
topological map, including additional elements of the theory (regions, boundary relations,

56
and metrical relations), along with more details and motivating examples. This theory pro-
vides a precise specification of the possible logical models (topological maps) that are con-
sistent with the axioms and the sequence of actions and views observed while exploring.
A prioritized circumscription policy (expressed as a nested abnormality theory [Lifschitz,
1995]) specifies how distinct consistent logical models are ordered by simplicity. Further-
more, Savelli and Kuipers [2004] have developed the non-local planarity constraint, which
enforces the requirement that a topological map is a graph embedded in the plane. Algo-
rithm 3.1 presents the process for constructing all possible topological maps by generating
all possible sets Q.

3.2.4 The SSH Metrical Level

The SSH, often thought of as a framework for creating purely topological maps, has always
allowed for local metrical knowledge to be utilized at the Control Level (Figure 3.1, right
column). Additionally, the SSH Metrical Level has always supported a global metrical map
to be created after the topological map—it is our belief that such a global metrical map
is often unnecessary for navigation in and communication about large-scale environments.
However, the SSH theory has lacked a formal description of exactly how metrical informa-
tion influences the hierarchical abstractions of space. One contribution of this thesis is to
clarify the relationships between metrical and symbolic knowledge in a navigational agent.
In work leading to the development of the SSH, Kuipers and Byun [1991] created
a “patchwork metrical map”. Their mapping implementation annotated topological places
and paths with metrical data gathered during exploration. Given a topological map hypoth-
esis, the global place layout was relaxed to minimize errors with respect to the annotated
metrical data before adding stored range information to create the obstacle map. This ap-
proach is similar to the techniques we define mathematically in Chapter 9.

57
 0. Perform initial action a0 that brings the robot to a place along a directed path.
Initialize the tree of maps with the map hypothesis hM0 , q0 i, where M0C contains the
single dstate q0 with its observed view v0 , and M0T contains the single place p0 and
path π0 .

After performing a new action a and observing the resulting view v, for each consistent
map hM, qi on the fringe of the tree:
1. If MC includes hq, a, q0 i in R and v0 = o(q0 ),
• if match(v, v0 ), then hM, q0 i is the successor to hM, qi, extending the tree;
• if not, then mark hM, qi as inconsistent.
2. Otherwise, MC does not include hq, a, q0 i in R. Let M 0 be M extended with a new
distinctive state symbol q0 and the assertions v = o(q0 ) and hq, a, q0 i. Consider the
k ≥ 0 dstates q j in M with v j = o(q j ), such that match(v j , v). Then hM, qi has k + 1
successors:
• hM 0j , q0 i for 1 ≤ j ≤ k, where M 0j is M 0 extended with the assertion q0 = q j .
0 , q0 i, where M 0
• hMk+1 0 0
k+1 is M extended with the k assertions that q 6= q j , for
1 ≤ j ≤ k.
3. Mark a new successor map inconsistent if it violates the axioms of topological
maps.
4. Define a preference order on the consistent maps at the leaves of the tree.

In the Basic SSH:

M = MT .
A view is a simple symbol.
match(v, v0 ) iff v = v0 .
Both a ∈ Turns and a ∈ Travels can reach step 2 and cause a branch.
Preference order from prioritized circumscription policy [Remolina and Kuipers,
2004].
Algorithm 3.1: Building the tree of topological maps in the SSH. This describes the
algorithm for building a tree of all possible topological consistent with a sequence of
actions and observations at discrete places.

58
3.3 Extending the SSH

The Spatial Semantic Hierarchy depends on the assumption that the environment naturally
decomposes into place neighborhoods, connected by path segments, which can then be ab-
stracted to a topological map. That is, it uses the sparse structure of man-made environments
(or man-made paths in natural environments) to define a small number of discrete places
and connecting paths. Obviously, topological structure may be imposed even in unstruc-
tured environments. Defining places at visually distinctive locations along a single path
(e.g., a water tower on the side of a highway) or even based on metrical path-integration in
wide-open spaces (as the Puluwat navigators do when piloting dugout canoes between dis-
tant islands [Gladwin, 1970]) are currently not handled by our SSH hill-climbing controllers
or the HSSH place detection methods. We believe these type of places can be represented
within the SSH framework, but we leave this problem for future work.
The basic SSH makes weak (i.e., very general) assumptions about the sensory ca-
pabilities of the navigational agent; thus, abstraction from continuous sensations to discrete
models of the environment depends on well-crafted control laws that move the robot reliably
between distinctive states. The Hybrid SSH makes stronger (i.e., more specific) assump-
tions about the types of sensors available to the agent, for example, range sensors. This
allows the HSSH to extend the basic SSH by using existing metrical mapping techniques to
create precise observational models of the local surround.
The HSSH has four major levels of representation that correspond to the four SSH
levels (Figure 3.4). At the Local Metrical Level, the agent builds and localizes itself in the
Local Perceptual Map (LPM), a metrically accurate map of the local space within its sensory
horizon. The LPM is used for local motion planning and hazard avoidance. At the Local
Topological Level, the agent identifies discrete places (e.g., corridor intersections, rooms,
etc.) in the large-scale continuous environment, and qualitatively describes the configura-
tion of the paths through the place—its local decision structure. At the Global Topological
Level, the agent resolves structural ambiguities and determines how the environment is best

59
described as a graph of places, paths, and regions. The Global Metrical Level specifies the
layout of places, paths, and obstacles within a single global frame of reference. It can be
built on the skeleton provided by the topological map. Figure 3.4 diagrams the basic flow
of data in the HSSH, from sensors, through the local metrical model of small-scale space
and the global symbolic model of the large-scale environment, finally creating the global
metrical model if desired.
Having small-scale space models of the local surround creates several advantages
when implementing the HSSH versus the basic SSH. First, the robot represents the local
environment using a local perceptual map (LPM) (Chapter 4). The robot can therefore use
algorithms for local metrical motion planning and obstacle avoidance instead of relying on
behavior-based controllers. Second, metrical localization can be done quickly after entering
a place neighborhood, rather than requiring physical hill-climbing to a distinctive pose.
A symbolic description of the local topology is extracted from this precise small-
scale-space model of the local surround via gateways (Chapter 5). Thus, the view of a
distinctive state no longer need be some user-defined function of the perceptual inputs. In-
stead, the method relies on the local topology abstracted from the LPM to describe places,
thus describing all distinctive states at each place. Using local topology to detect and de-
scribe places allows the robot to model more complicated intersections of paths than with
hill-climbing. Additionally, using local topology constrains the global topological model
search (Chapter 8), as branching in the tree of possible maps occurs only when arriving at
a place, not when visiting the various dstates of a place.
Stored metrical information along topological connections between places can be
used to efficiently obtain a global metrical layout of places (Chapter 9), which provides the
“backbone” for a global map if desired. The HSSH also improves navigational behaviors
and facilitates multi-modal human-robot interaction (Chapter 7).

60
 Effectors, Sensors Data Flow:
Control Flow:
Motor
Commands
Hardware Observations
Control
Motion
Commands

Local Metrical
Level

LPM with
LPM Pose
Robot Pose

Local Topology
Level

Place LPM, Small Scale Star

Travel, Turn Last Gateway, Action Type
Path Hazards/Metrical Annotations

Global Topology
Level

Topological Map,
Place LPMs of all Places,
Metrical Annotations
LPM Likelihood of
Pose Global Metrical Place Layouts
Level

Figure 3.4: HSSH description. The HSSH is an integrated framework of multiple, disparate rep-
resentations of spatial knowledge. Each level of abstraction uses its own ontology with concepts
motivated by human cognitive abilities and grounded to the environment via local metrical obser-
vations. The four major components here correspond to the four levels of the basic SSH shown in
Figure 3.1.

61
Chapter 4

Modeling Obstacles and Trajectory:

The HSSH Local Metrical Level

The knowledge of the world is only to be acquired in the world, and not in a closet.
Philip Dormer Stanhope, 4th Earl of Chesterfield, Letters to His Son (1774)

The critical difference between the basic SSH and the Hybrid SSH is the use of a local
metrical model to represent small-scale space surrounding the robot. We call this model
a local perceptual map (LPM). The LPM is currently built using sensor input from laser
range sensors that see walls, but the LPM could be built from visual sidewalk (or road)
detection or other sensor modalities. Similarly, the current LPM representation models
occupied, free, and unknown regions of space. Work by Murarka et al. [2006] investigates
incorporating semantic labels into the LPM to denote drop offs, pedestrians, and other types
of hazards.
This chapter discusses the characteristics and benefits of LPMs in the HSSH. One
important implementation detail is to ensure a quality local map even in the face of slip-
ping wheels, inaccurate odometry, or other factors that can affect localization performance.
Improper localization leads to an incorrect update of the local map, which can possibly
degrade the benefits of the LPM structure throughout the HSSH. We evaluate an adaptive

62
particle filter algorithm for improved localization in the face of odometry errors [Beeson
et al., 2006].

4.1 Local Perceptual Map (LPM)

The local perceptual map (LPM) is a bounded-size metrical description of the small-scale
space surrounding the agent. It functions as an observer, integrating sensor values over time
to determine the locations of obstacles and other hazards, for localization, motion planning,
and the derivation of local features for larger-scale mapping. The LPM represents the small-
scale space within the robot’s sensory horizon, not just what is currently in view. It is small
enough to avoid the problem of closing large loops. The frame of reference of the LPM is
local. Its relation with the world frame may be unknown, and will drift over time due to
accumulating errors.
When the agent travels from one place to another, the LPM acts as a scrolling map,
m̃, that describes the robot’s immediate surround. Information that scrolls off the LPM is
discarded, and new cells that scroll onto the map are initialized as unknown.1 Because the
LPM has a fixed, bounded size, the cost of updating it is constant in both time and space.
The full task of building metrical maps from exploration data can be described as
finding the joint posterior over maps m and trajectories x = (x, y, θ ) in P(x, m|z, u) with the
following symbol definitions.
1 Our rectangular LPM scrolls, horizontally or vertically, as needed to keep to keep the robot’s pose in a

central cell. Information in the occupancy grid is only shifted by integral numbers of cells to avoid blurring the
model by rotations or partial-cell translations.

63
 t The time-steps 0 ≤ t ≤ N of the agent’s experience.
x = x0:N The sequence of agent poses xt at each time-step t.
z = z0:N The sequence of observations zt .
u = u1:N The sequence of actions ut between time-steps.
m The set of map elements, which may be landmarks or occupancy
grid cells. m̃ (the scrolling local perceptual map) is a particular
example of a metrical map m.

The joint probability of the pose history x and the metrical map m can be decomposed as

P(x, m|z, u) = P(m|x, z, u) · P(x|z, u)

by the chain rule for probabilities.

For simple, local regions, the maximum-likelihood map can be estimated incremen-
tally given knowledge of x and z, so we really just need to solve for P(x|z, u). Additionally,
we are not concerned with the full distribution over pose trajectories, as we are updating the
map from the maximum-likelihood pose at each time step. Thus, for our online metrical
mapping we only need to determine the distribution over the current pose.

Bel(xt ) = P(xt |z0:t , u1:t )

Z
= η P(zt |xt , m) P(xt |xt−1 , ut )Bel(xt−1 ) dxt−1 ,

where η is a normalization constant. Figure 4.1 illustrates the basic structure of Markov
localization [Fox et al., 1999], which allows us to determine in an efficient and incremental
algorithm, the distribution of poses that best fit the current map.
Though multiple metrical mapping methods might be used for the LPM, we utilize
the well-known occupancy grid representation [Moravec, 1988; Elfes, 1989], along with
particle filter Markov localization [Fox et al., 1999] to overcome noisy odometry informa-
tion. Stated more plainly, we model the world as a discretized grid, where each cell contains

64
Figure 4.1: Markov localization. The standard graphical dynamic Bayesian network (DBN) for
Markov localization within a single frame of reference: combines belief about actions P(xt |ut , xt−1 )
and observation P(zt |xt , m). Simultaneous localization and mapping (SLAM) algorithms combine
localization, P(x|z, u), with one of a number of mapping methods to estimate P(x, m|z, u).

a probability of being occupied by an obstacle, as measured by a lidar sensor. Localization

is performed by comparing hypothesis poses to the current map, and the map is updated ac-
cordingly. This is a well-known version of simultaneous localization and mapping (SLAM)
[Thrun et al., 2005]. Discussions in this dissertation that refer to this implementation gen-
eralize to many SLAM implementations.

4.2 LPM Benefits

LPMs provide the HSSH with various information that allows both local and global abstrac-
tions of space. In Chapter 5, we discuss how the LPM supports the abstraction of a symbolic
small-scale space description of the local-paths in the surround. Chapter 9 discusses how
the local metrical information is used, along with the topological map, to find the global
metrical place layout of an exploration trace and, if desired, the entire global metrical map
on an explored environment. We forgo these details here; however, the LPM is also useful
for local control at the SSH Control Level.
Given a target pose in the LPM, the robot can compute a trajectory to reach the
target without colliding into obstacles. This can be done using the Vector Field Histogram
[Borenstein and Koren, 1991], the Dynamic Window approach [Fox et al., 1997], gradient

65
methods [Konolige, 2000], or even a simple search (using A∗ or RRTs [Kuffner and LaValle,
2000]) over the cells of the occupancy grid. The potential function (for gradient methods)
or the cost function (for A∗ ) reflects the distance of the agent from an obstacle or other
hazard represented in the LPM. Object tracking may be implemented at this level, but our
current robot implementation simply avoids obstacles by taking the first few steps along the
planned trajectory before replanning. We discuss the selection of target poses as they apply
to Causal Level Travels and Turns in Section 5.4.
In the basic SSH [Kuipers, 2000], an agent localizes itself in a place neighborhood
by hill-climbing to a distinctive state. Localization by physically moving to maximize a
“distinctiveness measure” requires very little knowledge about the nature of the environ-
ment or the sensors. In the Hybrid SSH, on the other hand, the agent uses an online SLAM
method to localize itself unambiguously within the local perceptual map. SLAM methods
depend on stronger knowledge about the relation between sensor input and the agent’s lo-
cation in the local frame of reference—P(zt |xt , m̃). In return for these stronger assumptions,
the agent does not need to move to a particular location to be adequately localized.
Finally, when the agent is in the neighborhood of a particular topological place p, a
snapshot of the LPM m̃ serves as a small-scale space description of the place neighborhood
that is stored as a place annotation m p in the topological map. When a place p is first
encountered, the local map m p for its neighborhood is initialized with the information from
the scrolling map m̃. The frame of reference defined for m p may be different from that of
m̃, appropriate to the characteristics of the place neighborhood. When the neighborhood of
p is encountered on subsequent occasions, the agent may localize itself with respect to the
stored map m p or update m p with the more recent information in m̃.

4.3 Building an Accurate Localization Algorithm

In order to ensure a good quality LPM, we investigate extensions to the traditional particle
filter algorithm for localization during mapping. Accurate localization is key to obtaining

66
a good maximum-likelihood pose estimate before incorporating range measurements into
the map. It is also important to keep a distribution over poses at each time step in order to
ensure proper localization on future SLAM iterations.
The most commonly used paradigm for incremental localization is probabilistic
localization [Thrun et al., 2005]. Here, the robot has some distribution of belief about its
pose x = (x, y, θ ). After performing a low-level control u and making an observation z, it
must then determine its new pose estimate x0 . Particle filters are an efficient approach to
estimating a continuous distribution with a finite set of samples.

Z
p(x0 |z, x, u, m) = η P(z|x0 , m) · P(x0 |x, u)p(x) dx

posterior ∝ likelihood · proposal

When performing probabilistic localization using a particle filter, a robot must have
a good proposal distribution in which to distribute its particles. Once weighted by their
normalized likelihood scores, these particles estimate a posterior distribution over the pos-
sible poses of the robot. We investigate the results of different algorithms that modify the
proposal distribution at each time step in order to obtain more accurate localization [Bee-
son et al., 2006]. Our resulting localization method gives a high quality LPM, but will also
improve localization for any metrical mapping algorithm.

4.3.1 Action Models

The proposal distribution is determined by an action model. The action model is a system
of equations that defines a probability distribution over resulting states x0 , given an initial
state x and an incremental motion u, measured by odometry sensors. Using the observation,
z, from the current location and a map of the local environment, m, the robot can create
a likelihood distribution—a distribution over possible robot poses based on an observation
model. Combining the likelihood and proposal distributions yields a posterior distribution
that represents the actual uncertainty distribution of the robot’s pose in the world.

67
Previous Work

An action model is important as it determines the location, shape, and span of each proposal
distribution. There has been previous research into different differential drive action mod-
els and into how to calibrate them. Some research utilizes specific, contrived environments
and constrained motion in order to disambiguate translational error from rotational error
[Borenstein and Feng, 1996; Rekleitis, 2003]. In addition to constraining both the environ-
ment and the motion taken, these methods are run offline, using the final localization error
to find the parameters of a static action model. If the underlying source of error changes
(e.g., motors wear out, new equipment changes the load, or the travel surface changes), the
researcher must run these experiments again to determine new model parameters.
Below, we discuss an approach that eliminates the need to re-tune the action model
for each new environment type. The idea is to adapt the proposal distribution at each step
based on observations, overcoming an incorrect action model. Before we discuss the adap-
tive particle filter, we define the action model we use.
Eliazar and Parr [2004] present a fairly complete sub-Cartesian action model, along
with a nice offline tuning method given ground-truth. A sub-Cartesian action model uses
the change in wheel shaft encoder counts between time i − 1 and time i to calculate the
incremental translational motion si and rotational motion φi .2 These values are used to
estimate a change in Cartesian space (x, y, θ ).
Every pair of si , φi values corresponds to a change in the estimated pose (x, y, θ ).3
The most common way to calculate the pose is to assume that both the translation and
rotation maintained constant velocity across the measured interval. For small time steps,
this arc is closely approximated by assuming the robot sequentially performs half of its
2 We assume that there is no opportunity for confusion between the values s, φ , and α used to describe the
action model and their use in the SSH and HSSH formal descriptions. Terminology defined in the action model
discussion has local scope.
3 We use x to refer to the robot’s pose in a planar model. We can decompose x into the (x, y, θ ) components

of the Euclidean space; thus, x is a dimension in the vector x and is distinguished in this discussion by using a
different font.

68
rotation, all of its translation, then the other half of its rotation [Wang, 1988].

φi
xi = xi−1 + si · cos(θi−1 + )
2
φi
yi = yi−1 + si · sin(θi−1 + )
2
θi = θi−1 + φi

Eliazar and Parr [2004] go one step further than this traditional model and add
another dimension of error that is normal to the direction of travel. This drift is represented
by δi . Additionally, instead of using scalar values for si and φi (estimating only the mode
of the uncertainty), this model tries to estimate the full distribution of error. They do this
by assuming that s̃i , δ̃i , and φ̃i are Gaussians. This makes the change in Cartesian pose a
distribution as well.

φ̃i φ̃i + π
x̃i = x̃i−1 + s̃i · cos(θ̃i−1 + ) + δ̃i · cos(θ̃i−1 + )
2 2
φ̃i φ̃i + π
ỹi = ỹi−1 + s̃i · sin(θ̃i−1 + ) + δ̃i · sin(θ̃i−1 + )
2 2
θ̃i = θ̃i−1 + φ̃i

The Gaussians s̃i , δ̃i , and φ̃i are dependent on both si and φi . The problem now becomes
how to determine the correct parameters, c0 , . . . , c11 , that define these Gaussians.

s̃i = N (si · c0 + φi · c1 , si 2 · c2 + φi 2 · c3 )

δ̃i = N (si · c4 + φi · c5 , si 2 · c6 + φi 2 · c7 )

φ̃i = N (si · c8 + φi · c9 , si 2 · c10 + φi 2 · c11 )

Due to the large number of parameters, the authors tune the model by gathering a
data set and calculating the ground-truth trajectory taken by the robot then solving a set
of linear least squared systems to determine the parameters. Ground-truth is computed

69
by overestimating the action model (to ensure good localization at each step) and running
SLAM using a Rao-Blackwellized particle filter. Here, they can use many particles to
reduce the chances of particle depletion, as this is done offline, where efficiency is not
important.
Once they have ground truth, they can solve the six independent linear least squared
systems. One system solves for the mean parameters of s̃ (c0 and c1 ) given the N localization
steps in the data trace. W is an NxN diagonal matrix where each diagonal element equals the
square-root of the likelihood of the winning particle at that time step. O is an Nx2 matrix,
where each row represents the measured si and φi for each time step. C is the column
vector [co c1 ]T . B is the Nx1 column vector representing the maximum-likelihood s value
for each time step (the s value that lead to the winning particle). They then solve for C
where WOC = WB. Using c0 and c1 , another system solves for the variance parameters of
s̃ (c2 and c3 ). Similarly, the other four linear systems solve for the parameters of δ̃ and φ̃
[Eliazar and Parr, 2004].

A New Action Model

The action model detailed above relies on having sub-Cartesian odometry information
(s, φ ). However, many localization implementations may not have access to this informa-
tion. Pre-compiled odometry modules, robot simulators, and shared log files usually only
provide Cartesian poses—(x, y, θ ). This can be a problem for some researchers wanting to
utilize the above models since there are many values of ∆(x, y, θ )i that yield no solution for
si , φi (e.g., odometry “jumps” sideways for some reason).
To overcome this problem, we use the approach proposed by Rekleitis [2003]. Sim-
ilar to the way the above action model approximates the constant velocity arc taken by the
robot with a “turn-travel-turn” approach [Wang, 1988], we can break any change in pose
into “turn-travel-turn” components—where the first turn α rotates to face the new loca-
tion, the robot travels straight for distance ρ, and the robot makes a final turn β to its final

70
Figure 4.2: Generalized action model. Given two successive odometry readings that give ∆(x, y, θ ),
we can model the motion by a turn α, a translation ρ, and a second turn β .

orientation (Figure 4.2).

∆y
αi = arctan( i ) − θi−1 ; αi ∈ [−π, π]
∆xi
q
ρi = ∆xi 2 + ∆yi 2

βi = ∆θ i − αi ; βi ∈ [−π, π]

When sitting still (ρi = 0), then αi = 0 and βi = ∆θ i . Because robots do not always translate
forward, but can move backwards, we also consider the case where αi0 = αi + π, αi0 ∈
[−π, π], ρi0 = −ρi , βi0 = βi + π, βi0 ∈ [−π, π]. We only use these values when |αi0 | + |βi0 | <
|αi | + |βi | (i.e., when less turning is needed when moving in reverse than when moving
forward).
Using α, ρ, and β in stead of s and φ , we can extend the action model of Eli-
azar and Parr [2004]. This is the action model we use below to compare changes to the

71
straightforward particle filter localization algorithm.

s̃i = N (αi · c0 + ρi · c1 + βi · c2 , c3 + αi 2 · c4 + ρi 2 · c5 + βi 2 · c6 )

δ̃i = N (αi · c7 + ρi · c8 + βi · c9 , c10 + αi 2 · c11 + ρi 2 · c12 + βi 2 · c13 )

φ̃i = N (αi · c14 + ρi · c15 + βi · c16 , c17 + αi 2 · c18 + ρi 2 · c19 + βi 2 · c20 )

π
x̃i = x̃i−1 + s̃i · cos(θ̃i−1 + αi ) + δ̃i · cos(θ̃i−1 + αi + )
2
π
yi = yi−1 + s̃i · sin(θ̃i−1 + αi ) + δ̃i · sin(θ̃i−1 + αi + )
2
θ̃i = θ̃i−1 + φ̃i

Notice, that we added constant parameters c3 , c10 , c17 into the variance formulas.
This ensures any “constant” localization error due to the discretization of the LPM can be
handled. Additionally, for testing purposes, it allows us a convenient way to create large,
overestimating models or small, underestimating models by simply setting all uncertainty
to be constant. Though this action model has 21 parameters, it can also be tuned by solving
six linear least squares systems.

4.3.2 Adaptive Particle Filter Localization

The particle filter is a well-known method that efficiently approximates these distributions
so that pose uncertainty can be quickly calculated by a robot moving through an environ-
ment. Particles are selected from the posterior distribution from the previous time step. The
action model, then “moves” each particle to create the proposal distribution (Figure 4.3).
Proposal particles are weighted according to their likelihood scores and normalized to sum
to one (Figure 4.4). These weighted particles represent the new posterior distribution over
poses—higher weighted particles are more likely to be considered on the next localization
cycle.
Many action models are hand tuned to generate proposal distributions that overesti-

72
Figure 4.3: Proposal distribution example: uncertainty in motion. The robot’s pose is estimated
by a set of particle (here we show x,y locations without illustrating orientation). After an action, the
robot’s new pose is estimated by a proposal distribution, which generally is quite spread out.

73
Figure 4.4: Likelihood weighting example: uncertainty in observations. The robot obtains an
observation at the new pose, and scores each particle in the proposal with a likelihood based on
matching its map of the world. The weighted particles represent the posterior distribution, modeling
the robot’s localization uncertainty.

74
Figure 4.5: Localization failure example: wheel slippage. In cases where the robot’s wheel slip,
either on slippery floors, or when rotating in place, the odometry estimate of the incremental motion
can be very far off from the actual trajectory the robot took. In this case, all particles may have low
likelihood scores, as the current observation projected from the proposed particles does not match
the world map.

mate the posterior. While this generally ensures high accuracy localization (the likelihood
distribution definitely intersects the proposal), many particles end up with negligible likeli-
hood scores. By tuning an action model, one can reduce these wasted particles. However,
if improperly tuned, proposals can become too small. Underestimating proposals can cause
inaccurate localization, as the proposal and likelihood distribution barely overlap. This is a
far more serious problem than having wasted particles. Similarly problems like wheel slip-
page can often lead to proposal distributions being far from the ground-truth location of the
robot. Again particles here have extremely low likelihoods, leading to incorrect posterior
distributions (Figure 4.5).
We investigate a method for adapting the proposal in order to achieve better local-
ization. The method adapts the proposal at each time step. The idea is to change each
proposal distribution based on the likelihood scores generated by an initial set of particles.
This allows a proposal to grow or shrink in order to place its remaining particles in a region
that will lead to higher accuracy localization.

75
KLD-Sampling

Fox [2003] provides a method by which a robot that uses a particle filter to localize can
be confident it has sufficiently sampled the proposal distribution. We utilize this method
in all of the tested algorithms below for two reasons. First, it provides an upper bound on
the number of samples needed for a given proposal distribution. This eliminates many re-
dundant particles when the proposal distribution is small. Second, it also provides a lower
bound on the number of particles. This ensures that large proposals are adequately sam-
pled, which permits a fair comparison of localization accuracy between various localization
algorithms.

Shrinking the Proposal

Suppose a proposal distribution overestimates the variance of the actual posterior distribu-
tion. Is there an intelligent way to shrink the proposal in order to use a small number of
additional particles to obtain better localization? We adapt a method proposed by Grisetti
et al. [2005] that shrinks the proposal distribution in such cases, therefore allowing more
accurate localization.
Grisetti et al. [2005] observe that localization with laser range sensors (lidar) usu-
ally produces likelihood distributions that are much more highly peaked (smaller variance)
than most proposal distributions—most modern-day planar lidar devices have ±1 cm of
error, which is much smaller than previous range-sensing devices. Given this assumption,
it is possible to treat the proposal distribution as being constant under the likelihood dis-
tribution. Given a constant proposal, the posterior distribution is equal to the likelihood
distribution. Thus, a new, better proposal would just be the estimated likelihood distribu-
tion. To obtain a closed form for the new proposal, the distribution around the likelihood
peak is approximated with a Gaussian.
We borrow this idea, treating the high-precision likelihood given by the lidar mea-
surements as a better approximation of the true posterior, than the combination of the like-

76
lihood and the (noisy) proposal. In our implementation the mean µ and covariance Λ of
the Gaussian of the new proposal are estimated by the weighted mean and covariance of
the particles {x0j } sampled from the original proposal and weighted by their normalized
likelihoods. The equations are as follows:

1
µ =
η ∑ x0j p(z|x0j , m)
j
1
Λ =
η ∑ p(z|x0j , m) (x0j − µ) (x0j − µ)T ,
j

where η = ∑ j p(z|x0j , m). For this new proposal, KLD-sampling should require far fewer
particles than in the original proposal due to the smaller variance. In the experiments below,
we refer to this algorithm as shrink. Figures 4.6 and 4.7 illustrate the potential improve-
ments of the shrink algorithm using a 1D example.

Growing the Proposal

Suppose that a proposal underestimates the actual posterior distribution. If the maximum
likelihood particle in the proposal has a high score, then localization will be largely unaf-
fected. The more serious case occurs if all sampled particles have low likelihood scores,
i.e. the proposal is not large enough to intersect the peak of the likelihood distribution.
A similar scenario occurs when the proposal’s mode is displaced from the actual posterior
(e.g., wheel slippage causes the action model to create a proposal far from the robot’s actual
pose). Is there an intelligent way to grow the proposal in order to “locate” an area of high
likelihood?
If all particles in a proposal distribution have low likelihood scores, there is no sig-
nal that tells us exactly how to grow the proposal to search for better likelihood scores. By
likelihood scores, we mean the unnormalized values of p(z|x0j , m). To handle such cases,
we set a lower threshold on the maximum likelihood score. In the event that all likeli-

77
 (a)

(b) (c)

Figure 4.6: Poor localization due to an overestimating proposal. Here we see a 1D example of a
single particle filter iteration. This distribution represents the kind of problem a robot encounters
quite frequently. (a) The proposal distribution is quite spread out due to large uncertainty (e.g., from
odometry sensors), while the likelihood is quite highly peaked (e.g., from precise lidar measure-
ments). (b) The algorithm approximates the proposal distribution by a set number of particles. (c)
The particles are weighted by their likelihood scores, and this weighted distribution serves as the
posterior distribution.

78
 (a) (b)

(c) (d)

Figure 4.7: Improved localization using the shrink algorithm. Here we perform one particle filter
iteration on the same example from Figure 4.6; however we use the shrink algorithm in our particle
filter. (a) We only use 100 of the alloted 200 particles to sample the proposal. (b) We weight these
particles based on the likelihood distribution. (c) We then approximate the weighted particles with
a new Gaussian. (d) The new Gaussian acts as another proposal for the final 100 particles, which
once weighted, approximate the posterior quite well. Compare this to Figure 4.6(c).

79
hood scores fall below this threshold, a new proposal is created that simply quadruples the
covariance matrix of the current proposal.
In our current implementation, we perform this quadrupling up to five times if nec-
essary, and the threshold is annealed to become more generous each time the proposal
grows. In the experiments below, we refer to this algorithm as grow. The proposal does not
always grow, only when required, due to a lack of accurate localization using the current
proposal. Figures 4.8 and 4.9 illustrate the potential improvements of the grow algorithm
using a 1D example.

Shrinking and Growing

The shrink and grow algorithms above should be considered mutually exclusive improve-
ments to the standard particle filter algorithm, which uses a single, static proposal distribu-
tion. In the experiments below, we also consider the shrink-and-grow algorithm. We expect
the shrink-and-grow algorithm to cover all cases covered by shrink or grow.
This combined algorithm performs shrink on the original proposal, using particles
wisely to get more localization precision. If no good particle is found, it then runs grow
(growing the original proposal, not the “smaller” proposal created by shrink). The shrink
algorithm is also performed on any new proposal generated by the grow algorithm, hoping
the extra precision will find a good particle. This combination of grow then shrink is re-
peated until a good particle is found or a maximum number of iterations is achieved. This
allows a proposal to grow until it intersects the likelihood distribution, then shrink down
to fit the likelihood with a highly peaked Gaussian. Figures 4.10 and 4.11 illustrate the
potential improvements of the shrink-and-grow algorithm using a 1D example.

Experimental Setup

We evaluate the three algorithms above in terms of their influence on the accuracy of a
particle filter in an actual robot localization setting. Here we describe our experimental

80
 (a)

(b) (c)

Figure 4.8: Poor localization due to an underestimating proposal. Here we see a 1D example of
a single particle filter iteration. This distribution represents the kind of problem a robot encounters
when it is in environments that are perceptually ambiguous in one or more dimensions (e.g., traveling
down a long, featureless hallway, the posterior should spread out along the length of the hallway).
(a) The proposal distribution is quite highly peaked here. (Note: The proposal and likelihood have
swapped from Figure 4.6). (b) The algorithm approximates the proposal distribution by a set number
of particles. (c) The particles are weighted by their likelihood scores, and this weighted distribution
serves as the posterior distribution.

81
 (a) (b)

(c) (d)

Figure 4.9: Improved localization using the grow algorithm. Here we perform one particle filter
iteration on the same example from Figure 4.8; however we use the grow algorithm in our particle
filter. (a) We only use 40 of the alloted 200 particles to sample the proposal, as the proposal has
such small variance. (b) Because no particle had a likelihood weight over the desired threshold, the
standard deviation is doubled, and another 40 particles are sampled. This continues until a good
particle is found or until a maximum number of iterations is reached. (c) After growing the proposal
three more times, a particle exists that has a high likelihood score, so the growing process terminates.
(d) The final 40 weighted particles are a much better approximation of the posterior than the 200
particles in Figure 4.8(c).

82
 (a)

(b) (c)

Figure 4.10: Poor localization due to offset proposal. Here we see a 1D example of a single particle
filter iteration. This distribution represents a wheel slippage or very bad odometry measurement,
where the action model proposes locations far from the ground-truth location of the robot. Normally
we see these errors most in orientation. (a) The mean of the proposal distribution is far from the
likelihood mean, which is a better approximation of the ground-truth location. (b) The algorithm
approximates the proposal distribution by a set number of particles. (c) The particles are weighted
by their likelihood scores, and this weighted distribution serves as the posterior distribution.

83
 (a) (b)

(c) (d)

Figure 4.11: Improved localization using the shrink-and-grow algorithm. Here we perform one
particle filter iteration on the same example from Figure 4.10; however we use the shrink-and-
grow algorithm in our particle filter. (a) We only use 33 of the alloted 200 particles to sample the
original proposal, as we expect to need the rest to overcome the offset. (b) The algorithm created a
new proposal using the shrink method, and sampled that with 33 more particles. Then because no
particle had a likelihood weight over the desired threshold, it doubled the standard deviation, and
sampled another 33 particles. (c) After shrinking and growing the proposal once more, a particle
exists that has a high likelihood score, so the growing process terminates. (d) shrink is run one
last time to get high likelihood values for the final 33 particles. The final 33 weighted particles
approximate the likelihood better than the 200 particles in Figure 4.10(c). For a robot with noisy
odometry but a high-precision lidar sensor, the likelihood is a better approximation of the posterior
than the combination of likelihood and (noisy) proposal.

84
setup.
First, the robot explores an environment to create a sensor log. The robot records an
alternating sequence of time labeled odometry and range observations. A S ICK LMS-200
lidar sensor is used as the range-sensing device. The log file used here has 3961 odometry
readings x j (and 3961 laser readings z j ) from a nine minute run. 418 time steps involve
no motion by the robot. For each pose x j in the robot’s trajectory, we calculate the relative
motion of the robot u j = x j+1 − x j .
The log file is then used to create a global metrical map via offline methods (Chapter
9). The global metrical map provides ground-truth of the actual location of the robot at
every laser reading, x̂ j . The global map is represented as an occupancy grid that has 5 cm
cells.
We ran localization experiments using three different action models. The parameter
values of these three models are shown in Table 4.1. The large action model has a constant
sized uncertainty that is centered on zero odometry error. The parameters, chosen before-
hand, were selected in an attempt to overestimate the uncertainty at almost every time step.
The small action model also has constant uncertainty centered on the odometry estimate.
These parameters were chosen to underestimate the posterior for a majority of the sensor
trace—when the robot is making very small movements, this model is sufficient for good
localization. The final action model is a least squares fitted model. These parameters were
calculated by finding the error between odometry and ground-truth (from the global map)
and solving a series of least squares systems to find the best parameters for the action model
as described by Eliazar and Parr [2004]. Thus, it is an action model created specifically for
this data set.
We tested five localization algorithms on these three action models: raw odometry,
standard particle filter, shrink, grow, and shrink-and-grow. For each localization algorithm
(excluding raw odometry, which only uses 1 particle), KLD-sampling is used to determine
the number of particles each proposal needs. Our KLD-sampling parameters [Fox, 2003]

85
 s̃ c0 c1 c2 c3 c4 c5 c6

large 0 1 0 1e-2 0 0 0
small 0 1 0 1e-6 0 0 0
least-squares -0.012 0.99 -0.012 2.6e-05 0 0.0052 0
fitted

δ̃ c7 c8 c9 c10 c11 c12 c13

large 0 1 0 1e-2 0 0 0
small 0 0 0 1e-6 0 0 0 0
least-squares 0.0014 -0.0016 0.0019 1.4e-05 0 0.0012 0
fitted

φ̃ c14 c15 c16 c17 c18 c19 c20

large 1 0 1 3e-2 0 0 0
small 1 0 1 3e-6 0 0 0
least-squares 0.98 -0.0048 0.98 4.1e-05 0 0.0093 0
fitted

Table 4.1: Action model parameters. large is an overestimating model. small is an underestimating
model. least squares fitted is a model fit using the ground-truth posterior distributions over the
entire data trace.

86
are set as follows: confidence=0.95, KLD error=0.5, bin size=(0.05, 0.05, 0.006). These
were tuned so that our robot does not use more than 500 particles, except in very rare
situations.
Comparing localization algorithms by performing Markov localization on the entire
trajectory is tricky, as one bad localization due to unlucky sampling can affect all future
localization. By restarting localization at ground-truth at each time step, we can compare
algorithms directly, eliminating the problem of cumulative error—for each laser reading, z j ,
we start the robot at x̂ j−1 and create a proposal distribution according to the action model
and the action u j−1 given by the raw odometry.
After the localization step terminates, the best (unnormalized) likelihood score of
each algorithm is recorded and used as our accuracy metric. A likelihood score is measured
by comparing how well the laser scan at a hypothesized odometry matches the laser scan at
the ground-truth odometry, given that lasers are discretized into a 15x15 meter occupancy
grid with 5 cm cells [Konolige, 1999]. Each score is a real value from 0 to 1.
The 5 localization algorithms, each run on the three action models, gives us 15
datasets (each with 3960 localization steps) to analyze (L1 -L5 , S1 -S5 , F1 -F5 in Table 4.2). If
the 5 localization algorithms were run separately, these datasets would be independent, but
to save time, we ran all 5 localization algorithms at the same time: we ran shrink-and-grow
and recorded the data for the other 4 algorithms. This means some of the algorithms had
proposals that were supersets of others (e.g., L5 had proposals that were supersets of the
proposals in L1 -L4 ). To be rigorous in our comparison, we created another 15 datasets from
a different run (i.e., different sampled particles due to different random number generation).
We compare these 15 datasets (L6 -L10 , S6 -S10 , F6 -F10 ) with their counterparts.

Results

The complete results of our analysis are shown in Table 4.2. The table shows the average
difference (over 3960 differences) between the best (unnormalized) likelihood scores of the

87
algorithms. The table also shows whether the set of 3960 maximum likelihood scores from
one algorithm is significantly different than the corresponding set of another algorithm: here
significance means p < 0.01 using the standard paired t-test. We expected no significant
difference between results collected from running the same algorithm on the same model
(the diagonals).
Note that the shrink algorithm outperforms the standard particle filter in all cases:
L3 over L7 , L8 over L2 , S3 over S7 , S8 over S2 , F3 over F7 , F8 over F2 . shrink outperforms
grow on the overestimating action model (L3 over L9 , L8 over L4 ). This result is expected
due to a more precise search space of the adapted proposal.
The grow algorithm also improves localization accuracy over the standard particle
filter for every action model. As expected, grow outperforms shrink on the underestimating
model (S4 over S8 , S9 over S3 ). This is due to the ability of the grow algorithm to search a
larger area in cases where the odometry was vastly incorrect.
In the case of the least squares action model, shrink and grow both outperform the
standard particle filter, with grow slightly outperforming shrink in localization accuracy.
This implies that even the least squares model incorrectly hypothesizes the peak of the
likelihood distribution.
Finally, as expected shrink-and-grow improves accuracy over all other algorithms
on all possible action models. In addition to these quantitative results, we have observed a
qualitative difference in the quality of the LPMs created when using shrink-and-grow over
a standard particle filter. Because we are using KLD-sampling to adequately sample each
proposal distribution, we often see little increase in the total number of particles needed—
the adapted proposals created by the shrink portion of the algorithm have small variance
and require few particles to represent. In situations where more particles are needed as
the adaptive proposal grows, we are willing to spend the extra cycles to ensure quality
localization; however, we use a maximum of 500 particles (100 particles for up to 5 possible
shrink-and-grow iterations) in order to ensure localization at 3 Hz.

88
 model large large small small least squares least squares
algorithm \ run1 run2 run1 run2 run1 run2
raw odometry L1 L6 S1 S6 F1 F6
standard PF L2 L7 S2 S7 F2 F7
shrink L3 L8 S3 S8 F3 F8
grow L4 L9 S4 S9 F4 F9
shrink-and-grow L5 L10 S5 S10 F5 F10

L6 L7 L8 L9 L10
L1 0.000 −0.098, ∗ −0.176, ∗ −0.100, ∗ −0.183, ∗
L2 0.098, ∗ 0.000 −0.078, ∗ −0.002 −0.084, ∗
L3 0.176, ∗ 0.078, ∗ 0.000 0.076, ∗ −0.007, ∗
L4 0.100, ∗ 0.003, ∗ −0.076, ∗ 0.001 −0.082, ∗
L5 0.182, ∗ 0.085, ∗ 0.006, ∗ 0.083, ∗ 0.000

S6 S7 S8 S9 S10
S1 0.000 −0.060, ∗ −0.067, ∗ −0.090, ∗ −0.099, ∗
S2 0.060, ∗ 0.000 −0.007, ∗ −0.030, ∗ −0.039, ∗
S3 0.066, ∗ 0.006, ∗ 0.000 −0.023, ∗ −0.033, ∗
S4 0.089, ∗ 0.029, ∗ 0.022, ∗ −0.001 −0.010, ∗
S5 0.099, ∗ 0.039, ∗ 0.032, ∗ 0.009, ∗ 0.000

F6 F7 F8 F9 F10
F1 0.000 −0.117, ∗ −0.134, ∗ −0.141, ∗ −0.158, ∗
F2 0.117, ∗ 0.000 −0.017, ∗ −0.024, ∗ −0.041, ∗
F3 0.133, ∗ 0.016, ∗ −0.001 −0.008, ∗ −0.025, ∗
F4 0.141, ∗ 0.023, ∗ 0.007, ∗ −0.001 −0.017, ∗
F5 0.158, ∗ 0.040, ∗ 0.024, ∗ 0.016, ∗ 0.000

Table 4.2: Localization accuracy of 5 different algorithms. Comparison of maximum likelihood

scores (which range from 0 to 1) between different localization algorithms. The scalar values show
the average difference in the maximum likelihood score of the row data minus the column data.
When the maximum likelihood scores of the two sets of scores are significantly different (p < 0.01),
there is an asterisk.

89
 As a result of utilizing this adaptive particle filter, we improve the quality of our
LPMs even in indoor environments where mobile robot wheels slip quite often. It is due to
having high quality LPMs that we are able to go from model of obstacles and free space to
symbolic structures of paths and places. This is discussed in Chapter 5.

90
Chapter 5

Symbol Grounding:
The HSSH Local Topology Level

It takes a wonderful brain and exquisite senses to produce a few stupid ideas.
George Santayana, The Life of Reason, Volume 5: Reason in Science (1905)

This chapter discusses how we abstract and utilize local topology models from the LPMs.
We begin by giving a formal description of local topology at places, which will be useful
for determining loop-closing hypotheses in the Global Topology Level (Chapter 8). Next
we discuss how local topology is abstracted from LPMs using gateways [Beeson et al.,
2003; Kuipers et al., 2004]. We illustrate how a robot with an LPM and a gateway detection
algorithm can robustly detect when it is at a place versus moving along a qualitatively
uninteresting path [Beeson et al., 2005]. We give an implementation for gateways, and
show the robustness of the algorithm under noisy conditions.

5.1 Formal Description of Local Topology

As the robot and its scrolling LPM move continuously through the environment, the robot
identifies a discrete set of isolated places and the path segments that connect them. In the

91
small-scale space of the LPM, a place neighborhood is an extended region. In the large-
scale space representation, a place is a node in the topological graph, and is connected by
paths to other places. These are the local elements from which a global topological map
is constructed. We abstract the structure of a place neighborhood to the local topology
description of the place. Just as a path describes the linear order of places on it, a place de-
scribes the circular order of directed paths radiating from it. We call this the local topology
S p of a place p, and describe the circular order with a structure called a star [Kuipers et al.,
2004]. This section discusses how this symbolic representation of a place (in large-scale
space) is grounded in the metrical description m p of the place neighborhood (in small-scale
space).
A local-path π̃ at a place p is the fragment of a topological path that is visible within
the stored local perceptual map, m p , of the neighborhood of p. A directed local-path is of
the form π̃ d , where d ∈ {+, −} represents the direction along π̃ moving away from p. Upon
arriving at a new place, a local-path and its directions may not yet have been matched with
a global topological path and its directions.
A star S is a set of directed local-paths such that π̃ + ∈ S ⇐⇒ π̃ − ∈ S. There are
two functions that describe stars.

next : S → S induces a clockwise circular order over π̃ d ∈ S. next(π̃ d ) is the

next element from π̃ d in the clockwise order.

α : S → {0, 1} associates an attribute value α(π̃ d ) to π̃ d ∈ S, where α(π̃ d ) = 1

means that travel is possible along π̃ d away from p, and α(π̃ d ) = 0 means
that travel away from p along π̃ d is not possible.

The star is naturally encoded as a sequence of pairs, where the sequence encodes the next
relation (next of the last element being the first element), and the second element of each
pair is the value of α applied to the first element. For example, consider the following local
topology (star) descriptions of familiar intersection types including local-paths π̃a , π̃b , and

92
sometimes π̃c .1 (For ease of visualization, the first directed local-path in the circular order
is the one directed upward.)

+ [hπ̃a+ , 1i, hπ̃b+ , 1i, hπ̃a− , 1i, hπ̃b− , 1i]

T [hπ̃a− , 0i, hπ̃b+ , 1i, hπ̃a+ , 1i, hπ̃b− , 1i]
L [hπ̃a+ , 1i, hπ̃b+ , 1i, hπ̃a− , 0i, hπ̃b− , 0i]
Y [hπ̃a− , 0i, hπ̃b+ , 1i, hπ̃c− , 0i, hπ̃a+ , 1i, hπ̃b− , 0i, hπ̃c+ , 1i]
K [hπ̃a+ , 1i, hπ̃b+ , 1i, hπ̃c+ , 1i, hπ̃a− , 1i, hπ̃b− , 0i, hπ̃c− , 0i]
ψ [hπ̃a+ , 1i, hπ̃b+ , 1i, hπ̃c− , 0i, hπ̃a− , 1i, hπ̃b− , 0i, hπ̃c+ , 1i]

An isomorphism φ : S → S0 between two stars S and S0 is a bijective function such

that

next(φ (π̃ d )) = φ (next(π̃ d ))

α(φ (π̃ d )) = α(π̃ d )

path(φ (π̃ d )) = path(φ (π̃ opp(d) )),

where path(π̃ d ) = π̃. An isomorphism means that the two stars have the same local topol-
ogy under a suitable rotation of the circular order. Note that two stars may have multiple
distinct isomorphisms. For example, there are four distinct isomorphisms between two +
intersections.
The local topology description provides a purely qualitative account of “left” and
“right”, avoiding the need to define them in terms of thresholds on some angular variable.
A particular directed local-path at a place p, π̃ + , and its opposite, π̃ − , partition the other
directed local-paths in the star into two groups. Those that are between π̃ + and π̃ − in the
clockwise direction can be described as being “to the right” of π̃ + . Those between π̃ +
and π̃ − in the counter-clockwise direction can be described as “to the left” of π̃ + . This
1 Notethat we do not have a fixed set of equivalence classes for local topology abstraction. Though there is
an upper bound on the number of paths that can fit into an LPM, this is determined by the path width and the
LPM size. Thus, many types of intersections can exist that cannot be “named” using a letter.

93
also defines the appropriate destination for a route instruction such as “turn right” when the
agent is at a place p, facing along a directed path π̃a+ . The pragmatics of natural language
requires that “turn right” must uniquely specify a directed local-path π̃bd that is “to the right”
of π̃a+ , such that α(π̃bd ) = 1 (i.e., π̃bd is navigable from p).

5.2 Grounding Local Topology in the Local Perceptual Map

We have illustrated how to describe a place symbolically as a circular order of directed

local-paths. Here we discuss how to use gateways to ground local-paths in the LPM. Gate-
ways allow the robot to ground large-scale actions in the small-scale metrical models, ab-
stract a symbolic local topology description from the small-scale model, and detect and
compare places in the environment. The term “gateway” is adapted from Chown et al.
[1995], who define gateways as the locations of major changes in visibility (Section 2.3.1).
Section 5.3 discusses detecting places via gateways and local-paths, Section 5.4 discusses
how motion between gateways abstracts to high-level motion. Chapter 6 addresses the is-
sues involved in implementing algorithms for reliably detecting gateways.

5.2.1 Gateways

We define a gateway as a boundary in the local perceptual map that separates the local
place neighborhood from the larger environment. That is, a gateway is the boundary where
control shifts between localization within the local place neighborhood and travel from one
place neighborhood to another. A gateway has two directions, inward (looking into the
place) and outward (looking away from the place), according to the direction of that shift.
The location, extent, and orientation of gateways at a place are saved as annotations of the
local place neighborhood map m p .
In much of human experience with large-scale environments (both natural and man-
made) local place neighborhoods are separated from each other (either by boundaries or
by distance), and they are connected by travel actions along paths. Navigation in large-

94
scale space is thus typically an alternation between motion along travel paths and motion
within place neighborhoods. The existence of gateways, as interfaces between the two
types of travel, is therefore a requirement for the abstraction from small-scale to large-scale
space. Certainly extreme situations occur, such as place neighborhoods that overlap or are
immediately adjacent, or environments with (apparently) no distinctive states at all. These
will be mentioned in appropriate sections below.

5.2.2 Local Topology Creation

Given an implementation for detecting gateways in a stored map of a place, m p , we can

ground the local topology concepts of local-paths in our small-scale model of the surround-
ing environment (Figure 5.1).

• For each outward-facing oriented gateway hg, outi, define a directed local-path π̃g+
that leads away from the current place.

• Initialize a circularly ordered star S p with a list (clockwise from an arbitrary starting
point) of associations between directed local-paths and oriented gateways, (hπ̃g+ , 1i ↔
hg, outi). Since these are traversable paths, each α(π̃g+ ) = 1. 2

• Test each pair of gateways, g and g0 , via a path continuity test, to determine whether
their directed local-paths belong to a single continuous path. If so, give both directed
local-paths the same path name (e.g., π̃a below and in Figure 5.1), and include the
inward oriented gateways in the association. For example, change

(hπ̃g+ , 1i ↔ hg, outi) (hπ̃a+ , 1i ↔ hg0 , ini, hg, outi)

to
(hπ̃g+0 , 1i ↔ hg0 , outi) (hπ̃a− , 1i ↔ hg, ini, hg0 , outi)
2 In future implementations, α : S → {M IDLINE, L EFT WALL, R IGHT WALL, D EAD E ND, N ONE}
should associate directed local-paths with attributes representing the control laws for traversing the path in that
direction. (L EFT WALL and R IGHT WALL imply coastal navigation scenarios. For terminating local-paths,
D EAD E ND means that further travel is blocked, while N ONE means that no control law is applicable.)

95
 • For each π̃g+ ∈ S p such that π̃g− 6∈ S p , insert the association (hπ̃g− , 0i ↔ hg, ini) into
the circular order of S p , in a position determined by its failure of the path continuity
test.

In our current implementation, the gateways g and g0 belong to a single continuous path if:
(1) a ray normal to the orientation of gateway g and centered at the midpoint of gateway
g intersects the line segment that defines gateway g0 ; and (2) vice versa for a ray from g0
towards gateway g. (Note that the failure of this test should determine a pair of gateways
that the non-traversable path continuation falls between.)
At this point, the star S p is a complete representation of the local topology of the
neighborhood described by the LPM. Since this representation is expressed completely in
terms of small-scale space (the gateways g and the directed local-paths π̃gd ), we refer to this
as the small-scale star. Binding the directed local-paths to directed paths in the topological
map of large-scale space implies the appropriate large-scale star. This binding is part of
the HSSH topological mapping process, and is discussed in Section 8.1.

5.3 Detecting Places

To explain local topology abstraction, we provided examples where the robot was already
at a place. Perhaps surprisingly, the method for constructing the local topology of a place
neighborhood does not actually depend on being at a place neighborhood. Gateways can
also be defined along paths (Chapter 6), as they separate the space immediately surrounding
the robot from the “frontier” of the LPM. Therefore, if we recalculate gateways and local
topology at each time step, we can very easily detect places.
Places are important in many different ways. For spatial reasoning, people use
places to define both bounded and unbounded regions of space: “in Texas” or “near the
statue”. Places are further used for metaphorical high-level reasoning such as mathematical
ranking and social ranking: “first place” or “overstepping one’s place”. Places are even

96
 (a) (b)

Small-scale star
description
((hπ̃a+ , 1i ↔ hg4 , ini, hg1 , outi)
(hπ̃b+ , 1i ↔ hg2 , outi)
(hπ̃c− , 0i ↔ hg5 , ini)
(hπ̃d+ , 1i ↔ hg3 , outi)
(hπ̃a− , 1i ↔ hg1 , ini, hg4 , outi)
(hπ̃d− , 0i ↔ hg3 , ini)
(hπ̃c+ , 1i ↔ hg5 , outi)
(hπ̃b− , 0i ↔ hg2 , ini))
(c)

Figure 5.1: Identifying gateways and local topology in the LPM. The local perceptual map (LPM)
is implemented as a bounded occupancy grid. The robot is shown as a circle in the center of the
LPM. (a) The gateways separating the place neighborhood from the paths are defined. In our im-
plementation this is done using a pruned Voronoi skeleton. (b) Gateway locations and directions are
used to identify the directed local-paths and to determine which pairs satisfy the path continuity re-
quirements. (c) The small-scale star enumerates directed local-paths in clockwise order, describing
their traversability and association with gateways. Note: the robot entered the place via g5 ; thus, it
arrived on directed local-path π̃c− . This environment has five gateways and four paths.

97
used to refer to normal roles or functions: “the place of the media” or “something is out of
place”.
The broad use of places is a specific example of a crucial point in the study of intel-
ligent agents—spatial knowledge is a foundation for high-level common-sense knowledge
[Lakoff and Johnson, 1980]. We believe that by studying the problem of grounding con-
tinuous sensory experience to low-level, symbolic spatial constructs, we are developing a
foundation for the high-level common-sense knowledge necessary for an intelligent agent
to act in environments with people and other animals.
Here we discuss the power of using gateways and local topology to detect and de-
scribe places over other popular approaches. We then detail and evaluate our current imple-
mentation for finding gateways in the LPM.

5.3.1 Place Criteria

A topological navigation system depends on the initial low-level abstraction of places from
continuous sensory experience. Here we attempt to define these places using criteria neces-
sary for topological map-building behaviors. We define places at locations in the environ-
ment that satisfy the following two criteria.

1. Places occur at qualitative changes along paths.

The majority of the time, this criterion defines places at intersections, or decision points.
The importance of path intersections in topological representations cannot be overstated.
The recent work of Guilford et al. [2004], demonstrates that pigeons, long considered to
use Earth’s magnetic field to fly long distances, follow man-made highway networks in
well-known environments, using familiar highway intersections as markers for specific turn
actions. It has long been known that humans use intersections as a basis for building spa-
tial representations in unknown environments [Siegel and White, 1975]—preferring first to
build structural models of novel environments prior to detailed visual models [Stankiewicz
and Kalia, 2007].

98
 Occasionally people may define places that are not at intersections. Dead-ends
are one example of important places that are not necessarily at an intersection of two or
more paths. Less useful places (from a topological mapping perspective) can be defined by
using salient landmarks along paths. Such landmarks are useful for tracking progress along
long paths and thus define places based on regions of visibility. Common examples are
unusual buildings along highways such as water towers. From the perspective of topological
map-building, places that are not at intersections or dead-ends do not yield any additional
structural information. In fact, as people do start to remember landmarks, they are biased
to first learn landmarks at critical path changes along routes [Heft, 1979; Aginsky et al.,
1997].

2. Places must be reliably detectable.

Although this criterion seems simple, it is extremely important. Qualitatively interesting

locations should be considered as possible places when they exist, but their persistence
needs to be taken into account.
Any place detection algorithm needs to provide a stable set of places. Intersections
are not only common but extremely stable (they rarely disappear). Thus, this criterion is
mostly needed for places that are not at intersections. For example, a bright pink car parked
on the side of a highway, while salient, does not define a reliable place. The chances that
the car will remain at the location for an extended period of time are very low.
Places that are not at intersections depend on higher-level knowledge about stability,
saliency, and informativeness that may not be available for most robotic implementations.
For this reason, the algorithms and implementation for mobile robots detailed in the fol-
lowing sections focus exclusively on decision points (places at intersections, doorways, and
dead-ends), although future robotic implementations with knowledge of objects [Modayil,
2007], semantically labeled LPMs [Murarka et al., 2006], or human-defined places may be
able to bootstrap place detection at other interesting locations.

99
5.3.2 Places from Local Topology

We define the robot to be on-path when the local topology of the LPM contains exactly
two gateways and exactly one path (e.g., Figure 5.2(a)). When the agent is on-path, it
is selecting and executing control laws (and hence primitive motions) to perform a travel
action (Section 5.4). The LPM scrolls as the agent moves, keeping the agent near its center
cell, and serving as an observer for the local small-scale space.
When the agent is not on-path, it is in a place neighborhood. In this situation, the
agent establishes a fixed correspondence between the LPM and the structure of the place
neighborhood. Here, the LPM serves as a local metrical map m p of the place neighborhood
(and does not scroll with the agent’s motion within the place neighborhood). Thus, the
number, location, and structure of places in an environment depends in part on the prede-
termined size of the LPM3 , though places are not sensitive to small changes in LPM size.
When the robot is not on-path, it either has more than one local-path (Figure 5.2(b)),
which occurs at intersections or open doorways, or one local-path with only one gateway,
which occurs at dead-ends (Figure 5.2(c)). These are all places. There is a degenerate
case where no gateways exist. Due to our implementation of gateways using an LPM,
this situation means there is no way out of the current location (the robot is completely
surrounded by obstacles), so the robot’s entire world is simply modeled by a single place
and LPM.4
When traveling along a path, the robot may see multiple unaligned gateways and
suspect it is at a place. Sometimes, false gateways appear in the LPM due to the boundary
between observed free space (i.e., white cells in the occupancy grid) and unknown space
(i.e., gray cells in the grid). This is often the case when the robot’s sensors do not provide
a 360◦ field of view, as with S ICK-brand lidars. Before the robot commits itself to being
at a place, it must perform some local exploration in the fixed map of the potential place
3 One interesting avenue of future research is to try adapting the LPM size by environment characteristics.
4 There is another degenerate case when the robot is in the middle of a featureless environment. As men-
tioned in Section 3.3, the HSSH currently does not handle these types of environments.

100
 (a)

(b) (c)

Figure 5.2: Gateway examples. (a) Here we see that gateways should be defined along paths, where
the robot is not at a place. Any gateway implementation needs to ensure that two aligned gateways
exist along paths. (b) At intersections, gateways defined multiple, unaligned local-paths. Thus the
robot knows it is at a place. (c) At dead-ends, there is a single path; however, there should only be a
single gateway, so that dead-ends are places.

101
to eliminate any false gateways. We have found that for a robot with a 180◦ field of view,
simply rotating in place eliminates most false gateways. This concept is illustrated as part
of a real-world navigation example in Figure 6.24(g,h).

5.3.3 Gateways versus Voronoi Junctions

As discussed in Chapter 2, researchers use the generalized Voronoi graph (GVG) both as
a way to discretize the continuous environment into a finite set of places and as a way to
define paths [Choset and Burdick, 2000; Choset et al., 2000; Choset and Nagatani, 2001].
Whenever the robot is near a junction in the Voronoi graph, it is at a topological place.
Junctions are points on the GVG equidistant from three or more closest obstacles in the 2D
metrical model of local environment: where the graph forks into multiple branches. The
branches that emanate from a junction define the paths that the robot can travel along to
leave the place.
Using the local topology defined by gateways allows the robot to detect places more
reliably than when using methods that simply look for Voronoi graph junctions. First,
non-pruned Voronoi graphs can have many junctions. This is especially true given noisy
occupancy grids, but even occurs in the face of no noise at small alcoves and other common
architectural features (Figure 5.3(a)). Similarly, complex intersections can have multiple
junctions. The gateways and local topology can see one place, whereas a junction-based
approach (including Delaunay triangle approaches [Silver et al., 2004]) must define multiple
strangely connected places. Figure 5.1(a) shows three junctions in one place.
Pruning is often performed on the Voronoi skeleton, either by fixed-depth prun-
ing [Choset and Nagatani, 2001], distance-based methods of eliminating short branches
[Wallgrün, 2005], or by using pixel-based approximations of the Voronoi skeleton, like a
thinning-based skeleton [Choi et al., 2002]. These remove many spurious junctions, but
in general not all unnecessary branches are removed (e.g., spurs still exist in the pruned
skeleton in Figure 5.3(e)). In Section 6.1.1, we discuss a way to prune the Voronoi graph

102
 (a) (b) (c)

(d) (e) (f)

Figure 5.3: The reduced generalized Voronoi graph (RGVG). (a,b,c) Many places in the GVG are
created by “spurs” that are defined by noise or concave corners. (d) The RGVG removes branches
that terminate with only two nearest obstacles, eliminating many false positive places. (e) Un-
fortunately, this fixed-depth pruning does not remove all undesired branches. (f) The RGVG also
removes branches that create important junctions, which leaves important places undefined when
using algorithms that equate Voronoi junctions with places.

103
in LPMs that ensures the elimination all spurious junctions. Given pruned skeletons, there
exist important places detected via the gateway approach, like L intersections, that may
contain no junctions at all after pruning the Voronoi graph (Figure 5.3(f)).

5.4 Selecting Local Motion Targets

Instead of relying on the dynamical system approach to motion used in the basic SSH, we
introduce gateways as an alternative approach. Gateways provide a geometric method for
controlling motion—where midline or coastal navigation along paths is applicable. The
motion of the robot in large-scale space can be adequately captured by noting which ori-
ented gateways the robot passes through. Figure 5.4 illustrates the Hybrid SSH approach to
large-scale motion where turns and travels correspond to moving towards gateways.
As discussed in Section 4.2, local motion planning consists of selecting a target
pose in the LPM, computing a safe trajectory to it, executing the first step of that trajectory,
sensing the environment, updating the LPM, and repeating the cycle. The selection of target
poses for local motion control corresponds to the action or goal being pursued. There are
three distinct cases.

• If the agent is not in a place neighborhood, it is on-path, in which case it is moving

along the local-path in the LPM toward one of the two gateways. Just beyond the
forward gateway, in the outward orientation, is an appropriate target for local motion
planning; however, a more robust approach with respect to obstacle avoidance is to
aim at a point well beyond the gateway, like the edge of the LPM. As the LPM scrolls,
the gateway location is constantly refreshed. The robot never reaches the gateway
until its location becomes stable (which only happens when the agent arrives at a
place).5

• If the agent is in a place neighborhood, the LPM is fixed to the local environment, so
5 Lee [1996] calls control algorithms that continuously re-plan for a moving point ahead as “red wagon”

controllers. One could also think of them as “carrot-on-a-stick” controllers.

104
 (a) (b)

(c) (d)

Figure 5.4: Grounding control using gateways. (a) The example from Figure 3.3 is further ex-
amined in a simulated 3D office environment with obstacles. The gateways are found in real-time
(using the anchor-based algorithm discussed in Section 6.3.2), with arrows representing the out-
ward orientations that leave the current area. The gateway associated with the robot’s past motion
is depicted using an arrow pointing in the inward orientation. (b) Traveling along directed path πa+
corresponds to aiming beyond an oriented gateway, e.g., hg2 , outi, in the appropriate direction. The
gateway is continually recomputed, which keeps moving the local motion target along the path, un-
til it becomes stable at the entrance to a place. (c) Arriving at dstate q2 at place p2 corresponds to
arriving at a gateway hg3 , ini associated with a directed local-path π̃a+ in the LPM for place p2 . The
turn action from dstate q2 to q3 corresponds to local motion within the LPM through outward-facing
oriented gateway hg6 , outi on directed local-path π̃b− . (d) In calculating local topology, “island” ob-
stacles that are surrounded by free space are removed to ensure reliable gateway detection. Planning
to move through a gateway requires consideration of these obstacles. Once the robot moves past
gateway hg6 , outi, two new pairs of aligned gateways appear that will flank the robot throughout the
next travel action, as in image (b).

105
 motion planning is confined to the small-scale space of the place neighborhood. The
agent may have a pragmatic destination within the place neighborhood, for example
an intelligent wheelchair may have the goal of bringing its driver to her desk after
entering her office, in which case the local motion target is a pose associated with
that destination. Such motion targets can also be generated when exploring the fixed
LPM of a potential place.

• The agent may be executing a turn action as part of a route through large-scale space.
In this situation, the LPM is fixed in the local frame of reference, and a large-scale
turn action corresponds to moving from an inward-facing oriented gateway to a lo-
cation just beyond an outward-facing oriented gateway. After passing through the
outbound gateway, the robot is in position to begin following another path. Note, that
the TurnAround action simply corresponds to traveling past the same gateway the
robot entered the place through, facing the outward instead of the inward orientation.
Continuing along a path that passes through a place (no turn) also falls into this case.

In certain scenarios, such as two large rooms connected by a doorway, it may be

possible for an agent to move directly from one place neighborhood to another, moving
between two distinct local topologies, without ever being significantly on-path. The SSH
can accommodate this transition with a dummy travel action whose effect is simply to tran-
sition between the reference frames of two adjacent, or even slightly overlapping, places.
Taking this idea to an extreme, the Atlas system [Bosse et al., 2003] creates new frames of
reference based on feature counts, building a “patchwork” map of overlapping frames of
references. However, if the entire environment is described in terms of overlapping place
neighborhoods, the benefit of the topological map as a concise description of large-scale
space is decreased. Likewise, the local and global distinctiveness of places is sacrificed.

106
Chapter 6

Implementing Gateways

The longest part of the journey is said to be the passing of the gate.
Marcus Terentius Varro, On Agriculture, Book 1 (1st Century BCE)

We have detailed how gateways can be used to determine the local topology in
the nearby surround, facilitating place detection and description. We have also discussed
how gateways are used at place neighborhoods to facilitate local motion that corresponds
to turns in large-scale space, and how large-scale travel actions correspond to repeatedly
moving towards a gateway along a local-path. Some of the figures used to illustrate these
ideas (Figures 5.1 and 5.2) have gateways that were generated using various methods we
have explored over the years, including some of the techniques discussed in Section 2.3
such as looking for hallways of a certain size, looking for large disparities using radial
ray-casting form the robot’s position, and looking for “critical points”.
Here we discuss two attempts at a more robust gateway algorithm, independent of
metrical thresholds. Our gateway algorithms rely on a Voronoi skeleton computed from the
free space in the LPM (previously illustrated in Figure 5.1(a)). Table 2.1 compares our two
gateway algorithms with others found in the literature. Our initial gateway implementation
is a constriction-based method that works reasonably well in corridor environments [Bee-
son et al., 2005]. We discuss this algorithm in detail, and we illustrate a specific problem we

107
have encountered when using constrictions. The second implementation is an anchor-based
method that generates similar gateways to the constriction-based method in most scenarios,
while overcoming the problems of the constriction-based method. Specifically, we demon-
strate that this algorithm creates more stable gateways at certain places. This alternative
implementation is also necessary to adequately handle coastal navigation (wall-following)
scenarios.

6.1 Algorithm Overview

Both of our gateway algorithms are composed of four main stages. This process is illus-
trated in Figure 6.1. First, a Voronoi-like skeleton must be computed that encodes the free
space of the LPM as a 1D graph. This includes any pruning necessary to remove branches
caused by noise or branches that represent notches or alcoves but do not connect free space
to a new frontier.1 Second, the algorithm must determine the core of the LPM, using the
skeleton to decide the rough extent of the local place neighborhood. (Note that the same
algorithm defines the core of an LPM along a single local-path; thus, a core is not restricted
to places.) Third, using the core, the algorithm establishes which portions of the skeleton
it will utilize in the search for gateways. Finally, using specific rules, the algorithm places
line segments at locations along the graph to represent gateways.
Both of the gateway implementations discussed below use the same steps to gener-
ate the skeleton and determine the core of the LPM (though we will discuss an extension
to the skeleton in order to handle coastal navigation scenarios with the second implemen-
tation). The difference between the two implementations lies in the rules used to detect
gateway locations. Below, we discuss the first two stages of finding the minimal skeleton
and core of an LPM. Section 6.2 discusses specifics of constriction-based gateways, and
Section 6.3 discusses anchor-based gateways.
1 We use the term frontier to coincide with the ideas of Yamauchi [1997] about exploring occupancy grids
by eliminating unknown regions (gray cells) that border free regions (white cells).

108
 (a) (b)

(c) (d)

Figure 6.1: Overview of gateway detection process. Here we show the four major stages of finding
gateways from an LPM. (a) The algorithm calculates a pruned Voronoi skeleton. Here we use
a thinning-based approximation of the Voronoi skeleton [Zhang and Suen, 1984], which we have
found to be much faster to calculate on slower processors. (b) The algorithm then determines the
“core” of the local region. (c) It then uses the core to determine where on the skeleton to search
for gateways. In this case, the branches outside the core are used. (d) The algorithm looks for
reasonable gateway locations. The robot’s position is denoted by a dark circle.

109
6.1.1 The Voronoi Skeleton of an LPM

The first step in finding gateways is to create a high-quality skeleton of free space from
the current snapshot of the LPM. All gateway processing (object removal, skeletonization,
pruning, etc.) on the occupancy grid representation of the local surround is done in a copy
of the LPM. For example, if we state that free cells are “flipped” to occupied cells, we are
referring to a copy of a recent LPM snapshot, which can be modified without affecting the
actual LPM or the underlying SLAM algorithm.
Additionally, this work uses a discrete approximation of the Voronoi graph. In
discussions on pruning graphs, finding the core of LPMs, and creating gateways, if we refer
to Voronoi graphs, we mean any Voronoi-like skeleton. Similarly, when we discuss Voronoi
junctions or Voronoi branches, we are referring to equivalent junctions and branches of the
Voronoi-like skeleton.

Various Skeleton Algorithms

Multiple skeleton algorithms exist that give some approximation of the continuous Voronoi
graph. Computing the precise, continuous Voronoi graph assuming point obstacles can
be computed in O(n log n) time, where n is the number of obstacles in the world [For-
tune, 1992]. For discretized models, Voronoi approximations can be computed by so-called
“brush-fire” algorithms: imagine fires at all occupied cells that burn inward through free
(fuel) cells at a constant speed such that a skeleton is created at all points where two or
more fires meet. Such algorithms can be computed in O(n) time, where n is the number of
cells.
We prefer to utilize these discretized algorithms for several reasons. First, they
are easy to implement. They also allow us to gather related metrical information during the
skeletal search (e.g., which obstacles are closest to the currently considered pixel) with little
overhead. Finally, they are easy to extend to modified Voronoi graph definitions (Section
6.3.1).

110
Figure 6.2: The thinned skeleton of a global metrical map. Using an occupancy grid, the skeleton
of free space can be computed by using any occupied cells as obstacles. Junction points include
intersections in corridor environments. For visual clarity, portions of the graph that are close to
obstacles are removed. Compare this to the GVG in Figure 2.3.

Specifically we use a skeleton created by thinning2 [Zhang and Suen, 1984]. Thin-
ning is similar to the “brush-fire” algorithm, though it has very specific rules about whether
a cell is flipped from “fuel” to “fire” based on its neighboring pixels. Thinning roughly
approximates a subset of the Voronoi graph (Figure 6.2). A thinned skeleton has all the im-
portant skeletal branches, but allows efficient post-processing (e.g., the pruning discussed
below), as many spurious terminal branches are never generated. However, thinned skele-
tons tend to contain long straight linear segments, thus they are quite dependent on the
orientation of paths. Since our methods can properly detect gateways on these thinned
skeletons, regardless of orientation (Section 6.3.3), we are confident that they can handle
skeletons that better approximate the true Voronoi graph, which is independent of the ori-
entation of point obstacles.
Discrete skeleton implementations like thinning or brush-fire algorithms flip free
2 Our thinning implementation is publicly available on the web [Beeson, 2006]. It provides pre/post-
processing options for removing obstacles, frontier-based pruning, and setting minimum and maximum dis-
tances from obstacles. In addition to being used in our work for determining gateway locations, this particular
thinning implementation is also used for converting graphical representations of chemical structures in elec-
tronic documents to text-based chemical names [Filippov, 2007].

111
cells to occupied cells until a thin set of connected free cells remain; however, in the LPM,
cells can be in one of three states: occupied, free, or unknown. It is important that we
consider unknown cells in the LPM (gray cells in the images) to be free cells when com-
puting the skeleton (Figure 6.3(b)). If unknown cells are used as obstacles, small patches
of unknown cells (unseen by the rays of the lidar) can greatly affect the algorithm, creating
skeletons that stay away from frontiers. Such skeletons are not useful for an exploring robot
that needs to hypothesize unknown space as potentially free. Additionally, as we show be-
low, efficient pruning makes use of exits where the skeleton touches the edge of the LPM
or where free cells border unknown cells.

Pre-processing Obstacles

Any skeleton algorithm is going to be affected by noise and other non-architecture obsta-
cles. We perform pre-processing on detected obstacles prior to generating the skeleton.
This improves the quality of the generated free space skeletons.
First, we remove “island” obstacles from the LPM (Figure 6.3(a,b)). Groups of
occupied and unknown cells that are completely surrounded by free cells are assumed not
to affect the structure of the paths and places in the local region; thus, they are considered
free cells in computing, pruning, and examining the local skeleton. Pedestrians, boxes
in hallways, and thin posts are common examples of obstacles (often with neighboring
unmapped regions due to the occluding obstacle) that can be easily removed from the LPM
before computing a skeleton of the surrounding free space.3
We are able to utilize an obstacle removal algorithm that is linear (instead of the
straightforward O(n2 ) algorithm) due to the fact that we are using an LPM with bounded
size. We find all non-free cells on the border of the LPM. We then recursively add neigh-
3 Pathological cases exist where wall segments between nearby open doors (or large, load bearing columns)
can be seen as “island” obstacles. Similarly, many kitchens contain “island” counters that may be appropriate
for a service robot to consider architecture. We assume such problems will be overcome as robotic perception
improves, as these are obviously permanent architectural obstacles. In the meantime, we could simply keep
obstacles that are larger than a predetermined size in order to avoid these problems; however, such a metric
assumes outside knowledge about the environment, which we are trying to avoid in this work.

112
 (a) (b)

(c) (d)

Figure 6.3: Finding the skeleton of LPM free space. (a) The algorithm starts with an LPM the
robot has built as it moves through the environment. (b) “Island” obstacles are removed, and the
remaining obstacles in the LPM are increased by a predetermined size to fill in gaps too small for
the robot to pass through. The skeleton of non-occupied space is then found, treating unknown cells
as free cells. This ensures the skeleton does not have artifacts due to treating unknown regions as
obstacles. (c) The algorithm then determines the closest point on the local, free space skeleton to
the robot (the star in the image). It crawls the skeleton, following branches only until the skeleton
finds exits (circles in the image) or until branches terminate (squares in the image). (d) The skeleton
(prior to final pruning) is simply a subset of the original skeleton examined by this skeletal crawl.

113
boring non-free cells, essentially finding the collection of all such cells that have a path to
the edge of the LPM via non-free cells. All cells not in this set are set to free. This is an
effective and efficient “island” removal method for LPMs.
Obstacles that move can be detected by simple extensions to the normal occupancy
grid [Modayil and Kuipers, 2004]. If previously free cells become occupied, then obsta-
cles that reside in those cells can be classified as “moveable”, and the occupied cells can
be changed to free before creating the skeleton. For example, if the robot notices a door
closing, it will classify the cells occupied by the closed door as free for purposes of creating
the skeleton. Future work on visual obstacle detection should allow us to remove stationary
doors, furniture, and pedestrians from the LPM, while keeping walls and other architecture.
Finally, we “bleed” obstacles by a safety-radius—usually half the width of the robot
plus some small safety margin (Figure 6.3(b)). This closes small, irrelevant gaps between
obstacles, which reduces spurious skeletal branches and helps speed up post-processing
steps like pruning. This step might seem unnecessary at first glance since branches near to
obstacles can be pruned after the skeleton is fully computed. However, we have found this
step convenient for two reasons: (1) it can be added as a side effect in the initialization of
a discrete skeletonization algorithm, so it does not greatly increase the overall runtime of
the algorithm; (2) it serves to smooth the discretized skeletons, as the “bled” obstacles have
softer, less noisy corners.

Frontier-based Skeleton Pruning

The RGVG [Choset and Nagatani, 2001] was introduced as a way to eliminate some of the
“weak” junction points caused by spurious branches; however, the algorithm cannot ensure
all spurious branches are removed. The problem with pruning the Voronoi graph is that it
is often difficult to determine what branches are part of the necessary, “base” skeleton and
which branches are non-essential. In order to determine the necessary portions of a Voronoi
skeleton, we again leverage the fact that the robot is computing the skeleton of a bounded

114
local perceptual map (LPM). We use the frontiers in the LPM to determine which branches
to keep, ensuring that all spurious branches are removed.
Given the small, manageable size of the LPM, the robot should never find itself
completely enclosed by occupied cells in large-scale environments. Exceptions could be
elevators or small rooms where doors get closed. In these instances, our LPM should detect
the dynamic doors [Modayil and Kuipers, 2004], allowing the algorithm to ignore these as
obstacles for the purposes of building the skeleton. The frontiers at the edge of the LPM or
at the boundary between free and unknown regions of the LPM allow us to easily prune the
graph, keeping only useful branches in the LPM.
The algorithm for pruning the skeleton in the LPM is a straightforward three step
process. First, the algorithm determines the closest skeletal point to the robot that has a
path to the robot through free space—here, it does not consider portions of the skeleton that
overlay onto unknown regions (gray cells). Determining this closest point is necessary as
occasionally multiple non-connected branches may exist after skeletonization: e.g., cases
where older portions of the map are fully separated from the locally reachable free space
by obstacles. We use this closest point to provide information on which of the disjoint,
non-connected branches must be kept during pruning.
The algorithm then “crawls” the graph from the closest skeletal point, looking for
exits (circles in Figure 6.3(c)) and for locations where the skeleton branches terminate
(squares in Figure 6.3(c)). Exits are points on the skeleton that neighbor the edge of the
LPM or neighbor unknown space (essentially skeletal points that lie on a frontier). The
algorithm stops crawling a branch when it detects exits, as graph points in unknown regions
will be pruned away. If we consider only the points considered by this crawling algorithm,
the pruned skeleton would be the graph of connected points that overlay free cells in the
local map (and have a path through free space to the robot’s location) (Figure 6.3(d)).
There may be some spurious branches that do not terminate at exits. Instead of
actively pruning away these branches (which is difficult, as it is hard to determine how far

115
to prune branches), our approach is to compute the minimal spanning tree (MST) within the
skeleton between all exits in the LPM (Figure 6.4). We keep the MST, removing any other
graph points from the skeleton. This ensures that no spurious junctions or branches exist in
the LPM. In the case of dead-ends, only a single exit exists. In this special case scenario,
the MST that connects the exit with the closest skeleton point to the robot is used (Figure
6.4(f)).

6.1.2 Defining the LPM Core

Once a fully pruned skeleton of free space is found, the gateway algorithm must look for
locations to put gateways. As part of this process, we must first define the base points and
the core of the local area, using the pruned Voronoi skeleton. Simply stated, we define a
set of base points nearby the robot (usually neighboring Voronoi junctions), and the core is
the set of skeletal points within the radius of the base points. In Algorithm 6.1, we formally
describe the method by which we determine the core of the LPM.
The core serves several purposes. It gives a coarse approximation of the area the
place encompasses. Similarly it gives an environment-based method for defining when to
aggregate together multiple areas into a single, complex place and when nearby areas should
belong to different places. Finally, it gives rough locations (near the edge of the core) to
start searching for gateways. Figure 6.5 illustrates the concept of the core of an LPM with
some examples.

6.2 Constriction-based Gateways

Once the robot has generated a minimal skeleton of the LPM and has computed the core
of the local region, it can begin to look for gateways along the skeleton. Our first attempt
at determining gateways examines the skeleton near the edges of the core, looking for con-
strictions. Algorithms 6.2 and 6.3 detail the process that finds gateways by looking for
constrictions in specific regions of the skeleton. In this work, we define a constriction on a

116
 (a) (b) (c)

(d) (e) (f)

Figure 6.4: Pruning a Voronoi graph using the LPM frontiers. Pruning is performed by keeping the
minimal spanning tree that connects all exits of the skeleton. (a,b,c) Examples of Voronoi graphs
at common location. (d) The final pruning step for the skeleton from Figure 6.3. This removes
portions of the skeleton that do not connect exits. (e) Skeleton algorithms that are sensitive to noise
may create many branches, even in qualitatively simple LPMs. These spurious branches are properly
pruned using the LPM exit method. (f) Dead-ends are a special case where we keep the minimal
branch that connects the single exit to the skeletal point closest to the robot.

117
 l is the location of the physical robot w.r.t. the LPM.
P is the set of points that constitute the pruned (MST) skeleton.
dist(x, y) is the Euclidean distance between two points.
c ∈ P is the skeletal point closest to l that also has a path through free space to l.
(Re-computed after finding the MST that connects the skeletal exits).
r p is the “Voronoi radius”.
(The distance from Voronoi point p ∈ P to the closest obstacle).
J is the set of Voronoi junctions j.

Define the base points:

Two junctions j and j0 are neighbors if dist( j, j0 ) ≤ max(r j , r j0 ).
Determine the nearby junctions K = { j ∈ J : dist( j, c) ≤ r j }.
(c is within the “radius” of the junction point).
Define the base points F of the place neighborhood as the equivalence class of
neighboring junctions that uses K as the starting set of junctions.
(asterisks in Figure 6.5)
If F ≡ 0,/ then F = {c}.
(If no nearby junctions exist, use c as the base point.)

Define the core:

Define the core as the set of skeletal points Q = {q ∈ P : ∃ f ∈ F [dist( f , q) ≤ r f ]}.

Algorithm 6.1: Finding the core of an LPM from a pruned free space skeleton.

118
 (a) (b)

(c) (d)

Figure 6.5: Examples of the base points and the core of different local environments. The robot is
always assumed to be exactly in the center of the images and is shown as a circle. Skeletal points
that define the base points (Voronoi junctions at places and the closest point to the robot along paths)
are shown as asterisks. Skeletal points that lie inside the core are shaded lighter than skeletal points
outside the core. For visual clarity, the figures also shade non-skeletal points within the radius of the
base points.

119
 skel dist(x, y) is the graph distance between two skeleton points.
on path(w, x, y) determines whether point w lies on or between points x and y on the
skeleton.
(Because the pruned skeleton is an MST, cycles cannot exist, so this is a
deterministic function.)

Define the gateway search spaces:

(This assumes functions and sets defined in Algorithm 6.1.)
Define the set of skeletal points on the edge of the core
E = {e ∈ Q : ∃ f ∈ F [dist( f , e) = r f ∧ ∀ f 0 ∈ F [dist( f 0 , e) ≥ r f 0 ]]}.
For each e ∈ E, define a search space on the skeleton branch
Se = {s ∈ P : dist(e, s) ≤ re ∧ ¬on path(c, s, e) ∧ ∀ j ∈ J [¬on path( j, s, e)]}.
(Collect points close to the core edge. Unlike with constrictions, these can be inside
the core. Junctions, and points separated from the core edge by junctions, are
excluded. Likewise, the closest skeletal point to the robot c and points separated
from the core edge by c are excluded.)

Find a gateway location hypothesis for each search space Se :

Define the set of minima Me = {m ∈ Se : ∀s ∈ Se [rm ≤ rs ]}.
he ∈ Me is a constriction, where ∀m ∈ Me [skel dist(e, he ) ≤ skel dist(e, m)].
The set H = {he } hypothesizes the locations of the gateways.

Algorithm 6.2: Finding constrictions of an LPM. We assume the robot already has
computed base points and a core of the LPM. See Algorithm 6.1.

set of skeleton points P as a minimum in the Voronoi radii R = {r p : p ∈ P} (i.e., the same
rule that defines the “critical points” of Thrun and Bücken [1996]). At constrictions we
define gateways as line segments that connect nearby obstacles.
Thrun and Bücken [1996] specify the size of regions in which to look for constric-
tions. In order to create a scale/environment independent method, our algorithm’s search
areas are a function of the skeleton of free space. We define the portions of the skeleton that
lie near the core edge, within the Voronoi radii of the skeleton points on the edges of the
core, as the constriction search areas. Figure 6.6 illustrates the need to limit the search areas
for constrictions, as detailed in Algorithm 6.2. Looking too far away from the core can find
local minima that are well behind the place neighborhood boundaries that we would like
to define, reducing there usefulness in defining the boundaries and qualitative structure of
places.

120
 Calculate the gateway line segment for each gateway location he :
(This assumes functions and sets defined in Algorithms 6.1 and 6.2.)
ne = he .
REPEAT
(In general, the two closest obstacles to a skeletal point are not necessarily collinear
with the point. This both forces collinearity and finds the exact Euclidean
constriction.)
Find the line segment ze that connects the closest two obstacles on each side of
the skeleton branch to he .
Determine the skeletal point n̂e ∈ P where ze intersects the skeleton.
IF ¬on path(c, n̂e , e)
(Make sure the robot is not between the proposed gateway location and the
edge.)
THEN ne = n̂e .
UNTIL ne becomes stationary.
Define the “inner-neighbors” Te = {T ∈ P : dist(ne ,t) ≤ rne ∧ ∀ f ∈
F [skel dist( f ,t) < skel dist( f , ne )] ∧ ∀ j ∈ J [¬on path( j,t, ne )]}.
(Collect points close to the gateway location, but closer to the core than the gateway
location. Junctions, and points separated from the gateway location by junctions,
are excluded.)
Compute the headings Be from ne to each point t ∈ Te , and the distances
De = {dist(ne ,t) : t ∈ Te }.
Define the “outer-neighbors” Te0 = {t 0 ∈ P : dist(ne ,t 0 ) ≤ rne ∧ ∀ f ∈
F [skel dist( f ,t 0 ) > skel dist( f , ne )] ∧ ∀ j ∈ J [¬on path( j,t 0 , ne )]}.
(Collect points close to the gateway location, but further from the core than the
gateway location. Junctions, and points separated from the gateway location by
junctions, are excluded.)
Compute the headings B0e to ne from each point t ∈ Te0 , and the distances
D0e = {dist(ne ,t) : t ∈ Te0 }.
Calculate the weighted average ve across {Be ∪ B0e }, where {De ∪ D0e } provides the
weights.
Generate a line segment ge , centered at ne , of size 2 · rne (twice the Voronoi radius).
The orientation of the segment is normal to ve .
The set G = {ge } represents the gateways of the LPM.

Algorithm 6.3: Finding constriction-based gateways. We assume the robot already

has computed search areas and constrictions near the edges of the LPM core—see Al-
gorithm 6.2.

121
 (a) (b)

(c) (d)

Figure 6.6: Defining the search space for constriction-based gateways. Here we show constriction-
based gateways on the environment used in Figure 5.1(a). (a,b) We find reasonable constriction-
based gateways when using the search areas we define in Algorithm 6.2. (The search areas are
within the Voronoi radii of points at the edge of the core, which is lightly shaded). (c,d) Suppose we
do not limit our search areas to be near the edge of the core (but we still limit the search area to not
include junctions). The danger here is that we can find minima far away from the core that do not
adequately represent the place boundaries.

122
 Once “critical points” are found, Thrun and Bücken [1996] simply connect the two
closest obstacles that define the Voronoi skeleton at that point. In our algorithm this would
be akin to defining gateways as the set {ze } from Algorithm 6.3. However, because noise
and obstacles can create constrictions, it is possible that the line segments z have strange
orientations with respect to the actual path. Figure 6.7(a,b) illustrates how this simple rule
can fail, creating gateways that could potentially be unaligned along a path due to small
obstacles near the path boundaries. This motivates the need for a more complex method
for determining line segments G from {ze } that is described in Algorithm 6.3—essentially,
neighboring skeletal points are used to determine a reasonable orientation of the path at the
gateway location (Figure 6.7(c,d)).

6.2.1 Discussion

This constriction-based gateway algorithm works reasonably well in indoor office environ-
ments. A recursive version of this algorithm was implemented that runs quickly enough
to recompute gateways in real-time—averaging ∼2 Hz for a 300x300 cell occupancy grid
(usually we use 15 meter grids with 5 cm cells) on a Pentium III 450 MHz research robot
computer that is also running the computationally intensive SLAM algorithm to maintain
the underlying LPM. Examples of gateways using this implementation are shown in Figure
6.8.
This implementation was also used to generate the gateways used for autonomous
place detection and description in our large-scale map-building environment from Figure
2.1; however, we used an artificially large robot radius in computing the skeleton to prevent
skeletal branches between office doors and cubicle opening. This allowed the robot to detect
gateways in hallways and intersections, but prevented our robot from exploring people’s
offices or desk areas. Chapters 8 and 9 illustrate global topological and metrical map-
building using places and local topologies grounded using this gateway implementation;
Figure 8.2 shows the places detected by using this constriction-based method.

123
 (a) (b)

(c) (d)

Figure 6.7: Defining gateways by connecting nearby obstacles. (a,b) A simple rule once a con-
striction is found is to simply connect the closest obstacles on both sides of the skeleton. This is
easy to implement but often leads to gateways that become oriented with respect to noisy walls or
obstacles against walls. In image (a), the simple path continuity criteria (Section 5.2.2) fails be-
cause of misaligned gateways. (c,d) Algorithm 6.3 gives a more complex way to determine the line
segment once a constriction is found. This method yields very good results in practice, resulting in
high-quality local topology abstraction. Figure 6.8 provides more examples of constriction-based
gateways as defined by Algorithms 6.2 and 6.3.

124
 In Table 2.1, we compared a variety of gateway approaches, and stated that our
constriction-based approach performs well when considering a variety of desired properties.
First off, this approach is relatively scale independent. Suppose we have a 15x15 meter
LPM with 5 cm cells that models a T intersection of 3 meter wide hallways. Suppose
we also have an equivalent grid that is actually a 150x150 meter LPM with 50 cm cells
that models a T intersection of 30 meter hallways. We expect to see very similar Voronoi
skeletons, cores, constrictions, and gateways. That is, our approach does not have built-in
thresholds about hallway size. Additionally, constrictions are well defined at doorways and
at the beginnings of hallways, so this approach finds the kinds of gateways we want in the
locations we expect. Because this approach uses a Voronoi skeleton, we also do not assume
rectangular intersections, and can find gateways in a variety of intersections. This algorithm
also works well in defining gateways along paths.

6.2.2 Problems with the Constriction Algorithm

The constriction approach works well in most indoor environments, which makes it equal to
or better than previously published gateway approaches discussed in Section 2.3; however,
as implied by Table 2.1 in certain situations, we found the constriction-based locations of
gateways to be unreliable. Below we illustrate these problems with specific examples.
Figure 6.9 illustrates issues that occur at places where no junctions exist (i.e., L
intersections). In these situations, constrictions can exist on portions of the skeleton where
the branch is about to curve sharply. As a result, our method for determining the orientation
of gateways can create strangely oriented gateways.
The strange orientation issue is mostly aesthetic, as the local topology derived from
the gateways in Figure 6.9(b) does in fact correspond to an L intersection; however, we
still consider this an issue to be fixed, as it may impact the comparison between two places
(Section 9.3.1 discusses how we can use gateways to quickly align two LPMs in creating
a globally consistent metrical map). Strangely oriented gateways may also affect human-

125
 (a) (b) (c)

(d) (e) (f)

Figure 6.8: Examples of constriction-based gateways in a variety of (real-world and pathological)

environments.

126
 (a) (b) (c)

Figure 6.9: Unstable gateways near L intersections. (a) Here we see a strange gateway orientation
because the constriction lies close to the sharp bend in the skeletal branch. (b) Here we see another
strange gateway after to robot has moved further into the intersection. Here the local topology
corresponds to an L intersection. (c) Once the robot moves further into the interior of the place,
gateways look more appropriate, showing that gateways at L intersections change with respect to the
robot’s location in the place.

robot interactions in future work.

More important than gateway orientation is the issue that the place gateways change
as the robot moves through the place. This has the potential to create mapping problems as
the boundaries of places change with respect to the robot’s location in the place. In Figure
6.9(b), the robot is at a place, and chooses a gateway to pass through to begin navigation
along a path. Once it passes through the gateway, the robot detects that it is still at a
place (Figure 6.9(c)). At this point it will consider this another place in the HSSH Global
Topology Level. In order to overcome this problem, we must add special knowledge to
handle such a scenario for L intersections, which is undesirable.
One might think that changing the search area to only look for gateways outside the
core might fix these problems. Figure 6.10 demonstrates this change to the algorithm. We
see that even here, we can get strangely oriented gateways, and that the boundaries of the
place can be unstable. Even worse, with this change to the search area, the robot will detect
places, even when it is well into a hallway (Figure 6.10(a,c)).
Additionally, once we extend our skeleton to handle coastal navigation (wall-following)

127
 (a) (b) (c)

Figure 6.10: Unstable place boundaries with constriction-based gateways. Here we show the
results of looking for constrictions only outside of the core. (a) We still see strangely oriented
gateways as in Figure 6.9(a,b). (b) Here the robot is at a place with reasonable gateways. At this
point, if the robot decides to leave the place via the gateway on the right of the image, it uses the
gateway as a local motion target (Section 5.4). (c) After the robot completes the local motion to pass
through the rightmost gateway from image (b), the robot finds itself still at a place. At the HSSH
Global Topology Level, the robot must try to determine whether this is a new place or the same
place. A new problem arises when looking outside the core, as the robot considers this a place even
though it is well into the hallway.

scenarios (Section 6.3.1), constrictions will not exist (in theory)—points along the skeleton
will be equidistant from the nearest obstacles. However, for discretized skeletons, con-
strictions can exist, though they are not visible because the change in radius is extremely
small. These constrictions are artifacts and may be at locations unsuitable for gateways
(Figure 6.11(a)). When no constriction is found in a search area along a branch, a gate-
way is located at the edge of the core by default. Here, the gateways become dependent
on the robot’s location, and we observe problems of badly oriented gateways and unstable
gateways at places throughout coastal navigation (Figure 6.11(b)).

6.3 Current Gateway Algorithm

The intuition behind the constriction-based method is that the architecture in the environ-
ment creates “cinch points” or constrictions in the Voronoi radii near doorways and at the
intersection of hallways. As we demonstrated above, this works well in the majority of situ-

128
 (a) (b)

Figure 6.11: Unstable gateways in coastal navigation. In coastal navigation scenarios, using the
extended Voronoi graph (Section 6.3.1), there are theoretically no constrictions. (a) However in
discretized skeletons, small local minima do arise, which can create poor gateway locations. (b) In
cases where no gateways exist, the gateways are places at the core edges by default. Thus, in coastal
navigation of convex corners, the problems shown in Figures 6.9 and 6.10 occur often. We introduce
anchor-based gateways as a solution to these problems.

129
ations, but constrictions do not always lead to nicely oriented gateways or stable gateways.
This fact becomes magnified in coastal navigation scenarios, where no constrictions exist
and gateways are simply placed at the edge of a core that scrolls with the robot (or artifact
constrictions from the discretized skeleton create incorrect gateways).
The fundamental problem with using constrictions is that we are searching a 1D
encoding of free space, with Voronoi radius annotations, for signals about the architecture
instead of looking at the architecture itself. Below we present a method that looks at the re-
lation between the skeleton and the local architecture. Our approach maintains the positive
aspects of our constriction-based method (non-rectangular intersections are handled and no
environment dependent metrics are needed), while also handling the problematic cases of
constriction-based gateways.
Specifically, we motivate this new approach to handle coastal navigation scenar-
ios, where walls exist on only one side of the path. So, we begin the discussion of the
architecture-influenced, anchor-based gateways by defining a new skeleton that facilitates
wall-following. We then detail a new gateway algorithm that allows us to define places at
corners in coastal navigation scenarios, where no Voronoi junction exists, and the skeleton
is a constant distance from nearby obstacles.

6.3.1 The Extended Voronoi Graph (EVG)

Although the Voronoi graph is well defined in corridor environments, the Voronoi graph no
longer becomes a useful skeleton when the robot transitions into non-corridor environments.
This is illustrated in Figure 6.12(a,b). Here, the robot moves into a room that is larger than
the LPM size. At this location, there exists a large qualitative change in the environment,
but the Voronoi graph simply ignores this fact.
Two related problems exist in this example. First, no place is defined. This is true
of any skeleton-based approach to place detection. This place is ignored by junction-based
approaches to place detection, and by our place detection scheme. The related problem is

130
that of navigation control, which depends either on the actual skeleton for Voronoi-based
navigation approaches or on target gateways in our control schemes (Section 5.4). The
appropriate control in this scenario is coastal navigation, keeping at least one obstacle in
sight at all times [Kuipers and Byun, 1991; Roy et al., 1999]. Instead of defining paths
parallel to the walls, the Voronoi graph defines a path that will quickly bring the robot to a
location where no visible obstacles exist. Once this occurs, the Voronoi graph can no longer
be computed from local sensor data, and the robot is simply lost.
To handle coastal navigation scenarios, the robot needs a different skeleton than the
traditional Voronoi graph—one that seamlessly transitions from a “midline” skeleton to a
“wall-following” skeleton when necessary (Figure 6.12(c)). In the example, the natural and
meaningful spatial configuration includes a place at the opening and gateways that inform
the robot that it can perform travel actions by leaving the place through the bottom gateway
or by leaving through the gateways to the left or right (Figure 6.12(d)). To obtain a graph
that exhibits the appropriate behavior, we introduce an extension to the 2D version of the
generalized Voronoi graph.
We define the extended Voronoi graph (EVG) as the subset of all points in the 2D
GVG closer than a threshold M units from any obstacle, combined with the set of all points
exactly M units away from the closest obstacle. More formally, given that the GVG in a
planar model is the set of all points p equidistant from two or more closest obstacles O, we
define the set of points P in the EVG as follows:

P = {p : ∃o, o0 ∈ O [o 6= o0 ∧ dist(p, o) = dist(p, o0 ) ∧

dist(p, o) ≤ M ∧ ∀o00 ∈ O [dist(p, o00 ) ≥ dist(p, o)]]}

∪ {q : ∃o ∈ O [dist(q, o) = M ∧ ∀o0 ∈ O [dist(q, o0 ) ≥ M]]},

where M is a preset maximum threshold. Because LPMs limit the sensory horizon of the
robot, and we want deterministic place detection regardless of how the robot enters a place,
it is practical to make M = side size(LPM)/c, where c ≥ 4: e.g., for the 300x300 cell LPMs

131
 (a) (b)

(c) (d)

Figure 6.12: Entering a room that is larger than the LPM. Voronoi graphs are not useful in non-
corridor environments due to the robot’s limited sensory horizon in the LPM. (a) Here, the robot
should begin to follow a wall, but the graph continues into unknown territory. (b) Gateways, which
are defined using the skeleton, simply define a single, local-path, forcing the robot to continue travel-
ing forward until all obstacles disappear from the LPM. At this point, the robot will be lost. (c) Here
the extended Voronoi graph (EVG) defines a transition from corridor-following to wall-following,
which is what we want. (d) Even constriction-based gateways work here, properly defining this as a
T intersection.

132
Figure 6.13: The extended Voronoi graph (EVG) of a global metrical map. Using an occupancy
grid, the extended Voronoi graph can be drawn by using any occupied cells as obstacles. Given a
reasonable maximum distance from wall, the graph now approximates “coastal navigation” routes
in large rooms. Compare to Figure 2.3. Note: This image illustrates the idea of a coastal navigation
graph. This is not an LPM, thus no pre-processing of the grid and no pruning of the skeleton are
performed.

we use in this work, c = 5 and M = 60 cells.

Although this extension to the traditional Voronoi graph is simple, we were unable
to determine a sub-quadratic algorithm for computing the continuous EVG. The continuous
EVG of the same pre-computed global metrical map from Figure 2.3 is shown in Figure
6.13. For discretized brush-fire or thinning approximations, it is straightforward to keep
track of the distance to the nearest obstacle when deciding whether to flip a cell and simply
stop the “fire” or “wavefront” at that point. The linear time thinning approximation of
the EVG, which we utilize in this work, is illustrated in Figure 6.14. Given this skeletal
approximation of the EVG, the same LPM pruning discussed in Section 6.1.1 can be used
to approximate the reduced extended Voronoi graph (REVG).

6.3.2 Anchor-based Gateways

Given a minimal skeletal of free space that adequately handles coastal navigation scenarios,
we need a gateway implementation that can overcome the problems of the constriction-

133
Figure 6.14: The thinning approximation of the EVG. The thinning algorithm can be extended to
approximate the EVG skeleton in an efficient fashion. Compare to Figure 6.13.

based method to give stable, reliable gateways in all situations. Here we detail our anchor-
based method which improves upon the state-of-the-art.
The intuition behind our anchor-based gateway approach is straightforward. In cor-
ridor environments, each skeletal point has at least two nearest obstacles (one on the left of
the branch and one on the right) that define that point as belonging to the skeleton. At con-
strictions like doorways, these nearest obstacles lie at the gateway line-segment endpoints.
Examining these constriction-defining obstacles, we notice that they not only determine the
skeletal point at the constriction, but they also help define many other nearby skeletal points.
This is illustrated in Figure 6.15(a). However, we see in Figure 6.15(b) that in coastal navi-
gation scenarios there also exist points that contribute heavily to defining the skeleton. We
term these points anchor points.
The idea of our anchor-based gateway method is to find these anchor points and
use them to help define reasonable gateway locations. Algorithm 6.4 details the process
for finding anchors from a pruned skeleton of free space and nearby obstacles. It defines
search areas near the edge of the core in the same fashion as the constriction method. The
obstacle that contributes to defining the most skeletal points in this search area is chosen at

134
 (a) (b)

Figure 6.15: Intuition behind skeletal anchors. (a) Constrictions occur where the skeleton gets
close to obstacles (Algorithm 6.2). The nearest obstacles to skeletal points at constrictions (circles
in the image) are also the closest obstacles to many other skeletal points (the emphasized portions
of the skeleton). These important obstacles will be used as anchor points to help determine gateway
locations. (b) In coastal navigation scenarios, we also notice that there exist obstacles that define
many skeletal points. These anchor points lie at locations where we expect gateway line segments
to end.

135
 O is the set of obstacles in the LPM.
le f t o f (s, o) determines whether an obstacle o ∈ O is on the left side of the branch
containing s ∈ Se at point s.
(“Left” is determined calculating the orientation at s at its neighboring point that is
further from the edge of the core e.)
closest obs(s, o, O) determines whether o is the closest obstacle to s in O.
(o ∈ O ∧ ∀o0 ∈ O [dist(s, o) ≤ dist(s, o0 )])

Define the gateway search spaces Se :

(Use the same process from Algorithm 6.2.)

Find the anchor for each search space Se :

Define the obstacles that are to the left of the points in the search area
O0e = {o0 ∈ O : ∃s ∈ Se [le f t o f (s, o0 )]}.
Define the obstacles that are to the right of the points in the search area
O00e = {o00 ∈ O : ∃s ∈ Se [¬le f t o f (s, o00 )]}.
Count the number of times each obstacle o0 ∈ O0e is the closest left obstacle to the
points s ∈ Se , and create a similar count for right obstacles O00e .
Determine the anchor ae ∈ {O0e ∪ O00e } that is the obstacle with the highest count.
(This obstacle is the closest left or right obstacle to the most skeletal points in the
search area. If multiple obstacles are valid, use the one closest to the core edge e.)
The set A = {ae } defines the anchors of the LPM.

Algorithm 6.4: Finding anchors of an LPM. We assume the robot already has com-
puted base points and a core of the LPM—see Algorithm 6.1, which applies even to the
extended Voronoi graph.

the anchor for that branch. An illustration of this search process is given in Figure 6.16.
Once anchors are found, the next step is to determine reasonable gateway location
hypotheses. In the constriction-based algorithm, the closest obstacles to constrictions were
used to create a line segment representation of a gateway. A naı̈ve approach might be to try
something similar with anchors, by looking for the closest skeletal point. However, in cases
where convex corners exist, multiple skeletal points are equidistant from an anchor. Sim-
ilarly, multiple branches may share an anchor point, so one anchor point might determine
multiple gateways (Figure 6.17).
Algorithm 6.5 details how we propose gateway location hypotheses on the skeleton
for four cases. These cases are illustrated by Figure 6.18. Basically, if the anchor point is
the closest “left” or closest “right” obstacle to the point at the edge of the core along the

136
 (a) (b) (c)

Figure 6.16: Searching for anchors. (a) The first step in finding anchors is to determine the core
of the LPM from the pruned skeleton. From the core, search areas are created that consist of points
that lie within the Voronoi radii of the points at the edge of the core (points separated from the core
edge by junctions or the closest point to the robot are excluded). (b) Next, the algorithm determines
the closest obstacles to the skeletal points in the search areas. (c) For each search area, the obstacle
involved in defining the largest number of points is found. That is, the anchor.

(a) (b)

Figure 6.17: Anchors and gateways in coastal navigation scenarios. (a) Anchors where a wall-
following path intersects a corridor-following path. Each of the three search areas have an anchor.
Sometimes the same anchor point is found for multiple branches. (b) From two anchor points, we
get three distinct gateways. This illustrates that a simple rule cannot be used to find the location of
the gateways with respect to the anchors.

137
branch, the algorithm looks for skeletal points in the search area where the anchor is not
the closest left or right obstacle. Skeletal points outside the core are examined only if none
exist inside the core that meet the above criterion. Of these points the closest to the edge is
used as the gateway hypothesis. However, if the anchor point is not the closest left or right
obstacle to the edge point, then the algorithm looks for a skeletal point in the search area
where this is the case. Again points in the interior of the core are preferred, and of these,
the one closest to the edge is chosen.
Once gateway location hypotheses are determined, the same process from Algo-
rithm 6.3 may be used to find the line segment that represents the gateway. This is only
true for cases where two obstacles define the skeleton at the gateway hypotheses. In coastal
navigation scenarios (where the EVG creates skeleton points based on only one nearby ob-
stacle), the gateway line segment is simply defined at the hypothesized location, with an
orientation determined by the line that connects the gateway location point with the anchor.

Discussion

This anchor-based gateway algorithm works as intended. The algorithm is simple to imple-
ment, and does not add any noticeable compute time to the real-time robot navigation: local
topology can still be updated at approximately 2 Hz for a 300x300 cell occupancy grid. In
fact, finding the skeleton (including pre-/post-processing) is the main bottleneck in the local
topology computation.
Examples of gateways using our implementation are shown in Figure 6.19 and
can be compared to the constriction-based gateways of Figure 6.8. Comparing examples
of constriction-based gateway with anchor-based gateways, we notice that anchor-based
gateways of corridor intersection are virtually identical to the gateways we obtain with
constriction-based gateways. The cases we used to show the constriction problems in Fig-
ures 6.9, 6.10, and 6.11 are shown to be solved by the anchoring solution in Figures 6.20
and 6.21.

138
 Find a gateway location hypothesis for each search space Se and anchor ae :
(This assumes functions and sets defined in Algorithms 6.1 and 6.4.)
IF closest obs(e, ae , O0e ) ∨ closest obs(e, ae , O00e )
(ae is either the closest left obstacle or the closest right obstacle to edge point e.)
THEN
Find the set of search points inside the core whose closest left obstacle and
closest right obstacle is not the anchor
Ie = {i ∈ Se : [¬closest obs(i, ae , O0e ) ∧ ¬closest obs(i, ae , O00e )] ∧ ∀ f ∈
F [skel dist(i, f ) < skel dist(e, f )].
IF Ie ≡ 0/ THEN
Find the set of search points outside the core whose closest left obstacle and
closest right obstacle is not the anchor
Ie = {i ∈ Se : [¬closest obs(i, ae , O0e ) ∧ ¬closest obs(i, ae , O00e )] ∧ ∀ f ∈
F [skel dist(i, f ) > skel dist(e, f )].
IF Ie ≡ 0/ THEN
Ie = {i ∈ Se : ∀s ∈ Se [skel dist(i, e) ≤ skel dist(s, e)]}.
(Choose the point in Se closest to the core edge.)
ELSE
(ae is not the closest left obstacle and not the closest right obstacle to e.)
Find the set of search points inside the core whose closest left obstacle or closest
right obstacle is the anchor.
IF Ie ≡ 0/ THEN
Find the set of search points outside the core whose closest left obstacle or
closest right obstacle is the anchor.
he ∈ Ie is the gateway location hypothesis, where
∀i ∈ Ie [skel dist(e, he ) ≤ skel dist(e, i)].
(Use closest point in Ie to edge e.)
The set H = {he } hypothesizes the locations of the gateways.

Calculate the gateway line segment for each gateway location he :

IF ∀o0 ∈ O0 [dist(he , o0 ) > rhe ] ∨ ∀o00 ∈ O00 [dist(he , o00 ) > rhe ]
(coastal navigation scenario)
THEN
Generate a line segment ge , centered at he , of size 2 · rhe (twice the Voronoi
radius). The orientation of the segment is the same as the orientation between
ae and he .
The set G = {ge } represents the gateways of the LPM.
ELSE
Calculate gateway line segment ge from he using the same process as the
constriction-based algorithm (Algorithm 6.3).

Algorithm 6.5: Finding anchor-based gateways. We assume the robot has already
determined the search areas and anchors near the edges of the LPM core—see Algo-
rithm 6.4. Once hypothesis gateway locations are determined, gateway line segments
are computed in the same fashion as in Algorithm 6.3 for non-coastal-navigation situa-
tions.

139
 (a) (b)

(c) (d)

Figure 6.18: Finding gateway location hypotheses from anchors. (a) If the anchor is the closest
obstacle to the core edge point, crawl the graph from the edge into the core looking for a skeletal
point where this is not true. (b) If no points are found in (a), crawl the graph from edge away
from the core looking for points where this is not true. (c) Sometimes, the anchor is not the closest
obstacle to the edge of the core. In this case, crawl the graph into the core looking for a point that
does have the anchor as its closest point. (d) Otherwise, crawl the graph away from the core, looking
for a point where the anchor is becomes the closest obstacle.

140
 (a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 6.19: Examples of anchor-based gateways in a variety of (real-world and pathological)

environments. Compare (a-f) to Figure 6.8. Compare (g-i) to Figures 6.6(b), 6.7(c), and 6.7(d)
respectively.

141
 (a) (b) (c)

Figure 6.20: Stable anchor-based gateways near L intersections. Compare with Figures 6.9 and
6.10.

(a) (b)

Figure 6.21: Stable anchor-based gateways at coastal L intersections. Compare to Figure 6.11.

142
 Though the anchor-based solution solves the gateway stability problem illustrated
in Figure 6.10 for L intersections where a convex corner exists, we have observed that
this stability problem is still an issue for coastal navigation at concave corners (Figure
6.22(a,b,c)). At noise-free concave corners, there is no obstacle to provide an anchor, so
the gateways are placed at the edge of the core by default, again causing the gateways of
a place to become dependent on the robot’s location in the place. A simple solution is to
change any free cells that are beyond the Voronoi radius of all pruned EVG skeletal points
into occupied cells. This is done after creating the pruned skeleton, but before looking for
anchors. This creates virtual obstacles on all side of the skeletal branches, and provides
anchors in situations where they are needed (Figure 6.22(d)). Performing this operation on
every LPM should work fine in theory (Figure 6.22(e,f)); however this operation is worst-
case O(n2 ) where n is the number of cells in the LPM (slower than the skeleton and the
gateway algorithm combined). Currently the addition of this small amount of code swells
local topology compute time in the full HSSH mapping suite from ∼2 Hz to ∼0.5 Hz in
coastal navigation scenarios using a 300x300 cell LPM. Finding an efficient solution to this
problem should be a reasonable short-term goal for future work.
Overall, the anchor-based gateway implementation works as intended. It generates
gateways similar to the constriction-based method, while overcoming the unstable issues
discussed in Section 6.2.2. This algorithm is reasonably independent of pose—at intersec-
tions with skeletal junctions, gateways do not change while the robot is within the Voronoi
radii of the base points, and at intersections without junctions, gateways are stable due to
using anchors. This algorithm is also independent of scale—though the robot’s radius and
the LPM size will determine the location of gateways, thus the location and description of
places, the algorithm does not expect any particular sized paths or doorways. Additionally,
we have shown examples where this algorithm produces stable gateways in non-rectangular
intersections and coastal navigation situations.

143
 (a) (b) (c)

(d) (e) (f)

Figure 6.22: Handling environments without good anchors. Concave corners are examples of
places where good anchors or constrictions may not exist. (a,b,c) Here, the robot’s gateways are
unstable because they are placed at the core edges, due to a lack of useful information. (d) By
changing all free cells that are outside the Voronoi radii of all skeletal points in the pruned EVG into
occupied cells, we can create virtual obstacles that provide anchors. Compare to image (a). (e,f)
This does not unduly affect gateways in other coastal navigation situations or non-coastal, corridor
environments.

144
6.3.3 Evaluating Anchor-based Gateways

The anchor-based gateway detection and subsequent local topology abstraction have proven
to be quite reliable in our experience. Chapter 7 details a human-robot interaction experi-
ment that uses this anchor-based gateway implementation to detect and describe places with
a simulated robot. The algorithm’s long-term reliability can be inferred from error-free per-
formance in this experiment. In the more than 360 places encountered and the hundreds of
thousands of LPMs examined during this experiment, there were no places missed or in-
correctly described in terms of local topology. Additionally, we had no false positive place
detections despite randomly placed stationary obstacles in the environment.
We have also experienced high-quality place detection and local topology abstrac-
tion on real-world robots that explore dynamic indoor environments using noisy sensors;
however no long-term quantitative data has been gathered. Samples of LPMs and local
topologies generated by our Magellan Pro mobile robot (Figure 2.1(b)), equipped with a
S ICK LMS lidar, are provided in Figures 6.23-6.26.
Figure 6.23 illustrates the quality of the gateway algorithm at places in a univer-
sity building. Gateways are reliable and stable at places despite multiple visits, entering
places from different directions, and having various LPM rotations with respect to the path
directions. We observe deterministic place detection at intersections in university hallways
and near-deterministic local topology abstraction. Figure 6.23(h,i) illustrates an example of
local topology ambiguity, due to a boundary case of the simple ray-casting gateway align-
ment operation (Section 5.2.2). Here a small shift in the LPM affects the skeleton enough
to change the detected local topology change from that of a T intersection to that of a Y
intersection.
Figure 6.24 shows the performance along paths. The algorithm handles hallways of
various sizes (Figure 6.24(a,b)) and even handles paths that contain “non-island” obstacles,
as seen in Figure 6.24(c-f). Figure 6.24(g,h) illustrates how the robot stops at a potential
place, but, after more deliberate examination of the local region, it determines that it is

145
 (a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 6.23: Gateway (and place) detection in a university hallway. (a,b) This 45◦ bend in the
hallway is seen as an L intersection. This place is detected by the robot when traveling in either
direction. (c) At this place, the Voronoi skeleton causes the rightmost gateway to be slightly angled,
creating a Y local topology. This is actually a reasonable assumption, especially since most people
navigating this environment actually turn left here. (d,e) & (f,g) We illustrate the stability of gate-
ways for the same intersection visited on separate occasions (and entered from different directions).
(h,i) Due to an almost imperceptible shift in the LPM, the topmost gateway in these two images
changes from being aligned with the lower gateway (image (h)) to being “left of” the lower gateway
(image (i)). This demonstrates that boundary cases in our simple gateway alignment criterion can
affect local topology abstraction.

146
 (a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 6.24: Traversing paths in a university hallway. (a,b) Our scale independent gateway al-
gorithm allows the robot to handle various sized paths. (c) Even when encountering “non-island”
obstacles, the anchor-based gateway algorithm produces aligned gateways that represent a single
path. (d-f) Here we see the same location as in image (c), but with the robot traveling in a different
direction. (g,h) Our robot can only see in front of it, so unmapped portions of the LPM may create
exits in the skeleton, which cause gateways. By examining potential places, the robot can remove
incorrect gateways to avoid false positive place detections. (Note: the lidar sees “through” the wall
a bit in image (h) because the metal legs of a chair reflect the instrument’s laser beams.) (i) Here
the rightmost gateway should not exist—smooth rubber tires slipping on the linoleum floor creates
a localization failure in the LPM that moves the hallway from its initial position.

147
 (a) (b) (c)

(d) (e) (f)

Figure 6.25: Detecting doorways in a university hallway. (a) Here the robot detects a place in front
of some double doors (the doorway is the middle gateway). Actually, the doors here are closed;
however, the robot saw them as open when it passed through them. Because the robot can remove
obstacles that reside in cells that it has previously seen as free, the doors are ignored by the gateway
detection algorithm. (b) A single open door to an office at an L intersection. (c,d) When passing this
alcove the first time, the robot did not consider it a place. The next time, someone had opened the
door, so the robot decided this is a place. (e,f) This is another scenario where the robot is traveling
past a door without stopping. Here, we stop the robot, and open the door. Upon seeing this, the
robot immediately decides that this is a place. (Notice the metallic chair legs cause the robot to
again assume free space beyond the wall—see Figure 6.24(h). The robot still sees the wall behind
the chair legs—notice the obstacles defining the wall still exist near the perceptual mistake—so no
gateway is detected at this spot.)

148
 (a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 6.26: Dealing with furniture. (a-d) In this computer lab, the furniture is mostly large desks
that divide the lab into distinct regions. The robot parses this space fairly well: detecting a T de-
cision, then a Y, then a small path in image (c), followed by another T. (e,f) When traveling in the
other direction, the robot detects the same places (compare images (d) and (e) and also images (b)
and (f)). (g) Desks and chairs often create occlusions that produce gateways. These false gateways
are often hard to eliminate because chairs prevent physical exploration under the desks. Such ex-
ploration is also time intensive, so it is not desired. We expect visual object recognition to help
overcome this issue. (h,i) Furniture with shiny surfaces can lead to false free space in the LPM (the
top of image (h) corresponds to the bottom of Figure 6.24(h)). If real free space intersects this false
free space, important pieces of architecture can be seen as “islands”, which affects gateways.

149
still on a single path. Figure 6.24(i) shows how a simple localization failure in the LPM
at the Local Metrical Level can be problematic at the HSSH Local Symbolic Level. A
reasonable piece of future work would be detecting and recovering from these localization
failure scenarios.
Figure 6.25 shows how the gateway algorithm handles doors. Figure 6.25(c-f) il-
lustrates how regions that are seen as paths when doors are closed are seen as places when
doors are opened. Not being able to detect closed doors with range sensors is a problem
that affects almost all modern-day indoor mobile robots.
Finally, Figure 6.26 shows how the algorithm behaves in rooms with lots of furni-
ture. Large furniture cannot be distinguished from architecture. Sometimes this does not
actually present a problem, like in Figure 6.26(a-f). However, many times furniture cre-
ates “frontiers” in the LPM, where multiple false gateways are proposed (Figure 6.26(g)).
Additionally, office furniture that has metallic legs or slick, black surfaces will reflect the
laser beams of the lidar device, creating false free space in the map. Once obstacles are sur-
rounded by free space, regardless of whether the LPM is correct or not, they are removed
for the purposes of computing gateways by the “island” removal pre-processing. Figure
6.26(h,i) shows a bad case where a wall is seen as an “island” obstacle because reflected
lidar measurements from metallic chair legs corrupt the LPM.
Because it is difficult to find non-rectangular or complex intersections in university
hallways, we also include snapshots from a simulated robot exploring an environment with
more complex places in Figures 6.27 and 6.28. Here we can verify the performance of local
topology abstraction in adjacent, non-rectangular places and places at sharp turns. In all
cases, we obtain deterministic, correct local topology abstraction.
Though we believe the figures in this chapter are compelling examples of our gate-
way algorithm, we want to thoroughly evaluate the anchor-based gateway algorithm to en-
sure it is robust to various conditions that we suspect could be problematic. Below, we
evaluate the robustness of the anchor-based gateway implementation under certain noise

150
 (a) (b) (c)

(d) (e) (f)

Figure 6.27: Gateway (and place) detection using a simulated robot. (a) We created a simulated
environment that approximates a 40 meter wide environment with complex and non-rectangular in-
tersections. We illustrate gateway detection and local topology abstraction using simulated robot
with low-noise actions and perception. (b) When starting, robots with only forward-facing obser-
vations will create gateways that may not be correct (here a single path). (c) Once the robot fully
observes its local surround, it abstracts the correct local topology. (d) The robot chooses to go
straight, and moves out of the place onto a path. (e) When the robot arrives at the next place, it
gets the local topology correct; however, the LPM is not fully explored. (The light asterisk in some
images denotes the local motion target of the robot in the LPM). (f) Only after filling the observa-
tions for the local region does the robot decide it is actually at a place. The robot’s exploration is
continues in Figure 6.28.

151
 (a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 6.28: Gateway (and place) detection using a simulated robot. (a) The robot turns right
in the previous place and arrives at the non-rectangular intersection in the bottom right of Figure
6.27(a). (b) Arriving at the first place again, the same local topology is found. (c) The complex
middle portion of image (a) is considered two places by our algorithm. (d) The robot travels left, to
the other T intersection. (e) It then travels down to the dead-end. (f,g,h,i) The robot makes it way
up to the top left of the environment before returning to the middle. It detects all non-rectangular
places deterministically and correctly, despite how it enters each place.

152
conditions by observing the performance of the local topology abstraction (which uses the
anchor-based gateway implementation and the simple gateway alignment criterion from
Section 5.2.2) as we alter a simple LPM in various ways. Specifically, we investigate the
issues of LPM rotations, LPM resolution/size, and uniformly distributed noise in order to
evaluate the sensitivity of the gateway algorithm to conditions known to affect the shape of
free space skeletons.

Evaluation Setup

To evaluate various properties of our anchor-based gateway algorithm, we created a corpus

of images that provide reasonable approximations of common LPMs. We used a high-
resolution image of a + intersection (Figure 6.29(a)) and created a corpus of 12000 images
from transformations of this initial image. We chose to use a + intersection for several
reasons: (1) these intersections are extremely common (consider road networks); (2) a + is
a sort of superset of both L and T intersections; (3) due to the symmetry, we only need to
rotate the image up to 45◦ to exhaust the rotational search space.
To generate test images, we first rotated the image from 0◦ to 45◦ in 5◦ increments,
creating 10 possible rotations (Figure 6.29(b)). As noted several times in this dissertation,
thinning skeletons are rotation dependent. By examining how rotation affects the accuracy
of local topology abstraction, we can determine whether using efficient, thinned skeletons is
reasonable or whether we should be utilizing continuous Voronoi diagrams that are rotation
independent.
Next, we scaled the rotated images from high-resolution images to 60 different res-
olutions from 9x9 cells to 599x599 cells in 10 cell (per side) increments (Figure 6.29(c)).
Because the gateway detection process is linear in the number of cells, doubling LPM sizes
(in the number of cells per side) increases the time of the local topology abstraction quadrat-
ically. For many robotic platforms, a minimum-sized area needs to be modeled in order to
facilitate good localization and control; however, often small mobile robot computers do

153
 (a) (b)

(c) (d)

Figure 6.29: Gateway evaluation conditions. (a) We start with a simple, high-resolution + inter-
section. (b) We rotate the image from 0 to 45◦ in 5◦ increments. This image is rotated by 45◦ . (c)
We scale the image to be between 9x9 and 599x599 cells in 10 cell (per side) increments. This image
is 39x39 cells. (d) We add random noise by probabilistically flipping a free cell to occupied. This
image shows a 25% noise probability.

154
not posses the speed required to process large grids in real-time. By halving the resolu-
tion of cells (each cell represents a larger area of the local space) the robot can model the
same sized region using quadratically fewer cells, decreasing the run time of LPM-based
algorithms. However, decreasing resolution can overestimate the size of small objects and
create noisy obstacle boundaries (e.g., see the 0.1 meter cells in Figure 6.8(a) as opposed to
the 0.05 meter cells in Figure 6.8(c)). Because the region modeled by the images is a con-
stant, reducing the image size is akin to decreasing the resolution of LPM cells; therefore
we can examine how our local topology abstraction works in LPMs with a small number of
cells and low resolution.
Rotations can create gray cells by averaging over white and black cells. Likewise,
scaling can create gray cells from black and white cells due to anti-aliasing. If we classify
gray cells as unknown, we can potentially get false exits, which will cause unwanted false
gate gateways. To avoid this problem, we classify all cells as either free or occupied, using
a gray-scale value of 127 as the separating threshold. Additionally, as we scale the image,
the robot’s size can affect the size and location of places it will consider. To avoid this
problem, we set the robot’s radius to zero.
Finally, we add noise to the images by flipping each free cell to an occupied cell
with some probability (Figure 6.29(d)). This probability ranges from 0 to 95 in 5% incre-
ments. Several factors exist that can create noise in a real-world LPM. First is the accuracy
and precision of localization. We discussed how to handle this using our shrink-and-grow
localization scheme in Chapter 4. Next is the noise from the sensors. Modern-day planar
lidar devices have only ±1 cm of error with an 80 meter range; thus many SLAM imple-
mentations simply do not consider sensor noise when building a metrical map with cells >2
cm. Finally, the environment itself can have posts, garbage cans, table legs, or pedestrians
that are essentially noise when trying to extract the structure of the nearby paths and places.
Though adding uniformly distributed noise over free space is not the most accu-
rate model of typical lidar-based LPM noise, we feel it is a reasonable starting point for

155
this evaluation. Additionally, we will show that most of this noise is actually filtered out
by the “island” obstacle removal process prior to generating the skeleton. As a result, in-
creasing the noise essentially increases the noise along the path boundaries. This may be
a reasonable model for vision-based LPMs, as outdoor path extraction (e.g., sidewalk or
streets) may have shadows, mud, tall grass, etc. that create noisy path boundaries. It is
certainly a reasonable model of the hallways near our robot lab, which often have physics
lab equipment along the walls.

Evaluation Results

Transforming the original image over rotation, resolution, and noise dimensions we obtain
(10 ∗ 60 ∗ 20) 12000 test images that we use to evaluate the performance of the anchor-
based gateway extraction. Rather than considering 4 detected gateways as the only criteria
for success, we must also consider if the 4 gateways lead to the correct local topology for
the + intersection. We use this more strict criteria to determine whether a particular run of
our algorithm on a test image was successful or not.

Rotations. Because thinned skeletons are rotation dependent, we expected to see a sig-
nificant effect of rotation of local topology abstraction. Figures 6.30-6.33 show the (lack
of) impact of rotation on our algorithm.
Figure 6.30 illustrates that successful gateway detection and subsequent local topol-
ogy abstraction are largely independent of rotation. The horizontal nature of the accuracy
graphs show that we get the same accuracy over each 1200 image subset of the data that
varies across image resolution and noise. Additionally, if we take each of these 1200 image
subsets and graph them with respect to image resolution and noise (calculating the mean
and standard deviation over the 10 rotation cases), we see that the standard deviation error
bars are reasonably small (Figure 6.31(a,b)). These error bars seem even smaller when we
compare these variances to the variances experienced when examining rotation accuracy
over various noise and image resolutions (Figure 6.31(c,d)).

156
Figure 6.30: Impact of rotations on gateways and local topology. The classification accuracy of
local topology abstraction is plotted over the different angles used in the test data. The accuracy is
expected to be rather low, as each angle uses 1200 images from all size and noise values. The fact
that the accuracy does not change with rotation demonstrates that rotation is largely independent of
classification accuracy.

157
 (a) (b)

(c) (d)

Figure 6.31: Variance of local topology classification across rotations. If classification accuracy is
largely independent of rotation, then we expect to see small standard deviation error bars when we
graph the accuracy over LPM size or LPM noise while using each rotational value as a separate set
of test images. Notice that in images (c,d) we have much larger variance with respect to the LPM
size and LPM noise than in images (a,b) where the error bars show the variance over rotations. (a)
We plot the classification accuracy of local topology abstraction over the LPM size. Classification
accuracy is the success rate for 20 different test images—one for each noise level. The error bars
show the standard deviation of classification accuracy across 10 different runs—one for each rotation
value. Notice how small these variances are when compared to image (c). (b) Similarly, we plot
the classification accuracy for 60 different test images (corresponding to all LPM sizes) over the 20
noise levels. The error bars are across the 10 different rotations. (c) Here, classification accuracy
is the success rate for 20 test images (different noise parameters) for each of the 10 rotations. The
error bars show the standard deviation of classification accuracy across 60 different runs—one for
each LPM size. (d) Accuracy for the 60 test images that correspond to different LPM sizes at each
of the 10 rotations. The error bars are for 20 runs across different noise levels.

158
Figure 6.32: Time performance of local topology abstraction. Though the gateway (and the entire
local topology) algorithm remains linear with respect to total image size, we observe that images
rotated by 45◦ use 50% more compute time than images of the same size that are not rotated. This
demonstrates that the thinning algorithm slows down when computing the skeleton along diagonals
(w.r.t. the grid axes). (Adding noise increases run time as well. Though the “island” obstacle
removal is linear, the full local topology algorithm takes almost twice as long at 50% noise as with
no noise.)

159
 (a) (b)

(c)

Figure 6.33: Stability of gateway orientations. For each image where the correct local topology is
found, we rotate the image back to the original orientation and record the orientation of the gateways
with respect to the image axes. (a) The histogram of the gateway orientations shows that we have
4 well separated clusters. (b) Here is a sample (299x299 cells, 25◦ rotation, 0 noise) where all
four gateways are approximately 4 degrees off from the theoretical orientations of 0, 90, 180, and
270 (after the image is rotated back to the original orientation). This is due to the type of skeleton
the thinning algorithm creates along diagonal paths. (c) The normalized gateway orientations can
be grouped into a single cluster (modulo 90◦ ). This illustrates that we have some images where
gateways are oriented perfectly with the paths, but most are slightly offset due to the phenomenon
in image (b).

160
 We do observe that using a thinning algorithm does create a slight computational
overhead when dealing with diagonals. Figure 6.32 shows that our gateway/local topology
abstraction algorithm is linear in the number of LPM cells. We see a ∼50% increase in
the time, which is largely due to examining the same cells multiple times when trying to
compute a skeleton along diagonals.
We are also interested in observing how image rotations affect the gateway orien-
tations with respect to the 4 straight path segments that emanate from the intersection. For
each image where the correct local topology is found, we rotate the image back to the orig-
inal orientation and record the orientation of the gateways with respect to the image axes.
Figure 6.33(a) illustrates that we get 4 distinct clusters of gateway orientations. This shows
that rotations are not largely affecting the skeleton, which is used to hypothesize gateway
orientation.
Figure 6.33(c) shows these angles modulo 90◦ . We observe an approximate Gaus-
sian with standard deviation 3.5295. Notice the mean and median deviate from 0 by 3.9407
and 4.3115 respectively. This is the only artifact of using the thinning algorithm on ar-
bitrarily rotated paths that we observe in this evaluation. Figure 6.33(b) illustrates why
this occurs—the thinned skeleton approximates angled paths with a collection of piecewise
linear segments. While this does not affect the functional outcome of the local topology
abstraction for the + intersection, it could easily cause two distinct local topologies to be
created for the same place at more complex intersections (e.g., where the gateways are fur-
ther apart with respect to the LPM size). As such, alternative gateway orientation and/or
gateway alignment methods should be investigated for future implementations.

LPM Resolution and Noise. We have determined that our algorithm is relatively robust
to rotations in this + intersection example. We now examine how reliable our algorithm
is under different LPM resolutions and noise conditions. We use the fact that classification
accuracy is largely independent of rotation in order to consider each of the 10 rotations as
independent samples for each of the 1200 resolution/noise pairs.

161
 As LPM size increases, we expect to see better classification performance until a
plateau is reached—at some point adding more resolution to an LPM will not enhance the
performance of local topology abstraction. This is shown by Figure 6.31(a) when all noise
levels are used, while Figure 6.34(a) graphs the accuracy of local topology abstraction over
LPM size in the cases where no noise exists. We observe that at a resolution of 39x39
cells, we achieve 100% classification accuracy for the 10 different rotations. Figure 6.34(b)
shows the gateways found for the single (noise-free) 29x29 case that failed—the resolution
is simply too low to properly align gateways to obtain the correct local topology.
The success of the 39x39 cell case provides a lower bound on the resolution of an
LPM. This shows that in order for our algorithm to work as intended, a path in the LPM
should be at least 13+δ cells wide—13 is the width of the free space paths in the 39x39 cell
grid, and δ is the number of cells (always rounded up) that equals the width of the robot.
We can compute lower bounds on resolution and LPM size as follows:

arg max max path width/x − drobot width/xe ≥ 13.

x

For example, if we have a 30 cm wide robot with very limited processing power, and we
know that a path will never be wider than 3 meters, we can estimate the smallest usable
LPM: arg max 300 / x - d30 / xe ≥ 13; here, x = 20 cm/cell. For robustness we probably
x
want a side size that can fit the path and the robot: at least 17 cells (d330 / 20e).
This lower-bound estimate assumes noise-free LPMs; however no real-world en-
vironment or robot can facilitate noise-free metrical modeling. By examining the classi-
fication accuracy of various noise levels with respect to a large LPM size, we obtain an
upper bound on the amount of uniform noise that our algorithm can handle. Figure 6.35(a)
shows the performance of local topology classification over different noise levels using the
10 rotated images of size 599x599.
We observe that for high resolution LPMs, our algorithm can handle complete uni-
form noise in free space (50% probability of flipping a free cell to occupied). On first

162
 (a)

(b) (c)

Figure 6.34: Impact of LPM resolution on gateways and local topology. (a) Local topology classi-
fication accuracy over all LPM sizes when no noise is present. We see that in the noise-free LPMs,
all images ≥39 cells on a side abstract the correct local topology, despite rotation. (b) At 29x29
cells, there is one rotation (30◦ ) where the local topology fails. Here the LPM resolution is simply
too low to adequately orient the gateways; thus the algorithm finds 3 paths—the top gateway has
normalized angle of 106◦ (instead of 90), so it does not properly align with the bottom gateway. (c)
At 39x39 cells, the algorithm succeeds for all rotations when no noise is present.

163
thought, the idea that our algorithm can handle uniform noise in free space seems surpris-
ing; however, Figure 6.35(b,c) illustrates the reason why this occurs—the obstacle removal
we perform as pre-processing to the skeleton removes much of this noise.
So far, we have determined the best-case upper and lower bounds where our algo-
rithm works: 39x39 cell grids and 50% noise respectively. Of course, we do not expect our
algorithm to work robustly in situations with both low resolutions and high noise. Figure
6.36 illustrates the resolutions and noise levels where we get 100% accurate local topology
abstraction for all 10 rotations. We see that in general, our algorithm works at almost all
noise levels <50% when the LPM is at least 300x300.
We do notice occasional spots where lower resolution images reach 100% accuracy
for a given noise level, while higher resolution images at the same noise level do not. We
have found two separate cases for this. The first is the case where the remaining noise (after
removing “island” obstacles) is near the true anchors of the + intersection. This causes the
skeleton branches to curve, and we see a situation very similar to Figure 6.9(a), where the
gateway is located at a bend in the skeleton, thus its orientation is incorrectly determined
(Figure 6.36(b)). This can most likely be reproduced for lidar-based LPMs of real-world
environments by simply placing an obstacle in an intersection, touching an anchor point.
The second case is where obstacles along the path boundaries iteratively pull the
gateway towards the edge of the LPM. Figure 6.36(c) illustrates such a case, and highlights
two issues with the actual algorithm. First is the problem where the iterative portion of
Algorithm 6.3—meant to find three collinear points to define a gateway in corridor-style
environments—can fail in certain scenarios due to the discretized skeleton.4 We do not
believe this is an issue if a more precise, continuous Voronoi graph is utilized. Second
is the fact that our linear “island” obstacle removal does not remove obstacles that touch
the edge of the LPM. Neither of these algorithmic issues has been an issue outside of this
particular evaluation.
4 Thediscretized skeleton may only have√ one obstacle within the Voronoi radius r p of a skeletal point. So,
we need to look for obstacles within r p + 2. In rare scenarios (Figure 6.36(c)) this can lead to problems.

164
 (a)

(b) (c)

Figure 6.35: Impact of noise on gateways and local topology. (a) Local topology classification
accuracy over all noise levels for a 599x599 cell grid. We see that in high resolution LPMs, all
images with ≤50% uniform noise abstract the correct local topology, despite rotation. (b,c) Here
we show a test image (299x299 cells, 0◦ rotation) with 50% noise, and the final pruned skeleton and
gateways abstracted. The local topology derived from these gateways is correct despite the noise.
Note that because our “island” obstacle removal keeps obstacles that touch the edge, we get many
spurs near the LPM edges.

165
 Overall, the anchor-based gateway algorithm yielded 100% accuracy (3200/3200)
for all test images ≥200 cells on a side and ≤ 35% free space noise. In cases between the
lower bound of 39 cells per side and the upper bound of 50% noise, the algorithm achieved
91.72% accuracy (5751/6270). Of these 6270 images, 98.8% (6195) of them were found
to have 4 gateways. Again this implies that we should consider a more robust gateway
alignment technique in future work, while the anchor-based gateway detection algorithm
seems quite dependable.

166
 (a)

(b) (c)

Figure 6.36: Determining the bounds of successful local topology abstraction. (a) Here we show
all cases (of image size and noise level) where the anchor-based gateway algorithm yielded 100%
successful local topology classification for the 10 possible rotations. (b) Notice in (a) that for 45%
noise, <100% accuracy is obtained for images with side size of 309 cells. Here we show the single
incorrect classification, which has 30◦ rotation. The noise affects the top and bottom gateways
enough to make them unaligned by our simple alignment criterion. (c) For 20% noise, there is
another misclassification (20◦ rotation) for LPM of side size 179 cells.

167
Chapter 7

Human-Robot Interaction:
Evaluating the HSSH Local Levels

Never trust anything that can think for itself if you can’t see where it keeps its brain.
J. K. Rowling, Harry Potter and The Chamber of Secrets (1999)

One of the unique characteristics of the Spatial Semantic Hierarchy framework with respect
to other map-building approaches is that it is inspired by human spatial knowledge. Earlier,
we argued that this would not only benefit robot navigation (Chapter 4), exploration (Chap-
ter 5), and map-building (Chapters 8 and 9) but would also provide a useful framework for
human-robot interaction. Here we show an initial evaluation of an interface for the local
levels of the HSSH that sketches an optimistic picture of a human-robot interaction in the
near future.
Chapter 4 discussed the LPMs of the Local Metrical Level, and we provided evi-
dence that we could achieve robust localization, thus a high-quality LPM, during navigation.
Chapter 5 discussed the Local Topology Level, where place detection and local topology
abstraction are grounded by finding gateways in the LPM. We provided evidence of robust
gateway detection in commonly seen LPMs. Here we evaluate these two levels together, ex-
pecting robust robot navigation, and observing the performance of human subjects utilizing

168
the HSSH implementation [Beeson et al., 2007].

7.1 HSSH Interface for a Mobility Assistant

We evaluate the HSSH using a scenario where we believe the HSSH can improve on the cur-
rent state-of-the-art: a robotic transport, such as an intelligent wheelchair, an autonomous
vehicle, etc. Our specific domain is that of a transport that allows a person with a visual
disability to achieve the same level of navigation performance as a person with full visual
capabilities. Evaluation of the robotic transport is done in simulation, so that we can sim-
ulate a user with visual disabilities that also needs mobility assistance. Vizard [WorldViz
LLC, 2007] provides a high-fidelity 3D virtual reality environment in which we created a
simulated robot with range sensors.
The human-robot interaction levels for a robotic transport correspond nicely with
the distinct representations in the Hybrid Spatial Semantic Hierarchy (Figure 7.1 and Table
7.1). The higher levels of interaction require more intelligence on the part of the robot, but
they also require less effort for communication and supervision by the human driver. In
order to maximize human autonomy, the driver can shift freely between the different levels
at any time.

7.1.1 Local Metrical Control

Humans have reliable metrical models of their local surroundings [Golledge, 1999]. People
can navigate safely through complex small-scale spaces, and they can use verbal and graph-
ical descriptions to communicate spatial relationships [Skubic et al., 2004a,b]. The HSSH
emulates this functionality using the fixed-size local perceptual map (LPM) that follows its
motion and describes its small-scale surroundings. Obstacles in the LPM can be classified
as static or non-static [Modayil and Kuipers, 2004], which makes it possible to identify and
model dynamic hazards such as cars or pedestrians, and to identify structures such as doors
that can change the apparent topology of places.

169
 Effectors, Sensors Data Flow:
Control Flow:
Human Interface:

Hardware
Control Communication with
Human User

Local Metrical Control Low level motion

commands; safety
Level queries, e.g., Go
forward 5 meters

Local Topology Command Local Topology based

interaction, e.g., Turn
Level right at intersection

Global Topology/
Global Topology Goal Place interaction,
Level e.g., Take me to my
office

Global metrical map

Global Metrical based interaction, e.g,
Level Take me to the spot
clicked on the map
Goal
Figure 7.1: HSSH interface. The HSSH is an integrated framework of multiple, disparate repre-
sentations of spatial knowledge, each with its ontology motivated by human cognitive abilities. The
HSSH interface uses three levels, Control, Command, and Goal, that allow a user to interact with a
robot using the HSSH.

170
 Hybrid Spatial Semantic Hierarchy (HSSH) Human-Robot Interface

Local Metrical Level: Environment is mod- Control: Driver uses the joystick to give di-
eled as a bounded map of small-scale space rection and speed of motion. Robot checks
within the agent’s perceptual surround that commands against the local map for safety.
scrolls with the agent’s motion, without being Driver may specify a target or direction of mo-
tied to a global frame of reference. Useful for tion within the local map. Robot plans hazard-
“situational awareness” of the immediate sur- avoiding motion toward that target.
round.
Local Topological Level: Environment is Command: Driver specifies actions to take
modeled as a set of discrete decision points, at, and to move among, decision points.
linked by actions: turn selects among options Robot identifies decision points in the environ-
at a decision point, and travel moves to the next ment and interprets instructions appropriately,
decision point. translating actions into hazard-avoiding control
laws.
Global Topological Level: Environment is Goal: Driver specifies a destination place in
modeled as a network of places, on extended a topological map, by name or in a schematic
paths, contained in regions, allowing efficient diagram (like a subway map). Robot plans a
route planning by graph search. route to that goal, which is translated into a se-
quence of commands, which are translated into
hazard-avoiding control laws.
Global Metrical Level: Environment has a ge- Goal: Driver specifies a destination place in
ometric model in a single global frame of refer- the global metrical map, like a printed city
ence. Useful for route optimization when avail- map. Robot plans a route to that destination
able. in the topological map.

Table 7.1: Robot transport interface to the HSSH. The levels of our human-robot interface for a
robotic transport correspond to the levels of spatial knowledge as represented in the HSSH.

171
 In our domain, the human driver specifies a desired direction of motion using the
joystick. (This is the Control interface of Table 7.1 and the evaluation below.) The HSSH
trajectory planner ensures this corresponds to a safe trajectory in the LPM. The LPM ab-
straction also supports a GUI that allows the driver to specify a destination point in the local
region. Eventually, the Control interface should also support gestures or natural language
to specify pre-labeled destination poses in the local environment, such as “Go to the desk”.

7.1.2 Local Topological Control

Starting from the perceptual model of the local surrounding small-scale space, humans gen-
erate symbolic descriptions of the navigational affordances of the local space, and therefore
its qualitative decision structure. This is reflected in the language people use to describe
intersections by shape (e.g. “T” or “corner”) or number of paths (“three-way intersection”
or “dead end”) [MacMahon et al., 2006]. A robotic transport should understand terms the
human driver finds useful and comfortable, including qualitative decision commands that
presuppose knowledge of the local decision structure, such as “Turn right” or “Take the
second left”. Fortunately, these terms correspond well with the HSSH Local Topological
Level.
As the robot moves through the environment, maintaining the LPM as an accurate
metrical model of local small-scale space, it describes the local topology of that space in
terms of local-paths and gateways. The local place topology is described as a circular or-
der of directed path fragments and gateways, which translates directly to the large-scale
space description of a place as a node in a graph, connected to paths. In large-scale space
(where most route planning takes place), a command such as “Turn left” selects an out-
going directed path, given the incoming one. In small-scale space (where motion control
actually takes place), the same command translates to a desired trajectory of motion from
an incoming gateway to an outgoing gateway.
Thus, at the Command interaction level, both the driver and the robot represent

172
space as a graph of decision points, and a command specifies the option to select at the next
decision point. In the current implementation, the available commands are Forward, Right,
Left, Turn Around, and Stop. Compound commands such as “Take a sharp right at the
second intersection” will be implemented in future work by integrating the use of complex
verbal instruction [MacMahon et al., 2006] with the resolution of qualitative references to
small-scale space [Skubic et al., 2004b].
By instructing the robot at the Command level, the human driver is delegating the
autonomy for hazard-avoiding travel between places (the robot plans its own safe route
along the path using the LPM), for recognizing the decision structures of places, and for se-
lecting the intended option at that place. The human driver can always reassert his authority
by issuing a Control level command using the joystick or by pressing the Stop button.

7.1.3 Global Topological and Metrical Control

Human-drawn maps based on exploration tend to be topologically correct, even though they
often contain gross metrical errors [Lynch, 1960; Siegel and White, 1975; Golledge, 1999].
People tend to solve way-finding problems primarily based on their topological knowledge
of the environment. On the other hand, people often find it helpful to use metrically accurate
graphical maps in familiar environments.
The HSSH builds a global topological map from the agent’s travel experience, ex-
pressed as a sequence of places with local topologies, linked by travel actions. The global
topological map is found by creating a tree of all possible topological maps, pruning out
inconsistencies, and focusing on the most likely consistent map. A global metrical map in
a single frame of reference can then be built robustly on the skeleton provided by the global
topological map.
At the Goal level of human-robot interaction, the driver instructs the robot by spec-
ifying a destination. The intended destination could be specified by name (“the main li-
brary”), by description (“the corner of Speedway and 24th Street”), by GUI selection from

173
a topological diagram (like a subway map), or by GUI selection from a graphical metrical
map of the environment.
The robot can learn the global topological map from its experience, accumulating
metrical annotations, and constructing the global metrical map if resources permit. Such a
map, learned from experience or extracted from a graphical map, is necessary for Goal level
control to be possible. Assuming that the robot has already learned a sufficiently complete
topological map, planning a suitable route is a straightforward graph search. Carrying out
the plan consists of translating it first to a sequence of decisions to take at specific places,
and then to a sequence of hazard-avoiding control laws for traveling from place to place,
and from gateway to gateway in each place neighborhood.
Building a correct global map may require large amounts of exploration, maybe
using specific exploration strategies; however, when the robot is serving a human driver, its
opportunities for learning the (local and global) structure of the environment are constrained
by the needs and desires of the human driver. This potential conflict between the driver’s
authority and the robot’s need to explore the environment should be reduced. As such, we
have postponed working on the Goal interface until we explore this issue of human-in-the-
loop exploration versus exploitation in more detail. Future experiments that evaluate the
Goal interface need to measure human cognitive load or other useful information that allow
us to evaluate all of the advantages of a fully autonomous human transport system.

7.2 Empirical Evaluation

The bottom-line question is: How well does the robotic transport help people with dis-
abilities get around? Our goal for the robotic transport is for the robot’s own perception,
mapping, and decision-making capabilities to augment those of the human driver, under the
human driver’s ultimate authority. The robot will initially be most useful to subjects who
have perceptual as well as motor disabilities, so we evaluated the robot’s HSSH interface
by comparing subjects with normal and degraded visual stimuli.

174
7.2.1 Experimental Setup

The algorithms for robot navigation were implemented and tuned using physical robot plat-
forms; however, to simulate visual impairment in a manner safe for the human subjects,
we ran our initial experiments in high-fidelity virtual reality environments using Vizard
[WorldViz LLC, 2007]. These experiments were run in an indoor virtual environment for
two perceptual cases: Degraded, in which the subjects’ visual input was degraded by adding
fog to the environment so that objects farther than 3.0 meters from the camera could not be
seen, and Normal, which had no visual degradation (Figure 7.2(a,b)). The simulated robot
itself perceives the virtual environment using omnidirectional range sensors, unaffected by
the fog.
There were four experimental conditions—Normal:Manual, Degraded:Manual, De-
graded:Control and Degraded:Command—each defined by the subjects’ perception and the
level of autonomy available. With the Manual interface, the joystick controls the motors
directly, with no robotic intelligence. The Control interface requires the robot to build and
maintain an LPM in order to avoid collisions while following the instructions given by the
user via the joystick—speed is assumed to be a predetermined constant (1 meter/s) that is
instantaneously achieved.1 The Command interface requires the robot to detect places and
describe their local topologies in order to execute the turn and travel decisions specified by
the driver via a GUI with command buttons (Figure 7.2(d)); however the buttons on the joy-
stick are easily mapped to these button presses, which will be needed for future experiments
with actual visually impaired subjects.
The experimental task was to navigate between a sequence of known places. The
large-scale virtual environment consisted of seven hallway segments (Figure 7.2(c)) with
ten avatars (Figure 7.2(a)). Each subject knew the layout (and place labels) of the envi-
ronment but not the locations of the avatars (stationary obstacles), which were randomly
distributed for each trial.
1 One advantage in simulated environments is that we do not have to model acceleration and deceleration
based on robot/driver weight, surface properties, etc.

175
 (a) (b)

(c) (d)

Figure 7.2: Stimuli, map, and interface for the HRI experiment. (a) Subject’s environment view
in the Normal Vision condition. (b) Subject’s limited view in Degraded Vision (fog) condition. (c)
Environment layout used. Numbers correspond to goal states. (d) The local symbolic control GUI.
Individual buttons light up when applicable for the current environmental situation.

176
 Each trial started by placing the simulated robotic transport at the same starting po-
sition (the lower end of the center corridor). At the beginning of a trial, the subject was told
by the computer the current location and destination (e.g., “Position 3. Go to position 5”).
Depending on the condition of the trial, the subject used the joystick (Manual and Control
interfaces) or the Command level interface (Figure 7.2(d)) to travel to the specified goal.
The computer would indicate when they had reached each goal location by announcing the
goal location and would then give the subject another goal location (e.g., “Position 5. Go
to Position 2”).
Each trial consisted of a sequence of five goal locations. There were five prede-
termined sequences that were used for each of the four experimental condition; thus each
subject ran a total of 20 trials. To avoid learning effects, the 20 trials were randomly or-
dered for each subject. For these initial experiments we ran three subjects. We logged their
motion through the environment and when/where they collided with both avatars and walls.
This evaluation is intended to address three questions:

1. Does reducing the visual information by adding fog make the task more difficult?

2. Is there a performance benefit from adding the Control level (collision avoidance)
navigation aid?

3. Is there a performance benefit from adding the Command level aid (autonomous low-
level control and place detection/local topology abstraction)?

7.2.2 Results

The answer to all three questions is, “Yes”. Figure 7.3 shows sample recorded traces from
subjects in the four conditions. Figure 7.4 shows the mean distance traveled and the mean
number of collisions of a trial in each condition.
Not surprisingly, we see a strong effect from degrading the visual input by compar-
ing the performances in the Normal:Manual versus Degraded:Manual in Figure 7.4. There

177
 Normal:Manual Degraded:Manual

Degraded:Control Degraded:Command

Figure 7.3: Example experimental runs. Sample traces of the paths used by the human subjects in
the four different conditions, with collisions indicated by small squares.

178
Figure 7.4: HRI experiment results.Summary data for the evaluation of the four conditions. The
open (white) bars show the mean distance traveled to complete a sequence (left y-axis). The solid
(black) bars show the mean number of collisions (right y-axis). Error bars represent standard devia-
tions.

179
was a 37% increase in the distance traveled (from 136.53 to 187.62 meters) and a 936%
increase (from 1.47 to 13.73) in the number of collisions.
To evaluate the benefit of adding Control level navigation, we compared perfor-
mances for the Degraded:Manual versus the Degraded:Control conditions. Adding control
level navigation showed little difference in the distance traveled (from 187.62 to 181.29 me-
ters). However, there was a dramatic change in the number of collisions: a 97% decrease
from 13.73 to 0.47.
To evaluate the benefit of adding Command level navigation, we compared per-
formances for the Degraded:Manual versus the Degraded:Command conditions. In this
case, adding Command level navigation did reduce the distance traveled: a 23% decrease
from 187.62 to 144.31 meters. Furthermore, there was a 100% decrease in the number of
collisions (from 13.73 to 0.0).
If we compare performance for the Degraded:Command condition versus the Nor-
mal:Manual condition, we can evaluate how well autonomous navigation does relative to a
driver with full visual abilities. There was a benefit for Degraded:Command navigation in
terms of the number of collisions (Normal:Manual = 1.47; Degraded:Command = 0.0). We
also find that there is a small difference in the distance traveled (Normal:Manual = 136.53
meters; Degraded:Command = 144.31 meters). This difference in distance was mainly due
the lack of “corner cutting” by the current autonomous control system (Figure 7.3(d)).
These results demonstrate distinct benefits from the Control and Command level
navigation aids. In visually degraded conditions, Control level navigation improves ob-
stacle avoidance, but does not reduce the distance traveled. By contrast, Command level
navigation (which includes Control) reduces both the number of collisions and the distance
traveled. Our preliminary data suggests that the resulting performance is similar to that of
the condition with no visual degradation.
In future tests of the Goal interface, we expect to get similar results with respect to
autonomous navigation. The environment used here was small enough that the human sub-

180
jects made very few Causal Level mistakes in commanding the robot. Larger environments
will undoubtedly demonstrate that the using the Goal interface reduces the human user’s
cognitive load and should yield optimal large-scale plans.

181
Chapter 8

Closing Loops:
The HSSH Global Topology Level

I believe the future is only the past again, entered through another gate.
Arthur Wing Pinero, The Second Mrs. Tanqueray (1893)

The next two chapters will address the problems of building a global topological map to
describe the qualitative structure of large-scale space and building a global metrical map
to describe its geometric structure within a single global frame of reference. We describe
these two map-building problems separately, but their solutions benefit from each other and
should be interleaved in future research (Section 10.1.2).
The first problem is to identify the best global topological map consistent with ex-
ploration experience. The process of generating possible topological maps from experience
and testing them for consistency can provide formal guarantees that the correct map is gen-
erated and never discarded [Dudek et al., 1993]. A logic-based theory of topological maps
[Remolina and Kuipers, 2004] makes explicit the assumptions upon which those guarantees
depend.
If the robot knows it is in an environment with no loops, creating a topological map
is quite easy. This is especially true given deterministic actions, as the robot simply moves

182
deterministically between known places when it revisits parts of the environment. Even
with non-deterministic actions, creating the topology of such environments is still possible
[Tomatis et al., 2002]. The difficulty in map-building arises from closing loops: determining
when a newly-encountered place is the same as a previously-experienced place, and creating
a hypothesized new loop in the topological map. When large loops in the environment
result in structural ambiguity, a topological representation can concisely represent the loop-
closing hypotheses by generating a single topological map for each qualitatively distinct
alternative.

8.1 From Small-Scale to Large-Scale Star

In small-scale space, the LPM is used for the detection of gateways, local-paths, and places,
and to create the local map m p that is stored at places. The small-scale star describes both
a circular order on the set of directed local-paths in a place neighborhood and also the
correspondence between directed local-paths and oriented gateways. A place p in large-
scale space is associated with the local map m p , a model of the place neighborhood in
small-scale space. At a place, each directed local-path π̃ d in small-scale space corresponds
to a directed path π d in large-scale space. This allows us to determine the large-scale star
that describes the circular ordering of topological directed paths at the place.
Assimilating the local topology of a place into the global topological map requires
a 1-1 mapping between the directed local-paths in the small-scale star and a set of directed
paths from the global topological map. In Figure 8.1, we illustrate such a mapping between
the local-paths, π̃a , π̃b , π̃c , and π̃d , and the corresponding global topological paths π1 , π2 ,
π3 , and π4 , respectively. To keep this example simple, we specified + and − on the directed
paths to correspond consistently, but of course this need not be true in general.
In large-scale space, a distinctive state q corresponds uniquely to a place, a path, and
a direction on that path (Equation 3.1). Thus, the dstate q is at a particular place p, and there
is a bijective association between a dstate and a directed local-path: ψ p (q) = π̃ d where π̃ d ∈

183
 (a)

Small-scale star An example large-scale

description star abstraction
((hπ̃a+ , 1i ↔ hg4 , ini, hg1 , outi) ((hπ1+ , 1i ↔ q1 )
(hπ̃b+ , 1i ↔ hg2 , outi) (hπ2+ , 1i ↔ q2 )
(hπ̃c− , 0i ↔ hg5 , ini) (hπ3− , 0i ↔ q3 )
(hπ̃d+ , 1i ↔ hg3 , outi) (hπ4+ , 1i ↔ q4 )
(hπ̃a− , 1i ↔ hg1 , ini, hg4 , outi) (hπ1− , 1i ↔ q5 )
(hπ̃d− , 0i ↔ hg3 , ini) (hπ4− , 0i ↔ q6 )
(hπ̃c+ , 1i ↔ hg5 , outi) (hπ3+ , 1i ↔ q7 )
(hπ̃b− , 0i ↔ hg2 , ini)) (hπ2− , 0i ↔ q8 ))
(b) (c)

Figure 8.1: Binding local-paths to distinctive states. (a) Gateway locations and directions are used
to identify the directed local-paths and to determine which pairs satisfy the path continuity require-
ments. (b) The small-scale star enumerates directed local-paths in clockwise order, describing their
traversability and association with gateways. Note: the robot entered the place via g5 ; thus, it arrived
on directed local-path π̃c− . (c) The large-scale star replaces local-paths with topological paths from
the global topological map, and defines a distinctive state for each directed path at this place. This
environment has five gateways, four paths, and eight distinctive states.

184
S p . This implies that in the case where the directed local-path passes through the place, the
distinctive state q will correspond with two different oriented gateways, one hg, ini entering
the place neighborhood, and the other hg0 , outi departing from it.
An isomorphism φ : S → S0 between two stars implies a bijective mapping between
the associated dstates as well. We will extend φ to write these implied mappings as φ (q) =
q0 . For a topological map M T , and the set P of places in M T , we can now define the set of
local place maps,
M P = {hp, m p , S p , ψ p i : p ∈ P}

associating each place p ∈ P with its local metrical map m p , its local topology S p , and ψ p ,
the association between dstates and directed local-paths in the local topology.
Assuming that the LPM is sufficiently well explored, the set of directed local-paths
and gateways in the small-scale star is complete, so the description of the distinctive states
and directed paths in the circular order of the large-scale star is also complete. A turn action
in large-scale space corresponds to motion in small-scale space within a place neighborhood
from the inward-facing oriented gateway the robot arrived upon to an outward-facing ori-
ented gateway (Figure 5.4(c)). Thus, for every pair of dstates qi and q j at the place, a causal
schema for the turn action hqi ,turn, q j i is implicitly defined. Exploration experience can
now be described as an alternating sequence of travel actions and place neighborhoods,
which simplifies construction of the global topological map (Algorithm 8.1).

8.2 The Tree of Possible Topological Maps

The topological map-builder maintains a tree whose nodes are pairs hM, qi, where M is
a topological map (augmented below for the HSSH) and q is a distinctive state within
M representing the robot’s current position. The leaves of the tree represent all possible
topological maps consistent with current experience [Dudek et al., 1993]. Algorithm 8.1
reiterates the procedure for growing the tree of possible topological maps. It also details the

185
differences between the basic SSH and the HSSH.
After each action a and resulting view v, we extend each map hypothesis at a leaf of
the tree. If the current action moves within known territory, the map hM, qi will predict the
resulting dstate q0 and the view to be observed, so the hypothesis can be updated or refuted
according to whether the prediction was correct or not. If the current action explores new
territory, then either the resulting dstate is also new, or the action closes a loop and connects
with a previously known dstate. Since there may be multiple possibilities that all match
view v, the tree of topological map hypotheses will branch. For purposes of generating and
testing candidate topological maps in the HSSH, we will extend the basic SSH topological
map M T with M P = {(p, m p , S p , ψ p ) : p ∈ P}, the set of local metrical maps and local
topologies of individual place neighborhoods.

M = hM T , M P i

In the SSH, a view v is an abstracted description of the agent’s perception of the

local environment from a distinctive state q. We select the level of description to ensure that
the view is a deterministic function of the dstate (v = o(q)), although we allow perceptual
aliasing (different states with the same view) [Kuipers and Beeson, 2002]. In the basic
SSH, a view is a symbol, abstracting away the nature of the perceptual system, and views
are matched only for equality. In the Hybrid SSH, we define a view to be the local topology
S p of the current place p and the current directed local-path the robot is on; thus, the new
view description is derived from the local topology, which is grounded in local perceptual
map m p .

v = hS p , π̃ d i where d ∈ {+, −}

= hS p , ψ p (q)i

Given two views v and v0 , we say that match(v, v0 ) holds iff there is an isomorphism φ :

186
 0. Perform initial action a0 that brings the robot to a place along a directed path.
Initialize the tree of maps with the map hypothesis hM0 , q0 i, where M0C contains the
single dstate q0 with its observed view v0 , and M0T contains the single place p0 and
path π0 .

After performing a new action a and observing the resulting view v, for each consistent
map hM, qi on the fringe of the tree:
1. If MC includes hq, a, q0 i in R and v0 = o(q0 ),
• if match(v, v0 ), then hM, q0 i is the successor to hM, qi, extending the tree;
• if not, then mark hM, qi as inconsistent.
2. Otherwise, MC does not include hq, a, q0 i in R. Let M 0 be M extended with a new
distinctive state symbol q0 and the assertions v = o(q0 ) and hq, a, q0 i. Consider the
k ≥ 0 dstates q j in M with v j = o(q j ), such that match(v j , v). Then hM, qi has k + 1
successors:
• hM 0j , q0 i for 1 ≤ j ≤ k, where M 0j is M 0 extended with the assertion q0 = q j .
0 , q0 i, where M 0
• hMk+1 0 0
k+1 is M extended with the k assertions that q 6= q j , for
1 ≤ j ≤ k.
3. Mark a new successor map inconsistent if it violates the axioms of topological
maps.
4. Define a preference order on the consistent maps at the leaves of the tree.

In the Basic SSH:

M = MT .
A view is a simple symbol.
match(v, v0 ) iff v = v0 .
Both a ∈ Turns and a ∈ Travels can reach step 2 and cause a branch.
Preference order from prioritized circumscription policy [Remolina and Kuipers,
2004].

In the Hybrid SSH:

M = hM T , M P i.
A view is a structure hS p , π̃ d i, where p = place(q), consisting of a local topology and
the directed local-path the robot arrived upon.
match(v, v0 ) iff there exists an isomorphism φ : S p → S0 where φ (q) = q0 .
Only a ∈ Travels can reach step 2 and cause a branch.
Future work: Preference order from map probabilities (Section 10.1.2).

Algorithm 8.1: Building the tree of topological maps in the HSSH. This describes the
algorithm for building a tree of all possible topological consistent with a sequence of
actions and observations at discrete places. The different instantiations. Note that the
HSSH algorithm is identical to the basic algorithm from Algorithm 3.1. The difference
is in the well-defined views and in the branching factor, both of which depend on the
local topology abstraction.

187
S → S0 such that φ (q) = q0 . That is, from the perspectives of the specified dstates, the local
topologies match.
As exploration progresses, the map M is extended with new information. For ex-
ample, after an exploration step that closes a loop in the map, the resulting map M 0 is M
extended with a new dstate q0 and assertions hq, a, q0 i, v = o(q0 ), and q0 = q j . A new version
of MC = hQ, A,V, S, oi is created, and the implications of the loop-closing assertion q0 = q j
propagate through new versions of M T and M P to unify place and path labels as necessary.
Because we are matching complete local topologies in the HSSH, the tree of maps only
branches on travel actions. Turn actions are already fully described by the large-scale star.

8.3 Topological Mapping Example

We applied an implementation of the Hybrid SSH map-builder to an exploration of an office

environment with multiple nested large loops. This office had a large number of cubicles
and office doorways. To respect student and faculty privacy, we prune the Voronoi skeleton
so that Voronoi branches, thus gateways, were defined only for large hallway intersections,
not at doorways or cubicle openings. The environment, as defined by the robot, contained
6 paths and 9 places with 4 distinct local topologies. Figure 8.2 shows the exploration
route as a sequence of place visits, the sequence of LPMs observed at successive place
neighborhoods, and the unique simplest topological map that resulted from the mapping
algorithm, with LPMs overlaid at corresponding places in the correct topological map.
After an exploration consisting of 14 travel actions, the topological mapper finds
83 possible configurations of the environment that are consistent with the observed local
topologies and the topological axioms—that is there exist 83 leaves in the tree of maps.
The prioritized circumscription [Remolina and Kuipers, 2004] on this set of maps produces
4 minimal models. All but one of these can be eliminated with further exploration or by
simply matching LPMs using the alignments specified by the four minimal maps. This final
map model is the correct topological representation of the environment.

188
 (a)

1 2 3 4

5 6 7 8

9 10 11 12

13 14 15
(b) (c)

Figure 8.2: Finding the topological map of an environment with multiple nested loops. In the CAD
drawing (a), we show the path traveled between places in the environment. We enumerate the order
of places in the exploration taken by the robot. (This exploration trace was also used for Figure
2.1(d).) In (b), we show the LPMs created at the places during the travel. We specifically tuned
the gateway algorithm to ignore open office doors and cubicle openings to ensure places only at
hallway intersections. The stars generated from these LPMs are used to search through the space of
consistent topological maps. In (c), we show the unique topological map generated after matching
local stars and LPMs. The map is overlaid with the LPMs generated at the places, with the gateways,
and with the connections between gateways which lie on the same path.

189
 If we assume planarity of the environment, we can use a more sophisticated ver-
sion of the topological map-building algorithm [Savelli and Kuipers, 2004] that rules out
many more models as inconsistent. Here, there are only 46 consistent configurations of the
exploration experience, and the circumscription policy produces a single minimal model,
which is the correct topological map of the environment (Figure 8.2(c)). Our current imple-
mentation can build the complete tree of maps for this exploration trace and determine the
unique minimal map of this environment in ∼200 ms on the robot’s Pentium III 450 MHz
processor. The results presented on this office environment would be unchanged if the path
segments were longer or even very convoluted, as the number of places and paths would
not change. Additionally, the tree of maps ensures the correct map is never discarded.
The complexity of the map computation is the following. Let n be the number of
poses in the exploration trajectory; let m (m  n) be the number of topological places;
let k (k < m) be the maximum number of places matching an observed view; and let l be
the maximum number of directed local-paths in the local topologies (often l ≤ 4). For
example, for the environment in Figure 8.2, n ≈ 7300, m = 9, k = 4, and l = 4. The
maximum branching factor in the tree of maps is k + 1. Branches only occur when the robot
travels between two connected places for the first time, which can only happen at most ml/2
times. This means the size of the tree of maps is O(km ); thus, computing the tree of maps is
exponential in m (not in n). The exponent decreases by at least a factor of 3 compared with
the basic SSH version due to branching only on travels, not on turns, and matching local
topologies of places.1 Savelli and Kuipers [2004] also show that the planarity constraint
gives an additional improvement in the branching factor k by rejecting many loop-closing
hypotheses.2
1 There are at most ml/2 unique travel actions, and there are at most l turns at each of the m places; thus, in
the worst case environment, we have ml turns and ml/2 travels, resulting in 3ml/2 actions in the basic SSH.
2 Savelli and Kuipers [2004] also point out that for each map m in the tree to be expanded, the reduction
i
of the branching factor ki due to the planarity constraint is proportional to the number of closed loops already
present in mi . In other words, “the more loops [that] have been closed, the more topologically compact the map
must be, and therefore the fewer ways there are to close new loops while preserving planarity,” which reduces
the branching factor further.

190
8.4 Levels of Spatial and Temporal Granularity

At this point, we summarize the three different levels of granularity, with different ontolo-
gies, that we are using to describe space and time.
The agent’s experience is a trajectory through the environment. At the SSH Control
Level and in the LPM, the trajectory is represented using a fine-grained representation for
time t, pose x, motor signal u, and sensory image z. These are used both for control laws,
and for simultaneous localization and mapping to build the LPM. Expanding Figure 4.1,
the agent’s exploration experience is described by

··· ut−1 ut ut+1 ··· uN

↓ ↓ ↓ ↓
x0 → · · · → xt−1 → xt → xt+1 → · · · → xN
↓ ↓ ↓ ↓ ↓
z0 ··· zt−1 zt zt+1 ··· zN

At the SSH Causal Level (which is part of the topological map), exploration expe-
rience is described by an alternating sequence of actions and distinctive states, with each
distinctive state associated with a view.

q0 a1 q1 a2 q2 · · · qn−1 an qn
| | | | |
v0 v1 v2 · · · vn−1 vn

In both the basic and hybrid versions of the SSH, distinctive states q correspond to being at
a place, facing along a directed path. In the basic SSH, the distinctive states q are grounded
by isolated distinctive states x̄ where hill-climbing control laws terminate. In the Hybrid
SSH, dstates are grounded by a directed local-path extracted from the LPM of a place
neighborhood.
At the SSH Topological Level, a particular place p j can correspond to several dis-

191
tinctive states, say qi−1 and qi and the turn action ai between them. A travel action ai+1
from qi at p j to qi+1 at a different place p j+1 can be used to infer the displacement λ j+1 ,
which is the pose of place p j+1 in the frame of reference of place p j . This lets us abstract
the sequence of distinctive states and actions to an alternating sequence of place p j and
displacements λ j .

p0 λ1 p1 λ2 p2 · · · pm−1 λm pm

As described in Chapter 9 (illustrated by Figure 9.1), in order to define the λi , each

place neighborhood must have its own frame of reference and we must select a set of dis-
tinguished time-points 0 ≤ t0 < t1 < · · · ≤ tN = N such that adjacent time-points belong to
different place neighborhoods, and the pose xti at each time-point ti can be unambiguously
localized in its place neighborhood. To fit this into the SSH causal framework, we select
distinguished time-points at the termination of each travel action: in the basic SSH, this is
after hill-climbing terminates, and in the hybrid SSH, this is after a place is detected. In the
Hybrid SSH, the dividing poses are near the incoming gateways in place neighborhoods.
The net effect of the turn and travel actions between these dividing points are used to esti-
mate the displacements λi between the frames of reference of adjacent place neighborhoods
connected by path segments.

192
Chapter 9

Drawing Maps:
The HSSH Global Metrical Level

A map is not the territory it represents, but if correct, it has a similar structure to the territory,
which accounts for its usefulness
Alfred Korzybski, “A Non-Aristotelian System and its Necessity for Rigour in Mathematics
and Physics” (1931)

This chapter details the process of building a global metrical map of an environment from
a symbolic, global topological map. The topological map identifies a discrete set of places,
each with its own local metrical map within its own frame of reference. The topological
map also encodes decisions about how loops are closed and which aliased local neighbor-
hoods represent the same places. The global metrical map is built on the structural skeleton
provided by the topological map [Modayil, Beeson, and Kuipers, 2004]. The steps in build-
ing the global metrical map are: (1) describe the displacements λ = {λi }, each describing
the change in pose from one place neighborhood to the next in the frame of reference of the
first; (2) describe the layout χ = {χ p }, specifying the poses of places in a global frame of
reference; (3) describe the trajectory x = {xt } of robot poses within the global frame; and
(4) create the global map m∗ from sensor readings given the trajectory.

193
9.1 Terminology

The global topology τ, used below, consists of the set M P = {hp, m p , S p , ψ p i : p ∈ P} of

places with their local information, the set of distinguished time-points 0 ≤ t0 < t1 < · · · <
tn ≤ N that divide the fine-grained sequence of exploration experience into segments corre-
sponding to travel between adjacent place neighborhoods, and the relation place(ti ) = p j
between them. It is convenient to relabel the variables x, z, and u, defining xi, j ≡ xti + j . At
each distinguished time-point ti , where place(ti ) = p j ∈ P and place(ti+1 ) 6= place(ti ), the
agent is localized in the local metrical map m p j .
Much of our metrical inference consists of defining an appropriate set of reference
frames, and estimating the values of local and non-local metrical quantities. Many of these
concepts can be simply understood by examining Figure 9.1.

[x] p The coordinates of the pose x in the frame of reference of

place p.
Op The pose x such that [x] p = (0, 0, 0).
Li ≡ [xti ] place(ti ) The coordinates of the pose xti in the reference frame of
place(ti ).
m̃i The scrolling map that models the agent’s surroundings
between distinctive time-points ti and ti+1 . The map’s
origin is defined as the agent’s pose at time ti . That is,
Om̃i = xti .
λi ≡ [O place(ti ) ] place(ti−1 ) The location of O place(ti ) in the reference frame of
place(ti−1 ), estimated using the experience from ti−1 to
ti .
χ p ≡ [O p ]m∗ The pose of O p in the global reference frame of m∗ .
m∗ The global metrical map.

194
 ~ i at ti +j
m
x i,0
x i+1,0

Li x i,j L i+1

place(t i+1 )
place(t i )

Figure 9.1: Defining local frames of reference. The agent creates the local scrolling map m̃i
when traveling between places. The agent’s poses at the distinguished time-points ti and ti+1 are
Li = [xi,0 ] place(ti ) and Li+1 = [xi+1,0 ] place(ti+1 ) . The displacement between the two place frames of
reference is λi+1 = Li ⊕ [xi+1,0 ]m̃i ⊕ ( Li+1 ).

9.2 The Theory of the Global Metrical Map

To build a global metrical map m∗ , we want to find the maximum-likelihood path the robot
traveled, using the topological skeleton in addition to odometry. As discussed in Section
4.1, the joint probability of the pose history x and the global map m∗ can be decomposed as

P(x, m∗ |z, u) = P(m∗ |x, z, u) · P(x|z, u)

by the chain rule for probabilities. This decomposition is valuable since P(m∗ |x, z, u) (map-
building given accurate localization) can be computed analytically and incrementally for
popular map types, so we can focus our attention on P(x|z, u) (pose estimation).
To include the effect of possible global topologies τ on pose estimation, we marginal-
ize over the space of topologies. If we assume that the correct global topology τ̄ has been
identified, only one topological hypothesis τ = τ̄ has nonzero probability.

P(x|z, u) = ∑ P(x|z, u, τ) · P(τ|z, u)

τ
= P(x|z, u, τ̄)

On the other hand, suppose there are multiple topologies τ with significantly non-zero val-
ues of P(τ|z, u). While the weighted sum provides a mathematically correct characteriza-

195
tion of the probability distribution P(x|z, u), it can easily lead to a nonsensical metrical map
due to the dramatic qualitative impact of topological structure on the metrical map. Thus
the summation should be regarded as describing a disjunction over topological maps, with
P(τ|z, u) being the likelihood of each map. This is exactly the tree of possible topological
maps we have already constructed. Therefore, even in the case where there are multiple
plausible topological maps, we will construct global metrical maps for each one individu-
ally.
Given a particular topology τ̄, we can marginalize over the poses of all topological
places χ = χi and their estimated displacements λ .

Z Z
P(x|z, u, τ̄) = P(x|χ, λ , z, u, τ̄) · P(χ|λ , z, u, τ̄) · P(λ |z, u, τ̄) dλ dχ

Because x is conditionally independent of λ given χ, and χ is conditionally independent of

z, u given λ , we can simplify this equation.

Z Z
P(x|z, u, τ̄) = P(x|χ, z, u, τ̄) P(χ|λ , τ̄) P(λ |z, u, τ̄) dλ dχ

We divide this equation into simpler components, defining the following functions repre-
senting probability distributions over their arguments.1

F(λ ) = P(λ |z, u, τ̄)

Z
G(χ) = P(χ|λ , τ̄) F(λ ) dλ
Z
H(x) = P(x|χ, z, u, τ̄) G(χ) dχ

Thus, we use the topological map τ̄ to factor the localization term P(x|z, u) = H(x)
into three separate probability distributions: place-to-place displacements F(λ ) derived
from local metrical maps; the metrical layout G(χ) of places in the global topological map;
1 We assume that there is no opportunity for confusion between these probability functions F, G, and H, and

the dynamical system functions F, G, and Hi used in Section 3.2.1.

196
and the global metrical layout H(x) of the robot’s pose trajectory. Finally, we can combine
the pose trajectory with P(m∗ |x, z, u) to define the joint distribution P(x, m∗ |z, u).

9.3 Global Mapping Example

Here we detail each step of creating the global map and discuss our current implementation,
which runs offline. Figures 9.2 and 9.3 demonstrate the stages of creating an accurate global
metrical map of a large, complex environment using these methods.

9.3.1 Estimating F(λ )

Given the topology τ, we can compute F(λ ). Each λi corresponds to a single experience of
a path segment. Since closing large loops is not a problem when considering a single path
segment, traditional SLAM methods may be employed to estimate F(λ ) by decoupling it
into a set of independent probabilities.

Di = zi,0 , . . . , zi,ni , ui,1 , . . . , ui+1,0

Fi (λi ) = P(λi |Di−1 , Li−1 , Li )

n
F(λ ) = ∏ Fi (λi )
i=1

See Figure 9.1 to understand L. Our current implementation is an incremental maximum-

likelihood method [Fox et al., 1999], modeling each Fi (λi ) as a Gaussian.
Using the notation of the compounding operator [Smith et al., 1990], we compute
the distribution of λi by composing three uncertain vectors: the vector Li−1 from O place(ti−1 )
to xi−1,0 ; the vector [xi,0 ]m̃i−1 from xi−1,0 to xi,0 ; and finally the vector −Li from xi,0 to
O place(ti ) .2
Fi (λi ) = P(λi = (Li−1 ⊕ [xi,0 ]m̃i−1 ⊕ ( Li )))
2 Giventwo poses a and b, we write [b]a for the coordinates of b in the frame where a defines the origin
[Smith et al., 1990]. Then, [c]a = [b]a ⊕ [c]b . The inverse operator is [b]a = [a]b .

197
The essential connection is that the pose xi,0 at the end of a path-segment is described in the
frame of reference of place pi−1 by the expression Li−1 ⊕ [xi,0 ]m̃i−1 , and simultaneously in
the frame of reference of place pi by Li .
The problem of estimating [xi,0 ]m̃i−1 is relatively simple along individual path seg-
ments, since loops cannot be involved. The more difficult problem arises from determining
Li after a loop closure. Here we need to align the map m pi with a previously stored map
m ph in order to determine [O ph ] pi , which allows us to solve Li . Matching maps can be
expensive and can lead to false positives due to local minima (e.g., two LPMs of a + inter-
section can be matched four ways). To eliminate this problem, we first align the LPMs based
on the locations of corresponding gateways, consistent with mh , Sh , ψh and mi , Si , ψi , before
refining the alignment using the obstacles and free space of the LPMs. (Figure 9.2(c) omits
this gateway alignment step in order to better illustrate the process of LPM alignment.)

9.3.2 Estimating G(χ)

The layout χ = {χ p } represents the poses of the places in the topological map, with respect
to the frame of reference of the global metrical map m∗ . G(χ) is a probability density
function over possible layouts χ. Among other things, it reflects the distortion in the place
layout due to a loop-closing hypothesis, compared with the observed displacements λ .
Given the topological map, which specifies the data association between observa-
tions and places, we can evaluate G(χ) for an arbitrary distribution of F(λ ). For a particu-
lar value of χ, P(χ|λ , τ̄) will only be non-zero for a single value of λ , namely when each
λi = ( χ place(ti−1 ) ) ⊕ χ place(ti ) . Hence, P(χ|λ , τ̄) is a Dirac delta function, which gives us a
simple expression for G(χ).

Z
G(χ) = P(χ|λ , τ̄) F(λ ) dλ
n
= ∏ Fi (( χ place(t i−1 )
) ⊕ χ place(ti ) )
i=1

198
 1 2 3 4

5 6 7 8

9 10 11 12

13 14 15
(a) (b)

(c) (d)

Figure 9.2: Finding the global place layout. (a) The sequence of local place maps m p experienced.
(b) The unique topological map consistent with topological and planarity constraints. (c) We deter-
mine λi for loop closures by finding the offset between the current pose and the place origin (defined
on the initial place visit). (d) The layout χ derived from the topological map and the place-to-place
displacements λ .

199
 (a)

(b) (c)

Figure 9.3: Creating a global metrical map. Figure 9.2 described the process of obtaining the
global place layout χ, shown in (a). Once this is known, we can quickly compute an estimate of the
global metrical map. (b) The pose trajectory x(t) anchored at points where the robot is localized in
place neighborhoods in the layout χ. (c) Given the localized pose trajectory x(t) in the global frame
of reference, the global metrical map m∗ is created accurately and efficiently. Compare with Figure
2.1(d).

200
 When F(λ ) is represented as a Gaussian, an Extended Kalman Filter (EKF) is a
simple way to approximate G(χ). The idea is to consider place p to be a landmark with pose
χ p . These landmarks are observed one at a time, linked by actions λi . This is essentially the
classic approach of Smith et al. [1990]. Given Gaussian uncertainty along each action ui
connecting the n robot poses, along with constraints that give Gaussian uncertainty between
poses taken from multiple visits to the same place (to associate poses after loop closures),
we can solve for H(x) in time O(n log n) using the sparse matrix methods of Konolige
[2004]. However, often we may only want G(χ), which can be computed in O(m log m)
time for m places, where m  n.
In our current implementation, we utilize a hill-climbing search to quickly converge
to a local maximum of G(χ) (Figure 9.2(d)). The Levenberg-Marquardt algorithm for non-
linear optimization [Press et al., 1992] treats the λi as “springs” between the poses of the
places pk in χ, and relaxes their configuration to reach a local minimum-energy configura-
tion. Efficient estimations of this non-linear optimization also exist [Olson et al., 2006]. A
good initial layout χ for this hill-climbing search can be derived from the displacements λi ,
which represent SLAM-corrected odometry from the scrolling map. We use the term χ̂ to
denote the maximum-likelihood estimate of G(χ).

9.3.3 Estimating H(x)

An extended Kalman filter can be used to estimate H(x) using G(χ) and individual pose
covariances from the experienced trajectory. Alternatively, if accurate pose covariances are
not available, a simple method can estimate the maximum-likelihood trajectory through the
environment. We calculate the independent trajectory Hi (x) along each path segment, as
each place location is fully determined by a global layout χ. In most cases, there will be
some discrepancy between the measured distance λi along the path segment and the fixed
distance between the places in χ. We scale the experienced motion along the path segment
to fit the global path segment distance. This process is similar to methods of distributing

201
odometry error after closing a loop in a global metrical map [Thrun et al., 2000a].
For each trajectory between adjacent places pi and pi+1 , we transform the relative,
incremental displacements ∆(x, y, θ ) from the pose estimates in the scrolling LPM m̃i into
relative displacements ξi, j in the global frame of reference. This uses a simple affine trans-
formation T .

[xi,0 ]m∗ ≡ T ([xi,0 ]m̃i )

= χ placeti ⊕ Li

ξi, j ≡ [xi, j ]m∗ − [xi, j−1 ]m∗

We compute the maximum-likelihood trajectory by satisfying the constraint that

the travel experience between the places must fit the globally defined distance between the
places. First, we find the discrepancy δ between the measured trajectory and the distance
given by the global place layout.

[xi+1,0 ]m∗ = χ placeti+1 ⊕ Li+1 = [xi,ni ]m∗

ni
δ = [xi+1,0 ]m∗ − [xi,0 ]m∗ − ∑ ξi, j
j=1

We assume that the global displacements are distributed with Gaussian uncertain-
ties. The means, µ, are simply the calculated displacements above, and we assume each Λ j
is a diagonal covariance matrix that is proportional to the absolute value of the mean (for
each component in the pose vector ξi, j ).

ξi, j ∼ N (µi, j , Λi, j )

We then calculate the maximum-likelihood displacements in the global frame of reference

202
for each component c in x, y, θ (the pose dimensions).

|µi,c j |
ξˆi,c j = µi,c j + ni c | ·δc
∑h=1 |µi,h

From our maximum-likelihood displacements (ξˆ ), we can estimate the trajectory x in the
global frame of reference.
j
[xi, j ]m∗ = [xi,0 ]m∗ + ∑ ξˆi,h .
h=1

9.3.4 Creating a map m∗

The maximum-likelihood trajectory above can be used as a starting trajectory for gradient
descent methods to align the pose positions with map estimates to converge upon a locally
optimal map [Lu and Milios, 1997; Thrun et al., 2000a]. A more principled approach is
to run a Rao-Blackwellized particle-filtering algorithm, using the maximum-likelihood tra-
jectory as the mean of a proposal distribution: P(x, m|z, u) = P(m|x, z, u) · H(x). However,
we have found that in practice the x values defined by the above scaling method adequately
approximate the mode of the posterior [Modayil, Beeson, and Kuipers, 2004]; thus the
global map can be built by projecting the recorded range measurements from poses in the
new global coordinates. The final map produced from the topological skeleton is shown in
Figure 9.3(c). Compare this to Figure 2.1(d) to see the improved map.

203
Chapter 10

Conclusion

Many places you would like to see are just off the map and many things you want to know are
just out of sight or a little beyond your reach. But someday you’ll reach them all, for what you
learn today, for no reason at all, will help you discover all the wonderful secrets of tomorrow.
Norton Juster, The Phantom Tollbooth (1961)

This chapter concludes the dissertation with various discussions about related future work
and a summary of the thesis. This dissertation describes the Hybrid Spatial Semantic Hier-
archy (HSSH) in detail, formally describing the interaction between small-scale and large-
scale representations of space, and gives preliminary results at all four levels of the hybrid
extension to the SSH: improved metrical localization and mapping in local regions, ro-
bust place detection and classification, more efficient topological map-building, and high-
precision global metrical mapping from topological skeletons.
As noted throughout this dissertation, the current implementation can be improved
upon at all four of the abstraction levels. In addition, we would like extend the human-
robot interaction studies to include experiments that test the entire HSSH interface on a
large population of subjects across various navigation domains. Finally, we plan to relax
the assumptions of deterministic actions, allowing rare undetected places or misclassified
places to be handled gracefully by our logical topological map-building paradigm.

204
10.1 Future Work

10.1.1 The Local Metrical and Topology Levels

The current implementation of the local perceptual map (LPM) uses planar lidar sensors to
build a metrically accurate model of the local surround. This is a typical SLAM formula-
tion. There has been recent work in vision-based SLAM [Sim and Little, 2006], detecting
paths from monocular vision [Posner et al., 2007], and fusing laser data with stereo-vision
data in occupancy grids [Murarka et al., 2006]. Advancements in these fields should make
it possible to create LPMs for sidewalk networks, for streets, and for non-planar environ-
ments, increasing the navigational abilities of a robot using the HSSH.
LPMs that can classify obstacle properties (architecture, doorway, furniture, pedes-
trians, pavement, gravel) as well as annotating regions based on their safety properties (over-
hang, drop-off, ramp) should significantly improve control. Annotated LPMs should also
prove useful at the Local Topology Level, for reasoning about non-structural paths (e.g.,
painted crosswalks) and discriminating between places with aliased local topologies. As
pointed out by Adams et al. [2000], furniture and other obstacles that humans abstract away
may cause the robot to mark gateways at locations where humans do not.
Similarly, one drawback of the current LPM implementation is that we cannot detect
closed doors with range sensors. Currently, if the robot observes a door being opened or
closed, the doorway is seen as free space in the copy of the LPM used for finding gateways
[Modayil and Kuipers, 2004]. This a gateway to be generated for the doorway. However, if
the door remains closed as the robot moves past it, there is no way for a robot utilizing only
range-sensing devices to distinguish the door from the wall. Schröter [2006] investigated
a visual door detection algorithm that he used to define gateways. Incorporating a robust
visual door detection algorithm into the either the LPM or directly into the local topology
abstraction would allow our robot to detect places in front of each doorway and generate
the proper local topology.

205
 Achieving this level of robustness is one of our goals; however, it raises questions
about whether doorways represent a different type of gateway than the types of gateways at
hallway or sidewalk intersections. Doorways seem to be “conditional gateways” while hall-
way intersections seem to be “unconditional gateways”. This may give rise to conditional
versus unconditional places: e.g., when following directions through a building, places in
front of doors are usually not considered; however when looking for a specific office, an
agent may begin to look for doorways until the correct one is found. Similarly, there may
be paths that a robot wheelchair might travel on (e.g., sidewalks), but there may exist other
paths that it models but should not travel on (e.g., streets). (A robot car would have the
reverse situation.) These are examples of the kinds of interesting work in robot navigation
and cognitive map research that remain to be explored.
Another important improvement in classifying places should be relatively straight-
forward to implement in the near future. We should extend the current star representation
of a local topology to better discriminative qualitatively different places that have similar
topology. We have already discussed how using local topology is more descriptive than
simply counting the number of paths leaving a place, or storing the area of a place neigh-
borhood. The current local topology representation cannot distinguish between a T inter-
section at intersecting corridors and one created by a corridor leading into a room larger
than the LPM (i.e., a coastal navigation or wall-following scenario). These two places
share the same topology but have drastically different LPMs. Annotating the local topology
with knowledge about the type of path (i.e., α : S → {M IDLINE, L EFT WALL, R IGHT WALL,
D EAD E ND, N ONE}), will allow us to keep deterministic place recognition, while reducing

perceptual aliasing.
Other useful improvements we hope to add to the LPM implementation are: effi-
cient planning algorithms (e.g., adaptive A∗ [Koenig and Likhachev, 2006]), high-fidelity
control in tight spaces, dynamic obstacle tracking, action models that adapt to surface prop-
erties, and multi-core or GPU parallelism. Other useful improvements we hope to add to

206
the local topology implementation are: efficient extended Voronoi graph implementations
(i.e., a modified version of the dynamic brushfire algorithm [Kalra et al., 2006]), efficient
generation of “virtual” obstacles in (concave) coastal navigation scenarios, robust gateway
alignment algorithms, intelligent exploration of potential places, and semantic labeling of
places. Additionally, improvements to the gateway algorithm (and to gateway alignment)
should be verified using more sophisticated models of local environments. This includes
better models of the types of LPM noise a real robot might experience. We also plan to
stress-test gateways on more places, simple and complex, during evaluation. Perhaps this
requires a large virtual reality world that contains pathological examples that are rare in
university buildings.

Extending Places

The HSSH is a first attempt at integrating small-scale space into a framework for large-scale
knowledge about the environment. By looking at decision points, we can begin to bootstrap
a useful symbolic abstraction from local metrical models built by robots. We believe there
is sometimes a need for larger, more complex places, defined by function or some other
criteria than just gateways.
We plan to begin investigation on how to group metrically separated places together
to form places at a higher-level topology. One good starting point is to group places in
buildings by floor (adding in elevator control as non-deterministic actions along paths).
Another level would be to group all places in a building (or network of buildings), using
sensors to detect when the robot leaves a particular building, or moves outdoors.
Another related area is that of medium-size spaces that have a bounded structure
but are larger than an LPM. A good example is a large research laboratory. Large rooms
are often complex in shape, larger than our typical LPM size, yet still enclosed by walls,
with doors (conditional gateways) defining the exits of the place. Defining where these
scenarios exist and how to handle these scenarios (e.g., adapting the LPM size, clustering

207
places defined by unconditional gateways, etc.) are all important tasks that remain to be
examined.

10.1.2 The Global Topology and Metrical Levels

Currently, the tree of maps contains every topological map consistent with exploration ex-
perience and the topological axioms. This guarantees soundness, which is useful in the
case where observations refute the current best map and the next best map must be iden-
tified. However, there remain two related problems that need to be addressed in future
work. First is the need for a reliable method to identify the best candidate among a set of
possible topological maps, given odometry and perceptual information [Ranganathan et al.,
2006]. Second, is the need to reduce the tree of maps from a “breadth-first” search to a
more focused search that tracks a small number of maps at a time.
In Section 8.3, we identified the “best” map as the simplest one based on a pri-
oritized circumscription policy [Lifschitz, 1995] over the models generated by the non-
monotonic theory of topological maps [Remolina and Kuipers, 2004]. This is sufficient
for the example environment and exploration we have used throughout this dissertation.
Although metrically large, this environment appears to have a simple topology; however,
the simplicity of the topology is deceptive—there is quite a bit of non-local, topological
(graph) symmetry as well as local perceptual aliasing of the places, both of which make
map-building much more difficult. Nonetheless, with the addition of a planarity constraint
[Savelli and Kuipers, 2004], the correct topology is identified as the unique simplest con-
sistent topological map.
In Section 8.3, we also discussed the size of the tree in our example: 46 final maps
consistent with the topological axioms. Savelli and Kuipers [2004] describe larger environ-
ments where extreme symmetry and aliasing cannot so easily be resolved by purely quali-
tative methods because the tree of maps grows too large to maintain in real-time. These are
not entirely unrealistic examples, since large grid-structured neighborhoods in real cities

208
provide opportunities for vast topological ambiguity [Lynch, 1960].1 We would like to in-
vestigate methods for expanding a small set of maps at a time, comparing leaves at different
levels in the tree to identify the best model that explains the robot’s current experience: i.e.,
a “best-first” search approach.
One obvious improvement that will limit the number of map hypotheses is to per-
form active exploration, that occasionally exploits knowledge of the environment to elim-
inate entire branches from the tree of maps. Such strategies are similar to the localization
procedures advocated by proponents of DFA-style maps [Kuipers and Byun, 1991; Dean
et al., 1995; Rekleitis et al., 1999]. Dudek et al. [1991] propose an exploration algorithm
that finds the correct topological structure in only O(n3 ) travel actions (n is the number
of places), but this requires that the robot drop markers and backtrack to determine which
loop-closing hypothesis was correct.
Outside of exploration strategies, we would like reduce the tree of maps by drawing
on perceptual information currently unused in the topological map-building process. We
should be able to use observational data to define weights on the tree of maps. These
weights should allow us to have a quantitative ordering on the map hypotheses, and should
allow best-first style expansion of the tree that focuses on a limited number of highly ranked
candidates at a time, allowing the robot to map larger environments including those with
large amounts of symmetry and perceptual aliasing.

Incorporating the Global Metrical Layout

The tree of maps is an exhaustively enumerated set of topological map hypotheses consis-
tent with the axioms for topological maps and with the agent’s exploration experience. The
maps along a path from the root to the fringe of the tree are not mutually exclusive, but
represent the growth of the topological map during exploration experience. Given the ex-
1 Ranganathan and Dellaert [2005] show that because (in the worst case) the number of aliased places grows

with the amount of exploration experience, the number of possible topological maps is given by Bell’s number,
which grows hyper-exponentially with the number of perceptually aliased places.

209
ploration experience from which the tree of maps was constructed, the maps hM, qi on the
fringe of the tree are an exhaustive set of mutually exclusive topological map hypotheses.
A global topological map provides a structural skeleton, on which a global metrical
map is built. The use of a topological skeleton helps prevent multi-modal distributions in
the metrical uncertainty. At the same time, metrical inference about the layout of places in
the global frame of reference should help discriminate among topological map hypotheses.
Similar to ideas published by Hähnel et al. [2003b] and Ranganathan and Dellaert [2005],
we should be able to compute beliefs for topological map hypotheses by interleaving the
computation of the global place layout with the search for the global topological map.
We believe that we should be able to weight the leaves on the tree of maps by
the likelihood of the global place layout for each map hypothesis: P(χ̂|M,Y ), where Y =
[q0 , . . . , qi ]. Obviously, there needs to be some discount for maps hypotheses that simply
add a new place after each travel. Ranganathan and Dellaert [2005] utilize an odome-
try based penalty for places too close together in their Bayesian topological framework.
Savelli [2005] discusses how, and how well, Bayesian probabilistic reasoning can express
qualitative preferences of topological maps.

Improved Place Recognition

Another compelling idea is to weight the leaves on the tree of maps by the latest observation.
One idea is to replace our current local topology matching with probabilistic matching,
e.g., where a Y intersection might be perceived as a T intersection due to noise. Allowing
non-deterministic place descriptions would make finding the correct global map even more
challenging [Basye et al., 1995]; however, by first comparing the symbolic local topology,
we forgo having to match high-resolution images or LPMs of the current place with all
previously visited places.
If local topology matching and/or the global place layout has a high probability, map
hypotheses could be scored by matching the stored small-scale place maps m p : p ∈ P, or

210
even by using a robust visual place recognition system. In previous work [Kuipers and
Beeson, 2002], we examined learning discriminating features in lidar measurements from
known places. Recently, similar work has been examined by others, investigating visual
place recognition and even semantic labeling of places [Mozos et al., 2005; Tapus and
Siegwart, 2005; Newman et al., 2006; Friedman et al., 2007].
Just as when considering the global layout, one important characteristic is that the
algorithm must have a prior probability for the case where the robot is at a new place. For
example, Cummins and Newman [2008] present a robust visual recognition system that
gives a reasonable prior to visual observations based on the database of images used for
training. Incorporating this visual place detection system into the map-building process is
a likely candidate for extending the current Hybrid SSH implementation, and we expect
by using the HSSH place detection we can show large improvements in properly detecting
loop closures from vision.

10.2 Summary

We have presented a hybrid metrical/topological framework that processes information at

both small-scale and large-scale abstractions. Our Hybrid Spatial Semantic Hierarchy is
inspired by human cognitive maps; thus, it represents the environment using human-like
concepts, such as places and paths, which support hierarchical navigation, human-robot in-
teraction, and logical reasoning. Specifically, we focused on the problem of map-building—
discussing how the HSSH builds metrical representations for local small-scale spaces, finds
a topological map representing the qualitative structure of large-scale space, and constructs
a metrical representation for large-scale space in a single global frame of reference by build-
ing on the skeleton provided by the topological map.
Unlike many robotic implementations that attempt to build a monolithic, Cartesian
global metrical map, we propose an alternative approach that handles closing large loops by
hypothesizing symbolic place matches. This ensures all possible loop closures are consid-

211
ered, not just ones where the robot, with accumulated odometry error, happens to be near
some older portion of the map. The minimal topological map that results from large-scale
exploration is sufficient for navigation and necessary for efficient planning, especially to
rule out alternative topological structures during exploration.
The thrust of this dissertation has been to formally describe how concepts of large-
scale space can be grounded in the robot’s low-level observations. This problem has hin-
dered topological map-building research, as it is an example of the hard AI problem of
symbol grounding [Harnad, 1990]. Our innovation has been to utilize metrical approaches
to model the immediate, local surround of the robot in order to ground gateways in small-
scale space. Gateways provide the robot with local motion targets that facilitate control
along paths. They also provide a local topology description of the local surround, useful for
detecting and describing places and the paths that emanate from places.
Additionally, we detailed improvements to the traditional incremental localization
schemes in order to ensure high-quality local metrical models. We demonstrated the robust-
ness of local topology abstraction from gateways over various transformations and noise
characteristics of the local metrical models. We validated the Local Topology and Local
Metrical Levels of the HSSH with a human-robot interaction experiment for visually im-
paired drivers of a robot.
At the Global Topology and Global Metrical Levels of the HSSH, we provided
insight into the computational benefits of using local topology when building the tree of
possible topological maps, and we demonstrated that a global layout of places is easily
achieved given a topological map hypothesis. We presented an implementation of HSSH
global topological map-building within an environment with fairly large, nested loop clo-
sures. The results support our claims of efficient, online map-building in the presence of
multiple loop closures. We also detailed an efficient algorithm for achieving a full global
metrical map by filling in exploration experience along the path segments that connect
places in the environment.

212
Bibliography

W. Adams, D. Perzanowski, and A. Schultz. Learning, storage, and use of spatial informa-
tion in a robotics domain. In Proceedings of the International Conference on Machine
Learning Workshop on Machine Learning of Spatial Knowledge, pages 23–27, Palo Alto,
California, 2000.

V. Aginsky, C. Harris, R. Rensink, and J. Beusmans. Two strategies for learning a route in
a driving simulator. Journal of Environmental Psychology, 17:317–331, 1997.

G. L. Allen, K. C. Kirasic, A. W. Siegel, and J. F. Herman. Developmental issues in cogni-

tive mapping: The selection and utilization of environmental landmarks. Child Develop-
ment, 50:1062–1070, 1979.

D. Angluin. On the complexity of minimum inference of regular sets. Information and

Control, 39:337–350, 1978.

D. Appleyard. Why buildings are known: a predictive tool for architects and planners.
Environment and Behavior, 1(2), December 1969.

D. Appleyard. Styles and methods of structuring a city. Environment and Behavior, 2(1),
June 1970.

K. Basye, T. Dean, and L. P. Kaelbling. Learning dynamics: system identification for

perceptually challenged agents. Artificial Intelligence, 72:139–171, 1995.

213
K. Basye, T. Dean, and J. S. Vitter. Coping with uncertainty in map learning. Machine
Learning, 29(1):65–88, 1997.

P. Beeson. EVG-Thin software, 2006. URL http://www.cs.utexas.edu/users/qr/

software/evg-thin.html.

P. Beeson, M. MacMahon, J. Modayil, J. Provost, F. Savelli, and B. Kuipers. Exploiting

local perceptual models for topological map-building. In Proceedings of the International
Workshop on Reasoning with Uncertainty in Robotics (RUR), pages 15–22, Acapulco,
Mexico, 2003.

P. Beeson, N. K. Jong, and B. Kuipers. Towards autonomous topological place detection

using the extended Voronoi graph. In Proceedings of the IEEE International Conference
on Robotics and Automation (ICRA), pages 4373–4379, Barcelona, Spain, April 2005.

P. Beeson, A. Murarka, and B. Kuipers. Adapting proposal distributions for accurate, effi-
cient mobile robot localization. In Proceedings of the IEEE International Conference on
Robotics and Automation (ICRA), pages 49–55, Orlando, Florida, 2006.

P. Beeson, M. MacMahon, J. Modayil, A. Murarka, B. Kuipers, and B. Stankiewicz. Inte-

grating multiple representations of spatial knowledge for mapping, navigation, and com-
munication. In Proceedings of the Symposium on Interaction Challenges for Intelligent
Assistants, AAAI Spring Symposium Series, pages 1–9, Stanford, CA, 2007. AAAI
Technical Report SS-07-04.

R. Biegler and R. G. M. Morris. Landmark stability: Studies exploring whether the per-
ceived stability of the environment influences spatial representation. Journal of Experi-
mental Biology, 199:187–193, 1996.

J.-L. Blanco, J.-A. Fernández-Madrigal, and J. González. Toward a unified bayesian ap-
proach to hybrid metric-topological SLAM. IEEE Transactions on Robotics, 24(2):259–
270, April 2008.

214
J. Borenstein and L. Feng. Measurement and correction of systematic odometry errors in
mobile robots. IEEE Transactions on Robotics and Automation, 12(6):869–880, Decem-
ber 1996.

J. Borenstein and Y. Koren. The Vector Field Histogram—fast obstacle avoidance for mo-
bile robots. IEEE Transactions on Robotics and Automation, 7(3):278–288, 1991.

M. Bosse, P. Newman, J. Leonard, M. Soika, W. Feiten, and S. Teller. An Atlas framework

for scalable mapping. In Proceedings of the IEEE International Conference on Robotics
and Automation (ICRA), pages 1899–1906, Taipei, Taiwan, 2003.

P. Buschka. An Investigation of Hybrid Maps for Mobile Robots. PhD thesis, Örebro
University, 2005.

Y.-T. Byun. Spatial Learning Mobile Robots with a Spatial Semantic Hierarchical Model.
PhD thesis, The University of Texas at Austin, 1990.

C.-H. Choi, J.-B. Song, W. Chung, and M. Kim. Topological map building based on thin-
ning and its application to localization. In Proceedings of the IEEE/RSJ Conference on
Intelligent Robots and Systems (IROS), pages 552–557, Lausanne, Switzerland, 2002.

H. Choset and J. Burdick. Sensor-based exploration: the hierarchical generalized Voronoi

graph. International Journal of Robotics Research, 19(2):96–125, 2000.

H. Choset and K. Nagatani. Topological simultaneous localization and mapping (SLAM):

toward exact localization without explicit localization. IEEE Transactions on Robotics
and Automation, 17(2):125–137, April 2001.

H. Choset, S. Walker, K. Eiamsa-Ard, and J. Burdick. Sensor-based exploration: incremen-

tal construction of the hierarchical generalized Voronoi graph. International Journal of
Robotics Research, 19(2):125–148, 2000.

215
E. Chown. Making predictions in an uncertain world: Environmental structure and cogni-
tive maps. Adaptive Behavior, 8(1):17–33, 1999.

E. Chown. Gateways: An approach to parsing spatial domains. In Proceedings of the In-

ternational Conference on Machine Learning Workshop on Machine Learning of Spatial
Knowledge, pages 1–6, Palo Alto, California, 2000.

E. Chown, S. Kaplan, and D. Kortenkamp. Prototypes, location, and associative networks

(PLAN): Towards a unified theory of cognitive mapping. Cognitive Science, 19(1):1–51,
1995.

K. J. W. Craik. The nature of exploration. Cambridge University Press, London, 1943.

M. Cummins and P. Newman. FAB-MAP: Probabilistic localization and mapping in the

space of appearance. International Journal of Robotics Research, 27(6):647–665, 2008.

M. Cummins and P. Newman. Probabilistic appearance based navigation and loop clos-
ing. In Proceedings of the IEEE International Conference on Robotics and Automation
(ICRA), pages 2042–2048, Rome, Italy, 2007.

T. Dean, D. Angluin, K. Basye, S. Engelson, L. Kaelbling, E. Kokkevis, and O. Maron. In-

ferring finite automata with stochastic output functions and an application to map learn-
ing. Machine Learning, 18(1):81–108, 1995.

B. N. Delaunay. Sur la sphère vide. Bulletin of Academy of Sciences of the USSR, 7:

793–800, 1934.

R. M. Downs and D. Stea. Image and Environment. Aldine Publishing Company, Chicago,
1973.

T. Duckett and U. Nehmzow. Exploration of unknown environments using a compass,

topological map and neural network. In Proceedings of the IEEE International Sympo-

216
 sium on Computational Intelligence in Robotics and Automation (CIRA), pages 312–317,
Monterey, California, 1999.

T. Duckett and A. Saffiotti. Building globally consistent gridmaps from topologies. In

Proceedings of the International IFAC Symposium on Robot Control (SYROCO), pages
357–361, Vienna, Austria, 2000.

G. Dudek, M. Jenkin, E. Milios, and D. Wilkes. Robotic exploration as graph construction.

IEEE Transactions on Robotics and Automation, 7(6):859–865, 1991.

G. Dudek, P. Freedman, and S. Hadjres. Using local information in a non-local way for
mapping graph-like worlds. In Proceedings of the International Joint Conference on
Artificial Intelligence (IJCAI), pages 1639–1647, Chambéry, France, 1993.

A. Elfes. Occupancy Grids: A Probabilistic Framework for Robot Perception and Naviga-
tion. PhD thesis, Carnegie Mellon University, 1989.

A. Eliazar and R. Parr. DP-SLAM: Fast, robust simultaneous localization and mapping
without predetermined landmarks. In Proceedings of the International Joint Conference
on Artificial Intelligence (IJCAI), pages 1135–1142, Acapulco, Mexico, 2003.

A. I. Eliazar and R. Parr. Learning probabilistic motion models for mobile robots. In
Proceedings of the International Conference on Machine Learning (ICML), volume 69,
page 32, Banff, Canada, 2004.

I. Filippov. OSRA: Optical structure recognition, v.0.9.9, 2007. URL http://osra.

sourceforge.net/.

S. Fortune. Voronoi diagrams and Delaunay triangulations. In D.-Z. Du and F. Hwang,

editors, Computing in Euclidean Geometry, volume 1 of Lecture Notes Series on Com-
puting, pages 193–234. World Scientific, 1992.

217
D. Fox. Adapting the sample size in particle filters through KLD-sampling. International
Journal of Robotics Research, 22(12):985–1003, 2003.

D. Fox, W. Burgard, and S. Thrun. The dynamic window approach to collision avoidance.
IEEE Robotics & Automation Magazine, 4(1), March 1997.

D. Fox, W. Burgard, and S. Thrun. Markov localization for mobile robots in dynamic
environments. Journal of Artificial Intelligence Research, 11:391–427, 1999.

S. Friedman, H. Pasula, and D. Fox. Voronoi random fields: Extracting the topological
structure of indoor environments via place labeling. In Proceedings of the International
Joint Conference on Artificial Intelligence (IJCAI), pages 2109–2114, Hyderabad, India,
2007.

N. Gale, R. G. Golledge, J. W. Pellegrino, and S. Doherty. The acquisition and integration of

route knowledge in an unfamiliar neighborhood. Journal of Environmental Psychology,
10(1):3–25, 1990.

C. R. Gallistel. The Organization of Learning. MIT Press, Cambridge, Massachusetts,

1990.

T. Gladwin. East is a Big Bird: Navigation and Logic on Puluwat Atoll. Harvard University
Press, Cambridge, Massachusetts, 1970.

E. M. Gold. Complexity of automaton identification from given sets. Information and

Control, 37:302–320, 1978.

R. G. Golledge, editor. Wayfinding Behavior: Cognitive Mapping and Other Spatial Pro-
cesses. The Johns Hopkins University Press, Baltimore, Maryland, 1999.

R. G. Golledge, T. R. Smith, J. W. Pellegrino, S. Doherty, and S. P. Marshall. A conceptual

model and empirical analysis of children’s acquisition of spatial knowledge. Journal of
Environmental Psychology, 5(2):125–152, 1985.

218
J. L. Gould. Honey bee cognition. Cognition, 37:83–103, 1990.

G. Grisetti, C. Stachniss, and W. Burgard. Improving grid-based SLAM with Rao-

Blackwellized particle filters by adaptive proposals and selective resampling. In Pro-
ceedings of the IEEE International Conference on Robotics and Automation (ICRA),
pages 2443–2448, Barcelona, Spain, 2005.

T. Guilford, S. Roberts, D. Biro, and I. Rezek. Positional entropy during pigeon homing II:
navigational interpretation of Bayesian latent state models. Theoretical Biology, 227(1):
25–38, March 2004.

D. Hähnel, W. Burgard, D. Fox, and S. Thrun. An efficient FastSLAM algorithm for gen-
erating maps of large-scale cyclic environments from raw laser range measurements.
In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS),
pages 206–211, Las Vegas, Nevada, 2003a.

D. Hähnel, S. Thrun, B. Wegbreit, and W. Burgard. Towards lazy data association in SLAM.
In Proceedings of the International Symposium on Robotics Research (ISRR), pages 83–
105, Sienna, Italy, 2003b.

S. Harnad. The symbol grounding problem. Physica D, 42:335–346, 1990.

J. Haugeland. Artificial Intelligence: The Very Idea. MIT PRess, Cambridge, Mas-
sachusetts, 1985.

H. Heft. The role of environmental features in route learning: Two exploratory studies of
way finding. Environmental Psychology and Nonverbal Behavior, 3:172–185, 1979.

N. Kalra, D. Ferguson, and A. Stentz. Incremental reconstruction of generalized Voronoi

diagrams on grids. In Proceedings of the International Conference on Intelligent Au-
tonomous Systems (IAS), pages 114–123, Tokyo, Japan, 2006.

219
L. E. Kavraki, P. Svestka, J.-C. Latombe, and M. H. Overmars. Probabilistic roadmaps for
path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics
and Automation, 12(4):566–580, August 1996.

R. M. Kitchin. Cognitive maps: What are they and why study them? Journal of Environ-
mental Psychology, 14(1):1–19, March 1994.

B.-Y. Ko, J.-B. Song, and S. Lee. Real-time building of a thinning-based topological map
with metric features. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots
and Systems (IROS), pages 797–802, Sendai, Japan, 2004.

S. Koenig and M. Likhachev. Real-time adaptive A∗ . In Proceedings of the International

Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 281–
288, Hakodate, Japan, 2006.

S. Koenig and R. G. Simmons. Unsupervised learning of probabilistic models for robot

navigation. In Proceedings of the IEEE International Conference on Robotics and Au-
tomation (ICRA), pages 2301–2308, Minneapolis, Minnesota, 1996.

K. Konolige. Large-scale map-making. In Proceedings of the National Conference on

Artificial Intelligence (AAAI), pages 457–463, San Jose, California, 2004.

K. Konolige. Markov localization using correlation. In Proceedings of the International

Joint Conference on Artificial Intelligence (IJCAI), pages 1154–1159, Stockholm, Swe-
den, 1999.

K. Konolige. A gradient method for realtime robot control. In Proceedings of the IEEE/RSJ
Conference on Intelligent Robots and Systems (IROS), pages 639–646, Takamatsu, Japan,
2000.

D. Kortenkamp and T. Weymouth. Topological mapping for mobile robots using a combina-
tion of sonar and vision sensing. In Proceedings of the National Conference on Artificial
Intelligence (AAAI), pages 979–984, Seattle, Washington, 1994.

220
J. J. Kuffner and S. M. LaValle. RRT-Connect: An efficient approach to single-query path
planning. In Proceedings of the IEEE International Conference on Robotics and Automa-
tion (ICRA), pages 995–1001, San Francisco, California, 2000.

B. Kuipers. The Spatial Semantic Hierarchy. Artificial Intelligence, 119:191–233, 2000.

B. Kuipers. An intellectual history of the Spatial Semantic Hierarchy. In M. E. Jefferies and

W.-K. Yeap, editors, Robotics and Cognitive Approaches to Spatial Mapping, volume 38
of Springer Tracts in Advanced Robotics, pages 243–264. Springer, Berlin, Germany,
2008.

B. Kuipers and P. Beeson. Bootstrap learning for place recognition. In Proceedings of

the National Conference on Artificial Intelligence (AAAI), pages 174–180, Edmonton,
Canada, 2002.

B. Kuipers, J. Modayil, P. Beeson, M. MacMahon, and F. Savelli. Local metrical and global
topological maps in the Hybrid Spatial Semantic Hierarchy. In Proceedings of the IEEE
International Conference on Robotics and Automation (ICRA), pages 4845–4851, New
Orleans, Louisiana, 2004.

B. J. Kuipers. Modeling spatial knowledge. Cognitive Science, 2:129–153, 1978.

B. J. Kuipers. The ‘Map in the Head’ metaphor. Environment and Behavior, 14(2):202–220,
1982.

B. J. Kuipers and Y.-T. Byun. A robot exploration and mapping strategy based on a semantic
hierarchy of spatial representations. Journal of Robotics and Autonomous Systems, 8:47–
63, 1991.

G. Lakoff and M. Johnson. Metaphors We Live By. The University of Chicago Press,
Chicago, 1980.

221
A. Lankenau, T. Röfer, and B. Krieg-Brückner. Self-localization in large-scale environ-
ments for the bremen autonomous wheelchair. In Spatial Cognition III, volume 2685 of
Lecture Notes in Artificial Intelligence, pages 34–61. Springer-Verlag, Berlin, Germany,
2002.

W.-Y. Lee. Spatial Semantic Hierarchy for a Physical Mobile Robot. PhD thesis, The
University of Texas at Austin, 1996.

J. Leonard and P. Newman. Consistent, convergent, and constant-time SLAM. In Pro-

ceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages
1143–1150, Acapulco, Mexico, 2003.

V. Lifschitz. Nested abnormality theories. Artificial Intelligence, 74:351–365, 1995.

F. Lu and E. Milios. Globally consistent range scan alignment for environment mapping.
Autonomous Robots, 4:333–349, 1997.

K. Lynch. The Image of the City. MIT Press, Cambridge, Massachusetts, 1960.

M. MacMahon. Understanding and Executing Natural Language Route Instructions. PhD

thesis, The University of Texas at Austin, 2007.

M. MacMahon, B. Stankiewicz, and B. J. Kuipers. Walk the talk: Connecting language,

knowledge, action in route instructions. In Proceedings of the National Conference on
Artificial Intelligence (AAAI), pages 1475–1482, Boston, Massachusetts, 2006.

M. J. Mataric. Integration of representation into goal-driven behavior-based robots. IEEE

Transactions on Robotics and Automation, 8(3):304–312, 1992.

E. Menegatti. Omnidirectional Vision for Mobile Robotics. PhD thesis, The University of
Padua, 2002.

222
D. Miller. A spatial representation system for mobile robots. In Proceedings of the IEEE
International Conference on Robotics and Automation (ICRA), pages 122–127, St. Louis,
Missouri, 1985.

J. Modayil. Robot Developmental Learning of an Object Ontology Grounded in Sensori-

motor Experience. PhD thesis, The University of Texas at Austin, 2007.

J. Modayil and B. Kuipers. Bootstrap learning for object discovery. In Proceedings of the
IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), pages 742–747, Sendai,
Japan, 2004.

J. Modayil, P. Beeson, and B. Kuipers. Using the topological skeleton for scalable, global,
metrical map-building. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots
and Systems (IROS), pages 1530–1536, Sendai, Japan, 2004.

M. Montemerlo, S. Thrun, D. Koller, and B. Wegbreit. FastSLAM: a factored solution

to the simultaneous localization and mapping problem. In Proceedings of the National
Conference on Artificial Intelligence (AAAI), pages 593–598, Edmonton, Canada, 2002.

M. Montemerlo, S. Thrun, D. Koller, and B. Wegbreit. FastSLAM 2.0: an improved particle

filtering algorithm for simultaneous localization and mapping that provably converges.
In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI),
pages 1151–1156, Acapulco, Mexico, 2003.

H. P. Moravec. Sensor fusion in certainty grids for mobile robots. AI Magazine, pages
61–74, Summer 1988.

A. C. Morris, D. Silver, D. Ferguson, and S. Thayer. Towards topological exploration of

abandoned mines. In Proceedings of the IEEE International Conference on Robotics and
Automation (ICRA), pages 2117–2123, Barcelona, Spain, 2005.

O. M. Mozos, C. Stachniss, and W. Burgard. Supervised learning of places from range data

223
 using Adaboost. In Proceedings of the IEEE International Conference on Robotics and
Automation (ICRA), pages 1742–1747, Barcelona, Spain, 2005.

A. Murarka, J. Modayil, and B. Kuipers. Building local safety maps for a wheelchair robot
using vision and lasers. In Proceedings of the Canadian Conference on Computer and
Robot Vision (CRV), page 25, Quebec City, Canada, 2006.

P. Newman, D. Cole, and K. Ho. Outdoor SLAM using visual appearance and laser rang-
ing. In Proceedings of the IEEE International Conference on Robotics and Automation
(ICRA), pages 1180–1187, Orlando, Florida, 2006.

E. Olson, J. Leonard, and S. Teller. Fast iterative optimization of pose graphs with poor
initial estimates. In Proceedings of the IEEE International Conference on Robotics and
Automation (ICRA), pages 2262–2269, Orlando, Florida, 2006.

M. A. Paskin. Thin junction tree filters for simultaneous localization and mapping. In Pro-
ceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages
1157–1166, Acapulco, Mexico, 2003.

J. Piaget. The Construction of Reality in the Child. Ballantine, New York, 1954.

I. Posner, D. Schröter, and P. Newman. Describing composite urban workspaces. In Pro-

ceedings of the IEEE International Conference on Robotics and Automation (ICRA),
pages 4962–4968, Rome, Italy, 2007.

W. H. Press, S. A. Teukolsky, W. T. Vitterling, and B. P. Flannery. Numerical Recipes in C:

The Art of Scientific Computing. Cambridge University Press, London, second edition,
1992.

A. Ranganathan and F. Dellaert. Data driven MCMC for appearance-based topological

mapping. In Proceedings of Robotics: Science and Systems (RSS), pages 209–216, Cam-
bridge, Massachusetts, 2005.

224
A. Ranganathan, E. Menegatti, and F. Dellaert. Bayesian inference in the space of topolog-
ical maps. IEEE Transactions on Robotics, 22(1):92–107, February 2006.

N. S.V. Rao. Robot navigation in unknown generalized polygonal terrains using vision
sensors. IEEE Transactions on Systems, Man, and Cybernetics, 25(66):947–962, 1995.

I. Rekleitis. Cooperative Localization and Multi-Robot Exploration. PhD thesis, School of

Computer Science, McGill University, Montreal, Quebec, Canada, 2003.

I. M. Rekleitis, V. Dujmovic, and G. Dudek. Efficient topological exploration. In Proceed-

ings of the IEEE International Conference on Robotics and Automation (ICRA), pages
676–681, Detroit, Michigan, 1999.

E. Remolina. Formalizing the Spatial Semantic Hierarchy. PhD thesis, The University of
Texas at Austin, 2001.

E. Remolina and B. Kuipers. Towards a general theory of topological maps. Artificial

Intelligence, 152:47–104, 2004.

R. L. Rivest and R. E. Schapire. Inference of finite automata using homing sequences. In

Proceedings of the ACM Symposium on Theory of Computing (STOC), pages 411–420,
Seattle, Washington, 1989.

N. Roy, W. Burgard, D. Fox, and S. Thrun. Coastal navigation – mobile robot navigation
with uncertainty in dynamic environments. In Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA), pages 35–40, 1999.

F. Savelli. Topological Mapping of Ambiguous Space: Combining Qualitative Biases and

Metrical Information. PhD thesis, University of Rome ”La Sapienza”, 2005.

F. Savelli and B. Kuipers. Loop-closing and planarity in topological map-building. In Pro-

ceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), pages
1511–1517, Sendai, Japan, 2004.

225
D. Schröter. Region & Gateway Mapping: Acquiring Structured and Object-Oriented Rep-
resentations of Indoor Environments. PhD thesis, Technical University of Munich, 2006.

D. Schröter, T. Weber, M. Beetz, and B. Radig. Detection and classification of gateways for
the acquisition of structured robot maps. In Proceedings of the Symposium of the Ger-
man Association for Pattern Recognition (DAGM), pages 553–561, Tübingen, Germany,
2004.

H. Shatkay and L. P. Kaelbling. Learning topological maps with weak local odometric in-
formation. In Proceedings of the International Joint Conference on Artificial Intelligence
(IJCAI), pages 920–929, Nagoya, Japan, 1997.

A. W. Siegel and S. H. White. The development of spatial representations of large-scale

environments. In H. W. Reese, editor, Advances in Child Development and Behavior,
volume 10, pages 9–55. Academic Press, New York, 1975.

D. Silver, D. Ferguson, A. C. Morris, and S. Thayer. Feature extraction for topological mine
maps. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems
(IROS), pages 773–779, Sendai, Japan, 2004.

R. Sim and J. J. Little. Autonomous vision-based exploration and mapping using hybrid
maps and Rao-Blackwellised particle filters. In Proceedings of the IEEE/RSJ Conference
on Intelligent Robots and Systems (IROS), pages 2082–2089, Beijing, China, 2006.

M. Skubic, S. Blisard, C. Bailey, J. A. Adams, and P. Matsakis. Qualitative analysis of

sketched route maps: translating a sketch into linguistic descriptions. IEEE Transactions
on Systems, Man and Cybernetics – Part B: Cybernetics, 34(2):1275–1282, 2004a.

M. Skubic, D. Perzanowski, S. Blisard, A. Schultz, W. Adams, M. Bugajska, and D. Brock.

Spatial language for human-robot dialogs. IEEE Transactions on Systems, Man and
Cybernetics – Part C: Applications and Reviews, 34(2):154–167, 2004b.

226
R. Smith, M. Self, and P. Cheeseman. Estimating uncertain spatial relationships in robotics.
In I. J. Cox and G. T. Wilfong, editors, Autonomous Robot Vehicles, pages 167–193.
Springer Verlag, New York, 1990.

B. J. Stankiewicz and A. A. Kalia. Acquisition of structural versus object landmark knowl-

edge. Journal of Experimental Psychology: Human Perception and Performance, 33(2):
378–390, April 2007.

P. Stopka and D. W. Macdonald. Way-marking behaviour: an aid to spatial navigation in

the wood mouse (Apodemus sylvaticus). BMC Ecology, 3(3), 2003.

A. Tapus and R. Siegwart. Incremental robot mapping with fingerprints of places. In Pro-
ceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), pages
2429–2434, Edmonton, Canada, 2005.

S. Thrun and A. Bücken. Integrating grid-based and topological maps for mobile robot
navigation. In Proceedings of the National Conference on Artificial Intelligence (AAAI),
pages 944–950, Portland, Oregon, 1996.

S. Thrun, S. Gutmann, D. Fox, W. Burgard, and B. J. Kuipers. Integrating topological

and metric maps for mobile robot navigation: A statistical approach. In Proceedings
of the National Conference on Artificial Intelligence (AAAI), pages 989–995, Madison,
Wisconsin, 1998.

S. Thrun, W. Burgard, and D. Fox. A real-time algorithm for mobile robot mapping with
applications to multi-robot and 3d mapping. In Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA), pages 321–328, San Francisico, Cali-
fornia, 2000a.

S. Thrun, D. Fox, and W. Burgard. Monte Carlo localization with mixture proposal dis-
tribution. In Proceedings of the National Conference on Artificial Intelligence (AAAI),
pages 859–865, Austin, Texas, 2000b.

227
S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics. MIT Press, Cambridge, Mas-
sachusetts, 2005.

E. C. Tolman. Cognitive maps in rats and men. The Psychological Review, 55(4):189–208,
1948.

N. Tomatis, I. Nourbakhsh, and R. Siegwart. Hybrid simultaneous localization and map

building: Closing the loop with multi-hypotheses tracking. In Proceedings of the IEEE
International Conference on Robotics and Automation (ICRA), pages 2749–2754, Wash-
ington, DC, 2002.

L. G. Valiant. Robust logics. Artificial Intelligence, 117(2):231–253, 2000.

A. C. Victorino, P. Rives, and J.-J. Borrelly. Safe navigation for indoor mobile robots. Part
I: a sensor-based navigation framework. International Journal of Robotics Research, 22
(12):1005–1018, 2003.

G. Voronoi. Nouvelles applications des paramètres continus à la théorie des formes quadra-
tiques. Journal for Pure and Applied Mathematics, 133:97–178, 1907.

J. O. Wallgrün. Autonomous construction of hierarchical voronoi-based route graph repre-

sentations. In Spatial Cognition IV. Reasoning, Action, Interaction, volume 3343 of Lec-
ture Notes in Artificial Intelligence, pages 413–433. Springer-Verlag, Berlin, Germany,
2005.

C. M. Wang. Location estimation and uncertainty analysis for mobile robots. In Proceed-
ings of the IEEE International Conference on Robotics and Automation (ICRA), pages
1231–1235, Philadelphia, Pennsylvania, 1988.

WorldViz LLC. Vizard virtual reality toolkit v3.0. http://www.worldviz.com/

products/vizard/index.html, 2007.

228
B. Yamauchi. A frontier-based approach for autonomous exploration. In Proceedings of
the IEEE International Symposium on Computational Intelligence in Robotics and Au-
tomation (CIRA), pages 146–151, Monterey, California, 1997.

M. Yannakakis and D. Lee. Testing finite state machines. In Proceedings of the ACM
Symposium on Theory of Computing (STOC), pages 476–485, New Orleans, Louisiana,
1991.

F. S. Yates. The Art of Memory. The University of Chicago Press, Chicago, 1966.

W.-K. Yeap. Towards a computational theory of cognitive maps. Artificial Intelligence, 34:
297–360, 1988.

W.-K. Yeap and M. E. Jefferies. Computing a representation of the local environment.

Artificial Intelligence, 107(2):265–301, 1999.

G. M. Youngblood, L. B. Holder, and D. J. Cook. A framework for autonomous mobile

robot exploration and map learning through the use of place-centric occupancy grids.
In Proceedings of the International Conference on Machine Learning Workshop on Ma-
chine Learning of Spatial Knowledge, pages 13–20, Palo Alto, California, 2000.

T. Y. Zhang and C. Y. Suen. A fast parallel algorithm for thinning digital patterns. Commu-
nications of the ACM, 27(3):236–239, March 1984.

U. R. Zimmer. Embedding local metrical map patches in a globally consistent topological

map. In Proceedings of the International Symposium on Underwater Technology (UT),
pages 301–305, Tokyo, Japan, 2000.

229
Vita

Patrick Beeson was born in Baton Rouge, Louisiana in 1977. He received a B.S. in Com-
puter Science from Tulane University in 1999. Since that time, he has been working on his
doctorate at the University of Texas at Austin. In addition to the hybrid spatial knowledge
framework described in his thesis, Patrick is interested in robot learning, creating efficient
implementations of common robotic algorithms, and human-robot interaction. In 2007, he
was a lead programmer on the Austin Robot Technology team, supervising undergraduate
programmers to produce an autonomous vehicle. This vehicle placed among the top 21
teams in the semi-finals of the DARPA Urban Challenge. After graduation, Patrick plans
to hone his teaching skills by continuing to supervise undergraduate research on the au-
tonomous vehicle platform.

Permanent Address: 3840 Far West Blvd

#206
Austin, TX 78731

This dissertation was typeset by the author.

230