You are on page 1of 62

PSfrag replacements

o [ ]
i
i
[ ]
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

Quadruped robot control and variable leg


transmissions

JOHAN INGVAST

Doctoral Thesis
Stockholm, Sweden 2006

TRITA MMK 2006-02


ISSN 1400-1179
KTH, School of Industrial Engineering and Management
ISRN KTH/MMK/R--06/02--SE
SE-100 44 Stockholm
ISBN 9171782575
Sweden
Akademisk avhandling som med tillstnd av Kungl Tekniska hgskolan framlgges
till offentlig granskning fr avlggande av teknologie doktorsexamen i maskinkonstruktion fredagen den 10 februari 2006 klockan 10.00 i M2, KTH, Brinellvgen 64,
Stockholm.
Johan Ingvast, Januari 2006
Tryck: Universitetsservice US AB

Till
Monica
Moa och Jakob

Abstract
The research presented in this thesis regards walking of quadruped robots,
and particularly the walking of the WARP 1 robot. The motivation for the
robot is to provide a platform for autonomous walking in rough terrain.
The thesis contains six papers ranging from development tools to actuation of robot legs. The first paper describes the methods and tools made for
control development. These tools feature: programming of the robot without low level coding (C-code); that the controller has to be built only once
for simulation and experiments; and that names of variables and constants
are unchanged through the chain of software Maple Matlab Simulink
Real Time Workshop xPCTarget.
Three controllers, each making the robot walk are presented. The first
controller makes the robot walk using the crawl gait. The method uses
static stability as method for keeping balance and the instantaneous trunk
motions are given by a concept using the so called weight-ratios. A method
for planning new footholds based on the positions of the existing footholds
is also proposed and the controller experimentally verified.
The second walking controller shows that the robot also can walk dynamically using the trot gait. The method proposed uses information from ground
contact sensors on the feet as input to control balance, instead of, which is
common, inertial sensors. It is experimentally verified that WARP 1 can trot
from level ground onto a slope and turn around while staying balanced.
The main ideas of these two walking controllers are fused in the third
which enables smooth transitions between crawl and trot. The idea of using
the ground contact sensors from the first controller is here used to estimate
the position of the center of mass. This controller uses weight-ratios in the
gait crawl as well as in the dynamic gait trot. Hence, the method of using
weight-ratios is not only useful for static stability for which it was originally
intended. The controller is experimentally verified on WARP 1.
The WARP 1 robot weighs about 60 kg, has 0.6 m long legs with three
actuated joints on each. The speed and strength is sufficient only for slow
walking, even though the installed power indicates that it should be enough
for faster walking. The reason is that a walking robot often needs to be strong
but slow when the feet are on the ground and the opposite when in the air.
This can not be achieved with the motors and transmissions currently used.
A transmission called the passively variable transmission (PVT) is proposed which enhance motor capabilities of robot joints. It is elastic, nonlinear
and conservative. Some general properties for elastic transmissions are derived such that they can be compared with conventional transmissions. The
PVT gives strong actuation at large loads and fast actuation at small loads.
The proposed transmission is compared to a conventional transmission for a
specific task, and the result is that a smaller motor can be used.
v

Preface
PhD studies have been a fantastic experience for me, of course lots of hard work,
late hours and desperation, but also great friends, parties and enthusiasm. The
CAS research group was pretty large when I joined in 1999. Head of the research
was Tom, who later went to Canada and now works with banking. Lennart handed
in his Licentiate thesis just a few months after I arrived and then left for the
industry. Freyr received his PhD a few years later and went home to Island, where
he now works with banking. Christian is still in the Stockholm area and sticks to
mechatronics, designing satelite equipment. Others have also been involved in the
research group: PhD students from other departments; technicians Micke, Johan
W. and Andreas; master students . . . most of them I dont know where they are
today. Thanks all, it has been enjoyable to work with you. Well see where I go
after finishing abroad?1
At the moment Im writing this, I have Johan T. in the same room, and Fredrik
in the next where also Magnus should be. Thanks for being there. Some occasions
related to my years at the department are more memorable than others:
Lunches at Restaurant Cypern. Thanks Payam, Peter and Christian for making those lunches so special
Johan T. organized a ski weekend in Kittelfjll. We had some days of great
skiing Johan stayed home being ill.
To all the friends that I should have cared about but not had the strength now,
Ill be less involved, please accept me as a friend again.
Jan, my supervisor, thanks for letting me follow my own research compass, and
for always taking the time, but most of all thanks for letting me do research on
walking robots.
Christian, we have written a number of papers together. It has been a pleasure
to work with you and I hope we will do it again.
Mom and dad, thanks for all the support.
My largest appreciation goes to my family: Moa, Jakob and particularly Monica.
Thanks for coping with me during these years.
Stockholm, January 2006
Johan Ingvast

Acknowledgment
This work was supported by the Foundation for Strategic Research through the
Centre of Autonomous Systems (CAS) at the Royal Institute of Technology.
1 No,

I could not stand working at a bank.

vii

List of published papers


Christian Ridderstrm, Johan Ingvast, Freyr Hardarson, Mats Gudmundsson,
Mikael Hellgren, Jan Wikander, Tom Wadden, and Henrik Rehbinder. The basic
design on the quadruped robot Warp1. In Professional Engineering Publishing
Limited, editor, 3rd International Conference on Climbing and Walking Robots,
pages 87 94, Madrid, Spain, October 2000.
Johan Ingvast, Christian Ridderstrm, Freyr Hardarson, and Jan Wikander. Improving a trotting robots gait by adapting foot reference offsets. In Professional
Engineering Publishing Limited, editor, 4th International Conference on Climbing
and Walking Robots, pages 711718, Karlsruhe, Germany, September 2001.
Christian Ridderstrm and Johan Ingvast. Combining control design tools
from modeling to implementation. In Int. Conf. on Robotics and Automation,
pages 13271333, 2001.
Christian Ridderstrm and Johan Ingvast. Quadruped posture control based on
simple force distribution a notion and a trial. In IEEE/RSJ Int. Conf. on
Intelligent Robots and Systems, pages 23262331, Hawaii, USA, 2001.
Johan Ingvast. Derivation of ground interaction models for plastic soils specially
suited for walking robots. Technical Report TRITA-MMK 2002:14, Dept. of Machine Design, Brinellv. 83, 100 44 Stockholm, Sweden, June 2002.
Johan Ingvast. A two dimensional bidirectional ground interaction model. In
Professional Engineering Publishing Limited, editor, 5th International Conference
on Climbing and Walking Robots, pages 547 554, Paris, France, September 2002.
Johan Ingvast and Jan Wikander. A passive load-sensitive revolute transmission.
In Professional Engineering Publishing Limited, editor, 5th International Conference on Climbing and Walking Robots, pages 603 610, Paris, France, September
2002.
Christian Ridderstrm and Johan Ingvast. Warp1: Towards walking in rough
terrain smooth foot placment. In Professional Engineering Publishing Limited,
editor, 6th International Conference on Climbing and Walking Robots, pages 467
474, Catania, Italy, September 2003.
Johan Ingvast, Christian Ridderstrm, Freyr Hardarson, and Jan Wikander.
Warp1: Towards walking in rough terrain control of walking. In Professional
Engineering Publishing Limited, editor, 6th International Conference on Climbing
and Walking Robots, pages 197 204, Catania, Italy, September 2003.
Johan Ingvast. PVT The passively variable transmission. Technical Report
TRITA-MMK 2005:17, Dept. of Machine Design, Royal Institute of Technology,
S-100 44 Stockholm, Sweden, December 2005.

Contents
Introduction
1
The science of walking
2
The WARP 1 robot . .
3
Methods and tools . .
4
Transmissions . . . . .
5
Summary of papers .
6
Discussion . . . . . . .
Bibliography . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

1
3
20
25
26
38
44
47

A Combining control design tools from modeling


tation
A 1 Introduction . . . . . . . . . . . . . . . . . . . . . .
A 2 Development tools and methods . . . . . . . . . . .
A 3 Models . . . . . . . . . . . . . . . . . . . . . . . . .
A 4 Results . . . . . . . . . . . . . . . . . . . . . . . . .
A 5 Discussion . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

53
54
55
60
62
64
67

B The basic design of the quadruped robot WARP 1


B 1 Introduction . . . . . . . . . . . . . . . . . . . . . .
B 2 Robot hardware . . . . . . . . . . . . . . . . . . . .
B 3 Control design and implementation tools . . . . . .
B 4 Control structure . . . . . . . . . . . . . . . . . .
B 5 Experiments . . . . . . . . . . . . . . . . . . . . . .
B 6 Discussion . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

69
70
70
73
75
76
77
79

walking
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

81
82
84
87
88
90

C Towards walking in rough


C 1 Introduction . . . . . . .
C 2 Control . . . . . . . . .
C 3 Experiments . . . . . . .
C 4 Discussion/Future work
Bibliography . . . . . . . . .

terrain
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
xi

Control of
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .

to implemen-

xii

CONTENTS

D Improving a trotting robots gait by adapting


offsets
D 1 Introduction . . . . . . . . . . . . . . . . . . . . .
D 2 The robot and control structure . . . . . . . . . .
D 3 Controllers . . . . . . . . . . . . . . . . . . . . .
D 4 Experiments . . . . . . . . . . . . . . . . . . . . .
D 5 Discussion . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . .
E The trunk follows the feet
ped robot walk and trot
E 1 Introduction . . . . . . .
E 2 Method . . . . . . . . .
E 3 Experiments . . . . . . .
E 4 Discussion . . . . . . . .
Bibliography . . . . . . . . .

foot trajectory
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

91
. 92
. 92
. 93
. 97
. 99
. 101

an approach for making a quadru.


.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

103
104
104
114
117
119

F The PVT, an elastic conservative transmission


F 1 Introduction . . . . . . . . . . . . . . . . . . . . .
F 2 Theory of conservative elastic transmissions . . .
F 3 The passively variable transmission . . . . . . . .
F 4 The load-sensitive CVT . . . . . . . . . . . . . .
F 5 Torque control . . . . . . . . . . . . . . . . . . .
F 6 Prototype . . . . . . . . . . . . . . . . . . . . . .
F 7 Results . . . . . . . . . . . . . . . . . . . . . . . .
F 8 Discussion . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

121
122
125
130
136
139
140
141
149
150

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

Introduction

Mechanical robots have long been a dream of humans. Around the time of the
birth of Christ2 there lived a scientist named Heron of Alexandria, who wrote a
number of books. In one of them, Automata, he described various ways of creating
wonders for temples, including automatically opening doors and dolls that moved
Sfrag replacements
by themselves. Reproduction of such an automatically moving doll is illustrated in
oFigure
[ ]
1.
iOne would think that these sophisticated ideas of their time would have been
i
used
for production of some kind, but they were instead used for religious purposes.
[
] we would call the inventions of Heron toys. The fact that such advanced
Today
T
o
technologies of the past were used for a non-productive purposes is similar to the
koo
kio2 The exact date is unknown: estimates range from a few hundred years before Christ to a few
hundred
years after.
nv
To
kc
Space for weight
ko
with sand below
Time [s]
Wire connected to wheels
[rad/s]
Weight
i [rad/s]
o [rad/s]
To [Nm]
Wheels
[rad]
[]
Sand runs
Ts [Nm]
through hole
o [ ]
T [Nm]
Sand
P [W]
Figure 1: A reproduction of a so-called self-moving stand, invented by Heron of
Alexandria. It was used in various configurations to move objects and
dolls on automated theatre stages around the time of the birth of Christ.
The pictures come from the website of the Museum of Ancient Inventions
at Smith College, Northampton, Massachusetts.

PSfrag replacements
o [ ]
i
i
[ ]
To
2
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION

Figure 2: AIBO robot dog, a toy from Sony.

situation today. Much high technology can be found in toys and in equipment
for sports. For example, large quantities of carbon fiber are found in golf clubs,
bicycles, and sports cars, and extremely powerful CPUs are advertised for use in
the PS3 game console manufactured by Sony.
Scientists working on walking robots are often asked of what use such walking
machines are. The answer frequently involves the fact that 50% [33]3 of the earths
landmass cannot be traversed by wheeled vehicles, while animals have proved that
legs are a practical way to cross such areas. This alone should thus be a good
reason for developing legged robots. However, after more than 50 years of serious
attempts to build agile walking robots, still no robot has proved to be much better
at traversing terrain than wheeled vehicles are.
But still, walking robots are produced in large quantities, not for any productive
purpose but for entertainment. Sony has sold more than 100 000 AIBO robot
dogs (see Figure 2). Entertainment is likely to be the largest market for walking
robots for some time to come.
The research presented in this thesis does not deal with the problem of what a
robot is to be used for: the main purpose is rather to build our understanding of
how quadruped robots can walk and run. The research stem from the need of the
Centre for Autonomous Systems (CAS) at the KTH where research on autonomous
robots for indoor and outdoor use is done. The contribution to the CAS activities,
from this thesis, regards walking robots for outdoor use, with the aim of blind
locomotion in difficult terrain. Other parts of the center study the methods for
vision, localization, mapping, and planning, such that autonomy can be achieved.
3I

have not been able to find the original report: the reference comes from [36].

1. THE SCIENCE OF WALKING

I believe it is important that the controllers of walking robots should be able to


function without the use of complex, computationally heavy algorithms. Therefore,
and partly because of the crude possibilities of the test platform, only simple models
of the robot were used in deriving the controllers. However, complex models of the
robot platform are used for simulation and analysis.
To build useful agile walking robots, not only control is important, they have
to be light-weight, strong, fast and energy efficient. The motors and transmissions
that actuate the legs affect all these properties, and therefore research in this area
is needed.
The thesis encompasses six papers, some already presented at conferences and
others not yet accepted. A summary of those can be found in section 5. Preceding
the papers is an introduction to the issues addressed. Section 1 presents a general
overview of walking and running, how they can be classified and common control
strategies for balancing. The next section describes the WARP 1 robot which is
the platform used in the included papers dealing with walking. Section 3 briefly
describes the tools and methods used in the research. The last section provides an
introduction to transmissions, particularly different kinds of variable transmissions.

The science of walking

Long before research into walking robots had started, people were already interested
in the motion of animals. At the end of the nineteenth century, there was discussion
among horse specialists as to whether all the feet of trotting horses were in the air
at the same instant. The issue was settled by the photographer E. Muybridge,
who took photographs showing that trotting horses definitely have all their feet
above the ground at a particular instant. His discovery and the reaction to it
must have encouraged him, because in following years he became known for taking
photographs of animals and humans in motion. He used a series of cameras, each
taking a single still photo; when the photos were combined, it was possible to see
how the animals moved, something that had never before been possible. His best
known publication is Animals in motion [15], which includes pictures of all kinds
of animals including man in motion. In fact, before the work of Muybridge,
very few paintings or sculptures showed realistic poses of horses in action [3] 4 .

1.1

Gaits

The sequence of the lifting and the placing of the feet and the relative time between
these movements is here called gait. In classifying gaits, Muybridges material was
the best available for a long time, until Milton Hildebrand in the 1960s used film
4I

have not been able to find the original article: The claim is made in Hildebrand [6].

INTRODUCTION

to capture the motion of animals. In the paper Symmetrical gaits of horses [5], he
classified some of the gaits of running and walking horses. The parameters he used
in the classification will, with minor changes, be used here.
We will assume a gait in which the motion is cyclic and within each cycle each
foot is lifted once. A stride is the interval from lifting a reference foot until lifting
it again. The distance the trunk advances during one stride is called the stride
length (represented by the stride vector r S ), and the distance a foot advances when
lifted is called the step length. If there is no slip of the foot on the ground, the
step length and the stride length are equal. Each instant in a cycle has a particular
phase value, . When lifting the reference foot, the phase is zero and when the
stride is completed it is one. The relative phase of a leg is the phase difference from
that of the reference leg. A foot lifted just after the reference foot is said to have a
positive relative phase. A foot is said to be in support if it is on the ground and in
transfer if it is in the air.
Gaits can be distinguished as either symmetric or asymmetric: in a symmetric
gait the left and right legs differ by half a stride, while in an asymmetric gait this
is not the case. For a thorough analysis of asymmetric gaits see Hildebrand [6].
Which leg to choose as the reference leg is of course arbitrary, and we have
chosen to use the front left leg. Thus, for a symmetric gait, the relative phase of
the front right leg is 0.5. The relative phase of the rear right leg will be denoted
by , resulting in the relative phase 0.5 + for the rear left leg.5,6 Thus, for a
symmetric gait for which is known, the relative phases of all the legs are known.
The next important factor is the duration of the phase during which the legs
are in stance. This is called the duty factor and is denoted by . If all legs are
assumed to have the same duty factor, each gait can be described in terms of
and . Hence, each symmetrical gait corresponds to a point in a diagram such
as Figure 3. The diagonal and vertical lines in the diagram correspond to gaits in
which one foot is lifted at the same instant as another foot is placed. The horizontal
lines correspond to gaits in which two pairs of legs are placed and lifted at the same
instance. Hence, in all gaits located in the same triangle, the feet are placed and
lifted in the same order. Hildebrand refers to gaits with at least two feet in support
at all times as walking gaits and the others as running gaits. The division between
these the two kinds of gaits occurs at a duty factor = 1/2.
Triangles 1 and 3 in Figure 3 are of particular interest for this thesis. Figure 4
depicts the gaits located in triangle 1, which are called crawl gaits.7 These gaits
allow for great stability and are preferred by animals that move very slowly. Crawl
5 Frequently, calculations of phase differences can be larger than one or less than zero. However,
the reader should bear in mind that only the fractions of the phase are of importance. Hence, a
phase difference of 1.3 has the same meaning as one of 0.3 and a phase difference of 0.2 has the
same meaning as one of 0.8.
6 This is nearly the same parametrization as Hildebrand used. He used the rear left leg as the
reference leg and denoted the phase difference of symmetrical gaits by the relative phase of the
front left, leg which is the same as 0.5 .
7 The crawl gait is sometimes also referred to as the walk gait (it would sound strange to say
that a horse crawls).

Phase difference,

}
}

PSfrag replacements
o [ ]
i
i
1. THE SCIENCE
OF WALKING
[ ]
To
Running
Walking
koo
gaits
gaits
kio
0.5
} Pace
nv
2
To
kc
4
1
0.25
ko
Time [s]
[rad/s]
3
i [rad/s]
Trot
0
o [rad/s]
5
To [Nm]
[rad]
6
-0.25
[]
Ts [Nm]
o [ ]
} Pace
-0.5
T [Nm]
0
0.25
0.5
0.75
1
P [W]
Duty factor,

Figure 3: A classification of symmetric gaits.

gaits are also very common among walking robots for the same reason. In the
included Paper C the crawl is used as the basis of analysis.
A trot occurs when the diagonal legs move nearly as a pair, hence the duty factor
lies near the = 0 line (independently of the phase, ). Here, and particularly in
the included Paper E the gaits that fall in triangle 3 of Figure 3 are also regarded
as trot. A trot from triangle 3 is depicted in Figure 5 . In this gait there are two-,
three-, and four-legged stances.
Some other important areas of the gait diagram (Figure 3) are:
The sequence of lifting and placing the feet in triangle 6 is the reverse of that
in triangle 1 and is thus useful for walking backwards.
In triangle 5, the gait is similar to the trot in triangle 3, with the difference
that the rear foot is lifted before front foot. Such a gait can thus be useful
for trotting backwards.
Near = 0.5 and = 0.5 is the pace. In such a gait, the legs on each side
move in unison. This is generally a fast gait and is preferred by long-legged
animals such as camels and giraffes. This gait allows for a large step length
without interference between the legs.
The most common asymmetric gaits are the gallop, pronk, and bound. The pronk
occurs when all feet move together in a sort of hopping. In the bound, the front
legs move as a pair, as do the rear legs. Hence, the movements in a bound are more

1
2

2
1

1 .5

+
0 .5

1 .5

0 .5

INTRODUCTION

g replacements
o [ ]
6
i
i
[ ]
Liftoff
To
Touchdown
koo

kio
nv
To Left front #1
kc
ko Right front #2
Time [s] Left rear #3
[rad/s] Right rear #4
i [rad/s]
o [rad/s]
To [Nm]
Transfer
[rad]
Support
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

2
1

3
1
3

4
3

3
4

Figure 4: Depiction of the crawling gait located in triangle 1 of Figure 3. The top
line is the time line on which each event of lifting or placing a foot is
marked. The thick horizontal lines in the middle denote the periods of
stance. The bottom figures show the positions of the feet. An open circle
denotes a lifted foot and a filled circle a foot placed on the ground. The
gray shaded areas represent the polygon of support. The starting point
of an arrow indicates where the foot is when lifted from ground and its
endpoint where the foot is to land; the arrow length represents the stride
vector.

or less two dimensional. The gallop is similar to the bound, only that there is small
phase difference between the two front feet and a small phase difference between
the two rear feet. If the phase difference between the front left foot (LF) and right
front foot (RF) has the same sign as the phase difference for the rear feet (LR and
RR) the gait is called a transverse gallop. If the sign is different, the gait is a called
rotary gallop. Typically, the sequence of feet hitting the ground is LFRFLRRR
for a transverse gallop and LFRFRRLR for a rotary gallop. Contrary to bound,
the gallop cannot be explained by a two dimensional model.
The duty factor is strongly connected to the speed. One reason for this is that
with respect to the trunk, each foot has to be moved between two extreme points,
one when placing the foot and the other when lifting it. For the support part of the
gait the time available is T , where T is the time needed to complete one stride; for
the transfer part, the time available is (1 ) T . Hence, if the duty factor is large,

2
1

1 .5

1 .5

0 .5

0 .5

+
1

g replacements
o [ ]
1. THE SCIENCE OF WALKING
i
i
[ ]
Liftoff
To
Touchdown
koo

kio
1
1
4
nv
To Left front #1
kc
ko Right front #2
Time [s] Left rear #3
[rad/s] Right rear #4
i [rad/s]
o [rad/s]
1
1
To [Nm]
Transfer
2
2
[rad]
Support
[]
1
3
3
3
Ts [Nm]
o [ ]
4
4
T [Nm]
P [W]

2
1

3
3

4
3

3
4

Figure 5: Depiction of a trot gait from triangle 3 of Figure 3. For explanation for
how to read the figure see the Figure 4

a very short time will be available for transferring the foot compared to the time
that the foot is on ground. Therefore, the transfer of the feet is what usually limits
the speed. When the speed of the foot in transfer limits the speed, a relatively
simple calculation [21] shows that the maximum speed of the trunk relative to the
ground is limited by
1
vlegmax
(1)

where vlegmax is the maximum speed of the foot relative to the trunk in transfer.
A result of this relation is that a quadruped using the crawl gait ( = 3/4) cannot
move faster than one third of the highest speed of the leg.

1.2

Strategies for balancing

Different gaits require different balancing strategies. Some gaits, such as the crawl
gait for quadrupeds, allow for so-called static balance, while other, presumably
faster gaits, require dynamic balance. So far, walking has been discussed without
distinguishing between the walking of animals and of robots. This section will treat
only the walking of robots.
Generally, the contact between the ground and the foot is not a point. However,

INTRODUCTION

most quadrupeds have small feet in comparison to the distances between the feet.
Therefore, we will assume that the contact between the ground and a foot can be
reduced to a point.8 An important concept for walking is the area of support. The
area of support is the convex hull of the feet in contact with the ground. When
the foot contact areas are reduced to points, the edge of the support area outlines
a polygon, so the support area is sometimes referred to as the support polygon.
Sometimes only the vertical projection of the contact points is considered, then
the support area is always well defined. However, if all contact points lie in one
plane the contact area is well defined also for the case when three dimensions are
considered.
There is a point on the ground where all the vertical forces from the ground to
the robot can be replaced by one force. This point is called the center of pressure,
CP and can be calculated as follows:
r CP = P

1
lL

fzPl

rPl fzPl

(2)

lL

where L is the set of all feet/legs on the robot, r Pl is a vector from some origin to
the foot of leg l, and fzPl is the vertical force acting on the robot at foot l, and it is
assumed to always be lifting the robot. The vector r CP points from the origin to
the center of pressure.9 In a vertical projection, the center of pressure does always
belong to the support area.
Note that, normally when defining the center of pressure, only the horizontal
components are considered. There are however reasons (see section 1.4), to bring
the vertical component into the definition.
1.2.1 Static balance
Static balance means that the projection of the robots center of mass on the ground
is always kept within the area of support and that the robot moves sufficiently slow.
Just as a chair with two legs cannot stand by itself, a robot is not statically balanced
if fewer than three legs are on ground at a time. Therefore, in order to move the
feet into new positions, a statically balanced robot needs at least four legs.
When a robot walks sufficiently slow, the inertial forces can be ignored. Hence,
the vertical force the ground exerts on the robot has to be as large as the weight of
the robot. Furthermore, the vertical projection of the center of mass, CM , coincides
with the center of pressure.
McGhee [12] defines the static stability margin as the smallest horizontal distance from the center of mass to the border of the support polygon. If the static
8 This assumption is common for quadrupeds, however, most bipeds have large feet which
normally not can be reduced to point contacts.
9 Vectors between two points will frequently be denoted in a format like that of r AB , which
denotes the vector from point A to point B. When r A is written, the implicit meaning is r OA ,
where O is the origin.

kio
nv
To
kc
ko
1. THE SCIENCE
OF [s]
WALKING
Time
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]Pinpoint foot
P [W]

Robo

CM

F = mg
l
Ground

Figure 6: A model of a robot with two (point) feet on the ground

stability margin is positive, then the robot is statically stable. The margin is a
measure of stability.
There are alternative measures of the stability margin such as the energy stability
margin [13] and a normalization of it, the normalized energy margin. These two
stability margins provide better measures of stability when the the robot is on an
inclined terrain, as they indicate how much energy is required to disturb the robot
such that it falls over an edge of the support area. The static stability margin is
independent of the vertical position of the trunk, whereas the two energy stability
margins are not: for these measures the stability decreases with height, which much
better reflects the intuitive meaning of static stability.
1.2.2 Dynamic balance
Dynamic balance is a very wide characterization. Dynamically balanced robots
range from those that move more or less in a statically balanced state, to those
with aerial phases. In between are most bipeds (two-legged robots), which walk
with at least one foot on the ground at all times (bipeds mostly have large feet
which provide an area of support), and quadruped robots that trot (having only a
line of support).
A special case of the trot treated in the included Paper D, namely, trotting in
place lifting the feet as in trot but with zero stride length, will be analyzed to
gain an understanding of some of the dynamics of trotting. In this gait, the robot
has either four or two legs on the ground. When two legs are lifted, they are lifted
straight up and no special measures are taken to maintain balance, so the robot
will start to fall. Hence, the robot can be modeled as an inverted pendulum, as
depicted in Figure 6.
Assuming a linear system ( small), the equation of motion is

Jrobot + ml2 = mgl

where the meanings of l and are as shown in the figure. The total mass of the
robot is denoted m and the gravity g. The inertia of the trunk, Jrobot , is often

10

INTRODUCTION

much smaller than ml2 and is neglected. The solution to this system is
(t) = 0 et

(3)
p
where = g/l and 0 is angle (t) at the moment the feet leave the ground.
Equation (3) shows that angle increases exponentially with time. Therefore, so
that the robot does not fall too far when the legs are lifted, the initial angle, 0 ,
should be as small as possible. The balance controllers presented in Papers D and
E try to minimize this angle. As previously noted, the energy margin for statically
balanced robots decreases when the vertical position of the center of mass increases.
For dynamically balanced robots, it is tempting to say that the opposite holds, since
the dynamics becomes slower (i.e., is smaller) when the vertical position of the
center of mass increases,10 as equation (3) indicates.
A frequently used concept in dynamic balancing is so-called zero moment point
(ZMP) first introduced by M. Vukobratovic and D. Juridic in 1968 [35].11 There
has been much misunderstanding of the ZMP concept, but I feel that the concept
is sufficiently clarified in Vukobratovic and Borovac [34]. This confusion has arisen
because the ZMP is 1) sometimes used in planning robotic movements and 2) is
closely related to the center of pressure.
According to Vukobratovic [34] for a foot to be stable on the ground, the center
of pressure must be within the support area. If the center of pressure lies within
the support area, it is called the ZMP; when the center of pressure is on the edge
of the support area, no ZMP exists.
Realization of the planned motion of all joints of a robot can be tested by
checking whether the ZMP exists at all times. This can be done by taking all
gravitational and inertial forces into account, and determining at what point on a
plane (the ground) a force must act on the foot so as to accomplish the motion. If
this point lies within the support area, the ZMP exists and the motion is stable.
Vukobratovic refers to this point which may lie outside of the support area
as the fictitious ZMP. The distinction between the ZMP and the fictitious ZMP is
often ignored.
One leg running Most larger mammals have aerial phases when running at
the highest speed, and it is also likely that the fastest gaits of robots will have
aerial phases. A special kind of robot exploit aerial phases to their extreme, such
that it cannot balance when stationary. Raibert started his analysis with a onelegged hopping robot [18] (see Figure 7), and later extended his original idea to
be applicable in two-legged running and in four-legged trotting [19], bounding, and
pacing [18] robots.
Surprisingly, the controller for the one-legged hopping robot is simple. There are
three states: attitude, jumping height, and forward velocity. The last two states are
10 Considering

the stability of dynamic walking is pointless, however in terms of control theory


the eigenvalue moves to the left in a plot of the poles.
11 I have not been able to find the original article; it was cited in [34].

PSfrag replacements
o [ ]
i
i
1. THE SCIENCE OF[ ]WALKING

11

To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]
Figure 7: The one legged hopping robot developed at MIT.

controlled only once in each cycle, and all states are controlled separately. During
flight (the transfer phase) the center of mass follows a ballistic motion. Hence,
none of the states can be controlled. When the foot is on the ground the robot
bounces on a springy leg and a hydraulic actuator additionally pushes to control
the jumping height. The touchdown point of the foot controls the forward velocity.
For each speed there is one touchdown point that maintains the speed. This point
is called the neutral point. At zero speed the neutral point is directly under the
center of mass and at a high speed it is more forward.12 To increase the speed, i.e.
accelerate the robot, the touchdown point is made to the rear of the neutral point,
and to decrease the speed the reverse is the case. The touchdown point is controlled
by positioning the foot (the angle between the leg and trunk) to the appropriate
point before touchdown. During the support phase, the leg angle is controlled such
that the trunk attitude is horizontal.
Gallop/four legged running with air-phases Raiberts ideas have been further developed to improve the one-legged running, for example in Brown and Zeglin
[2] and Zeglin and Brown [38], and others have used the ideas to develop the fourlegged gallop gait.
The highest running speed of Raiberts quadruped was 2.9 m/s. To make robots
really useful, they need to move faster, and the fastest gait is supposed to be the
12 Interestingly, the neutral point is independent on the height of the jump, which means that
the dynamics of the forward speed is independent of the ground elevation, and of the ground
irregularities.

12

[ ]
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION

Figure 8: The pronking, bounding and galloping robot Scout II.

gallop. Though many galloping simulations have been performed, to the best of
my knowledge only one robot has so far succeeded in gallop Scout II developed
at McGill University (see Figure 8). Scout II has proven to be a good bounder,
and has in that gait run as fast as 1.3 m/s. Using the rotary gallop, Smith and
Poulakakis [28] experimentally demonstrated that Scout II can perform a rotary
gallop at 1.4 m/s. This is not such a high running speed, but considering the
simplicity of the robot (only one actuator per leg) it is amazing that it can run at
all.
Poulakakis et al. have pointed out [17, 16] that the natural dynamics of Scout II
might account for why it works well with such a simple controller. They back this
up by simulations of an unactuated ideal Scout that performs stable running, and
as they state [16],
. . . there exists a regime where the model stabilizes itself without the
need of any control action. This might explain why simple controllers, as
reported in [32], are adequate in stabilizing a complex dynamic task like
quadruped running. Self-stabilization can facilitate the design of control
laws for dynamically stable legged locomotion by designing controllers
that expand the domain of attraction of that behaviour.

1.3

Plan of foot landing position

Where to place the feet is not as trivial a matter as might first seem. For a robot
walking at a constant speed and step length, placing the feet a fixed distance in
front of the hip will maintain a specific pattern of the feet on the ground. However,
with the same placement strategy, if the robot keeps its main heading constant
(the stride length constant), but instantly, the trunk moves sideways or changes its

13

1. THE SCIENCE OF WALKING

forward speed, the pattern will be different, and will depend on the instantaneous
trunk movements.
One way of solving this problem is to introduce a virtual vehicle [4] that moves
at a constant speed, and then place the feet with respect to this virtual vehicle.
For this approach to work, the distance between the virtual vehicle and the actual
robot must not be too large, otherwise the robot will try to place its feet outside
its available reach.
The feet can also be placed with respect to the liftoff point. In this way, the
pattern of the feet on the ground is independent of the trunk motion. However,
there are a few problems with this method: 1) If the intended foot position is not
useful and another has to be chosen, then this other position will become the new
reference point for next placement an the pattern will not return to the desired. 2)
If one foot slips when on the ground, the same thing happens, i.e., the reference
point for the next transfer will move. 3) If the commanded step length is different
for each foot for some reason, each foot will move in a direction of its own. This
can easily cause the desired foot landing position to be out of reach for the robot.
The last problem is cited by Yoneda et al. in [37] to justify another approach.
Their robot, TITAN VI, has four legs and in this study the gait examined is a trot,
so two legs have to be placed at a time. They introduced two virtual legs, one
for each diagonal pair of legs. The average points of the pair of actual legs is the
position of the virtual leg. They noted that the distance between the virtual legs
should be the same as half the stride length, so in their algorithm they place the
new virtual leg at a distance equal to half the stride length from the other virtual
legs position. In this way they overcame the above three problems.
In the included Paper C a similar algorithm is proposed. The gait examined in
this case is not a trot, so the algorithm Yoneda used could not be applied, but the
idea of placing the feet with respect to the actual foot positions is used. The method
is applicable in any gait in which at least one foot is on the ground at all times, and
for a trot the result is exactly the same as with the method Yoneda proposed. The
method works as follows: Assume a normal pattern of foot positions. The normal
pattern corresponds to the foothold pattern when the robot walks in place, and
should be aligned in the same direction as the robot is. Denote each foot position
in this pattern by Pl ; let L denote the set of all legs (or feet) and LC the set of all
legs in support. The average point of the feet on the ground is called PC and the
corresponding point for the feet in the normal pattern is PC , hence
1 X Pj
1 X Pj
r and r PC =
r
r PC =
(4)
|LC |
|LC |
jLC

jLC

where |LC | is the size of this set, i.e. the number of feet in support. The method
builds on the proposition that for walking in place (r S = 0), the vector between the
,
average foot on the ground, PC , and the planned landing position of foot l, Pplan
l
is equal to the difference between PC and Pl , i.e.,
plan

r PC Pl

= r PC Pl

(5)

14

i [rad/s]
To
o [rad/s]
kc
To [Nm]
ko
[rad]
Time
[s]
[]
[rad/s]
[Nm]
iTs[rad/s]
o [ ]
o [rad/s]
TTo [Nm]
[W]
P [rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION

(a) Walking in place

(b) Walking with stride

Figure 9: Examples of the evolution of support areas for the crawling gait. Each
frame shows the support patterns after either lifting or placing a foot.

To gain forward velocity, half the stride vector is added to the found position, which
results in the preferred foot placement point. Rearranging the previous expression
and adding the stride vector results in
plan

r Pl

1
= r PC PC + r Pl + r S
2

(6)

The normal pattern is preferably fixed in the case of the trunk, and a convenient
pattern being the same as that of the hips. Then the planned point of landing (6)
can be rewritten as
plan
1
1 X Hj Pj 1 S
r Hl Pl = r HC PC + r S =
+ r
r
(7)
2
|LC |
2
jLC

That is, the foot landing point with reference to the hip should be half the stride
vector plus the average of the position of the feet in contact with the ground relative to their respective hips. The algorithm does have some nice features such as
independence of foot lifting order and good convergence. Figure 9(a) shows how
the support polygon converges to the normal pattern when the initial foot positions are in a single point and the desired stride length is zero. Figure 9(b) shows
the same but with a stride vector added; both (a) and (b) depict two full strides.
The use of the name stride vector is in this section misleading because it does not
explictly control the trunk motion, however in steady state the real stride vector
will be equal to r S which therefore should be seen as the desired stride vector.

1.4

Trunk control

We have so far discussed when and where the feet should be placed, as well as,
different balancing strategies, but have not actually touched on how exactly the
robot should move to accomplish this.
In the case of a static walk, at each point in time there is an area within which
the robot must remain in order to stay statically stable. However, every time a foot

To
kc
ko
Time [s]

[rad/s]
1. THE SCIENCE
OF WALKING
i [rad/s]
o [rad/s]
To [Nm]
[rad]
S m = hi i
[]
Ts [Nm]
o [ ]
Cdes
M
T [Nm]
P [W] Area of support

15

hi

Foot i

Figure 10: Depiction of the relationship between the weight ratio, , and the static
stability margin, Sm .

is lifted, this area is reduced, hence the controller has to guarantee that the robots
center of mass is above the area remaining.
A crude method to solve this problem is to plan all joint motions beforehand.
This implicitly plans the trunk motion, so if the plan takes the center of mass
position into account, the robot can be statically stable. A problem with this
approach is that there has to be one plan for each stride length and every radius of
turning. Therefore, more dynamic planning is preferred.
Hardarson [4] proposed in his thesis that the trunk motion should be planned
by interpolation between the foot positions. For each foot, a cyclic function called
the weight-ratio, , is planned in advance. It must be continuous and lie in the
range [0, 1]. During transfer, the weight-ratio has to be zero and the sum of the
weight-ratios for all feet must be one. The weighted average of the actual foot
positions is calculated as follows:
X
des
r CP =
l r Pl
(8)
lL

where l is the weight-ratio for leg l. The point calculated is used as the horizontal
reference position of the trunk. For a slowly moving robot, the vertical projection
des
of the center of mass is the center of pressure. Therefore, r CP is called the desired
center of pressure. Since the weight ratios are continuous in time, the resulting path
of the desired trunk position is also continuous. Furthermore, since the weight-ratio
is constrained to the range [0, 1], the reference position will always be on the edge
of or within the support polygon: and when at least three of the weight-ratios are
positive, the reference point will be within the polygon. Note that when a weightratio is zero the desired center of pressure is independent of the corresponding foot
position.
When three feet are on the ground, there is a close relationship between the
static stability margin and the weight-ratio. Calculation reveals that the weightratio of a foot can be found as the relative distance from the opposite line of the
support polygon to the projection of the desired center of pressure onto the plane,
as Figure 10 depicts. Hence, if the support polygon is an equilateral triangle13 the

[ ]
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

16

INTRODUCTION

Figure 11: The adaptive suspension vehicle developed at the Ohio State University.

minimum weight-ratio value is proportional to the static stability margin. Using


the minimum weight-ratio is therefore an alternative to using the static stability
margin. With that in mind, planning the weight-ratios with a sufficient stability
margin is simplified.
In the included Paper C the reference for the trunk position is calculated using
the weight-ratios for a statically stable crawl gait. Further, in the included Paper E
it is also used for dynamic walking.
At a lower level, to control the trunk position and attitude, the feet that are in
stance must be controlled with respect to the trunk. Assuming that the feet on the
ground do not slip, there is a one-to-one relationship between the velocities of the
trunk and the velocities of the feet with respect to the trunk. This relationship is
used in the included Papers C and E, such that the trunk errors result in desired
velocities of the trunk that reduce the errors. These desired velocities are then
transformed via a rigid body transformation to corresponding leg velocities, which
are tracked by the leg controllers.
Force distribution Another approach is to let the trunk attitude and position
errors produce a desired torque and force on the trunk, which in turn decrease the
errors. The desired torque and force then has to be distributed among the feet in
some way, a task generally referred to as the force distribution problem. To the
force distribution problem there is usually an infinite number of solutions, often no
solution at all, and only occasionally one distinct solution.14
One method for solving the problem of force distribution is to use some sort of
optimization. The adaptive suspension vehicle (ASV) seen in Figure 11 presented
by Song and Waldron [29], is a large six-legged walking robot. The ASV uses
13 All

sides are of equal length.


is exactly one solution only on the rare occasion when one foot is on the ground and
the desired force happens to generate the desired torque.
14 There

17

1. THE SCIENCE OF WALKING

force as the control signal for attitude and position. To solve the force distribution
problem an objective function, the square of the forces on the ground, is minimized
using the pseudo inverse. Naturally, there are other objective functions 15 that
can also be useful. Jiang and Howard, [11], made a comparative study of various
objective functions that all can all be solved using the pseudo inverse. The functions
minimized were: the square of the forces on the ground, square of torque of the
joints, the square of the power of the motors at the joints, and the square of the
quotient of used to available friction. The results from their example indicated that
the differences between the methods were small in terms of the amount of power
or joint torque used. However, in terms of the fraction of used friction, minimizing
the square of torque performed the worst, producing values approximately five
times as large as those obtained by minimizing of the friction which naturally
performed the best. Minimization problems tend to be computationally heavy, so
other distribution methods are preferable.
The weight-ratios introduced for planning the trunk position can also help
with force distribution, something not suggested in other publications. The torque
around the desired center of pressure that the ground contact exerts on the robot
can be calculated as follows:

X
X des
des
r Pl r CP f Pl =
r CP Pl f Pl =
lL

lL

Pl

Pl

lL

X
lL

Pl

Pl

des

CP
r|{z}

lL

r l f s =

f Pl =

lL

l r Pl

Pl

| {z }
=fs

Pl

f Pl f s l

lL

lL

(9)

where f Pl is the force acting on the robot on foot l and fs is the sum of all the forces.
If fs is taken as the reference force, fdes , for the algorithm governing distribution
among the feet, we see that choosing
f Pl = fdes l

(10)

results in zero torque around the desired center of pressure.


Using this result, we can formulate an algorithm that distributes the desired
CM
, on the trunk to desired forces on the feet,
force, fdes , and desired torque, Tdes
Pl
fdes , as follows:
1. Move the acting point of the force fdes from the center of mass to the desired
center of pressure and then find the torque that results in the same torque at
CM
, i.e.,
the center of mass as Tdes
des

CM
CP
Tdes
+ rCM CP fdes
= Tdes
15 The

objective function is the function to be minimized.

(11)

18

INTRODUCTION
Pl

2. Find some forces, f , on the feet that give the correct torque around the
desired center of pressure, i.e. the forces fulfill
X des
Pl
CP
(12)
rCP Pl f = Tdes
lL

These forces do not have to fulfil any constraint on the total force.
Pl

3. Sum up the forces f from point 2 (12) and subtract those from the reference
force on the trunk, i.e.,
X Pl
fdes
f
(13)
lL

4. The resulting force from point 3 is applied on equation (10) and the chosen
forces from point 2 added, which results in the forces to apply on the feet,
i.e.,
!
X
P
P
i
l
Pl
fdes
= f + fdes
f
l
(14)
iL

To see that the proposed method actually applies the desired force, we make the
summation
!
X P
X Pl X
X Pi
l
l =
f +
f
fdes
=
fdes
lL

iL

lL

lL

Pl

fdes

Pi

iL

lL

l = fdes

(15)

lL

| {z }
=1

and similarly for the torque


X

CM Pl

Pl
fdes

X

CM Pl

Pl

lL

lL

X 

des

l r

CM Pl

lL

{z

des

=r CM CP

!
}

CM Pl

fdes

des

Pl

X
iL

lL

r CM CP + r CP

lL

fdes

Pl

X
iL

Pi

Pi

l =

19

1. THE SCIENCE OF WALKING

des

r CM CP

lL

+r

CM Cdes
P

Pl

X

Pl

lL

fdes

X
iL

des

r CP

des

{z

P
=Tdes

Pi

CM
CP
+ rCM CP fdes = Tdes
Tdes

Pl

=
(16)

The algorithm for finding the forces in point 2 is much simplified compared to
the full system, but no solution for it will be proposed here. One problem with the
algorithm is that it does not ensure that the vertical forces exerted on each foot will
point up, therefore the algorithm for solving point 2 should take the weight-ratios
into account and prefer forces on feet for which the weight-ratios are large.
Pl
If the foot forces resulting from the torque, f , are small relative to l fdes , two
things are noted:
1. The quotient between the horizontal and the vertical forces will be equal for
each foot, meaning that the available friction is used equally on all feet. This
resembles the minimization of friction utilization, as Jiang and Howard [11]
did using the pseudo inverse.
2. As the weight-ratios are continuous, the vertical force for a foot about to be
lifted will approach zero, arriving at zero at the moment when it is supposed
to be lifted. Hence, a foot that is to be lifted does not first have to release
the vertical force before it leaves the ground.
The algorithm is not restricted to use in robots that are statically stable, or even
in robots that have their center of mass vertically above the center of pressure.
However, if the robots center of mass is not located vertically above the center
Pl
of pressure or if the robot is moving fast, the f forces are unlikely to be small
relative to the l fdes forces.

PSfrag replacements
o [ ]
i
i
[ ]
20
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION

Figure 12: The walking robot, WARP 1.

The WARP 1 robot

WARP 1, seen in Figure 12, is a four-legged walking robot developed at the department of Machine Design, KTH. It is the platform used to validate the hypothesis
regarding walking and comprises the basis of the papers included in this thesis. A
detailed description of it is given in the included Paper B: the rest of this section
recapitulates and emphasizes some of the content of this earlier article and discusses
features ignored there.
The WARP 1 robot is a complex system, its four legs each being equipped with
three joints, motors, sensors, drivers, electronics, computers and communication
buses. To be able to operate the robot, it has a user interface and tools for developing controllers. Systems that combine mechanics with digitized controllers are
called mechatronic systems, and WARP 1 is a good example of such a system.
Since WARP 1 was built, it has been the basis for testing various design ideas,
beginning with the trot gait (Paper D) and then later the crawl gait (Paper C)
and combinations of the crawl and trot gaits (Paper E). We have also developed
controllers for more specific hypotheses such as how to control the posture using a
simple heuristic force distribution [23].
Hardarson [4] developed and implemented a walking controller using weightratios in planning the gait. Contrary to what was done in the included Papers C
and E, the weight-ratios were not planned beforehand, but parameterized and the
parameters calculated on-line by minimizing a quadratic goal function with additional constraints. The controller was very successful, but computationally heavy.
The WARP 1 robot contains an attitude sensor module. It is equipped with an
inclinometer, three rate gyros, and a three-axis accelerometer. The sensor values

PSfrag replacements
o [ ]
i
i
[ ]
2. THE WARP1 ROBOT
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

21
Motor

Harmonic drive
Steel wire

Thigh fixed pulley


Thigh

Figure 13: The hip flexion extension joint of WARP 1. The figure depicts the two
stage gearing.

are fused in an algorithm developed by Rehbinder [20]. The attitude values are
used in the walking controller devised by Hardarson [4] and also in the controllers
presented in the included Papers E and C.

2.1

The legs and joints

Each leg has three joints, each controlled by a motor. The first joint, starting from
the robot trunk, moves the legs sideways and is called the abduction/adduction joint
(a/a, a term borrowed from physiology, meaning away from/toward the center of
the body). The next joint, which axis crosses that of the first at 90, is the hip
flexion/extension (f/e) joint. The f/e joint drives the 290 mm long thigh forward
and backward. At the end of the thigh is the knee which, contains a f/e joint.
This joint drives the shank (300 mm), at the which end there is a small spherically
shaped rubber foot.
The motors angular velocity is reduced in two stages before it is applied to the
joint. The first stage uses Harmonic Drive gearing with a 100to1 reduction,
and the next stage is a wire reduction, as shown in Figure 13. All drive lines from
motor to the leg sections are similar, differing only in terms of the reduction ratio
of the wire reduction, which for the hip a/a joint is 2.50:1 and for the two f/e joints
is 2.85:1.
The use of two-stage transmission makes the robot robust. The Harmonic Drive
is the main torque amplifier, which is protected from shocks by the wire transmis-

22

INTRODUCTION

sion, since the wire is somewhat compliant. The wire also protects the Harmonic
Drive from overload, since at excessive torques the wire breaks.

2.2

Control

The control of the robot is a cascaded controller of several layers. In the outer loop
is the trunk controller, which establishes the positions/velocities/forces for the feet
to follow. These references are treated by the leg controller, which in turn gives
torque references to the joint controllers.
Most developed walking controllers of WARP 1 use, at some level, force control of
the feet. However, the force control is not very accurate, and the force can only be
approximately determined. This inaccuracy stems from the use of non-ideal drive
lines and the fact that the controller does not take gravity or inertia into account.
The following discussion shows how the force control works.
From force to torque The controller is of a feed-forward type. The reference
force is mapped to torques at the joints by multiplication with a matrix, called
a Jacobian, J . The foot position with respect to the hip is denoted r HPl (q) =
T
T
[rx , ry , rz ] , which is a function of the joint angles q = [q1 , q2 , q3 ] . The velocity of
the foot can be expressed in terms of the angular velocities, q,
as follows

rx
rx
rx
q

q1
q2
q3 1
r HPl dq
dr HPl

ry ry ry
HPl
=
= q
(17)
v
=
q = J q
dt
q dt
1 q2 q3 2
rz
rz
rz
q3
q1
q2
q3
Assume mass-less legs and no power losses, in which case the power put into the leg
through the actuators has to come out at the foot (since this is the only interaction
point with the environment). Hence, we can write
Pout = Pin f v HPl = T q
where f = [fx , fy , fz ] is the force acting on the foot and T = [T1 , T2 , T3 ] is the
torque acting on the joints. Now, the velocity at the l.h.s. can be replaced by the
expression of velocity from (17), as follows:
f J q = T q (f J T ) q = 0
This statement has to be valid for any combination of angular velocities, hence 16
T =f J

(18)

This is the map from the reference force, f des , to reference torque of the joints
T des . The inertia of the links is not accounted for, nor are the effects of gravity.
16 The torque and forces are often written as column matrices, in which case the relationship
then takes the form T = J T f .

23

2. THE WARP1 ROBOT

Torque control The actuators used are DC motors with permanent magnets. It
is assumed that the torque is linearly dependent on the coil current, as follows 17
Tm

Ikemf

(19)

where Tm is the motor torque, kemf the motor constant and I the electrical current
of the coil. The voltage connected to the motor is balanced by the resistive voltage
and the back electromotive force (back EMF) as described by the following equation
Um

joint kemf + RI

(20)

where Um is the voltage over the motor connections and R the total electrical
resistance. Hence, inductance, nonlinearities, and any other possible states are
ignored.18 To control the torque, a corresponding reference current is calculated
from the reference torque using equation (19). The current is then tracked using
a feed-forward part based on equation (20), and a PI control as follows (with s as
the Laplace operator):


KI
Uc = kemf joint + RIref + (Iref Isensed ) KP +
(21)
s
where KP and KI are proportional and integral gains of the controller, and Uc is the
voltage command sent to the motor driver. The reference current, Iref = Tdes /kemf ,
is limited to 5 A, which corresponds to Tdes = 90 Nm at an f/e joint.

2.3

A note on force and leg control

The control of the force exerted on the feet depends on a number of parameters
that are more or less accurately known; this control does not, however, compensate
for inertia, gravity, or friction. Yet, it is still possible to use it for control of the
trunk, using position (Paper D), velocity (Papers C) and force [23] as reference
for the leg controllers. This works because of robust design at a higher level, and
possibly because the tasks studied do not require high precision.
As the included Paper B points out, because the software limits the electrical
currents, it takes approximately 90 ms for the joints to reach maximum speed.
This is a major limiting factor for both the walking speed and cycle time. If it is
assumed that the hip f/e is the limiting joint and the hip f/e changes 1 rad during
support phase and the same angle during transfer phase, the shortest time in the
air can be calculated. The acceleration of the foot to its highest speed takes 90 ms
and uses 0.1 rad (the same as is needed to decelerate it). Transferring the last
0.8 rad takes approximately 300 ms. Altogether the whole process takes 480 ms,
17 It is assumed that the transmission is ideal, such that the gear ratio can be included in the
motor constant, which is scaled with the transmission ratio.
18 Such as the resistance dependence of temperature and commutation effects.

24

INTRODUCTION

which underestimates the time needed to transfer the foot, as before the foot can
be transferred it first has to be lifted. The lowest duty factor that can be used in
the crawling gait is 0.75, which means the foot is in the air three times as long as
it is on the ground. Hence, the time on the ground is approximately 1.5 s, and
assuming that the distance from hip to foot is approximately 0.5 m, the speed of
the robot is 0.3 m/s an upper estimate of the speed of WARP 1 in the crawling
gait.
At certain instances, the current limit is likely to be overprotective (the limit
was introduced to protect the mechanics, a task it has done well). However, when
accelerating the joint, most of the torque of the motor goes to accelerating the
rotor, not to the links of the legs. Hence, much of the torque generated does not
affect the mechanics. It should thus be possible to introduce a limit that instead
depends on the states and torque commands in a way, that increases the acceleration
possibilities and still protects the mechanics.

25

3. METHODS AND TOOLS

Walk using support ratios


$Revision: 1.2 $ $Date: 2005/11/28 15:38:07 $
Prerequisete: >>setpath; initpar;
Load & Run

User
Interface

Settings

Scopes

Topview
animation

Trunk
Observer

Trunk
Control

LCN:s

Warp1
(simulated)

See Doc. & help


for documentation.
Doc. &
help
Output
& busses

1
log

PSfrag replacements
o [ ]
i
i
[ ]
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

Init&update

Figure 14: The top level of the controller built in Simulink. The Warp1 subsystem
holds the interface with the robot and is the only subsystem that has
to be changed (a process that is automated) in order to switch between
simulation and experiments.

Methods and tools

The included Paper A describes the process used in developing and implementing
design ideas for the robot. The process involves several steps: creation of symbolic expressions of the robots properties19 exporting the expressions into Simulink
S-functions, and finally the simulation, visualization, and analysis of the results.
This process is of course iterative and is followed by building and uploading the
controllers to the robot computers, performing the experiments, downloading the
experimental data, and analyzing the results. The main strengths of the process
are the possibility of using symbolic manipulation to develop expressions and controllers, and the possibility of using exactly the same controller files for both experiments and simulations. Figure 14 shows a top view of the simulink model in
the simulation environment.
These and similar tools and methods developed for WARP 1 were also used when
working with the transmissions described in the next section.
19 The expressions include the position of a single point of the robot, the center of mass of the
trunk or the whole robot, a Jacobian, or simply the position of the foot with respect to the hip.

26

kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION

Figure 15: A clutch transmission developed at Tokai University

Transmissions

Actuators comprise a motor and a transmission, and designing them for robots is
a delicate matter. The results should ideally be both strong and fast, but since
speed and strength are in opposition to each other one characteristic has to be
given priority. Choosing a transmission with a low transmission ratio gives an
actuator that is fast but weak. On the other hand, a high transmission ratio gives
an actuator that is strong but slow. Making the trade-off between these extremes
is usually not easy, since robots in general, and walking robots in particular, often
operate in two different operation modes: fast under low loads or slow under high
loads.
Designers have tried to get around this problem by using more complex designs.
For example the advanced suspension vehicle (ASV) [36] (Figure 11 fig:kappa-ASV)
from Ohio State University successfully used hydrostatic transmissions that work
as continuously variable transmissions (CVT). At Tokai University in Japan, Taku
Takahama and Katsuhiko Inagaki proposed a robot with one engine driving all legs
[30]. The individual joints would be controlled by magnetic clutches, illustrated
in Figure 15, and by activating different clutches the direction of motion can be
reversed. The same idea could easily be used to switch between transmissions using
different ratios.
To get an idea of the importance of using the right transmission, three kinds of
transmissions will be compared: I) a transmission that can change the transmission
ratio infinitely (CVT), II) a transmission with a fixed transmission ratio, and III)
the passively variable transmission, (PVT) described in the next section. The
transmissions are connected to a motor with torque and velocity characteristics
restricted as illustrated in Figure 16(b). The motor and transmission drive an
inertial load by maximizing the output of the motor. The time it takes to transfer
the load a total of 45 is calculated for the different inertial loads and the results
are shown in Figure 16(a). Quite independently of the load, the CVT produces the
best results as it can always output the full power (0.5 W). Case II, the fixed-ratio
transmission is optimized for a load of approximately 20 kgm2 .

Time [s]

g replacements
o [ ]
i
i
[ ]
To 4. TRANSMISSIONS
koo
kio
nv
To
kc
10
ko
Fixed
PVT
transmission
Time [s]
[rad/s] 5
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[] 1
Ts [Nm]
o [ ]
T [Nm] 101
1
10
P [W]
(a)

PSfrag replacements
o [ ]
i
i
[ ]
To
koo
kio
nv
To
kc
ko
Tmax
Time [s]
[rad/s]

i [rad/s]
CVT
o [rad/s] max
To [Nm]
Pdesign
[rad]
[]
Ts [Nm]
o [ ]
2
T [Nm]
10
103
2
Inertial load [kgm
]
P [W]

27

Torque
Pdesign
max

Angular
vel.
Tmax

(b)

Figure 16: A comparison between different kinds of transmissions. I CVT, II


Fixed transmission and III - the passively variable transmission (PVT).

The PVT used (transmission type III) is tuned to have roughly the same transfer
time at a large inertia as the fixed-ratio transmission does. Around the inertia for
which the fixed transmission is optimized the PVT is slightly slower than the fixed
transmission is; however, at a lower inertia the speed will be higher. It seem as
if the PVT could be a useful replacement for fixed transmissions in situations in
which the load varies.
Normally when designing transmissions it is desirable to keep the stiffness high.
The reason for this is that it simplifies
position control. As a rule of thumb the first
p
eigenfrequency of the system k/m should be of an order higher than the highest
frequency of the expected resulting motion. However, a torque controller can be
simplified if the transmission is not so stiff (elastic). In particular the impedance of
the controller will be equal to the transmission stiffness for high frequencies, hence
low stiffness can be preferred. Furthermore, low stiffness decouples the inertia of
the motor from that of the actuated shaft. This results in that a robot equipped
with a transmission with low stiffness will appear less heavy to the surrounding and
can thus be more human friendly.
The included Paper F deals with nonlinear elastic transmissions in general and
the PVT in particular. A short description of the PVT follows in the next section.

rag replacements
o 28
[ ]
i
i
[ ]
To Lever
koo
kio
nv
To
kc o
To
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[] d
Ts [Nm]
o [ ] o
To
T [Nm]
P [W]

INTRODUCTION

Wire
d
Ti
i

Pulley

Ri
(a)

(b)

Ti
i
Ri
d
(c)

(d)

Figure 17: Illustration of the principle of the PVT.

4.1

The passively variable transmission

The passively variable transmission (PVT) is elastic and nonlinear. It is elastic in


the sense that the input and output shafts are not rigidly connected to each other.
It is nonlinear in many senses: for example the elasticity is not linear, and a fixed
velocity on the input shaft does not result in a fixed velocity on the output shaft.
The PVT has properties such as a high amplification of torque when the load
is high and that it reduces velocity little when the load is light. The main principle
is that depicted in Figure 17. A pulley of radius Ri acts as an input shaft with the
torque Ti and the angular velocity i . Similarly, a lever acts as the output shaft
with the output variables To and o . The pulley and the lever are connected with a
wire, such that when the pulley turns, the wire pulls on the lever. The orthogonal
distance between the wire and the lever pivot point is d. The relationship between
the output and input torque is equal to the relationship between d and Ri , as
follows:
To
d
=
(22)
Ti
Ri
Hence, if the wire is connected to the lever at a point far from the lever pivot
point, the torque magnification is large as depicted in Figure 17(a) and if it is
connected to the lever at a point near the pivot point, as in Figure 17(c) the torque
magnification is small. For the kind of mechanism depicted in Figure 17(a) and (c)

4. TRANSMISSIONS
g replacements
PSfrag replacements
o [ ]
o [ ]
i
i
i
i
i
Ti
[ ]
[ ]
To
To
koo
koo
Input/Motor pulley
Wire
kio
kio
nv
nv
To
To
kc
kc
ko
ko
Time [s]
Time [s]
Spring pulley
[rad/s]

[rad/s]
Spring
i [rad/s]
i [rad/s]
Rollero [rad/s]
o [rad/s]
To [Nm]
To [Nm]
[rad]
[rad]
[]
[]
Lever pivot point
Ts [Nm]
Ts [Nm]
To
o [ ]
o [ ]
Lever
T [Nm]
T [Nm]
o
P [W]
P [W]

29

Figure 18: The PVT.

the relationship between the input and output angular velocities is the inverse of
the relationship for the torque, hence we can write
d
i
=
o
Ri

(23)

As pointed out before, for a motor with limited angular velocity and torque,
the transmission ratio should be large when the load is high and small when the
load is low. If the orthogonal distance, d, by some means could become large when
the load is high and small when the load is low, then this would be accomplished.
One way to accomplish this would be to connect a spring between the lever pivot
point and the wire, as depicted in Figure 17(b) and (d). At high loads the tension
in the wire is large and the spring is displaced such that the distance between the
wire and the pivot point is large. On the other hand, when the load is small the
wire tension is also small, so its ability to displace the spring is small. Therefore,
the distance between the wire and the pivot point is short.
The PVT is built on this principle but uses two levers so as to be able to transmit
torque in both directions, as depicted in Figure 18.

kc
ko
Time [s]
[rad/s]
i [rad/s]
30
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION
Pi

Po
To

Ti
Motor

Load

Transmission
To

Ti

Figure 19: A transmission consisting of input and output shafts together with a
motor and a load. The diagram explains the sign conventions used for
the power, torque, and angular velocities. The important property is the
direction of the power, which is defined as going into the transmission,
and that the power is defined as Pi = Ti i and Po = To o .

4.2

Transmission basics

Transmission refers to the transfer of mechanical power from one point to another.
Sometimes the meaning emphasizes the spatial distance as in the drive-shaft of a
car. Here the emphasis is on the quality of the power transferred, not the distance.
The angular velocity (occasionally also the translational velocity) and the torque
(force) are the qualities of interest.
The transmission transfers power from one shaft to another as depicted in Figure 19. One shaft is called the input shaft and the other the output shaft. The
sign convention of the power used here is that positive power indicates flow into
the transmission.
Definition 1 A transmission for which the sum of the power flowing into it is zero
is called an ideal transmission.
The requirement can also be written as Pi = Po . The mechanical power flowing
into the transmission through one of the shafts is the product of two quantities:
angular velocity and torque, as follows:
Pi = Ti i and Po = To o

(24)

For some types of transmissions the velocity ratio, i.e., the ratio between the input
angular velocity and the output angular velocity, is well defined,20 as follows:
nv =

i
o

(25)

Transmissions consisting of gears (with teeth) usually have this property, and furthermore, the ratio for most such transmissions is fixed.
Definition 2 A transmission with a constant velocity ratio is called a fixed-ratio
transmission.
20 Some

authors choose to define the ratio inversely.

31

4. TRANSMISSIONS

One can also think of transmissions that similarly have the torque ratio well
defined, as follows:
nt =

To
Ti

(26)

The sum of the input and output power for an ideal transmission, Pi + Po = 0,
can also be written as i Ti = o To . This can be rearranged to give nv = nt . Thus
the velocity and torque ratios are the same in all ideal transmissions; when this is
the case, the ratio is simply called the transmission ratio.
In most cases, the torque coming directly from a motor is insufficient for the
application, so, the transmission ratio is commonly larger than one, meaning that
the input shaft moves faster than the output shaft and that the output torque is
larger than the input torque. For robotics, transmission ratios of a magnitude of 100
or more are common. The transmission ratios in cars are much lower: for example,
the manual stick-shift transmission of a BMW 320i has a ratio of nv = 4.32 in its
first (lowest) gear and nv = 0.85 in its sixth (highest) gear. Note that for cars, a
low gear number means a high transmission ratio, and vice versa.
4.2.1 The CVT
Most transmissions comprise a train of toothed gears. Each single connection between two gears has a fixed transmission ratio, so the gear ratio can really not
change for a given set of gears in action. For cars with stick-shift transmissions,
the set of gears active in the transmission is changed in order to change gear ratio.
Hence, such a transmission does not have a continuum of gear ratios; there are,
however, other ways of building transmissions that do.
The continuously variable transmission (CVT) has the ability to change its
transmission ratio continuously. The design of CVTs can thus not solely depend
on gears, and some other way of transferring the power has to be used. Two of the
most common methods are friction and fluids (preferably hydraulic oil). Examples
of the former are belt transmissions and the latter are hydrostatic transmissions.
The following sections describe both an old a newer type of friction CVT.
Belt CVT Since the 1950s some cars have been equipped with CVTs. The first
car manufacturer to use the CVT was the Dutch company DAF, which used a
belt design, called variomatic. Figure 20 shows the principal components of this
transmission: the input and output shafts have pulleys that are connected with
a belt, each pulley comprises two conical halves whose axial position on the shaft
can vary. By increasing the distances between the halves of a pulley, the effective
radius of the pulley decreases; the reverse happens when the halves are brought
together. The distance between the pulley halves on the input and output shafts
are changed such that the belt tension is always maintained. Since ideally the
velocity and torque ratio is the ratio of the radii of the pulleys, they are equal and
can be changed. However, they are equal only if there is no energy losses. In reality,

PSfrag replacements
o [ ]
i
i
32
[ ]
To
koo
kio
nv
To
kc
ko
Time [s]
[rad/s] Belt
i [rad/s]
o [rad/s]
Input shaft
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
T [Nm]
P [W]

INTRODUCTION
Output shaft

Cones

High ratio

Low ratio

Figure 20: Schematic of a belt CVT.

the belt slips such that the velocity ratio is not well defined (for a fixed distance
between the conical pulley halves). Torque is also lost in the interaction between
the belt and the cones, therefore the torque ratio is not well defined either.
Though car manufacturer DAF does not exist any more, other manufacturers
have picked up the idea of using belt CVTs. Audi has a transmission called the
multitronic transmission that builds on the same principle as the DAF transmission. However, Audi has managed to use a steel chain instead of a belt, which makes
the transmission much more durable than any of the DAF variomatic designs.
Sphere CVT In the so called sphere CVT, designed and built by Moore [14], a
sphere is constrained in a fixed position by four rollers, as shown in Figure 21. The
contact between the sphere and a roller can be regarded as like a bicycle wheel, in
that it allows forward but not lateral motion. Like a bicycle wheel, the roller can
turn on a single spot.
Two of the rollers have fixed orientation and they act as the input, i , and
output, o , shafts. The other two rollers which can change orientation, , determine
the direction in which the sphere can rotate.
To understand the principle, a similar transmission is depicted in Figure 22,
in which six rollers are used to hold the sphere in place. First consider the case
in which the angle of the steering roller, , is zero. Then the input shaft can run
freely together with the sphere, and the output shaft is locked from rotating. In
a sense this can be regarded as an infinite transmission ratio. The other extreme
is when the steering roller angle is 90. Then the opposite holds: the output shaft

To
koo
kio
nv
To 4. TRANSMISSIONS
kc
ko
Time [s]
PSfrag replacements
[rad/s]
o [1 ]
Steering
i [rad/s]
roller

i
o [rad/s]
i
2
Drive
To [Nm] 2
1
[ ]
roller
[rad]

To
[]
koo
Sphere
Ts [Nm]
kio

o [ ]
nv
T [Nm]
To
Top view
P [W]
Side view
k

33

o
Figure 21: The kdesign
of a sphere CVT. The illustrations are taken from Moore
Time
[s]
[14].
[rad/s]
i [rad/s]
i
o [rad/s]
Input shaft
To [Nm]
Steering roller
[rad]

[]
Ts [Nm] Output shaft
o [ ]
o
T [Nm]
Sphere
P [W]

Figure 22: The principle of the sphere CVT. There is another, hidden steering roller
on the back side of the sphere, the central axis of which has the same
direction as the one seen in the middle of the figure.

will be free to rotate while the input shaft is fixed, in a sense resulting in a zero
transmission ratio. For the 45 case, the angular velocities of the input and output
shafts will be the same. A kinematic study reveals that the transmission ratio can
be calculated as
i
= 1/ tan
(27)
n=
o
Moores experiments indicated that the practical lower limit of the transmission
ratio for such a transmission is approximately n = 0.35 while the upper limit is
n = 2.8.
The simplicity of the principle at first glance make the sphere CVT appear
attractive. However, there are a number of problems with the design, energy losses
being one and wear of the rollers and sphere likely being another.

Pivot points

i
i
[ ]
To
koo
34
kio
nv
To
kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm]
o [ ]
Motor
T [Nm]
P [W]

INTRODUCTION
Extreme position
Slow

Fast

Very fast

Output link
o

Slider
i
i
Crank
(Input link)

Figure 23: Schematic of the dual SMART drive.

4.3

Nonlinear transmissions

Apart from a CVT, in which it is possible to choose any transmission ratio, other
transmissions have transmission ratios that are not directly controllable, but rather
change as a consequence of the state. In some situations such transmissions can
have the same characteristics as those of a controlled variable transmission.
In robotics, a task frequently performed is moving an arm between two positions
where the starting and final velocities are zero. For such a highly specialized task
a transmission like the dual smart drive can be of use.
The dual SMART21 drive This transmission designed by Akinfiev et al. [1], is
intended for walking robots in which each cycle is much like another. The principle
is depicted in Figure 23. The input shaft drives a crank arm at the end of which
is a linear bearing sliding on a link, which serves as the output shaft. Turning the
crank moves the output link to different positions. The transmission ratio can be
referred to as variable for two reasons:
1. At different positions of the crank arm, a small crank displacement will cause
different displacements of the leg. The transmission ratio will be very high
near the leg extreme positions and smaller in the middle. This is a good
property when the transmission is used to drive a leg that starts and stops
close to its extreme positions: it uses the effective power of a motor well,
using a large transmission ratio when accelerating and a small one when
transferring.
21 SMART

stands for Special Mechatronic Actuator for Robot joinTs

To
kc
ko
Time [s]
[rad/s]
4. TRANSMISSIONS
i [rad/s]
o [rad/s]
To [Nm]
[rad]
[]
Ts [Nm] Output o
link
o [ ]
T [Nm]
To
P [W]

35
Flexible joints
Torsion spring
Input link
Fixed link

i
Ti

Figure 24: Schematic of a load-sensitive continuously variable transmission


(LSCVT) designed by Takaki and Omata [31] and the principal relationship between the input and output torques.

2. For a given output angle there are two possible crank angles at which the
transmission ratios are different. When the crank is near the output link
pivot point, the transmission ratio is lower than when the crank is in the
other direction. In this way the dual SMART drive would be ideal for a
walking robot, since high ratios could be used in the support stage and low
ratios in the transfer stage.
Switching between high and low transmission ratios is somewhat problematic, because the output link has to pass one of its extreme positions, but it is possible to
think of situations in which it is possible.

The load-sensitive continuously variable transmission The load-sensitive


continuously variable transmission (LSCVT) is what Takaki and Omata [31] call
a transmission they made to control the fingers of a robotic hand. It consists of
a five-bar linkage (see Figure 24), with a motor driving a crank arm consisting of
two linkages that are flexibly connected. The crank arm is connected to a rod that
drives another crankarm of fixed length, which is the output shaft. Torque will
cause the crankarm to fold, shortening the effective arm length. The torque ratio is
closely connected to the relationship between the crank arm lengths, and since this
ratio becomes larger when the transmission is loaded, the torque ratio will increase.
The effect will thus be for the relation between the input and output torques to be
non-linear. At the same time, the velocity ratio is kept low in the case of low-torque
case.
The adjustable power transmitting mechanism (APTM) The adjustable
power transmitting mechanism (APTM) is a transmission that Ishikawa et al. [10]
built for a robotic hand prosthesis. It is similar to the PVT (see Figure 25), and
can be analyzed using a half model of the PVT. This mechanism has many of the

kc
ko
Time [s]
[rad/s]
i [rad/s]
o [rad/s]
To [Nm]
[rad]
Finger outline
[]
Ts [Nm]
o [ ]
T [Nm]
P [W] Flexible joint

36

INTRODUCTION

Wire

Spring

Figure 25: Schematic of the adjustable power transmitting mechanism (APTM)


designed by Ishikawa et al. [10].

properties of the LSCVT, in that it increases the torque ratio with increasing load
and contains an internal elastic element.
The APTM described can only close the finger. In order to open the finger, a
small spring is used that exerts torque in the opposite (i.e. opening) direction.

4.4

Control of an actuator with compliance

D.W. Robinson wrote a PhD thesis entitled Design and Analysis of Series Elasticity
in Closed-loop Actuator Force Control [26]. He investigated the possibilities of
using linear series elasticity in an actuator that uses force as a reference. The
study concentrates on two issues: 1) force control when the output is at a constant
position, and 2) output impedance, i.e., how the output force changes when the
output is moved (not because of intentional, controlled movement, but because
something external is acting on the output). The saturation of the actuator is taken
into account, and simulations as well as experiments are performed to confirm the
theoretical results.
Robinsons main conclusions are that:
For a linear system with feedback control, mechanical elasticity can be fully
compensated for, with stiffness in the controller. Likewise, compliance can be
added to a relatively stiff system by adjusting the controller parameters.
The impedance is small at low frequencies, while at high frequencies the
impedance is as large as the serial physical spring.
In the case of sinusoidal force references of an amplitude as large as the
maximum force of the actuator, the bandwidth is limited. This is because the
motor has to rotate to compensate for the elasticity, becoming saturated at
higher frequencies. Having a stiffer serial elasticity increases the bandwidth.
Since the PVT is also elastic, these results are applicable in the PVT as well. In
the included Paper F torque control of the PVT is considered to some extent, and
at least the last point above is apparent.

4. TRANSMISSIONS

37

Robinson also discusses the impact tolerance resulting from series elasticity. In
the event of a collision, the impulse force will be dissipated over time, effectively reducing the peak forces. Hence, an actuator with series elasticity is better protected
from damage than a stiff actuator is.

38

INTRODUCTION

Summary of papers

This section gives a brief summary of the papers included in this thesis. The first
two papers are included to explain the platform on which research is done and the
methods used. The research contributions of the thesis are found in the Papers C
F.
Paper A:

Combining control design tools

from modelling to implementation

Christian Ridderstrm

Johan Ingvast

Jan Wikander

This paper describes the tools and methods used when modeling, simulating and doing control design for the robot WARP 1. Similar methods
are also used for the work with the PVT (Paper F).
The tools developed for modeling of the robot are based on the computer algebra system Maple which allows the models to be expressed
symbolically. Hence, not only code for a simulation model (Simulink)
is generated, but also auxiliary functions for control, analysis and evaluation. Furthermore, the symbolical model can be simplified, e.g. linearized symbolically, which makes it possible to use the model more
direct for analysis.
Control design, simulation, experiments and evaluation is done within
the Simulink/Matlab environment. A feature is that the controller
used for simulation also is used for experiments. The switching between
performing simulation and performing experiment is highly automated.
Another feature is that names of variables (signals) and constants are
the same through out the developed tools. Thereby, the values of constants only have to be defined in one tool and signal names only have
to be defined in Maple.
My contribution: Low level programming for communication; building
interfaces in simulink.
The paper was presented at the International Conference on Robotics
and Automation (ICRA) 2001, [22].

39

5. SUMMARY OF PAPERS
Paper B:

The basic design of the quadruped robot WARP 1

Christian Ridderstrm

Johan Ingvast

Freyr Hardarson

Mats Gudmundsson

Mikael Hellgren

Jan Wikander

Tom Wadden

Henrik Rehbinder

WARP 1 is a quadruped robot that weighs approximately 60 kg. Its four


legs are 0.6 m long and each leg has three actuated joints.
This paper describes WARP 1, its actuation system, sensors, and surrounding hardware. It also describes, the communication between microcontrollers and the computers that execute the algorithms.
The paper ends with experimental results of the strength and speed of
the legs.
Since the paper was written, some details on the robot have been
changed: the gear ratio of the hip abduction/adduction joint has been
changed (see section 2); the two computers for control have been replaced by one; and finally, the robot lead batteries are now NiMH batteries. They can power the robot while walking for half an hour.
My contribution: redesign of electrical and mechanical parts of the
robot; performing simulation and experiments.
The paper was presented at the International Conference on Climbing
and Walking Robots (CLAWAR) 2000, [25].

40
Paper C:

INTRODUCTION
WARP 1: Towards walking in rough terrain

Control of walking

Johan Ingvast
Freyr Hardarson

Christian Ridderstrm
Jan Wikander

This paper presents a crawl gait on the robot WARP 1. The controller is
based on the concept of weight-ratios proposed by Hardarson [4] to plan
the instantaneous motion of the robot. It also presents a method for
choosing new footholds based on the existing footholds. Together with
the controller for transfer of the feet, [24], it makes the robot capable of
walking on uneven ground. This is demonstrated in experiments where
the robot walks over an obstacle 0.11 m high.
My contribution: the idea behind foothold planning; analysis of the
weight-ratios; building of the controller; simulation and experiments.
The paper was presented at the International Conference on Climbing
and Walking Robots (CLAWAR) 2003, [9].

41

5. SUMMARY OF PAPERS
Paper D:

Improving a trotting robots gait


by adapting foot trajectory offsets

Johan Ingvast
Freyr Hardarson

Christian Ridderstrm
Jan Wikander

This paper describes a trot gait of the robot WARP 1. The paper shows
that even on ground with unknown slope, explicit sensing of the robot
attitude is not necessary for a dynamic gait like trot. This is made
possible by sensing when the feet hit the ground. In short, if the front
feet are supporting for a longer duration than the rear feet, it is an
indication that the feet should be more forward relative to the trunk.
This way, WARP 1 adapts to its environment.
In experiments, the robot is steered with a joystick and trots up a slope
and turns around.
My contribution: the idea behind using contact switches for balancing;
building controller; simulation and experiments.
The paper was presented at the International Conference on Climbing
and Walking Robots (CLAWAR) 2001, [8].

42
Paper E:

INTRODUCTION
The trunk follows the feet

an approach for making a quadruped robot trot and walk

Johan Ingvast

Jan Wikander

This paper describes how the method used in in Paper C for the crawl
gait can be extended to the trot gait and allow smooth changes between
the crawl gait and the trot gait. The main idea is to use weight-ratios
for both gaits and to slowly modify them to obtain the gait changes.
The idea from Paper D of using the time that the feet are in contact with
ground to improve balance was brought into the controller. However,
instead of directly changing the foot positions, the controller estimates
the position of the center of mass.
Each time a foot is about to be lifted from ground, the force from the
ground must first be reduced to zero. If the change of the force is large,
then the robot motion is disturbed. Therefore, a method is introduced
for adapting the vertical forces of the feet such that they approach zero
prior to lifting a foot. The method bases the references for the legs
vertical force on the weight-ratios.
The controllers are demonstrated in experiments. Smooth gait transitions as well as the convergence of the center of mass estimation are
shown experimentally.
My contributions: the idea behind using the weight-ratios for trot and
controlling the vertical forces to follow the weight-ratios; building controller; simulations and experiments.
The paper is intendend for submission to the IEEE International Conference on Intelligent Robots and Systems (IROS).

43

5. SUMMARY OF PAPERS
Paper F:

The PVT, an elastic conservative transmission

Johan Ingvast

Christian Ridderstrm

Jan Wikander

This paper presents the passively variably transmission (PVT). It has


some of the properties of a continuously variable transmission (CVT)
in that torque is amplified at large loads and velocity is amplified at
small loads. However, contrary to traditional transmissions, it is elastic.
Using the elastic property and an assumption of no losses, theory for
such transmissions is developed. The concepts of velocity ratio and
torque ratio are introduced to characterize the transmissions and they
allow for comparison to more traditional transmissions.
The PVT is compared to a similar transmission called the load sensitive
CVT which also increases torque at high loads and velocity at low loads.
A prototype of a PVT is built and presented. Experiments show that
it is well suited for use in a torque servo. Furthermore, in simulation
it is shown that by using a PVT instead of a traditional transmission
the motor size can be reduced when driving a joint subjected to loads
found at a joint of a horse.
My contributions: the idea behind the PVT and all of the work from
theory to experiments and writing.
The paper will be submitted to the International Journal of Robotics
Research.

44

INTRODUCTION

Discussion

The original aim of WARP 1 was to build a robot suitable for autonomous walking
in rough terrain. To be able to walk autonomously we had to learn how to make
the robot walk. Therefore, the robot is still connected to off-board computers. The
robot is also during most experiments electrically powered via an external cable.
However, the robot is sometimes used with batteries that have the ability to supply
all the power (excluding the off-board computer) for at least half an hour of walking.
Furthermore, it would be relatively easy to move the computers onboard since the
robot is strong enough and there is enough space in the trunk. By doing that,
the robot could walk outside of the lab and be steered by a joystick. To be able
to perform experiments under realistic conditions outdoors making the computers
on-board is almost necessary.
The original intention of WARP 1 was to make the robot modular and the control
highly distributed. This resulted in a design with two microcontrollers on each leg
and two on the trunk. However, from a control point of view, we found it more
practical, in the scope of walking research, to make the controllers centralized. The
two microcontrollers on each leg were reduced to one microcontroller which only
functions as an intelligent I/O node. At the time of writing Papers C and D the
control was distributed on two PCs, one taking care of attitude estimation and
the other everything from current control to balancing. This worked well from a
control point of view since the attitude estimation is well separated from the rest
of the control. However, this separation made experiments more complicated, so
the two tasks were later put into one computer. Even though the control hardware
is centralized, the control structure still has the original layout.
As noted earlier, the current control of the motors is done in the PC. The
implication of this is a rather poor performance in the current control loop. This
is mainly due to two reasons. The first reason is that the sample time is set by
the relatively long sample time of the execution of the code in the PC. The second
reason is the delays in the loop because of the communication. Therefore, the
overall performance would likely benefit from having the current control executed
on the microcontrollers instead of the PC.
The trunk of the robot comprise a frame of extruded aluminum profiles. These
are joined with brackets, nuts and bolts. The frame is therefore not as stiff as
intended and first expected. Furthermore, this flexibility is not modeled in the
simulation environment. An estimation of the twisting stiffness showed that it is
only about three times as stiff as the twisting stiffness resulting from the vertical
control of the feet thus a significant contribution of flexibility to the overall
stiffness. It is likely that the walking, particularly the static walking would benefit
from a stiffer trunk.
One implication of the trunk flexibility is that the feet are not were the joint
sensors indicate. The flexibility also affects how high a foot can be lifted. The knee
has a limited working range that restricts the distance between the hip and foot
to between 0.37 m and 0.60 m. In a normal stance the hip is about 0.52 m above

6. DISCUSSION

45

ground. Hence, when the foot is directly below the hip, it can not be placed on
something higher than 0.15 m above the ground. The ground contact sensor and
the flexibility of the trunk additionally reduces this. To be able to walk up stairs,
and in uneven terrain, the ground clearance should be larger.
Another important property of the robot is the leg weight. As stated in Paper E,
the acceleration and deceleration of the legs exert disturbances on the robot. If the
leg weight is reduced or moved closer to the trunk, much could be gained. The
moving parts of the legs, thigh and shank weigh 5.3 kg. Out of this the motor (for
the knee) and main parts of the transmission contribute only 1 kg. Hence, most
of the weight comes from bearings, wire transmission, and structure there is a
large potential to make the legs lighter.
WARP 1 trotted before it crawled, even though some unsuccessful attempts were
made to make it crawl earlier. This is an indication that the design is more suited
for dynamic than static walk:
the center of mass is high.
the compliances of the leg joints and trunk make the position control of the
center of mass with respect to the support area poor.
the controllers do not take into account the motion of the center of mass of
the robot resulting from the motions of the relatively heavy legs.
It is amazing that the controller for the trot in Paper D works as well as it does.
The vertical distance between the feet on the ground and the center of mass is
approximately 0.6 m, resulting in an eigenvalue for the dynamics of 4 s 1 (see
Figure 6 fig:kappa-inverted-pendulum and equation (3)). The feet can be in the air
for longer than 0.6 s, which implies that the angle , that represents the center of
mass position, is approximately 10 times larger when the feet hit the ground than
the angle when the feet leave the ground. Experimental evidence indicates that the
degree of falling is almost undetectable by the naked eye which must be considered
good.
The graphical design environment given by Simulink has many advantages. The
design becomes to some extent transparent so that it is easy to see what is done.
Furthermore, the design is easy to modify and the syntax much simplified. However,
as the complexity of the controller grows, disadvantages of the graphical environment become more obvious. The hierarchy of subsystems becomes deeper and a
lot of subsystems are needed only to organize the signals (select the appropriate
signals). Handling initial values and transitions between different controllers are
difficult in the environment we have built. There is a feeling that a few lines of Ccode could replace a number of graphical subsystems. Furthermore, sometimes the
graphical design becomes so complicated that the transparency, one of the reasons
for using graphical design, is lost.
Even so, as long as the controllers are simple, it is easy to use the graphical
approach in the environment around WARP 1. Occasionally, master students have

46

INTRODUCTION

made projects on WARP 1, and in getting to know the robot, they have implemented
simple controllers by themselves. This after being taught the system for only a few
hours.
The main inspiration for the work on the PVT was its possible use as a transmission of a leg joint on a walking robot. Since its capabilities of amplifying torque
at high loads and speed at low, it should be well suited. However, doing that is not
straight forward. In Paper F the PVT was demonstrated in a torque control loop
where it performed reasonably well. However, the results of [7] indicate that it is
not so useful for position or velocity control.
Control theory often separates regulators from servos. Regulators are controllers
that are good at keeping a value constant under disturbances, and servos are good
at following a reference signal that varies. In this sense, the PVT is an excellent
torque regulator and less good torque servo. Since velocity and position control
requires that the torque has to change, the PVT is not so suited for that. The
walking controllers presented in this thesis have position or velocity control of the
feet. Hence, those can probably not be used as they are when using PVTs for the
leg joints. However, to change the leg controllers in Paper C into controllers with
force as reference would probably not be too hard. The force distribution algorithm
proposed in section 1.4 could then be used.
The PVT is only demonstrated alone, without interference from dynamics of
other actuators and controllers. The nonlinearities, mainly nonconstant stiffness
and saturation of the motor, are likely difficult to handle when more than one PVT
is acting on the same leg. The alternative to use the PVT at one joint of each leg
(preferrably the knee) is more likely to be possible.
Having, the PVT at a knee joint would give more features than increased
strength and speed. Raiberts robots [18] have springs in the leg to be able to
store and retrieve the energy in the running gaits (with aerial phases). If the PVT
would be used in actuation of the legs, the spring characteristics would come for
free.
Waldron heads an ongoing project aimed at quadruped gallop at 10 m/s [27].
Their robot uses springs to store energy as a mean to reduce the actuator size. The
group has designed the springs in a way that the leg stiffness becomes larger as the
leg becomes shorter. This way they hope to avoid that the springs become completely compressed. The PVT at a knee joint produces this characteristic inherently
since its stiffness approaches infinity as the load grows.
These features should be enough to motivate a test of the PVT as a transmission
on walking robots.

47

Bibliography
[1]

T. Akinfiev, R. Fernndez, and M. Armada. Dual smart drive for walking


robots. In Professional Engineering Publishing Limited, editor, 5th International Conference on Climbing and Walking Robots, pages 595602, Paris,
France, September 2002.

[2]

Ben Brown and Garth Zeglin. The bow leg hopping robot. In Int. Conf. on
Robotics and Automation, pages 781786, 1998.

[3]

L. Brown. The importance of proper tripods. Curator, 11:732, 1968.

[4]

Freyr Hardarson. Stability analysis and synthesis of statically balanced walking for quadruped robots. PhD thesis, The Royal Inst. of Technology, 100 44
Stockholm, Sweden, June 2002.

[5]

M. Hildebrand. Symmetrical gaits of horses. Science, 150:701708, 1965.

[6]

M. Hildebrand. The quadrupedal gaits of vertebrates. BioScience, 39(11):766


775, December 1989.

[7]

Johan Ingvast. PVT The passively variable transmission. Technical Report


TRITA-MMK 2005:17, Dept. of Machine Design, Royal Institute of Technology, S-100 44 Stockholm, Sweden, December 2005.

[8]

Johan Ingvast, Christian Ridderstrm, Freyr Hardarson, and Jan Wikander.


Improving a trotting robots gait by adapting foot reference offsets. In Professional Engineering Publishing Limited, editor, 4th International Conference on
Climbing and Walking Robots, pages 711718, Karlsruhe, Germany, September
2001.

[9]

Johan Ingvast, Christian Ridderstrm, Freyr Hardarson, and Jan Wikander.


Warp1: Towards walking in rough terrain control of walking. In Professional Engineering Publishing Limited, editor, 6th International Conference
on Climbing and Walking Robots, pages 197 204, Catania, Italy, September
2003.

[10] Yasuhiro Ishikawa, Wenwei Yu, Hiroshi Yokoi, and Yukinori Kakazu. Development of robot hands with an adjustable power transmitting mechanism.
Intelligent Engineering Systems Through Neural Networks, 10:631636, 2000.
[11] W. Y. Jiang and D. Howard. Foot-force distribution in legged robots. In
Professional Engineering Publishing Limited, editor, 4th International Conference on Climbing and Walking Robots, pages 331 338, Karlsruhe, Germany,
September 2001.
[12] R. B. McGhee and A. A. Frank. On the stability properties of quadruped
creeping gaits. J. Math. Biosciences, 3:331351, 1968.

48

BIBLIOGRAPHY

[13] Domini A. Messuri and Charles Klein. Automatic body regulation for maintaining stability of a legged vehicle during rough-terrain locomotion. IEEE
Robotics and Automation Magazine, RA-1(3), September 1985.
[14] Carl Anthony Moore. Continuously variable transmission for serial link cobot
architectures. Masters thesis, Northwestern University, March 1997.
[15] Eadweard Muybridge. Animals in Motion. New Dover Edition, Dover Publications, Inc., New York, 1957. First published in 1899.
[16] Ionannis Pulakakis, Evangelos Papadopoulos, and Martin Buehler. On the
stable passive dynamics of quadrupedal running. In Int. Conf. on Robotics
and Automation, September 2003.
[17] Ionannis Pulakakis, James Andrew Smith, and Martin Buehler. On the dynamics of bounding and extensions towards the half-bound and the gallop gaits.
In International Symposium on Adaptive Motion of Animals and Machines,
March 2003.
[18] Marc H. Raibert. Legged robots that balance. The MIT Press, 1986.
[19] Marc H. Raibert, Michael Chepponis, and jr H. Benjamin Brown. Running on
four legs as though they were one. IEEE Robotics and Automation Magazine,
2(2):70 82, June 1986.
[20] Henrik Rehbinder and Xiaoming Hu. Nonlinear pitch and roll estimation for
walking robots. In Int. Conf. on Robotics and Automation, volume 3, pages
26172622, 2000.
[21] Christian Ridderstrm. Legged locomotion: Balance, control and tools from
equation to action. PhD thesis, The Royal Inst. of Technology, 100 44 Stockholm, Sweden, May 2003. See http://www.md.kth.se/~cas/publications.
[22] Christian Ridderstrm and Johan Ingvast. Combining control design tools
from modeling to implementation. In Int. Conf. on Robotics and Automation,
pages 13271333, 2001.
[23] Christian Ridderstrm and Johan Ingvast. Quadruped posture control based
on simple force distribution a notion and a trial. In IEEE/RSJ Int. Conf.
on Intelligent Robots and Systems, pages 23262331, Hawaii, USA, 2001.
[24] Christian Ridderstrm and Johan Ingvast. Warp1: Towards walking in rough
terrain smooth foot placment. In Professional Engineering Publishing Limited, editor, 6th International Conference on Climbing and Walking Robots,
pages 467 474, Catania, Italy, September 2003.

49
[25] Christian Ridderstrm, Johan Ingvast, Freyr Hardarson, Mats Gudmundsson,
Mikael Hellgren, Jan Wikander, Tom Wadden, and Henrik Rehbinder. The
basic design on the quadruped robot Warp1. In Professional Engineering Publishing Limited, editor, 3rd International Conference on Climbing and Walking
Robots, pages 87 94, Madrid, Spain, October 2000.
[26] David William Robinson. Design and analysis of series elasticity in closedloop actuator force control. PhD thesis, Massachusetts Inst. of Technology,
June 2000.
[27] J. P. Schmiedler and K. J. Waldron. The mechanics of quadrupedal galloping
and the future of legged vehicles. Int. J. of Robotics Research, 18(2):12241234,
1999.
[28] James Andrew Smith and Ioannis Poulakakis. Rotary gallop in the untethered
quadrupedal robot Scout II. In IEEE/RSJ Int. Conf. on Intelligent Robots
and Systems, volume 3, pages 2556 2561, 2004.
[29] Shin-Min Song and Kenneth J. Waldron. Machines that Walk. MIT Press,
Cambridge, MA, 1989.
[30] Taku Takahama and Katsuhiko Inagaki. A design method for practically used
walking machine. In IEEE/RSJ Int. Conf. on Intelligent Robots and Systems,
pages 414 419. IEEE, New York, NY, USA, 1997.
[31] Takeshi Takaki and Toru Omata. Load-sensitive continuously variable transmission for robot hands. In Int. Conf. on Robotics and Automation, pages 3391
3396, 2004.
[32] S. Talebi, I. Poulakakis, E. Papadopoulos, and M. Buehler. Quadruped robot
running with a bounding gait. In International Symposium on Experimental
Robotics, pages 281 289, 2000.
[33] Logistical vehicle off-road mobility. Fort Eustis, Va.: U.S. Army Transportation Combat Developments Agency, feb 1967. (Project TCCO 62-5) Not read
by the author.
[34] Miomir Vukobratovic and Branislav Borovac. Zero-moment point thirty five
years of its life. Journal of Humanoid Robotics, 1(1):157 173, 2004.
[35] Miomir Vukobratovic and D. Juricic. Contributions to the synthesis of biped
gait. In IFAC Symp. Technical and Biological Problem on Control, USSR,
1968.
[36] Kenneth J. Waldron, Vincent J. Vohnout, Arrie Pery, and Robert B. McGhee.
Configuration of the adaptive suspension vehcile. Int. J. of Robotics Research,
3(2):3748, 1984.

50

BIBLIOGRAPHY

[37] K. Yoneda, H. Iiyama, and S. Hirose. Intermittent trot gait of a quadruped


walking machine dynamic stability control of an omnidirectinal walk. In Int.
Conf. on Robotics and Automation, volume 4, pages 30023007, 1996.
[38] Garth Zeglin and Ben Brown. Control of a bow leg hopping robot. In Int.
Conf. on Robotics and Automation, pages 793798, 1998.

You might also like