Design of Robotic Quadruped Legs

by

Jacob Elijah McKenzie

Submitted to the Department of Mechanical Engineering ARCHIVES
in partial fulfillment of the requirements for the degree of MASSACHUSETS INSTITUrE
OF TECHOL0

Master of Science FEB 18 2012

at the
L BRARIES
MASSACHUSETTS INSTITUTE OF TECHNOLOGY

February 2012

© Massachusetts Institute of Technology 2012. All rights reserved.

A u th o r .............................................................
Department of Mechanical Engineering
January 20, 2012

Certified by...
angbae Kim
Esther and Harold E. Edgerton Assis nt Professor of Me anical Engineering
Thesis Supervisor

~-, /.

Accepted by .................... ....... ..............................

David E. Hardt
Ralph E. and Eloise E Cross Professor of Mechanical Engineering
Chairman, Department Committee on Graduate Theses
2
 Design of Robotic Quadruped Legs

by

Jacob Elijah McKenzie

Submitted to the Department of Mechanical Engineering

on January 20, 2012, in partial fulfillment of the
requirements for the degree of
Master of Science

Abstract
Prized for their performance on prepared surfaces, wheeled vehicles are often limited in mo-
bility by rough and unstructured terrain. Conversely, systems that rely on legs have shown
promising rough terrain performance but only a modest ability to achieve high speeds over
flat terrain. The goal of this thesis is to develop four robotic legs that are capable of robust
dynamic running over flat terrain. Demonstration of this ability is necessary to improve the
viability of robotic legs as a propulsion system. Achieving true dynamic running presents
many challenges, and the first step in prevailing over the difficulties this task presents is the
development of a sound mechanical system. The leg designs presented here are based on
the development of four design principles from both biological systems, dynamic simulations
and previous research. These principles suggest that a leg design should: minimize passive
mechanical impedance, minimize mass and inertia, maximize actuator strength and develop
a balance between leg kinematics and robot use. To bring these principles into reality sev-
eral unique design features were introduced including a doubly concentric actuator layout,
synthetic fiber tendons to reduce bending loads in the legs, polymer leg links and the use
of electric motors to their thermal limit. To accompany these technical features simulation-
based design tools were developed that provide an intuitive insight into how altering design
parameters of the leg may affect locomotion performance. The key feature of these tools is
that they plot the forces that the leg is capable of imparting on the body for a given set of
dynamic conditions. Single and multiple leg testing has shown that the legs perform well
under dynamic loading and that they are capable producing vertical ground reaction forces
larger than 800 N and horizontal forces larger than 150 N. Many of the design principles,
features and tools developed may be used with a large variety of leg structures and actuation
systems.

Thesis Supervisor: Sangbae Kim

Title: Esther and Harold E. Edgerton Assistant Professor of Mechanical Engineering
4
Acknowledgments

This research is supported by the Defense Advanced Research Projects Agency (DARPA) as
part of the Maximum Mobility and Manipulation (M3) project. I would like to aknowledge
the members of the Biomimetic Robotics Lab, and thank them for their support.

For EflflJ
6
Contents

1 Introduction 15
1.1 Motivation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 Background: Running in Biology . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.1 G aits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.2 Leg Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.2.3 Leg Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.2.4 Force Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3 Background: Dynamic Performance . . . . . . . . . . . . . . . . . . . . . 30

2 Design Principles 31
2.1 Operational Phases . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . 31
2.1.1 Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.2 Stance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.3 Swing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.2 Performance Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.3 Design Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3 Leg Design 47
3.1 Design Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3 Detailed Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.3.1 Front Leg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.3.2 Rear Leg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4 Modeling and Performance Estimation 71

4.1 Actuator Modeling ........ .............................. 71

4.2 Single Leg Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.2.1 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.2.2 Stance Phase Simulation . . . . . . . . . . . . . . . . . . . . . . . 82

4.2.3 Protraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

5 Design Tools 91

5.1 Assessing Leg Capability with Force Maps . . . . . . . . . . . . . . . . . . 91

5.2 Force Map Stride Planning . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2.1 Selecting a Stride Primitive . . . . . . . . . . . . . . . . . . . . . . 97

5.2.2 Calculating the Force Production Envelope . . . . . . . . . . . . . 98

5.2.3 Selecting Touchdown Configuration . . . . . . . . . . . . . . . . . 101

5.2.4 Calculating Torque Profiles . . . . . . . . . . . . . . . . . . . . . . 102

6 Testing and Design Iteration 105

6.1 Design Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

6.2 Single Leg Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3 Multi Leg Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

7 Summary/Conclusions 109

7.1 Design considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.2 Unique Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3 D esign tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

A Engineering Drawings 113

B Matlab Code Listing 137

List of Figures

1-1 Conceptual model of quadrupedal robot . . . . . . . . . . . . . . . . . . . 16

1-2 Workspace of wheel versus leg . . . . . . . . . . . . . . . . . . . . . . . . 18

1-3 Support phase diagrams for biped gaits . . . . . . . . . . . . . . . . . . . . 19
1-4 Footfall event diagrams for biped gaits . . . . . . . . . . . . . . . . . . . . 19
1-5 Definitions of stride related variables . . . . . . . . . . . . . . . . . . . . . 20
1-6 Common quadrupedal gaits . . . . . . . . . . . . . . . . . . . . . . . . . 21
1-7 Support phase diagram of the rotary gallop . . . . . . . . . . . . . . . . . 21
1-8 Support phase diagram of the transverse gallop . . . . . . . . . . . . . . . 22
1-9 The plantigrade, digitigrade, and unguligrade legs . . . . . . . . . . . . . . 23
1-10 Skeletal structure of the wolf . . . . . . . . . . . . . . . . . . . . . . . . . 24
1-11 Position of the shoulder throughout a stride . . . . . . . . . . . . . . . . . 26
1-12 Ground reaction forces of a galloping dog . . . . . . . . . . . . . . . . . . 28

2-1 Decoupling of body reaction forces . . . . . . . . . . . . . . . . . . . . . . 33

2-2 Model of leg extension and compression during impact . . . . . . . . . . . 35
2-3 Series compliance method of reducing impact impedance . . . . . . . . . . 35
2-4 Two dimensional, one legged hopping model . . . . . . . . . . . . . . . . 37
2-5 Effect of horizontal disturbance force (Fdb,) on one leg hopper . . . . . . . 39

2-6 Effect of vertical disturbance force (Fdby) on one leg hopper . . . . . . . . . 39

2-7 Effect of mass disturbance (md) on one leg hopper . . . . . . . . . . . . . . 40

2-8 Effect of horizontal velocity disturbance (ad) on one leg hopper . . . . . . . 40

3-1 Prismatic, articulated, and redundantly articulated topologies . . . . . . . . 49

3-2 Motion of the prismatic leg . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3-3 Motion of the articulated leg . . . . ... . . . . . . . . . . 51

3-4 Configuration gearing of the elbow . . . . . . . . . . . . . 51

3-5 Motion of the redundantly articulated leg . . . . . . . . . 52

3-6 Configuration possibilities with redundant leg . . . . . . . 52

3-7 Selected leg topology . . . . . . . . . . . . . . . . . . . . 53

3-8 Simple leg design, shoulder high inertia . . . . . . . . . . 55

3-9 Doubly concentric leg design, low shoulder inertia . . . . . 55

3-10 Link bending load during stance . . . . . . . . . . . . . . 56

3-11 Bending loads are reduced by a static tendon . . . . . . . . 56

3-12 Front leg, front view . . . . 57

3-13 Front leg, rear view . . . . . . . . . . . . . . . . 57

3-14 Allied Motion HT05001 motor . . . . . . . . . . 60

3-15 Shoulder unit of front leg (front view) . . . . . . . 61

3-16 Shoulder unit of front leg (rear view) . . . . . . . 61

3-17 Shoulder actuation . . . . . . . . . . . . . . . . 62

3-18 Elbow actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2

3-19 Schematic representation of gearbox . . . . . . . . . . . . . . . . . . . . . 6 3
3-20 Gear weight reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3-21 Planet carrier and planets assembled . . . . . . . . . . . . . . . . . . . . . 6 4
3-22 Closeup photo of the assembled gearbox . . . . . . . . . . . . . . . . . . . 6 4
3-23 Bearing placement in front leg . . . . . . . . . . . . . . . . . . . . . . . . 6 5
3-24 Ad/abduction mount on the shoulder unit . . . . . . . . . . . . . . . . . . 6 5
3-25 Assembled front leg links . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3-26 Foam composite leg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 6

3-27 Front leg Kevlar tendon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 6

3-28 Rear leg, front view . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3-29 Rear leg, rear view . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . 67

3-30 Gear reduction of the hip motor . . . . . . . . .. . . . . . . . . . . . . . 6 9

3-31 Motion of the knee motor . . . . . . . . . . . .. . . . . . . . . . . . . . . 6 9

3-32 Bearing placement in the rear leg . . . . . . . . . . . . . . . . . . . . . . . 6 9

4-1 Anticipated motor current in a fast gallop . . . . . . . . . . . . . 72
4-2 Winding temperature under 300 W load . . . . . . . . . . . . . . 73
4-3 Thermal model of motor . . . . . . . . . . . . . . . . . . . . . . 73
4-4 Kinematics of a two link serial chain . . . . . . . . . . . . . . . . 75
4-5 Revolute and prismatic joints in screw theory . . . . . . . . . . . 76
4-6 n-link kinematic chain . . . . . . . . . . . . . . . . . . . . . . . 77
4-7 Definition of geometry for two link serial chain . . . . . . . . . . 82
4-8 Stance phase program layout . . . . . . . . . . . . . . . . . . . . 83
4-9 Accessible body forces in a specific configuration . . . . . . . . . . 83
4-10 Available body forces at y = 35 cm . . . . . . . . . . . . . . . . . 85
4-11 Available body forces at y = 40 cm . . . . . . . . . . . . . . . . . 85

4-12 Available body forces at y = 45 cm . . . . . . . . . . . . . . . . . 85

4-13 Available body forces at y = 50 cm . . . . . . . . . . . . . . . . . 85
4-14 Available body forces at v, = 2 m -s- . . . . . . . . . . . . . . . 86
4-15 Available body forces at v2 = 5 m - s . . . . . . . . . . . . . . . 86
4-16 Available body forces at v2 = 10 m - . . . . . . . . . . . . . . . 86
4-17 Available body forces at v2 = 15 m - s . . . . . . . . . . . . . . 86
4-18 Three link leg for protraction analysis . . . . . . . . . . . . . . . 87
4-19 Protraction optimization results for shoulder torque . . . . . . . . 88
4-20 Protraction optimization results for knee torque . . . . . . . . . . 88
4-21 Leg motion during protraction . . . . . . . . . . . . . . . . . . . 89

5-1 Approximate hip (dashed) and shoulder (solid) trajectories for a running
ch eetah . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5-2 Workspace of the front leg with fixed foot . . . . . . . . . . . . . . . . .

5-3 Force production envelope for = 2m - s . . . . . . . . . . . . . . . .

5-4 Weighted force production envelope for i = 2m - s. . . . . . . . . . . .

5-5 Acceptable touchdown configurations . . . . . . . . . . . . . . . . . . .

5-6 Acceptable touchdown configurations as a function of speed . . . . . . .

5-7 Straight versus curved shoulder profile . . . . . . . . . . . . . . . . . . .
5-8 State dependent force envelope . . . . . . . . . .... . . . . . . . . . . . . 99
5-9 The bounds of the SDFE . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

6-1 Single leg testing slide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6-2 Video of single leg testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6-3 Single leg test data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

6-4 Jumping of four legged robot . . . . . . . . . . . . . . . . . . . . . . . . . 108

List of Tables

3.1 Weight budget for quadruped robot . . . . . . . . . . . . . . . . . . . . . 48

3.2 Front leg design specifications . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.3 Motor parameters for custom wound HT05001 . . . . . . . . . . . . . . . 60
3.4 Rear leg design specifications . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.1 Leg parameters used in dynamic model . . . . . . . . . . . . . . . . . . . 84

14
Chapter 1

Introduction

The research presented in this thesis documents the development of the front and rear legs of
a quadrupedal robot designed for high speed, dynamic locomotion (cf. Figure 1-1). Target
speeds over nominally flat natural terrain are 50 km - h- - by 2014. Attaining such speeds
requires substantial research in the areas of mechanical design, actuation systems, control
theory and manufacturing. The work presented within addresses these four areas and sug-
gests a unique, low mechanical impedance, leg design. The structure of the leg is based on
four design principles developed from analysis of biological systems and dynamic modeling.
A detailed design, covering the topics of leg geometry, actuator selection and component
level design, is presented alongside the development of simulation design tools, intended
to quantify the effects of design modifications in a manner that provides insight into the
performance of the locomotive system.

1.1 Motivation

Wheeled vehicles provide many practical advantages, and over time a highly refined set of
design tools and processes have been developed to aid their implementation. However, one
disadvantage inherent to the wheel itself, its inability to cope with uneven terrain, severely
limits the reach of all wheeled platforms [64]. To avoid this disadvantage the terrain is
often engineered, cleared of obstacles and surfaced, giving it a structure that complements
the wheel's strengths. Unfortunately, nature and the human environment rarely exhibit this
 Figure 1-1: Conceptual model of quadrupedal robot showing four legs, compliant spine and tail.

type of structural uniformity, limiting the reach of vehicles of a given size to a small portion of
their surroundings. Increasing the mobility of robotic platforms requires a new approach,

a fundamental change in the way the ground is used to generate locomotive forces. The
option explored in this thesis is the use of four robotic legs. This research is part of the
Maximum Mobility and Manipulation (M3) program initiated by the Defense Advanced
Research Projects Agency (DARPA).

Several features of wheeled vehicles have lead to their widespread adoption. First and
foremost among these is the simplicity of their fundamental structure. A four wheeled ve-
hicle of sound construction is statically stable and is thus capable of supporting itself and
remaining stationary on level ground. If a driving torque is applied to a laterally opposed
pair of wheels the resulting motions can be simply expressed, particularly at low speeds un-
der quasi-static conditions [52]. Under more extreme conditions the equations of vehicle
dynamics grow in complexity, however, a range of dynamic models of varying fidelity allow
the designer to select a model that suits the demands of the application. Another admirable
pair of features are the exceptionally high locomotive efficiencies possible on prepared sur-
faces and the excellent maneuverability that is possible with careful design. Advances in tire
construction, suspension systems and aerodynamics allow some vehicles to maneuver with
lateral accelerations larger than three times the acceleration of gravity [11].

A key characteristic of the environments that are prepared for wheeled vehicles is that
no abrupt terrain irregularities approach the wheel diameter [64]. If such an obstacle is
encountered the vehicle must be directed along an alternative route. This decision is likely
to increase the total transit time from origin to destination, or in the worst case, the inability
of the vehicle to traverse the obstacle may make the destination unreachable.

Increasing the mobility of wheeled platforms over unstructured terrain can be approached
in two ways: increasing wheel size or increasing the capacity of the wheel to deflect from
its nominal position, such a change is said to increase the workspace of the wheel. The first
option, enlarging the wheel, necessitates the proportional enlargement of the wheel well in
order to accommodate the largest anticipated wheel displacement. This may substantially,
and perhaps unacceptably, increase the overall size and mass of the vehicle, detracting from
its performance in other areas. The alternative approach, increasing the workspace of the
wheel, increases the ability of the system to handle terrain irregularities to as much as three
wheel diameters [56]. However, this ability often comes at the expense of increasing kine-
matic complexity and reducing dynamic performance.

With a legged system a step change in height is approached in a fundamentally different

way. The restrictions that a continuous path of contact imposes are removed, and therefore
the point of contact is no longer required to climb the side of a step change in height. Con-
sequently the foot may be made small, resulting in an increase in the available workspace
(cf. Figure 1-2). Thus for a given external dimension the legged system tends to exhibit a
superior ability to maneuver on stepped terrain. To illustrate this difference, suppose that an
objective lies on the far side of a fallen square column that is 75 cm high. Most humans could
simply step on to and then off of the column with relative ease, making the most of the large
workspace our legs provide. However, even the most capable off-road vehicles would have
difficulty traversing this obstacle and reaching the objective. In the human environment this
type of terrain irregularity is exceedingly common.

The large workspace of legs can also be used to limit, or eliminate, steady state orienta-
tion errors of the body. This is generally not possible with wheeled vehicles that implement
passive suspensions. Furthermore, because foot placement locations may be selected dis-
cretely during the traversal of uneven terrain, locations may be chosen to minimize the
pitch, roll and height dynamics of the body. This foot placement approach reduces the
apparent ruggedness of the terrain.
Figure 1-2: The workspace of a wheeled system and legged system of similar size are shown in a nominal
configuration (solid lines) and in a maximally deflected vertical position (dashed lines).

The capability of legged systems to traverse rough terrain is continually being advanced
by researchers [50]. However, before a legged robot can become a truly viable alternative

to a wheeled system, it must demonstrate comparable ability on relatively flat, structured

terrain as well. This is the focus of the present research. Once speed on level terrain has
been improved, the origin to destination transit time over mixed terrain may begin to favor
legged systems. Demonstration of this ability, and the development of the design methods
and tools that enable it, are intended to prompt wider adoption and improve the mobility
of robotic platforms.

1.2 Background: Running in Biology

Mixed, unstructured, terrain is abundant in the natural world, and one need look no further
than the animals that occupy it to witness the potential capabilities of legged systems to
perform admirably across its many terrain types. High speed locomotion across relatively
even terrain typically occurs in the form of running. The precise definition of running is
somewhat nebulous, however it is typically defined as the locomotion pattern of a legged
system that contains an aerial phase, during which none of the limbs are in contact with
the ground for some (generally short) period of time [6]. Many footfall patterns or gaits can
be found in nature that align with this definition. However, the operational range of many
legged robots confines them to gaits with quasi-static stability conditions, precluding the
execution of true running gaits. Moving legged robots into the realm of dynamic running
requires the development of a thorough understanding of the measures required to maintain
robustness, maneuverability and speed. i
1.2.1 Gaits

A gait is a particular temporal sequence of foot contact events that are repeated during the

steady locomotion of a legged system. Walking, running and hopping are all examples of
bipedal gaits (cf. Figure 1-3 and 1-4). In a human walking gait, support is transferred from
the right to the left foot with a short period of double stance, during this time both feet
are in contact with the ground [2]. For a running gait however, this overlap does not exist;
instead a period of flight, commonly referred to as the aerial phase, takes its place. During
hopping, the contact events of both feet fully overlap, followed by an aerial phase. The use
of a particular gait has far reaching implications on the contact forces and the dynamics of
the body and legs [42], and it is common for a variety of gaits to be used over the speed
range of a given species.

walking walking
L0000
R 0 0 0 0
running running
0 0 0 0
0 0 00
hopping hopping
0 0
0 0

Figure 1-3: Support phase diagrams for three Figure 1-4: Footfall event diagrams for three
biped gaits. Shaded areas indicate when each foot biped gaits. Shaded circles indicate that a foot
is in contact with the ground as time progresses is in contact with the ground. When a contact or
from left to right. liftoff event occurs a new pair of circles are plotted
to the right until the sequence repeats.

A stride is the shortest repeating footfall pattern, a single execution of a gait. The stride
is repeated with a stridefrequency, f (cf. Figure 1-5). The stride length, A, is the distance covered
in a single stride and thus the average forward velocity may be expressed as i= Af. The

duty factor for a single leg, 0, can be defined as the ratio of the leg's contact duration to the
total stride duration, 1/f. The combined duty factor, B, is the ratio of time during which
any leg is in contact to the stride duration.
 t=0 t=T
x=0 x=

Figure 1-5: Definition of stride related variables using a two legged model with a point mass body.

In nature, quadrupeds exhibit many different types of gaits (cf. Figure 1-6) [1]. Two
ubiquitous high speed gaits are the transverse and rotary gallops. The rotary gallop is ex-
ceedingly common in digitigrade carnivores such as cats and dogs. This type of gallop is
characterized by a footfall pattern in which the next limb to make contact is the closest limb
in either a clockwise or a counter clockwise direction (cf. Figure 1-6 and 1-7). This is to
say that if the front right foot is in contact and the animal is using a clockwise rotary gallop
the next foot to touch down will be the rear right. Between each footfall is an aerial phase.
The aerial phase is generally short or nonexistent between laterally opposed limbs and may
be very long between the front and rear limb pairs, depending on speed and species. The
rotary gallop is used by both greyhounds and cheetahs at their maximum speeds [18].

The rotary gallop executed by cheetahs has several unique characteristics at maximum
speed (close to 100 km - h 1 ) [55]. First, the combined duty factor, B, is very small, and
because there are no concurrent contact events, B = /3. Second, the stride length is
very long (A 7 m) and the stride frequency is quite fast (f 2.6 Hz) at 72 km - h-' [18].
In comparison, a horse, which has much longer legs, completes strides of lesser length at a
slower rate.

One major difference between the manner in which cheetahs and horses run is that
horses employ a different gait. The transverse gallop of a horse is used to achieve speeds

close to 60 km - h- [55]. For an animal that is approximately nine times heavier than a
 trot @0 00 00 00
00 000 0 00

bound 00 00 00 00
0 0000 0 00

rotarygallop 0 0 00 0 0 0 0 00 00 00 00
(fast) 00 00 00 00 00 00 00 00

transverse 0 0 0o 0 0 0 0 00 00 00 00
gallop 00 0 0 00 00 00 00 00 00

Figure 1-6: Four common quadrupedal gaits are the trot, bound, rotary gallop and transverse gallop, here
they are shown using footfall phasing diagrams. Each group of four dots represents one contact state.

LH
LF
RF
RH

Figure 1-7: Support phase diagram for the rotary gallop of the cheetah based on data from [18].

cheetah this is quite a substantial speed. The transverse gallop may be either right or left
leading (cf. Figure 1-8). It is worth noting that horses also differ substantially in leg structure
from cats and dogs, which is a topic that will be discussed shortly [63].

In selecting a target gait for a legged robot, consideration must be given to the quantity

and quality of available data and to the applicability of the gait to the desired locomotion
task. For the task of running quickly the rotary gallop is somewhat common in nature but
this does not imply that it will be the most effective gait when implemented with a robotic
platform. The principles of legged locomotion, together with the principles of specific gaits,
must be combined to drive the design and development of the system as a whole. The
leg design and truly the entire robot design must be balanced enough to perform under
maximum speed running conditions as well as the conditions of other gaits that may be
used throughout the speed range.
 LH
LF
RF
RH

Figure 1-8: Support phase diagram for the transverse gallop of the horse based on data from [18].

1.2.2 Leg Structure

Leg musculoskeletal structure varies widely in both geometry and kinematics throughout the
animal kingdom, and when combined with a compatible gait these two elements form the
principle mechanism of legged locomotion. Leg structure and maximum speed have been
shown to correlate with total animal mass, leg structure and many other factors [36]. And
although the precise interaction between load bearing skeletal members is often complex,

joint behavior can be approximated by the substitution of revolute and spherical kinematic
joints under most circumstances [67]. This simplification is essential to the development of
legged robots.

Three classes of leg structure are common throughout the natural world, they are: planti-
grade, digitigradeand unguligrade (cf. Figure 1-9). These structures differ in the elements used
to support the body during ground contact. The most familiar of these structures is planti-
grade, the class used by humans. In a plantigrade leg the tarsals, at the heel of the foot,
lie flat on the ground when standing. Proceeding proximally, the tarsals join the tibia and

fibula of the lower leg followed by the femur of the upper leg. Distally, the tarsals join the
metatarsals of the mid-foot, followed by the phalanges, the small bones of the toes.

A digitigrade leg structure is one in which the digits (toes) are used as the primary support
structure for ground contact. This structure is common among fast and nimble hunters such
as dogs and wolves, cats, birds and dinosaurs. In comparison to the plantigrade structure
of humans, the foot (composed of the tarsals, metatarsals and phalanges) is much longer,
and the weight during standing is borne by the equivalent of the human forefoot. The
unguligrade structure takes this a degree further, bearing the full weight on the tips of the
toes, which are often hoofed to support this immense load.
 plantigrade digitigrade unguligrade

Figure 1-9: Plantigrade, digitigrade and unguligrade hind legs shown omitting members of the skeletal system
that are in contact with the ground while standing.

By examining the leg structure of the wolf in detail, some of the key structural features of
the digitigrade leg become clear (cf. Figure 1-10). Starting at the pelvis, which is connected
to the femur and continuing to the more distal bones of the rear leg. The tibia and fibula
meet the femur at the knee, the tibia being the largest of the two. At the human equivalent
of the ankle the tibia and fibula meet the tarsals. The substantial protrusion to the rear of the
foot is the calcaneum, a tendon attachment point that plays a central role in muscle gearing
[48]. Continuing distally, the majority of the foot length is covered by the metatarsals. These
long and narrow bones each connect to one of the phalanges. At the end of each toe is a
claw. Cheetah claws, unlike those of most cats, are not retractable, with the exception of the
first digital claw of the front paws.

The rear leg is somewhat different than the front when they are compared directly. The
first obvious difference is the inversion of the upper limb structure. Specifically, the knee
and elbow articulate in opposite directions. It has been postulated that this arrangement
enhances the pitch stability of the body when compared to the three other possible configu-
rations [32]. Again starting at the body of the wolf the first skeletal element of interest is the
scapula. This large, flat bone is an attachment point for many muscles in the upper limb and

provides some freedom of positioning and orienting the shoulder joint. The first bone of the
front leg is the humerus, which connects to the radius and ulna at the elbow. Interestingly,
the radius and ulna are far more comparable in size than the tibia and fibula. Both of these
bones are connected to the carpals in the upper foot. Metacarpals make up the majority of
the foot's length, and finally the phalanges and claws constitute the contacting elements of
the foot.
 scapula pelvis
humerus femur
radius & ulna tibia & fibula
carpals tarsals
metacarpals metatarsals
phalanges phalanges

Figure 1-10: Labeled skeletal structure of the wolf showing the characteristics of front and rear digitigrade
legs, adapted from [37].

The muscular system of the cheetah is of indisputable importance to the cheetah's im-
pressive running speed. A cursory examination of feline species reveals relatively few funda-
mental skeletal differences, but the lean muscular build of the cheetah makes its appearance
unique. Muscles, as with any actuator, perform best when they are implemented in a way
that complements their strengths and minimizes their flaws. The muscle composition of
cheetahs tends to favor fast twitch muscle fibers [66] and the connections between the mus-
cle fibers and bone are located to meet the demands of both speed and force [48]. For
example, the placement of the teres major, which links the scapula to the humerus and is
active during stance, connects to the humerus at a location substantially further from the
shoulder than in the greyhound [22]. This muscle gearingincreases the mechanical advantage
of the muscle about the shoulder and decreases the maximum joint rotational velocity, due
to the fundamental limits of muscle contraction velocity [47]. In addition to muscle geome-
try the use of tendons to transmit motive forces produced by proximal muscles to distal joints
represents an admirable structural strategy. This approach reduces the amount of muscle
mass located far from the body, thus minimizing the leg inertia that would accompany such
a mass distribution.

The structure of the foot, as the sole medium through which all locomotive forces are
exerted, also plays a vital role in leg function. The majority of the load bearing of digitigrade
feet is accomplished by the plantar. This soft fibrous tissue has highly nonlinear vertical stiff-
ness, high shear stiffness and exhibits a substantial amount of damping [10]. The location
of the plantar is just aft and below the interface between the phalanges and the metacarpals
or metatarsals. This implies that under standing conditions, a moment is produced about
the phalange-metacarpus (or phalange-metatarsus) joint that forces the phalange tips and
claws into the ground.
In addition to the joints of the legs, the spine also plays a substantial role in locomotion
[4, 18]. Flexion of the spine allows the rear limbs to contact the ground ahead of the position

in which they would land with a fixed length body. This appreciably lengthens the stride and
increases the overall contact duration of the legs. The spine may also aid pitch correction
and enhance the force generating capabilities of the legs by functioning as an additional
redundant joint with very powerful actuation.

1.2.3 Leg Dynamics

The dynamics of quadruped legs have been studied by many biologists [57] and roboticists
[53, 49]. Several key features of the rotary gallop and its implementation by digitigrade
quadrupeds can be understood by close examination of the joint angle trajectories. Joint
angles are typically measured by placing reflective markers on the surface of the skin at
the approximate location of underlying joint. By recording the locations of these markers
with high-speed photography, the approximate joint angles can be extracted with minimal
obstruction of the natural gait [9].
Joint angle trajectories have a period of 1/f, the stride duration, and are repeated during
steady locomotion. In high speed gaits such as the rotary gallop the single leg duty factor
is low (# < 1/4); this implies that the limb is not touching the ground for the vast majority
of the stride. The contact interaction of each leg lasts for only #/f seconds. In this time
the body travels forward approximately A# meters while the foot remains stationary on the
ground. The changes in leg configuration that support this relative motion are captured in
the joint angle trajectories.
During the steady rotary gallop of a dog, at the instant the leading front foot contacts
the ground the limb is extended in a forward position, knee slightly bent and foot fully
extended. Upon contact the foot angle, then the knee angle close, and the shoulder begins
to move forward over the foot. Near mid stance (when the shoulder is directly above the foot)
the knee reaches its maximal flexion. As the stride progresses the knee extends to maintain
ground contact and to provide propulsive forces. Continuing the stride, the knee begins to
approach its maximum extension; flexion of the ankle also reaches a maximum, and as the
load is reduced on the foot the ankle angle opens rapidly.

The shoulder position over the course of the stance phase typically follows a U shaped,
relatively level, trajectory while each foot is in contact with the ground (cf. Figure 1-11) [18].

A profile of this type is common in most legged systems during running gaits, including those
of humans [44]. A typical approach used to model this type of response is the spring loaded
inverted pendulum (SLIP) model, in which a point mass rests atop a massless springy leg
[41].

LH
LF

RH RF

Figure 1-11: The position of the shoulder (solid) and hip (dashed) throughout a single stride of the cheetah
executing a rotary gallop. Trajectories are drawn to scale above the ground at y = 0 based on data from [18].

After a limb has completed the ground contact phase it must be repositioned to prepare
for the next ground contact at a distance A from the present foot location. The forward
movement of a limb (protraction)is often followed by a short dwell time and a rearward move-
ment of the foot, just prior to ground contact (retraction). The joint angle trajectories during
protraction show that it is common for the leg to be collapsed prior to rotation of the hip
or shoulder [9]. This decreases the inertia of the leg and allows the entire protraction event
to be completed more quickly. The degree of collapse depends on the relative masses and
inertias of the leg links, and generally differs from species to species and animal to animal
[14]. Dynamic simulations, presented in later chapters, show that the aggressive forward
swing of the leg may be stopped with relatively little effort on the part of the actuators by
allowing the natural dynamics of the limb to extend the elbow, and thereby increase the
inertia and slow the shoulder.
Because the limbs of many animals cannot be considered massless in comparison to the
body one must also consider the forces and torques produced by protraction and retraction
and their influence on the dynamics of the body. Clearly, unless the limb is extremely massive

(as it is in an elephant for example) these forces will be small in comparison to the forces
encountered during stance. Nonetheless, a well timed protraction may serve to neutralize
any pitch disturbances before they cause large changes in pitch angle.

1.2.4 Force Generation

During the support phase of a gallop, forces generated between the foot and ground are
transmitted to the body by the limb (cf. Figure 1-12). The contact event begins with a de-
celerating impact that generates forces opposing both the vertical (falling) and horizontal

(running) velocity of the body [62]. After this impact has settled, and the leg has reached
a suitable configuration, the limb begins to provide propulsive forces, typically near mid-
stance, where the shoulder is directly above the foot. This propulsive force must be of suf-
ficient magnitude to overcome the deceleration imparted by the earlier impact and any
drag forces in order to maintain steady state. Additionally, the vertical forces must provide
sufficient impulse to counteract the pull of gravity over the entire stride. As the shoulder
continues to move forward both normal and propulsive forces decrease smoothly, with the
final propulsive effort being exerted by rotation onto the phalanges and extension of the
foot.
Formulation of the required locomotive forces in an impulse framework provides some

insight into the scaling of contact loads with varying stride parameters. Over the entire
stride, gravity imparts a downward pull with a total impulse of -mg/f, where m is the total
mass, g is the gravitational acceleration and f is the stride frequency. This impulse must be

balanced by the vertical contact forces, which are exerted over B/f seconds. This implies
that the average contact force is
mg
(1B
 2.7-
2.4- Lead forelimb Lead hindlimb
218--
21 8- -.....-.

-~1.5-

Q) 1.2-

-e0.6-
0
0.3
Q) 0_-_
-0.3 -
0 -0.6 - Vertical
-0.9 Fore-aft
-1.2
Mediolateral
-1.51
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (percent of stance)

Figure 1-12: Ground reaction forces of a dog executing a rotary gallop. Adapted with permission from The
Journal of ExperimentalBiology [62].

In the rotary gallop, B = E #. Thus as the combined duty factor is decreased, average
vertical loads at the foot increase. If only aerodynamic drag is considered, the net propulsive
force may be similarly calculated as

- PV2CDA
FB (1.2)
B

where CD is the drag coefficient, A is the frontal area and p is the density of air.

The first phase of the ground reaction force profile is impact. During this time the foot
strikes the ground with some nonzero velocity. Relative compliances of the ground and leg
may cause the contact forces to oscillate briefly. This phase of the force profile terminates
when the foot firmly comes to rest on ground, which is often somewhat difficult to precisely
determine. The character of the forces generated during this time depend heavily on leg
morphology. The effective mass of the robot as seen from the foot, sometimes referred to
as the unsprung mass, is of particular importance [33], as are the stiffness and damping
characteristics of the footpad [10] and the configuration of the leg [20]. In humans, mod-
ulation of these impact forces by varying touchdown configuration is easy to experience.
Compare the initial impact of landing after a vertical jump with feet flat and legs straight,

to the impact that results when the knees are bent about ten degrees. Clearly not only does
the impact influence the initial forces that are generated, but it also influences the forces
required throughout the remainder of the stride to produce the desired dynamic response
(in this case bringing the body to rest). This likely occurs because altering the impact con-
ditions changes both the configuration and the net impulses imparted at the conclusion of
the impact.

As the previous experiment has shown, the impact of the foot provides a set of exit condi-
tions which then require precise muscle control to bring the stride to a successful conclusion.
This is the second phase of stance, often called support. The support phase is character-
ized by active control of the leg musculature to impart the desired forces on the body using
the foot-ground interaction. During running, the derivative of the fore-aft force generally
changes from negative to positive as the impact ends and shortly afterwards the leg begins to
generate forces in the forward direction at the foot [62]. In leg structures with joint redun-
dancy, such as the digitigrade structure, some discretion is required in distributing torques
amongst the joints in order to ensure that the joints remain near their optimal operating
orientations.

As the stride nears completion the knee joint approaches its maximum extension, and
several changes occur to maintain ground contact and to lengthen the stride. These changes
occur during the end of the support phase. The ankle begins to rotate away from the body
and increase the total length from the point of contact to the hip. This recruits the effort
of the lower limb muscles [4]. Both the vertical and horizontal forces begin to decay, the
vertical often more quickly than the horizontal. Ultimately, the phalanges lose contact with
the ground and rotate caudally as the limb enters its protraction phase.

The front and rear legs of quadrupeds appear to perform somewhat distinct tasks. The
front legs tend to exhibit more stabilizing and decelerating forces while the rear limbs provide
similar or larger propulsive effort with lower vertical loads [62]. Clues to this specialization
can be seen in the relative length of the aerial phases that follow the front and hind stances in
a rotary gallop [18] and in the differences in force profiles [62]. The action of the spine may
also contribute to this asymmetry. Higher vertical forces are experienced by the front legs
than the rear which may be due to the location of the center of gravity (CG). For quadrupeds
the CG is generally forward of the geometrical center between the hips and shoulders [30].
1.3 Background: Dynamic Performance

In searching for insight into the next steps required to advance the development and capa-
bilities of legged robotic platforms perhaps something can be learned from the early history
of motor vehicles. Early horseless carriages (ca. 1895) performed all of the same basic tasks
as modern vehicles, if they are considered under quasi-static operating conditions. The ve-
hicles turn to either side, move forward, stop and provide seating for the occupants. Over

the following thirty years the development of the automobile proceeded at a tremendous
pace, and a close examination of all of the changes made over this period show one overrid-

ing principle. Each change is intended to enhance the transient (or dynamic) performance
of the vehicle [43]. Since the road is not perfectly flat, better suspension systems were de-
veloped, which maintained more uniform loading on the tires, enhancing tractive forces.
Wider pneumatic tires also improved traction, and thus the capacity for acceleration. More
powerful engines make use of the tires' enhanced grip on the road. A wheeled vehicle is
an inherently dynamic device, and therefore the metrics used to measure the impact of a
specific change must also be dynamic in nature.
For autos the engine horsepower is one such metric, as is the frequency response of the
suspension. Both of these metrics include the vital temporal element that provides clues
towards the effect changes in these metrics have under dynamic conditions. Currently, no
such tools exist for the designer of a legged robot. This makes it difficult to develop a clear
understanding of how design changes will impact the rather complicated dynamics of legged
locomotion. My hope is that developing such tools, or merely pointing out the urgent need
for them, will result in better design of dynamic robotic systems. Several such tools are
proposed in Chapter 5.
Chapter 2

Design Principles

The ultimate goal of any leg is to harness interactions with the terrain to efficiently and
robustly generate the locomotive forces demanded by the application. In other words, the
leg must posses the ability to drive the body to a specific dynamic state based on the available
sensor data and control algorithms. The first objective of the analysis presented within
this chapter is to identify characteristics of the physical system that enhance the ability of
the leg to reach the desired dynamic state, particularly in the presence of uncertainties in
sensed or modeled parameters. This is approached through a detailed analysis of the three
distinct phases of a stride (impact, stance and swing) and the unique characteristics that they
each suggest. Once these desirable characteristics are identified, an effort can be made to
expand the range of attainable dynamic states by discerning the factors that limit dynamic
performance. Following the development of these robustness and performance imparting
characteristics, a set of simple and balanced leg design criteria are proposed.

2.1 Operational Phases

The dynamic behavior of an actuated mechanical system can be expressed in many ways.
The approach taken here is to quantify the dynamic characteristics of the leg in terms of
differences in mechanical impedance [45]. The mechanical impedance relates the dynamic be-
havior of the leg endpoints (one at the body and one at the foot) to the forces that develop
between them. Because the leg, under the influence of the leg controller, may be equiva-
lently represented by a virtual mechanical system [19], issues of physical hardware design
and control may be considered concurrently. Following the analysis of the impact, stance
and swing phases, the total impedance is strategically decomposed to yield insight into me-
chanical design considerations, the control system, and how these two vital elements interact

to achieve robust locomotion.

In the general case of a foot contacting the ground, the body will experience three in-

dependent reaction components (two forces and a moment) at the point where it joins the
leg. However, under the simplifying assumption that the foot-terrain interaction supports

negligible moments about the point of contact, only two of these reaction components due

to the contact loading remain independent. The force component along the leg and the
moment at the body are taken as the independent components, leaving the force perpen-
dicular to the leg dependent (cf. Figure 2-1). For the purpose of this qualitative impedance
analysis it is assumed that these two independent components may be decoupled, into two
one dimensional dynamic systems; one representing the extension and contraction of the
leg and another, rotationally analogous system, representing the fore and aft rotation of the
leg about the shoulder.

For each system the total mechanical impedance Zt is composed of two elements, the
passive (physical) impedance Zp, and the active (virtual) impedance Za [31]. The passive
impedance represents the physical reality of the manipulator. It is, in a sense, an expression
of the dynamic character of the leg. The active impedance on the other hand is due purely
to effort applied by the actuators, and it allows the manipulator to mimic the impedance
characteristics of alternative physical systems. However, because the active impedance is
the result of actuator, sensor and control system involvement, it inherits the limitations of
each of these systems. The two most substantial factors to consider in the establishment of
locomotion design principles are control loop time and actuatorsaturation. Each of these factors
limit the degree to which the desired active impedance can be achieved, thereby limiting
the total impedance to

Zt = Zp + 6cZa (2.1)

6c 6c(Za, W), (2.2)

where 6c is the degree of capability, a coefficient with values ranging from zero to one, and o
is the angular frequency of the input motion. As the limits of the controller and actuator are
approached, 6c decreases substantially from an ideal value of one, and the total impedance
begins to inherit the character of the passive impedance alone [31]. Thus, for the very high
frequency regime it is difficult to create the impression that the leg is anything other than
the physical system. This motivates the analysis of the first phase of contact, impact.

Figure 2-1: Decoupling of body reaction forces into two one dimensional systems.

2.1.1 Impact

During each stride, the leg transitions from the swing phase to the stance phase by under-
going an impact of the foot [62, 33]. This impact may be the intentional result of swing
leg retraction [27], or the unintentional result of premature leg actuation, perhaps due to
sensing errors. In either case, it is difficult to precisely match the velocity of the foot to the
relative velocity of the ground during dynamic locomotion, and thus an impact must be
considered unavoidable. Two considerations guide the impedance selection process during
this phase. First, the forces generated during the collision should not be directly transmitted
to the body. This demand stems from the desire to posses full knowledge of the forces ex-
erted on the body at all times. If forces are allowed to be transmitted, then their magnitude
cannot be predicted and predictions of the body state immediately following the impact will
be inaccurate. The second demand is that contact forces, specifically the magnitude of the
impulsive load on the foot, should be small. This is purely a practical consideration, aimed
at reducing the structural demands on the leg hardware.

To determine the impedance that achieves such characteristics consider the simplified
one dimensional impedance model that represents the compression and extension of the leg

(cf. Figure 2-2). During impact, an inherently high frequency event, the velocity of the foot
changes suddenly. Throughout this event the body is assumed to remain, for all practical
purposes, stationary. Applying (2.1) to this physical system and analyzing the high frequency
domain indicates that in order to achieve the design targets the following relation must hold:

lim Z (jW) = 0. (2.3)

w-+oo

In words, the total leg impedance in the high frequency regime should be as low as possi-
ble. Now, following from the discussion of (2.1) and (2.2), the total impedance during high
frequency motion will be predominantly due to the passive impedance of the leg (Zr). This
implies that in order to limit the transmission of impact loads to the body, measures must
be taken in the design of the physical system to reduce as much as practical the passive
leg impedance. Two methods of realizing this goal are reduction of mass, damping and
stiffness of the entire leg or the selective addition of a redundant, low impedance, degree of
freedom (cf. Figure 2-3). Implementation of the second approach is complicated by the high
force transmission demands of stance, which would cause large motions in a low impedance
element, such as a soft spring. A compromised solution is to employ a nonlinear compli-
ance at the foot. Such a nonlinear compliance would initially be soft, requiring little effort
to deflect, and would stiffen as loading on the foot increases, perhaps becoming effectively
rigid as it comes to bear the entire body weight. This approach has the distinct advantage
of selectively reducing the impedance during impact beyond the level that may be possible
for the remainder of the leg. However, the addition of a passive redundant joints generally
complicates control.
 XA nonlinear
stiffening spring

Figure 2-2: Model of leg extension and compres- Figure 2-3: A method of reducing impact
sion during impact. The body is effectively im- impedance is the introduction of a nonlinear
mobile and the foot cxperiences a step change in spring in series with the leg.
velocity

2.1.2 Stance

The stance phase is undeniably the most crucial period of leg operation and therefore must
be considered thoroughly. During this period the leg has access to a medium (the terrain)
that allows it to apply a limited range of forces at the foot. The foot-terrain interaction
of the most consequence to locomotion is the development of forces within the plane of
the ground. Although this interaction is generally quite complex, and an area of ongoing
research [8, 24], the Coulomb Friction model provides sufficient depth to develop a design
strategy for relatively rigid surfaces. The task of the leg is to use this interaction to precisely
control the dynamic state of the body at the end of stance. Two extreme approaches to
achieving this goal are to either attempt to control the trajectory of the body through space
or to attempt to control the forces applied to the body. Both of these methods are special cases
of impedance control [20], and each will produce fundamentally different responses when
subjected to common disturbances. The ideal leg impedance must be well balanced and
capable of mitigating the destabilizing effects of a wide array of disturbance types. However,
it is beyond the scope of this document to identify the specific impedance characteristics
that result in robust locomotion, instead the aim is only to characterize the demands on the
physical system.
 The selection of a suitable impedance relies heavily on a sound mathematical formula-
tion of the fundamental locomotive goals. During steady running the task of the controller
is to direct the evolution of a stride such that the same stride may be repeated indefinitely
to move forward across the terrain. However, the presence of disturbances complicates the
practicability of this ideal behavior. The apex state of the body captures the stride to stride
variations that lie at the center of this issue clearly and concisely. If the controller output for

a nominal stride is used in the presence of disturbances, then the apex state following the
stride will most probably differ from the initial apex state. As these differences aggregate,
the propensity towards failure grows. In this case, the control framework lacks the necessary

flexibility to respond to disturbances in a manner that maintains the elements essential to

steady locomotion. Assuming for a moment that the legs are massless, and therefore that

their precise position and motion has no bearing on the position or orientation of the body
at apex, the values that may differ from stride to stride are:

- dT difference in stride period

- dx difference in body horizontal position

- dy difference in body vertical position

- d# difference in body pitch angle

- di - difference in body horizontal velocity

- d# difference in body pitch rate

Of these the first two errors (dT and dx) vanish if the following stride is undisturbed. The
final four errors, on the other hand, will likely be preserved or amplified, even if the following
stride is undisturbed. This suggests that a reasonable approach to enhancing the robustness
of stance is to direct the effects of disturbances to either changes in stride period or horizontal
position.

The aptitude of the leg in dispensing with each disturbance type lies wholly in the magni-
tude and character of the leg impedance. To identify the influence of impedance character
(spring, damper, etc.) on stride to stride disturbance transmission a simplified two dimen-
sional model is employed. This model is similar to the spring loaded inverted pendulum
(SLIP) model, but it employs a leg of general mechanical impedance. The model consists of
a single massless leg supporting a point mass at its proximal end (cf Figure 2-4). The goal of

such a system being to move forward over the terrain with minimal stride to stride variation
in the apex state. Clearly this model provides no means of directly assessing the transmission
of pitch errors. However, this shortcoming is of little practical significance for quadrupeds
because pitch may be interpreted as vertical position differences between the front and rear
leg pairs. Furthermore the goal of this model is not to develop a precise control algorithm;
simply to understand how impedance relates to the qualitative behavior of the body during
stance. The disturbance types considered are:

- Fdbz- disturbance force on body in x direction

- FdbV- disturbance force on body in y direction

Sid- disturbance in x velocity

- md disturbance body mass

Figure 2-4: Two dimensional, one legged hopping model of locomotion with a leg of general impedance Zt
in series with a prismatic actuator.

A fixed body mass (m = 20kg) and single component impedance (either spring or
damper) is used to achieve the desired apex state. However, several constraints on the stance
forces must be imposed to arrive at a realistic behavior. The first of these is that the foot
forces must start and end at zero and be continuous. This ensures that the leg is not sub-
jected to any impulsive loading during stance. Second, the foot must only exert compressive
forces on the ground, otherwise the leg will lift off, signaling the end of the stance phase.
The forces for the spring and damper legs are given by

Fs = -k(l - le) (2.4)

Fd = - b(i - ), (2.5)

where k and b are the apparent stiffness and damping ratio of the leg, 1 is the leg length
and lc is the commanded actuator length. To achieve the desired response with a spring

leg constant values of le and k may be commanded. However, for a damper leg 1c must be
varied with time for a fixed b. The commanded velocity (le) is chosen to provide the same
force profile as the spring for a nominal stride. The leg touches down with an angle of 30'
to the vertical. The initial apex height is 1 m and the forward velocity is 3 m - s-1 .

Simulating the motion of the system from one apex to the next, the effect of the four dis-
turbance types was analyzed. The spring leg suffered less from disturbances to the horizontal
and vertical force on the body (cf. Figure 2-5 and 2-6). Forces in the negative x direction
tended to result in higher apex states while forces in the negative y direction tended to lower
the apex. When a disturbance mass is introduced to the body the damper leg yields less
variation in apex height (cf. Figure 2-7). This is primarily due to the velocity command of
the damper possessing a fixed frequency. This tends to reduce foot forces late in stance,
which in turn reduces variations in contact duration. On the other hand, increasing body
mass decreases the natural frequency of the spring-mass system, tending to increase the du-
ration of stance. Horizontal velocity disturbances had less of an effect on the spring leg,
this is possibly due to the fact that as forward velocity increases the fixed lc command of
the damper leg becomes less synchronized with the angular orientation of the leg on the
ground (cf. Figure 2-8). These results hint at the complexity of precise impedance selection
and indicate that an ideal impedance may combine the features of various pure impedance
types. Furthermore, the precise magnitude of the impedance must vary with forward speed
in order to meet the changing demands of total impulse and stride duration.

1.4 1.3

1.3 1.2

1.2 .
0 1.1-

0~. .........
09 -

0 7 00.
0.8 /0.7-

0.7 . 0 .

0.5
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
x position of CG (m) x position of CG (m)

Figure 2-5: Effect of horizontal disturbance Figure 2-6: Effect of vertical disturbance force
force (Fdb.) on one leg hopper with spring-like leg (Fdby) on one leg hopper with spring-like leg
impedance (black) and damper-like impedance impedance (black) and damper-like impedance
(red). (red).

This analysis identifies two desirable features of the mechanical system. First, because
the leg impedance is expected to change throughout the speed range, the majority of the
impedance during stance must be due to the contribution of the actuator in the form of Za.
Second, to maximize the range of impedances that may be achieved the actuator must be
made powerful.

2.1.3 Swing

Forces during the swing phase, unlike impact and stance, are governed purely by the inertial
characteristics of the mechanical system. Depending on the application, these swing phase
reaction forces may provide a desirable degree of control over body behavior, particularly in
the pitch control of bipeds [29]. However, large swing forces also tend to slow protraction

and increase the dynamic coupling between the body and legs. The lack of contact during
this period suggests that there is no specific impedance goal to speak of, simply the notion
that the inertial characteristics, together with a protraction trajectory, result in the generation

of a restricted range of reaction forces. Consequently, the relative benefits and drawbacks of
altering the mass and inertia of the leg must be thoroughly explored.
 1.3 1.4
1.2 - - 1.3

1.1 - 1.2

U0. U ,;
0.9 / . 0... .. . . .t . . .

0.7 0.8 E)

0.6 - 0.7--

0.5-
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
x position of CG (m) x position of CG (m)

Figure 2-7: Effect of mass disturbance (md) on Figure 2-8: Effect of horizontal velocity distur-
one leg hopper with spring-like leg impedance bance (.d) on one leg hopper with spring-like leg
(black) and damper-like impedance (red). impedance (black) and damper-like impedance
(red).

If the leg inertia is increased it becomes possible to apply larger forces to the body dur-
ing the swing phase. This may help to correct small errors in body orientation during flight.
However, the nature of the forces that may be applied is not arbitrary, and to a large extent

depends on the mechanical configuration of the leg [60, 46]. The initial and terminal leg
locations during the swing phase are generally fixed, requiring the leg to be moved from a
posterior to an anterior position. This leaves only the temporal evolution of the protraction
free to be altered, severely restricting the character of the body motions that may be gener-
ated. The mechanism behind the ability to alter body orientation during swing is the devel-
opment of stronger dynamic coupling between the leg and the body. From the perspective
of controller design such coupling is undesirable, and a source of substantial complication
because the controller is forced to consider both the leg and the body as a single mechanical
entity. On the other hand, decreasing the inertia will result in consistently lower body forces
and a reduction in the body to leg dynamic coupling, facilitating the separation of leg and
body dynamics.

In the case of quadrupeds the need for swing phase reaction forces is uncertain. The
availability of leg pairs fore and aft of the center of gravity allow pitch adjustments to be made
by altering the relative magnitudes of stance forces. Furthermore, the inherent complication
such coupling introduces into the control system diminishes its attractiveness, particularly
because the dynamic coupling will continue to exist during stance, forcing the actuators
to support larger dynamic loads. Given the intrinsic complexity of developing a dynamic
robot the appeal of decoupling the leg from the body is too great to merit sacrificing for the
relatively limited control authority available during protraction, particularly when an alter-
native method (such as a tail) may be introduced instead. Dynamic decoupling is a feature
that will come to lie at the center of the overall control strategy, and thus it is considered
essential to reduce the leg mass and inertia to the greatest possible extent.

mieg < mbody (2.6)

Iieg < Ibody (2.7)

2.2 Performance Limits

As the locomotive goal of this robot is to achieve the highest possible speed, particular atten-
tion must be payed to the factors that define the upper limits of performance. In biological
systems the force-velocity and force-length characteristics of the muscle groups employed
during gait execution form the core of the system that tends to limit running speed [25].
These characteristics are the biological equivalent of the actuator dynamics that limit the re-
sponse of robotic systems. The shape of the force-velocity and force-length curves for skeletal
muscle have a tremendous impact on gait, leg structure (kinematics) and muscle distribution
[65]. Together these components produce the performance that is recognized in the natural
world, but independently they hold little value. Their strength comes not from any single
element, but from the system as a whole. The challenge of achieving high performance in
legged robotics is to develop a unified system around a fundamentally different actuator.

Leg kinematics and the accompanying musculature work in concert to generate the de-
sired locomotive forces. The leg structure guides muscle distribution and the limits of mus-
cles guide leg structure. This interplay between actuation, kinematics and the dynamics
that they ultimately produce, must be realized for effective high-speed locomotion. If this
harmony is not achieved, an individual element of the system, instead of the system as a
whole, would define the performance limit. The element that most commonly shoulders
the burden of an unbalanced design is the actuator.

The principle consequence of limits on actuation forces is a reduction in the ability to

provide the propulsive forces required to maintain speed. At the maximum speed the net
propulsive impulse over a stride is equivalent to the impulse due to drag. Drag is often made

up of several components, such as joint friction, soil work and aerodynamic forces, which
will vary in magnitude depending on, among other factors, the speed and the gait. At high
speeds, aerodynamic drag may become a dominant component, along with the leg-motive

friction that results from the movement of the motors, bearings and joints. Although it is
not possible to completely eliminate either of these factors, they may be reduced to increase

maximum running speed.

Another form of friction has enormous implications on the performance limit: the fric-
tional interaction between the foot and the ground. Foot-terrain friction is the medium
through which all propulsive forces are exerted, and as such, performance is maximized
when the limits on friction forces are high. To attain the large friction forces required for
fast running it is often necessary to exert large vertical forces (normal loading). This re-
quirement places additional demand on the actuators. A high friction coefficient between
the foot and the ground can ease this type of secondary loading, but even if the Coulomb
friction coefficient (p) is one a unit increase in horizontal force requires that the total force
be increased by v/2. One method of circumventing this troublesome interaction is the use
of claws to mechanically grip the surface.

The weight of any dynamic system is of the utmost importance. Whether it is an auto-
mobile, aircraft or a piece of sporting equipment the overwhelming trend is that when high
dynamic performance is desired, weight must be kept to a minimum while power is pushed
to the maximum [43, 38, 15]. The cause of this trend is as simple as Newton's second law.
If it is possible to reduce the weight of a system without compromising its ability to generate
forces, then the system will be capable of achieving higher accelerations or better "dynamic
performance". The practical difficulty in applying this rationale stems from the fact that it is
not always possible to reduce the mass and maintain the same force generating capability. If
mass as well as force are reduced simultaneously, at what ratio does the performance begin

to decrease, and at what force level is the performance compromised?

 In the automotive industry the ratio of force generation to mass is typically supplied as
the power-to-weight ratio, and if dynamic performance is to be increased, this value should
be made as large as possible. However, in order to determine the optimal power level, and
thereby the precise weight, additional information on the operating conditions is required.
Assume for example that the design intent is to drive as fast as possible in a straight line.

Taking two identical cars with the same power-to-weight ratio, but one with ten times the
power and weight and accelerating them to their maximum speeds, several key differences
become apparent. Initially, the cars accelerate at a similar rate. However, the less powerful
car, slowed by aerodynamic drag, attains a lower maximum speed than the more powerful

car. This difference occurs because at the performance limit the power of the car is not
absorbed by the acceleration of its mass but by aerodynamic drag and rolling resistance.
If the same cars are compared in a dynamic performance test, such as maximum speed
through a slalom, often the lighter car will perform better due to nonlinearities in tire and
suspension behavior [43].

In a legged system the maximum speed behavior is a combination of these two extremes.
The actuator forces must counter any aerodynamic drag and produce fore and aft acceler-
ations of the legs. Thus not only must actuator power be high, the leg masses and inertias
must be very low, in order to limit the effort required to produce their reciprocation. This
ensures that work done throughout a stride goes into producing useful forces at the shoulder
as opposed to accelerating and decelerating the leg.

The limb and body must also adhere to strict structural limits. In biology the cost of
exceeding the loading limits on bone, muscle or tissue is injury or death. In robotic systems,
exceeding the structural limits will result in yielding or catastrophic failure. In either case
these are situations to be avoided. Robotic systems have a material advantage in their ability
to implement features of various chemical composition such as metals, plastics and compos-
ites. However, nature has a structural advantage in its ability to shape a limited selection of
materials with nano-scale precision [3]. The union of material and structure must be tai-
lored to perform under the anticipated loading conditions. Stated differently, if the material
is not bone, then it need not retain the physical structure of a bone to retain its function.
This structural constraint often defines the lowest possible degree of weight reduction.
2.3 Design Criteria

The analysis of the previous two sections highlights several of the key issues that limit the
effectiveness of a leg design. The reduction of this information into a set of simple design

criteria aids its introduction to the design process. The following four criteria are proposed,
with the intention of instilling a leg design with a balance of robustness and performance.

To achieve this the leg must exhibit:

- Low passive mechanical impedance

- Low mass and inertia

- Strong actuation

- Harmony between kinematics and use

The necessity for low passive mechanical impedance originates from the demands of
impact and stance. During impact, a large impedance amplifies small errors in the foot
velocity with respect to the ground into large body forces. These forces alter the state of
the body unpredictably, which makes the pre-impact planning of a stride impractical. Due
to the dynamic complexity of a running system it is computationally difficult to re-plan the
stride post-impact. This difficulty is avoided if the passive impedance is decreased, insulating
the body from rapid motions of the foot. Although the impedance must be increased during
stance this may be accomplished actively owing to the substantially longer duration of stance
relative to impact.
Mass and inertia of the leg affect the behavior of the system throughout all areas of leg
operation. But their influence is most clearly visible during protraction and in defining the
limits of dynamic performance. During the swing phase the leg is dynamically coupled to
the body and the degree to which motions of the leg influence motions of the body increases
with increasing mass and inertia. However, unless swing phase reaction forces are abso-
lutely essential to the application, these coupling reactions serve only to complicate control.
Furthermore, reduction of the leg mass and inertia increases the maximum acceleration and

deceleration of the leg for fixed maximum actuator loads.

 Strong actuation increases the range of realizable leg impedances and provides ample
effort to counter inescapable losses such as aerodynamic drag. This type of loss, along with
the losses inherent in the active acceleration and deceleration of the leg, ultimately come to
limit the maximum running speed. However, strong actuation alone is not an indicator of
success. The actuator merely sets the foundation for the remainder of the system to build
upon.
Balance between the form of the leg and its use ensures that every element of the system
operates to its full potential. The kinematics, dynamics, actuator and gait are heavily inter-
dependent. No individual element stands on its own. Thus when one component changes,
the remaining elements must be reevaluated and refined iteratively. Direct imitation of
specific aspects of a biological system in robotics disregards the fundamental differences be-
tween the two. Embracing these differences allows a design solution to develop organically
instead of forcing elements together that were not intended to fit.
46
Chapter 3

Leg Design

This chapter begins by delineating the design space, specifying particular targets for weight,
size and speed. Next, a conceptual framework is developed for the leg design, establishing
the ideas and design direction that enable the design goals to be met. After this framework
is put in place the detailed mechanical design is presented for both front and rear legs.

3.1 Design Space

The intent of this research project is to develop a quadrupedal robot capable of true dynamic
running in three dimensions with a maximum speed greater than 13 m - s- or 50 km - h-1.
This speed is well above the highest observed speed for legged robots (6 m - s- 1)which was
established in 1984 by Raibert et al. [51]. The robot is to complete this task standalone,
without tethers, electrical or hydraulic assistance over nominally flat, natural terrain, such
as a grass or dirt field. The ability to maneuver within the human environment, in a manner
similar to a dog or large cat, stands as a long term goal that imposes specific constraints on
size, shape and weight.

As a starting point, a fully extended leg length of approximately 50 cm and total weight
of 35 kg was defined. These values, particularly the leg length, are design parameters that
are flexible and are intended merely to guide the design of the robot to an appropriate size.
The aggressively minimized weight budget of 35 kg permits 3.5 kg for each front leg and
3.75 kg for each of the rear legs (cf. Table 3.1), these values include both actuation and the
 Component Weight (kg)
Front leg (ea.) 3.50
Rear leg (ea.) 3.75
Ad/abduction 1.25
Body structure 5.25
Computation 3.50
Batteries 10.50
Total 35

Table 3.1: Weight budget for the quadrupedal robot.

mechanical structure. The loose confines of the design space allow many possible directions
to be explored, each with its own set of advantages and disadvantages. The section that
follows details the identification a suitable design direction and set of design ideas that form
the basis for the detailed final design.

3.2 Conceptual Framework

The goals of this project can undoubtedly be approached in many different ways. Among
the highest level decisions to be made is the selection of a suitable leg topology. A leg topology
is the physical arrangement of the degrees of freedom that make locomotion possible. This
definition of leg kinematics has a powerful effect on gait generation, force production, and
how force production at the foot and body vary with increasing speed. Three leg topolo-
gies suitable for dynamic running are the prismatic, articulated and redundantly articulated
configurations (cf. Figure 3-1).

Possibly the simplest workable leg design combines a single prismatic leg link with a ro-
tational joint at the hip (cf. Figure 3-2). This topology was applied with impressive results

by Raibert et al. in the 1980s, and a robot of this type still holds the speed record for legged
robots. Several leg lengthening actuator options are available to the designer of such a sys-
tem. Specifically hydraulic actuators, linear motors, rack and pinion and lead-screw initially
appear suitable for the task. In comparison to the forces required to lengthen the legs, the
torques required at the body pivot are rather small [33].

This type of leg exhibits several extremely convenient dynamic characteristics that sim-
 prismatic articulated redundantly
articulated

Figure 3-1: Prismatic, articulated and redun- Figure 3-2: The prismatic leg consists of a revo-
dantly articulated leg topologies. Revolute joints lute joint at the body and a prismatic joint within
are indicated by circles and prismatic joints by the leg.
rollers.

plify controller implementation. During stance, if the hip joint is left free and the leg assumed
to be massless the ground reaction force vector points along the leg from the foot to the hip
and has a magnitude that is directly controlled by the leg length actuator force. In this type
of gait the leg is capable of producing propulsive forces after reaching midstance. The sim-
plicity of this decomposition, in which only hip angle and actuator force determine the force
exerted on the body, is very attractive from the perspective of control.

The use of a linear hydraulic actuator for the generation of leg lengthening forces was
also pioneered by Raibert [51]. This actuator selection has several practical advantages.
First, the actuator is relatively lightweight and high force, particularly if the hydraulic com-
pressor and accumulator are not located on the robot. Furthermore, the mass that exists
in the leg functions as both the actuator and leg structure. When the passive impedance

of hydraulic actuator is modulated by the addition of a serially located pneumatic spring,

the high speed and high thrust characteristics of hydraulics allow the leg to support a quick,

powerful and stable stance phase. If the pneumatic spring were omitted from this design, the
relatively small passive compliance of the hydraulic actuator would result in a large passive
impedance. Because of this, impact forces would be very high unless the velocity of the foot,
relative to the ground, could be precisely controlled. These impact forces and the resulting
fluctuations in body forces, would likely make the robot difficult to control. The combina-
tion of speed and force afforded by hydraulic actuation is difficult to obtain at similar mass
levels for alternative prismatic actuators.

A principle disadvantage of implementing a topology of this sort, utilizing hydraulic

actuators, is the fundamental inability of the actuator to cope with increasing leg exten-
sion speeds and the larger forces that accompany them. The force production capabilities

of hydraulic actuation are very good, particularly at high loads and intermediate speeds.
However, as speed increases fluid losses increase. This imposes a relatively strict limit on
maximum running speed which, although high, may be challenged by a leg topology that
offers greater actuation flexibility over the speed range. Another disadvantage of hydraulic

actuation is the difficulty of executing precise control, particularly when high reservoir pres-
sures are used at low speeds [23]. In addition, the actuation of more than one leg may
unpredictably change the forces generated by a given control input if the legs use a common
hydraulic accumulator.

An alternative to the strictly prismatic leg is the articulated leg. That is, a leg with a
single rotational joint used to control leg length, similar in function to the knee or elbow of a
quadruped (cf. Figure 3-3). This topology alters the actuator requirements from a combina-
tion of rotary and prismatic elements to two rotary elements, yet retains the same two planar
degrees of freedom as the prismatic leg. If the shoulder is left free this limb produces forces
in precisely the same manner as the prismatic leg. However, use of the rotational knee joint
as a leg lengthening mechanism has some key implications for the force generation capabil-
ity of the actuator. For a constant actuator torque at the knee the force generated between
the shoulder and foot depends on the knee moment arm (cf. Figure 3-4). This configuration
dependent gearing of the leg lengthening actuator is a key feature of any articulated leg de-
sign. Clearly, configuration gearing complicates the calculation of body forces because the
force now also depends on the knee angle. If the leg links are given finite mass, the hip and
knee degrees of freedom will interact non-trivially according to the passive dynamics of a
double pendulum. These dynamics complicate control strategies and somewhat cloud the
formation of an intuitive understanding of high speed locomotion.

The advantages of an articulated topology over the simpler prismatic layout lie within
the configuration dependent gearing and increased workspace. If the limb is nearly straight,
Figure 3-3: The articulated leg consists of two Figure 3-4: The force between the body and foot
revolute joints, one at the body and another be- for a fixed elbow torque depends on the amount
tween the body and foot. that the elbow is bent.

a relatively weak actuator can produce substantial extension forces. Under heavy loading
it is possible to alter the gait in order to operate the leg near this singularity. Thus vari-
ations on a fundamental gait may be employed to alter the region in which the actuator
operates. In humans, a common example of such a change is the adoption of a straight-leg
gait with shorter steps while carrying heavy loads [39]. As the leg is bent further, the maxi-
mum extension force decreases, but extension or compression speed is increased. Thus the
configuration dependent gearing aids in the production of both high load and high speed
behavior, even if such characteristics are not inherent to the actuator.

An extension of the articulated topology is the redundantly articulated leg, a leg pos-
sessing more than one degree of freedom between the contacting foot link and the body
(cf. Figure 3-5). Often redundantly articulated legs assume topologies that are geometrically
similar to the digitigrade or unguligrade legs observed in biology. A digitigrade structure
possesses two intermediate rotational joints, while an unguligrade possesses three. The draw
of introducing additional degrees of freedom lies in developing the ability to augment the
configuration dependent gearing, and thereby increase the performance of the leg as a sys-
tem. For a given distance between the foot and the shoulder many leg configurations are
possible, unlike the case a simple articulated leg (cf Figure 3-6). This additional freedom can
be used to design a stride trajectory that places the leg in configurations that dynamically
match the force and speed requirements at each instant in time. The additional joints also
allow the passive impedance to be altered during impact and throughout the stride. Unfor-
tunately this increase in leg ability comes at the expense of additional actuators. And increas-
ing number of actuators substantially increases controller complexity, sensing requirements
and structural demands, particularly if actuators are located far from the body. The addi-
tion of actuators make it difficult to directly compare the performance of a redundant leg to
either of the simple legs.

Figure 3-5: The redundantly articulated leg con- Figure 3-6: Three possible joint configurations
sists of three or more revolute joints. with the same leg length (from shoulder to foot).
Freedom to modify configuration may aid in
maintaining joints near optimal operating condi-
tions or alter the passive impedance.

Of these topologies the articulated leg was selected, as it appears to present the great-
est balance between complexity and ability. The use of prismatic legs, particularly those
actuated by hydraulics, showed relatively few prospects for the substantial performance im-
provement required to meet the design goals. An articulated leg on the other hand has a
dynamic and kinematic richness that, if harnessed effectively, may provide just that. The
two actuation technologies considered for this task were hydraulic actuators and electric
motors. The high force and high speed capabilities of a hydraulic system initially appeared
attractive, however the development a lightweight self-contained system, complete with ac-
cumulator, hydraulic fluid, compressor and compressor fuel (batteries or liquid fuel) was
anticipated to present substantial difficulties. This factor, combined with the relative diffi-
culty of controlling actuator impedance motivate the use of electric motors over hydraulic
actuators. Thus the overall leg topology consists of an articulated leg driven by battery pow-
ered electric motors. A compliant, redundant foot link is used to reduce passive impedance
during impact (cf. Figure 3-7).

Figure 3-7: The leg topology selected is a articulated leg with one passive redundant joint that is intended to
reduce impact impedance.

In pursuit of attaining the highest possible performance from a given motor one of three
common failure modes must be approached [61]. The first, and most belabored of these
is electrical failure, caused by the breakdown of the winding insulation. In low life motors
this is typically attributed to exceeding the thermal limits of the winding insulation. As
time over the thermal limit is increased, motor life decreases dramatically [7]. Thus to
preserve the integrity of the motor, limits on the core temperature of the windings must
be enforced. Next is the issue of magnetic failure. A magnetic failure is only possible in

permanent magnet motors and will occur when the magnetic field induced by the stator is
strong enough to demagnetize the permanent magnets. This generally occurs when loads
and/or magnet temperatures are high [28]. The final failure mode is mechanical failure of
the motor structure. Many application dependent factors may cause a mechanical failure,
however, bearing failure is quite common, particularly when bearings are worn or the load
rating is exceeded.
In the majority of DC motor applications input voltages are commanded to lie on or
under a constant voltage curve [35]. The slope of this curve is the motor torque constant,
which is generally assumed to be invariant of voltage. Such an implementation is safe, but
leaves a substantial amount of the actuator's true capabilities unutilized, particularly the
short duration, high-power pulses that make up much of the loading in running systems. A
redefinition of the motor limits to reflect the thermal, magnetic and structural failure modes
tends to impose strict limits only on motor current. Thus, removing the artificial limit on

voltage effectively increases the potential operating range to include much larger unloaded

speeds. In the high load regime the thermal limit of a given motor can be visualized by
the construction of a plot of time required to reach the thermal limit (from a reference

temperature) versus motor current commonly referred to as a thermal limit curve. The
torques that these currents induce are available at any speed, less friction or drag losses,
provided sufficient voltage is available from the supply. Logistically, use of the thermal limit

approach allows the designer to balance the proportions of relatively light batteries with
often heavy and large actuators.

Under short duration, high torque, loading the majority of the thermal energy dissi-
pated goes into increasing the temperature of the motor windings. This implies that one
method of increasing the short duration performance of the motor is to increase its thermal
mass. Equivalently, if provided with two seemingly identical motors one would expect the
motor with higher thermal mass to perform better under impulsive loading conditions. The
thermal mass of a given motor, which is largely due to the volume of the windings, can be
increased by ensuring good thermal contact between the windings, stator and housing, or
by placing heat sinks directly onto the exposed motor windings. Active methods of tempera-
ture management, such as water cooling or evaporative cooling, may also increase the high
load performance at the expense of additional mass and complexity.

The functional, yet slightly nafve, approach to integrating actuation and leg kinematics
is to place actuators at each joint, such that the actuator and joint axes coincide (cf. Figure 3-
8). This approach conveniently compartmentalizes leg design and has been shown to work
admirably under low speed, statically stable, conditions [26, 17]. However, for dynamic
locomotion this compartmentalization presents several difficulties. Primary among these is
that the placement of actuators far from the body has a tremendous impact on leg inertia,
increasing it to the point that maximum swing frequency may become severely limited.
A compartmentalized leg also goes against the ideal of developing the leg as a coherent

system, capable of reacting sensibly to external influences. If the leg is heavily segregated,
with unrelated control loops driving each joint, it becomes difficult to ensure that the system-
wide goal of body state control is not lost.
To address these issues, a new type of leg design was developed in which actuators for
the shoulder and elbow both lie concentric with the shoulder joint (cf. Figure 3-9). This ar-

rangement has the unique advantages of reducing the contribution of the elbow actuator to
leg inertia about the hip and modularizing the actuation unit. Actuator torque is transmitted

to the distal joint by a rigid tendon which is placed such that the presence of contact loads
at the foot put the tendon in tension. The precise leg geometry, based in part on dynamic
simulation results and kinematic constraints, is left flexible by actuator collocation because

the leg members themselves retain a minimal number of precision interfaces. This simplic-
ity of interface allows many different leg geometries and configurations to be explored and
greatly simplifies system maintenance.

motors --- 2x motors

Figure 3-8: Simple leg design with the actuators Figure 3-9: Doubly concentric actuator layout
at the joints. High shoulder inertia due to remote reduces shoulder inertia. Elbow is actuated by a
actuator mass limits swing speeds and increases four bar linkage with one bar being the humerus
shoulder actuator loads. and another an extension of the radius behind the
elbow.

Throughout much of the stride the leg links are under a combined state of compression
and bending (cf. Figure 3-10). In order to reduce the structural demands on the leg links,

and thereby open up the possibility of reducing leg weight, supplementary tendons are intro-
duced. These tendons are intended to take a portion of the bending load, thereby reducing
the moment loading on the leg link, in a manner similar to the guy lines of an articulated
crane (cf. Figure 3-11). Such a design enables the use of a wide range of leg link materials,
particularly plastics and composites.
 Ir

1lVIA
static tendon 1

Figure 3-10: Left is a typical loading situation Figure 3-11: Bending loads in the radius are
on a front leg during mid stance. This results in reduced by adding a lightweight Kevlar tendon
loading on the radius shown to the right. from the foot to the tricep.

The final consideration in developing the ideas behind a detailed design is the need for
abduction and adduction (henceforth referred to as ad/ab). This is the rotation of the leg
out of the sagittal plane either toward or away from the body. This degree of freedom is
accommodated by the inclusion of an additional rotational joint at the shoulder. The joint
is parallel to the spine (perpendicular to the shoulder joint) and located on the medial side
of the actuator unit.

3.3 Detailed Design

The following section develops the conceptual framework into tangible front and rear robotic
legs. The unique features specifically discussed are the:

- Concentric actuator layout

- Use of motors to thermal limit

- Use of composites in leg fabrication

- Use of tendons to reduce structural loads

3.3.1 Front Leg

The design of the front leg falls within the weight budget and is rated to transmit peak

compressive or extensive forces of up to 1000 N, equivalent to roughly three times the body
weight (cf. Table 3.2). It has two active joints, the shoulder and the elbow, and a passive
ankle (cf. Figure 3-12 and 3-13)

Figure 3-12: Front view of right front leg. Figure 3-13: Rear view of right front leg.

Specification Value
Max. load 1000 N
Shoulder torque (peak) 23 N- m
Knee torque (peak) 134 N - m
Shoulder range of motion 150'
Knee range of motion 1070
Max. leg length 480 mm
Min. leg length 238 mm
Weight 3.25 kg

Table 3.2: Front leg design specifications

Actuator Selection

The selection of a suitable electric motor begins with the unique demands of locomotion.
Analysis of biological data and dynamic simulation highlights several trends which enable
qualitative comparison of the shoulder and elbow actuation demands. In running gaits, in-
cluding the rotary gallop, the ground reaction force vector on the front feet generally points
from the foot to the shoulder, rotating slightly fore and aft of the shoulder throughout the
stride [62]. This suggests that for a lightweight leg, shoulder torque is relatively small in

comparison to the force along the leg (from foot to shoulder). Elbow torque may be esti-
mated by measuring the moment arm of the ground reaction force [5]. This moment arm
is generally several times larger than the moment arm of the contact load at the shoulder.
This trend implies that a stronger elbow actuator may be placed in series after a weaker
shoulder actuator without jeopardizing the dynamic performance. Although the actuators
are in series mechanically, they are effectively decoupled by joint placement and the dynam-
ics of locomotion. Using these approximations the desired elbow joint torque of 120 N - m
and shoulder torque of 20 N - m were established.

A review of available actuators indicated that these peak torques lied beyond the reach
of lightweight direct drive motors and thus some gear reduction (of ratio N) was required
for the elbow. Rotor inertia, when driven from the output through a gearbox scales with
N 2 , thus in order to limit the contribution of the motor to the passive leg impedance the
lowest possible gear ratio was employed. A low gear reduction also promotes the use of
the passive system dynamics during locomotion tasks. The direct drive torque motor type
suits the demands of locomotion because it provides large peak torques and possesses a large
thermal mass. The large thermal mass increases the high load, low duration performance.

The two major classes of permanent magnet DC electric motors considered were brushed
and brushless. A brushed motor uses several graphite brushes pressed into contact with a
commutator. This commutator rotates with the rotor and directs currents to the proper
windings based on rotor orientation. The principle advantage of this motor type is its sim-
plicity of operation. A constant voltage applied to the brushes will result in rotor rotation
and varying this voltage will alter rotation speed. The downside of this simple interface
is twofold. First, the brushes wear, are susceptible to arcing at high currents and increase
friction. Second, less control over the motor torque is available because the timing advance
or switching point of the windings is fixed. A brushless motor, as the name implies, has no
brushes and instead exposes the connections to each stator winding. These wires must be
driven in the correct manner by an external motor driver. The driver must sense the rotor
position and apply the input current to the appropriate windings. The advantage of this ar-
rangement is flexibility and reliability. Because there are no brushes to wear, maintenance is
reduced, and the driver provides precise control over the distribution and timing of winding

currents, enabling precise torque control. For this application, the precision and flexibility
of a brushless motor design was preferred.

The weight of a motor was also a key consideration in actuator selection. The peak
torque density or specific peak torque is a measure of the maximum available torque per
unit mass of the actuator, it is typically highest for permanent magnet motors [13]. This
quantity is important in legged robot design, because the actuators typically compose up
to forty percent of robot mass [16]. Thus a 10 percent increase in actuator mass will likely
result in robot that is more than 20 percent heavier. This additional mass must be matched
by additional force production capabilities. Often it is not possible to achieve the additional
force required for equivalent performance, therefore it is of the utmost importance to select
the lightest actuators capable of generating the design torque. The scaling of torque density
with typical motor parameters, such as diameter, is somewhat complicated due to the wide
range of motor topologies and construction methods. However, large gap diameter motors
generally provide more torque with a similar torque density to smaller motors, the cost that
accompanies this increase in diameter is an increase in rotor inertia.

In pursuing the maximum torque density every additional component of the actuator
design must be justified. A component that is often overlooked is the motor housing. Direct
drive motors are commonly sold in framed or frameless configurations. A framed motor is
an enclosed system with bearings between the rotor and stator while a frameless motor is
provided as simply the bare rotor and stator. Designing around a frameless motor allows
the motor housing to act as a stressed structural component. This is a design method that
is paralleled in the world of motor sport, in which it is common to mount suspension com-
ponents directly to the engine block or gearbox [58]. This reduces total mass and often
increases strength.

Combining all of these diverse requirements a custom wound version of the Allied Mo-
tion HT05001 frameless motor was selected (cf. Figure 3-14). This motor has a suitable
 Property Value
Peak torque (demagnetization) 23.2 N -m
Peak current (at peak torque) 69.5 A
Torque sensitivity 0.334N - m- A-
Terminal resistance 0.791 0
Number of poles 12
Number of teeth 45
Phase connection wye
Stator outside diameter 127.0 mm
Length over coil 40.3 mm
Rotor inside diameter 63.5 mm
Weight (rotor + stator) 1.23 kg
2
Rotor inertia 2.2 x 104 kg -m

Table 3.3: Motor parameters for custom wound HT0500 1 torque motor.

size, shape, weight, and with custom winding is capable of peak torques larger than 23 N - m
(cf. Table 3.3). This peak torque makes the design goal of 120 N - m possible with a gear
reduction of less than 6:1.

Figure 3-14: Allied Motion HT05001 torque motor, peak torque: 23.2 N - m, shown here with components
of the front leg.
Shoulder Unit

The shoulder unit forms the enclosure for the actuators, gearboxes and encoders of each leg
and provides a structural mounting interface to the body (cf. Figure 3-15 and 3-16). The unit

rests atop the humerus and actuates the radius via a steel linkage (the tricep). Functionally,
the shoulder may be decomposed into two sections, the first of these resides on the medial
half of the enclosure and actuates the shoulder joint (cf. Figure 3-17). This is accomplished
by connecting the rotor of the shoulder motor to both the humerus and the stator of the
elbow motor. The second section lies on the lateral half of the unit and actuates the tricep,
which is attached to the output of the gearbox (cf. Figure 3-18). Because the stator of the
elbow motor is driven with the motion of the shoulder, immobilization of this actuator results
in no rotation of the elbow joint.

Figure 3-15: The shoulder unit of the front leg Figure 3-16: Partial section view of the shoulder
encloses both actuators, gearboxes and sensors. unit of the front leg viewed from behind, the steel
tendon actuates the elbow joint.

The elbow motor stage is accompanied by a planetary gear reduction. A planetary

gearbox, incorporated into the housing, was selected for it's favorable strength, weight, size
and shape. The gearbox has a minimal impact on the overall dimensions of the shoulder
unit, increasing the width by approximately 1 cm. The gears, owing to their large teeth, have
a high maximum load and the planetary structure allows the motor to be driven from the
output with only a small increase in passive impedance. Some gear reduction technologies,
Figure 3-17: The shoulder motor rotates both Figure 3-18: The elbow motor actuates the tri-
the humerus and the stator of the elbow motor. cep, a steel linkage that is connected to a protru-
Components that move when the shoulder motor sion of the radius behind the elbow joint. Compo-
is actuated are highlighted in blue. nents that move when the elbow motor is actuated
are highlighted in blue.

particularly harmonic drive, do not have this capability and thus are unsuitable for the low
passive impedance demands of this project.

Gear sizing was approached from two perspectives, first the gears must be able to with-
stand the rather extreme loading of 120 N - m without excessive backlash and second, they
must be suitably sized to minimize the total volume of the shoulder unit. Based on these
considerations a gearbox with a tooth pitch of 48 teeth-per-inch (TPI)and ratio of 5.8:1 was
designed (cf. Figure 3-19). The sun gear functions as the gearbox input with 40 teeth and
is orbited by four planet gears each with 76 teeth. These planets mesh with a fixed internal
ring gear with 192 teeth. The dimensions of the gearbox, exclusive of the planet carrier are

120.6 mm outside diameter by 6.6 mm axial length. Gear weight was substantially reduced
by the removal of excess material (cf. Figure 3-20). This results in weight savings of approx-
imately 50 %, reducing the total weight from 562 g to 280 g and a reducing the reflected
2 4 2
inertia of the gearbox from 1.8 x 10-4 kg - m to 1.5 x 10- kg - m (cf. Figure 3-21 and
3-22).

A critical element in ensuring the proper function of the gearbox and the shoulder unit as

a whole is the bearings. Large diameter, thin cross section, ball bearings of angular and four
point contact type are employed to support the rotors and gear hardware (cf. Figure 3-23).
 PO 15, PA20
H.25.

-6X 8316THRU
fixed
R1 6X0 2X01.42

A A

- 0.0
P BC. 0.2505=.0002THRU

2X 0.39

C 4X0RHO4 (01.42)-
*013)

S ~ outputx
R

input s output K

SECTION A-A

Figure 3-19: Schematic representation of the Figure 3-20: Portion of engineering drawing for
planetary gearbox design. low weight planet gear. The planets, along with
the sun, ring and planet carrier were lightened ex-
tensively. Drawing dimensions are in inches.

Use of these bearings frees the internal cavity of the actuator unit for sensing and spreads
loads over a larger number of ball bearings. The path of loading through the shoulder
unit was designed to bypass the rotor-stator interface and utilizes the stators as reinforcing
structural elements. Loads are transferred from the humerus to the two large bearings, these
bearings are mounted directly to the outer housing.

The volume within the rotor of each motor is used for angular position sensing. Each
motor is equipped with a RLS RMB20 absolute angular position encoder. The use of slip
rings to transmit sensor information from the moving elbow encoder to the body is avoided
by the clock spring winding of the encoder wiring. The RMB20 encoder was selected for its

low mass and inertia, small size and unique mounting requirements. Unlike similar optical
encoders the RMB20 magnetic encoder has a small diameter and the sensor is mounted on
the axis of rotation.

Provisions for the introduction of abduction and adduction take the form of two bearing
mounts on the posterior and anterior faces of the shoulder unit (cf. Figure 3-24). Together
with a pull rod mount on the top surface of the shoulder unit. This degree of freedom is
actuated by a Dynamixel EX-106+ servomotor.
Figure 3-21: Assembled view of the planet car- Figure 3-22: Closeup photo of the gearbox as
rier and planet gears. assembled in the front leg.

Leg and Foot

The design of the front leg and foot is left flexible by the remote location of the actuators and
associated precision structures. The remaining function of the leg is merely to transmit forces
generated by ground contact to the body. Characteristics identified as desirable in a leg
design are ease of manufacture, strength, durability and lightness. These design objectives
were accomplished by employing manufacturing techniques and materials uncommon in
legged robotics. Each leg link contains two distinct structures, a polyester foam core and a
polyester resin shell forming a geometrically flexible structure that resembles a foam filled
box beam (cf. Figure 3-25 and 3-26). Use of plastics in the place of more common leg

materials such as aluminum or titanium allows legs to be manufactured quickly, cheaply

and easily, with design iteration time approaching 24 h. This extremely short iteration time
allows adaptive refinement of the leg geometry to an extent that is impractical with the
standard three week turnaround of contracted machining.

The leg links are manufactured by machining two molds from Delrin stock. The molds
are machined on a three axis computer numerical control (CNC) vertical milling station
using basic tooling and techniques. The Delrin molds machine easily and release the cast
polyester parts well. The first mold is machined slightly undersized of the final external link
dimensions to form the foam core, varying in deviation by 1.5 mm to 3.0 mm depending on
Figure 3-23: Placement and orientation of the Figure 3-24: Ad/abduction mount on the shoul-
bearings in front leg. der unit permits rotation about an axis parallel to
the spine shown here as a red line.

the anticipated loading in a particular region of the leg link. The second mold is machined
at the nominal external dimensions. The foam core is cast first and transferred to the second
mold after curing, where the shell material is cast around the core insert. The shell mold
is filled in a vacuum chamber and cured at the elevated pressure of approximately 420 kPa
gauge to inhibit the formation of voids.
The introduction of a Kevlar tendon and passive foot reduce the bending load on the
radius by converting a portion of the load moment to tension of the tendon and compression
in the leg link (cf. Figure 3-27). The tendon is a length of Kevlar webbing, anchored at one
end to the tricep using a ladder lock mechanism while the other end is embedded into the
foot. A compressive load on the foot, such as that due to ground contact, places the tendon
in tension and reduces the peak stress within the radius. The foot itself is a simple link, with
a silicone rubber pad. The foot reduces the passive impedance of the leg during impact and
provides the anchor necessary for the use of a tendon.
Figure 3-25: Front leg links assembled to the Figure 3-26: Internal structure of a foam com-
housing. posite leg link that was broken during single leg
testing. The yellow material is the lightweight
foam core and the white material is the plastic
shell.

Figure 3-27: Kevlar tendon of the front leg, shown here during assembly of the foot pad. The tendon is
connected to the tricep pin at the posterior end of the radius, from there it partially wraps the wrist joint and
is embedded into the foot.
3.3.2 Rear Leg

The rear leg design inherits many features of the front leg but retains a degree of uniqueness.
The most apparent of these differences is a three member leg structure instead of the simpler
two element leg of the front (cf. Figure 3-28 and 3-29). This modification was introduced
to increase the maximum leg length, while reducing the moment arm of the knee. The leg
has the following specifications (cf. Table 3.4).

Figure 3-28: Front view of right rear leg. Figure 3-29: Rear view of right rear leg.

Actuator Selection

The design loads of the rear leg, being very similar to those of the front leg, did not require
a substantially different actuator. Thus the same HT05001 was used in the rear leg to drive

Specification Value
Max. load 1250 N
Shoulder torque (peak) 134 N- m
Knee torque (peak) 134 N- m
Shoulder range of motion 150'
Knee range of motion 850
Max. leg length 500 mm
Min. leg length 232 mm
Weight 3.60 kg

Table 3.4: Rear leg design specifications

both the hip and the knee. This facilitates a degree of modularity in the leg design and
eliminates the expense of developing multiple custom motor windings. The use of the same
motor also facilitates the reuse of many components from the shoulder unit of the front leg,
decreasing per unit production costs.

Hip Unit

The most substantial difference between the shoulder unit of the front legs and the hip unit
of the rear legs is the introduction of a gear reduction to the first motor stage (cf. Figure

3-30). This stage drives the hip in the rear leg. The addition of a gear reduction increases
the capacity of the rear leg to produce propulsive forces that require moment loading of the
hip. The difference in width between the shoulder units with one gear reduction (front legs)
and two gear reductions (rear legs) is less than 1 cm.
Although much of the basic structure remains the same in the hip unit, the newly intro-

duced gear reduction complicates the sensing system for knee motion (cf. Figure 3-31). Two
small bearings were added to support the motion of a shaft that provides the encoder refer-
ence position to the knee position sensor. The knee sensor wiring can no longer be exposed
to the medial face of the unit and instead must exit on the lateral face. Bearing placement
in the rear leg is similar to the front with the exception that one bearing is modified from an
angular contact to a four point contact to support possible moment loads (cf. Figure 3-32).

Leg and Foot

The fabrication methods developed for the front leg are also utilized in producing rear leg
links. The major difference between the front and rear leg structures is the introduction of

coupling between the metatarsal link and the femur. This presents itself in the form of a
four bar linkage that constrains the angle of the metatarsal link to an orientation parallel
with the femur. This modification was introduced to reduce the moment arm of the knee
actuator while increasing the maximum leg length.
 lie,

Figure 3-30: Gear reduction of the hip motor Figure 3-31: The knee motor actuates the knee
was introduced to increase hip torque of the rear through a steel tendon. Elements that move when
leg. Elements that move when the hip motor is the knee motor is actuated are highlighted in blue.
actuated are highlighted in blue.

AW''aim

WIN

Figure 3-32: Bearing placement and orientation in the hip unit of the rear leg.
70
Chapter 4

Modeling and Performance

Estimation

4.1 Actuator Modeling

To understand the factors that contribute to motor heating, considered the most likely cause
of actuator failure, a thermal model of the HT05001 was developed. Taking the thermal
system as the volume containing motor windings heat is generated at the rate of Q = 12 R,
as governed by Joule's First Law, where Q is the thermal energy converted per unit time,
I is the winding current and R is the winding resistance. This energy is either stored as
a temperature increase of the motor windings or dissipated by conduction, convection or
radiation. The thermal load is anticipated to be the greatest during high speed running, a
task that is unlikely to continue for extended periods of time. Therefore, the thermal analysis
of the HT05001 is focused on timescales on the order of minutes.
Motor loading conditions for a high speed rotary gallop are estimated based on footfall
durations observed by Hildebrand for a cheetah running at approximately 72km - h-1 [18].
Under these conditions each leg is in contact for 60 ms to 70 ms of a 390 ms stride. During
contact the motor loading is assumed to resemble a half sine with a maximum value equal
to the motor demagnetization current of 69.5 A (cf. Figure 4-1). This results in a root mean
squared (RMS) motor current of 20.1 A over the entire stride and thus an average power
dissipation of 318 W using the nominal winding resistance of 0.791 0
 70
Motor current
60 RMS current

50

S40.. ..

20

10 -

0
0 0.1 0.2 0.3 0.4
time, t (s)

Figure 4-1: Anticipated motor current in a fast gallop.

To test the thermal characteristics of the windings under this type of loading the stator
was placed upright on a thermally insulating surface in still air with an ambient temperature
of 23.6 C. The stator contains three windings connected in a wye configuration with an
average terminal to terminal resistance of 0.791 Q. One pair of terminals was connected to
an Agilent 8757, 5.1 kW power supply, which was commanded to provide a constant power
dissipation of 300 W. The current and voltage provided to the winding were simultaneously
reported to LabView at 750 ms intervals. The winding temperature was determined by
utilizing the well established relationship between the temperature and resistance of copper
wire [12]. If the temperature coefficient of resistivity (a) is assumed to be constant over the
temperature range of interest the relationship between winding temperature and resistance
may be expressed as:

T= R-R + TO (4.1)
a Ro

where T and R are the instantaneous temperature and resistance of the winding and To and
Ro are the temperature and resistance of the winding at a reference state, immediately prior
to commencing the thermal test.

The 300 W test was run for 39 s resulting in an increase in temperature of 41.6 'C (cf. Fig-

ure 4-2). The maximum permissible winding temperature depends on the insulation class.
The F class winding insulation used in the HT05001 permits core temperatures up to 155 0C
before motor life is compromised [61]. The character of the thermal response shows that
despite the short timescales involved the effects of heat transfer from the windings is not
negligible. Much of the heat transfered from the active winding is assumed to have been
transfered by conduction to the inactive windings and the stator.
It was determined that a single time constant thermal model captures the thermal char-

acter of the motor to sufficient detail for short term operation. This model consists of a
heat capacity (C) and a lumped dissipation factor (H) (cf. Figure 4-3). For such a system the
governing differential equation and its solution are

C =I2 R - HO (4.2)
dt
2 R -tH43)

where 0 is the temperature rise above ambient (T - To). Fitting this model to the experi-
mental data results in an estimated heat capacity of 170.7J - C-1 and a dissipation factor
of 4.802W - 'C-1. Using the estimated dissipation of 318W together with the empirical
thermal model suggest that peak output running may be sustained for up to 5 min. This
result represents an acceptable maximum effort stamina for an untethered machine.

50

60 10 20.30 .4
* motor
C
i 50 .. . . .. . . . .. . . .

~40 -. 2011
. ... .. ..

S30. .. . .
Experiment
Model
20-
0 10 20 30 40
time, t (s)

Figure 4-2: Experimental and model of winding Figure 4-3: Thermal motor of the motor with
temperature under constant 300 W thermal load. heat generated by dissipation and rejected by con-
duction, convection and radiation.
4.2 Single Leg Dynamics

The first step in understanding the locomotion performance of a legged system is under-

standing how a single leg uses the ground to generate propulsive forces. This begins with
the construction of a representative kinematic model. After a kinematic model of the leg

is established, and thus the manner in which the leg may move is precisely defined, a sys-
tem of leg dynamics is put in place to connect the forces and torques that are applied to

the system with the motions that result. Once this dynamic framework is in place, what re-
mains is exploration. In this analysis, studies of the effects of mean shoulder height, forward
velocity and their relation to the force production capability of a front leg are presented.
Following this, an analysis of leg return trajectories identifies characteristics typical of rapid
protraction.

4.2.1 Mathematical Model

The kinematic model is based on the fundamental mechanical structure of the front leg
and therefore contains only two components, the humerus and the radius (cf. Figure 4-4).
The removal of the passive ankle joint eliminates joint redundancy and thereby simplifies
both the kinematic and dynamic analysis. This simplification is not expected to alter the
fundamental characteristics and responses of the leg during stance because the wrist reaches
a state of maximum flexion shortly after impact and remains in this position until just before
lift off. The effect of the foot on overall leg geometry is captured by incorporating it into the
radius dimensions.

The system dynamics are obtained by the use of screw theory [59, 40]. Screw theory
was selected because it facilitates the arbitrary extension of the dynamic solution to a general
serial chain, composed of revolute and prismatic joints, and it provides the equations of
motion in a format that is immediately compatible with numerical integration tools. The
following text provides a brief introduction to the use of screw theory in two dimensions
alongside the derivation of the dynamics of a general serial chain, the result of which is later
used in the dynamic simulator. The notation and approach employed are based on the work
of Zhao and Dai [68].
 T 2
c2

ci

Figure 4-4: Kinematics of a two link serial chain.

The motion of a rigid body in two dimensions is expressed in screw theory as a twist vector
(T) with three components:
-T
T= zo yo W (4-4)
where zo and yo are the Cartesian velocity of an imaginary point within the body that is
instantaneously coincident with the origin and w is the angular velocity. The twist vector is
related to the generalized coordinate (q) by a unit twist, written $, such that:

T = $(4.5)

where 4 is the time derivative of the generalized coordinate. For a revolute joint located at
the Cartesian coordinates (Xr, Yr) it may be shown that the unit twist is $= [Yr -X, 1]T.

While the unit twist for a prismatic joint inclined at an angle < from the horizontal is $=
[cos+ sin+ o ]T (cf. Figure 4-5).

To extend this analysis to a system of rigid bodies consider a serial chain composed of n
links joined by revolute joints (cf. Figure 4-6). If the generalized coordinates are combined

into a generalized coordinate vector such that q = [1 q2 - ]

n T the twist of link k may be
 Figure 4-5: Revolute and prismatic joints in screw theory.

written as:

ok

Tk = ok [$1 $2 '.. $n = J(

Wk

where J is a Jacobian matrix formed from the unit twists of each joint. This solution al-
lows the twist of each link to be determined based on the unit twists and time derivative of
generalized coordinates. However, because the twist provides the velocity of an imaginary
point within the body that is coincident with the origin the usefulness of the result in this
form is somewhat limited. To cast this result into a more intuitive form a coordinate trans-
formation is used to transform the twist to the velocity of the center of mass of each link.
Letting (xck, Yck) be the Cartesian coordinates of the center of mass of link k the coordinate
transformation matrix (rk) that converts the twist to a center of mass velocity (Vk) is:

1 0 -yck

U = 0 1 Xck (4.7)

0 0 1

and thus the center of mass velocity may be written directly as

Vk = TkTk = FkJ4. (4.8)

 Figure 4-6: n-link kinematic chain.

At this point it is convenient to define the first order influence coefficient matrix for link
k as
Gk = TkJ. (4.9)

The first order influence coefficient matrix is an n X n matrix that relates the generalized
coordinates to center of mass velocities such that

Vk = G4q. (4.10)

Differentiating Vk to get the acceleration of link k

d d
ak = dk = (GGk) = Gkq + Gkq (4.11)
dt dt

it becomes apparent that in order to find the acceleration of a link the time derivative of
the first order influence coefficient matrix must be computed. Expanding the total deriva-
tive and converting the result to an equivalent matrix multiplication shows that the time
derivative may be separated into two components.

DGk dql dGk dq2 aGk dq,

(4.12)
89q1 dt 8q 2 dt q dt
&Gk
aq1

aGk

k~ = 6 q2 .. en- (4.13)

aGk
_ &qn _

The first component, qT, is the transpose of the generalized velocities and G* is a block
matrix of the partial derivatives of Gk with respect to the generalized coordinates. Defining
column i of Gk as the column vector g, the second order influence coefficient matrix (G*)

may be equivalently defined as

a9k1 a9k2 a9kn

9
aql 8 1 aqi

9
( 9k 1

8g1
(4.14)
aq2,

gk2 ... agkn

aqn, 91q

Using this definition the acceleration of the center of gravity of link k may be expressed as

ak = G + q T G*q. (4.15)

Proceeding from this point using the Newton-Euler formulation of rigid body dynamics
the inertial force of each link (F'=[Fk Frk T[k ]T)may be written in terms of the diagonal
inertia matrix (Nk) and the center of gravity acceleration as

mk 0 0

- 0 mk 0 ak= -Nkak (4.16)

0 0 Ik
where mk and Ik are the mass and inertia of link k about its center of mass. This inertial
force on link k creates torques on the generalized coordinates equivalent to

GkF
=Tkn (4.17)
ri = 71 Tr2 ...

where Tk' = 0 for links after k in the serial chain (k + 1, k + 2, ... ,n) by virtual work.
Substituting F' from (4.16) and ak from (4.11) into (4.17) yields

r = GT Fr = G T(-N kak) = -GN k(Gk + Gk) (4.18)

ri = -(G TNkGAy+GGT Nkdk4). (4.19)

This form of the equation is simplified by defining a conjugated form of the inertia matrix

Mk= G NkGk, (4.20)

and using the definition of G* to rewrite the coefficient of the second term as

dql aG dq 2 + k(Gk dq,

G)TNGo = G TNk DGk
k Gq dq + OGqN + (4.21)
1t Dq2 dt k 89, dt
G TNkk
G TNk aq
q-2 --. n (4.22)

Nk

G TNk ak G G~k
N 9k2
aql GG~k
TNkan
agq,

G TNk ak
GNk G GTNkN2 k2 G TNkan
GkN
(4.23)
G TNk agk

G~k aq1 G~k agq 2 G TNk (9kn

= qi((GT Nk) G*) (4.24)

where 9 is used to denote the Kronecker product of two matrices. This equation may be
simplified further by defining

Hk = (GTNk)o G*. (4.25)

Summing the inertial contributions of all links of the serial chain the total inertia torque

vector is
n
r' =
- r = -(Ma + q THq) (4.26)
k=1

where
n n
M = M = S(GTNGk) (4.27)
k=1 k=1

and
n n

H = Hk =Z ((G Nk) 9 G* ). (4.28)

k-i k~
k=1 k=1

Expressing the dynamic model in equilibrium form the joint torques due to inertia must
balance those due to gravity, external forces and actuators, thus

r' + 7 + -7 + -ra = 0, (4.29)

where 79 is the joint torques induced by gravity, -re is the joint torques induced by exter-
nal forces and ra is the torques applied by joint actuators. Using the first order influence
coefficient matrix the torques due to gravity and external forces may be expressed as

n
7g + re = (GT(F + F'e)), (4.30)
k=1

where Fq is the force on the center of gravity of link k due to gravity and F" is the external
forces and torques applied at the center of gravity of link k. For a system in which gravity
acts in the negative y direction the force of gravity on link k is F9= mkg = mk o -g o]T,

where g is the gravitational acceleration. Combining these torque terms the equations of
motion for the serial chain may be written as

n
M4 + 4 T Hq - ((GT(Fg + Fe)) - Ta = 0. (4.31)
k=1
Or by taking the inverse of M and arranging for 4 as

q=M-1 ra + (G(F+F))+ k
- qTH ). (4.32)
k k
k=1

For the two link model used to study the contact phase of leg operation G 1 and G 2 may
be assembled in the following fashion. First, in cases where i < k the ith column of Gk may
be written as

Yri - Yck

9 i -ri + -Tck

where the position of the revolute joint for link i has the Cartesian position (Xri, yri). Then,
for columns where i > k the column vector is equivalent to the null vector, gi = 0. Thus
G1 and G 2 for the given geometry (cEf. Figure 4-7) are

Yri - Yci 0 -l1 sin qi 0

G1= -XrI + Xci 0 le cos qi 0 (4.33)

1 0 1 0

Yr1 - Yc2 Yr2 - Yc2

G2= -XrI + Xc2 -Xr2 + Xc2 (4.34)

1 1

-11 sin qi - 1e2 sin(q1 + q2 ) Ic2 sin(qi + q2 )

licos qi + 1c2 cos(qi + q2) 2 cos(q1 + q2 ) (4.35)

1 1

These definitions of Gk along with the inertia matrix of each link completely describe the
serial chain. From them the additional chain parameters, M and H, may be calculated.

If these parameters are combined with a configuration (q), velocity (q) and loading, the
acceleration of the generalized coordinates (4) may be determined.
 T 2
c2
/11 C1

Figure 4-7: Definition of physical quantities for two link serial chain.

4.2.2 Stance Phase Simulation

The intent of the leg itself is to produce controllable forces on the body, that may be used for
propulsion or maneuvering. Consequently, the intent of the single leg dynamic simulation is
to determine the precise manner in which the leg must be employed to access its full capacity
in this regard. Two variations in the character of a stride are studied: mean shoulder height
and forward velocity. Each of these stride modifications alters the sense and magnitude of
the body forces that may be generated throughout stance. Thus a thorough understanding
of the stride is essential to achieving the highest force production.

The single leg dynamic simulation is implemented in MATLAB using objects. The
parameters of each link in a serial chain are defined by creating n link objects, where n is
the kinematic chain length. These links are then arranged into a single chain object. The
chain object is passed to a preprocessor which derives the equations of motion by calculating
the M, H, Gk and Nk matrices. After these matrices have been determined for a specific
chain they are saved so that they need not be re-derived to perform dynamic calculations.

The structure of the dynamic simulation is unique in approach and somewhat resembles
the moment method used in aerodynamics [34] and the Milliken moment method used in
vehicle dynamics [43]. The shoulder trajectory and joint torques are specified as inputs
and for each instant in time the resultant body force (Fb., FY) and ground reaction force
(Fgx, Fgy) are calculated (cf. Figure 4-8). This approach allows the actuator torques to be
swept over their range while the leg is in a particular dynamic state, generating a region of
accessible body forces at that instant (cf. Figure 4-9). The factors that directly limit these
body forces are the maximum actuator torques and the limits of the contact interaction at
the foot.

Input Output

Figure 4-8: Program layout of the stance phase dynamic simulator.

-500 0 500
x direction body force, Fbx (N)

Figure 4-9: The range of accessible body forces in a specific configuration. Plotted in the Fb,-Fy plane.

In order to simplify the analysis and focus on the effect of shoulder height several simpli-
fications to the shoulder trajectory are introduced. First, based on biological data, variation
in shoulder height throughout the contact portion of a rotary gallop is small in compari-
son to changes in horizontal position [18]. Thus the shoulder trajectory is simplified to a
 Property Value
Humerus length 240 mm
Humerus mass 200 g
Humerus inertia 9.7 x 10-4 kg M2
Shoulder drive inertia 4.7 x 10-3 kg M2
Radius + foot length 320 mm
Radius + foot mass 180 g
Radius + foot inertia 1.5 x 10-3 kg M2
Knee drive inertia 9.9 x 10-4 kg M2

Table 4.1: Leg parameters used in dynamic model

straight horizontal line and the vertical velocity and acceleration of the shoulder are set to
zero throughout. Furthermore the forward velocity of the shoulder does not vary substan-
tially throughout stance, allowing the horizontal acceleration to be set to zero as well. Note
that this horizontal trajectory requires rapid acceleration and deceleration of the elbow and
shoulder joints even though the shoulder itself does not accelerate. The leg parameters used
in the dynamic simulation are based on those of the physical hardware (cf. Table 4.1).

For a specified shoulder height the range of accessible horizontal positions of the shoul-
der is broken into fifty equally spaced configurations. The possible body forces for each of
these configurations is computed and a contour plot in the Fb,-Fv plane is generated. The
contours in this plot show the forces that the leg is capable of generating with contours of
greater magnitude indicating that forces within this contour may be generated in a larger
number of configurations.

The results of these simulations show that two characteristics of the forces available at
the body change with shoulder height. First, a higher stride height reduces the range of
angles through which the leg may sweep due to the limits of leg length. This causes the
cone of available body forces to narrow as height increases (cf. Figure 4-10, 4-11, 4-12 and
4-13). Physically this implies that less deceleration and acceleration is possible. The second
feature is related to the first in that if a higher shoulder height is selected the elbow is less
bent throughout stance. This causes a increase in the forces that are generated along the leg
direction (from the foot to the shoulder). A balanced stride height avoids excessive compres-
sion of the elbow to maximize forces along the leg and avoids over extension of the knee to
maximize the capacity of the leg to provide propulsive forces. For the simulated parameters
a suitable stride height seems to be around 45 cm.

2500 2500

2000 - 2000

40
1500 i
40
1500

*0 >1
0

1000 . 1000

500 500

-1000 -500 0 500 1000 -1000 -500 0 500 1000

x direction body force, F, (N) x direction body force, Fbx (N)

Figure 4-10: Possible body forces for strides with Figure 4-11: Possible body forces for strides with
a shoulder height of y = 35 cm. a shoulder height of y = 40 cm.

2500

- 2000

0 1500

0 1000

500

O'
-1000 -500 0 500 1000 -1000 -500 0 500 1000
x direction body force, Fbx (N) x direction body force, Fbx (N)

Figure 4-12: Possible body forces for strides with Figure 4-13: Possible body forces for strides with
a shoulder height of y = 45 cm. a shoulder height of y = 50 cm.

The forward velocity of the shoulder was then simulated for speeds of 2 m - s1, 5 m - s- ,
10 m - S- and 15 m - s1 at a shoulder height of 45 cm. The results of these simulations show

that as speed is increased the body force production cone rotates about the origin away from
the first quadrant and translates in the negative Fz direction (cf. Figure 4-14, 4-15, 4-16 and
4-17). This implies that the capacity to produce propulsive forces is reduced. Furthermore
the boundary of the entire curve shrinks. This is due to the increasing amount of actuator
torque that must be used to accelerate and decelerate the masses and inertias of the leg.
The torque limits were set to the thermally limited torques for each joint, to achieve this the
battery voltage must be high enough to overcome the back electromotive force (EMF) that
would be present at this speed.

2500

-1000 -500 0 500 1000 -1000 -500 0 500 1000

x direction body force, Fbx (N) x direction body force, Fbx (N)

Figure 4-14: Possible body forces for strides with Figure 4-15: Possible body forces for strides with
a forward velocity of v, = 2 m* s- a forward velocity of v = 5 m - s- 1 .

2500

Z 2000

1500
40

1000

500

-1000 -500 0 500 1000 -1000 -500 0 500 1000

x direction body force, Fbx (N) x direction body force, Fbx (N)

Figure 4-16: Possible body forces for strides with Figure 4-17: Possible body forces for strides with
a forward velocity of v. = 10 m . s-1 a forward velocity of v. = 15 m - s-.
4.2.3 Protraction

Protraction is the movement of a limb from a rearward position to a forward position and
is a required part of any gait. Depending on the leg and actuators, protraction may limit

the maximum running speed. Thus it is essential to understand effective protraction strate-

gies, specifically the joint trajectories and the torques that induce them. The framework
for protraction simulation is somewhat different than the stance dynamic simulator and re-
lies on different assumptions. In the protraction simulator the shoulder node is considered
fixed (or at constant velocity) and the limb is protracted with no external loading at the foot
(cf. Figure 4-18). This makes the simulation substantially faster and less complicated than
the contact simulation. Shoulder reaction forces are calculated based on input shoulder
and elbow torques but are assumed to have negligible effect on the motion of the body. The
underlying equations of motion are identical to those derived for the stance simulation.

Figure 4-18: Three link leg with fixed shoulder node used to analyze protraction joint trajectories.

A optimization framework was built around this dynamic simulation to identify fast pro-
traction torque profiles (and the joint angle profiles to which they correspond). This frame-
work was implemented in MATILAB, and makes use of the fmincon function, a nonlinear
optimization tool. The objective function was defined as obj = t+w(effort) where w = 0.01
is a weighting factor and effort is a measure of the actuator loading defined as

effort jT +T 2dt.

87
Thus the objective function depends on both time and actuator thermal load, as approxi-
mated by the total torque impulse. If a purely time optimized objective function was defined

one would anticipate rail to rail actuator behavior. Because fmincon is a local optimization
tool the optimizer was seeded randomly and the optimal results for many iterations were
weighted and displayed simultaneously (cf. Figure 4-19 and 4-20).

10 10

5 5

0 0

-5 -5

-10 -10
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
time, t (s) time, t (s)

Figure 4-19: The shoulder torque that results Figure 4-20: The knee torque that results from
from protraction optimization. The results of protraction optimization. The results of many
many randomly seeded optimizations are over- randomly seeded optimizations are overlaid.
laid.

Typical joint angle profiles that result from this optimization have several interesting
characteristics and bear a great resemblance to the trajectories observed in biological sys-
tems. First, for fast protraction the leg is collapsed before protraction begins (cf. Figure 4-21).
This reduces the inertia as seen at the shoulder and allows the leg to protract more quickly
within the actuator torque limits. A notable asymmetry is present in the torque profiles that
is centered around the very small torques required to halt protraction of the limb in a for-
ward position. This is a consequence of the natural dynamics of the kinematic chain. If the
joints are allowed to run free the elbow extends, increasing the inertia and slowing rotation
about the shoulder.
 0.6

0.4

0.2

o 0

> -0.2

-0.4
4A
-0.4 -0.2 0 0.2 0.4 0.6
x position, x (m)

Figure 4-21: Movement of the leg through space when the protraction torques are commanded.
90
Chapter 5

Design Tools

5.1 Assessing Leg Capability with Force Maps

The force map is a design tool proposed to graphically illustrate how factors such as foot
friction, actuator limits and leg geometry influence the shoulder force production capability
of a particular leg design. It thus provides a quantitative metric for comparing and evaluating
leg designs and building an intuitive understanding of how changes made to the leg alter the
maximum forces that may be provided to the body.

Understanding the forces at the shoulder is a critical step in developing effective legged
locomotion [54]. The body is often far more massive than the legs, particularly in systems
designed for maneuverability and speed, and therefore the shoulder force becomes the vital
coupling element between the propulsion system and the central hardware. In a familiar
automotive system, the shoulder forces forces are analogous to the forces applied through
the suspension to the chassis. In a car, interaction at this joint is typically rather straightfor-
ward, due to the decoupling of propulsive (tractive) forces and support (suspension) forces.
However, in legged locomotion these two elements are inextricably linked.

Precise control of the forces generated at the shoulder allows the trajectory of the body
through space to be prescribed. Abstracting the interface between the leg and the body
allows the force production mechanism to be considered independently of the often compli-
cated body dynamics. With a lightweight leg dynamic coupling between the body and the
leg is weak and this separation of the locomotion task into low level (leg control) and high
level (body control) is possible. Separating the system in this way emphasizes the importance
of achieving a specific net impulse over the course of a contact event.

During steady state locomotion the net vertical impulse on the body due to all legs must

be equal and opposite to the impulse generated by the constant pull of gravity. Additionally,
the net propulsive impulse must be equal to any impulse caused by drag over a full gait
cycle. If these goals are not met the body will tend to change speed. Given these strict
demands on net impulse over the stride it must be asked: what are the limits of shoulder

force production? Certainly one of the limits is the configuration dependent static actuator
limits (typically computed from the Jacobian). However, during dynamic locomotion these
limits are altered by joint accelerations and the natural dynamics of the system. This effect
is somewhat nonintuitive but it is perhaps the most significant interaction in defining the
limits of ultimate running speed. As speed increases a larger increment of actuator torque is
devoted to acceleration of leg links and therefore less torque is available for shoulder force
production. Furthermore because maximum stride length is fixed by the leg geometry a
higher forward velocity leaves less time for impulses to be generated, necessitating higher
peak forces.

Another important limit to running speed is the contact interaction at the foot, specifi-
cally friction. If the standard Coulomb friction model is applied to the foot, it is clear that
vertical loading must be increased in tandem with longitudinal loading to avoid sliding. This
increased vertical force must be provided by the leg actuators. If the friction coefficient is
less than one, increasing longitudinal forces at the foot requires not only the ability to apply
these longitudinal forces but the ability to apply an increment in vertical force that is even
larger than the desired change in longitudinal force.

To quantify the forces that may be produced by the leg the individual joint accelerations
must first be identified, these depend heavily on the nature of the stride. Several trends seem
to develop in the character of a running stride across various species and speeds. First, the
shoulder trajectory relative to a fixed foot is generally a flat U shape (cf. Figure 5-1) [18].
This implies that shortly after touchdown the leg is primarily experiencing compression of
the knee. Next, as the stride progresses, the dominant motion becomes rotation of the limb
about the shoulder. After mid-stance the behavior of the leg changes once more to a knee
dominant posture, however in this case the knee is being extended. These three phases can
be observed across many species and many gaits within an individual species. At high speed
changes in the forward and vertical velocity of the shoulder are small in comparison to the
average forward speed. The simplification employed when creating a force map is that the
shoulder trajectory may be approximated as a horizontal line traversed at constant velocity.
Thus the vertical and horizontal accelerations as well as the vertical velocity are zeroed.
This condition, although simple, captures the majority of the joint accelerations induced by
high speed running gaits.

--------------------------------
.Y

LH
F

RH RF

Figure 5-1: Approximate hip and shoulder trajectories for a running cheetah, adapted from data in [18].

The idea of a force map is to plot schematically the range of body forces (Fb., FV) that
the leg is capable of generating at a specific forward velocity over the entire workspace. The
workspace is generated by sweeping all possible joint angles with a fixed foot and mobile
shoulder node (cf. Figure 5-2). The result of a force map is the velocity dependentforce
production envelope, or more simply the force envelope, which is the curve in the F,-Fv plane
that encompasses all feasible body forces (cf. Figure 5-3). As the forward velocity, friction,
actuator limits or leg geometry is modified the force production envelope changes size and
shape.

Interpretation of the boundary of the force production envelope in terms of the body
dynamics provides insight into the performance that can be expected when the leg is used
during locomotion. For example the area enclosed by the curve to the right of the ver-
tical line F, = 0 is populated by configurations and joint torques that result in propul-
sive forces being applied to the body. As this area decreases it becomes more difficult to
 2500

z 2000

2 1500-
0 0

0
1000 -500 0 500 1000
x direction body force, Fbx (N)

Figure 5-2: Workspace of the front leg with a Figure 5-3: Force production envelope for a front
fixed foot location. led with forward shoulder velocity i = 2 m s-I

provide a substantial propulsive impulse, and thus it is more difficult to maintain speed.
When this area vanishes altogether it becomes impossible to maintain the current veloc-
ity. This provides a rough estimate of the maximum running speed that is an overestimate
because each stance phase is expected to contain periods of deceleration and acceleration
even if the net propulsive impulse is zero. Similarly the limitations of vertical impulse gener-
ation may be estimated by taking the largest Fv intercept between the force envelope and
FbX = 0 and assuming that the leg generates a parabolic profile in vertical force and has a

stance duration of ts. This implies that the net vertical impulse is approximately given by

Ap = fo" Fbydt -= { Fbylmaxts.

The force map in the given form provides little context for how the forces are generated,
and often the boundary is defined by operating near dynamic singularities. However, the
force map may also be used to evaluate the relative ease or difficulty of obtaining a spe-
cific propulsive force by weighting each occurrence of a force during the sweep of the entire
workspace for a given forward velocity (cf. Figure 5-4). This difficulty metric provides ad-
ditional perspective into the ability of the leg to produce forces in multiple configurations.
Flexibility in force production is useful in locomotion because it allows the control system
to correct for disturbances or make sudden changes to the gait. If the leg is operating at
its maximum potential for a given configuration it is difficult to apply the additional forces

that may be necessary to correct for even the most minor disturbance, this decreases the
apparent performance of the device.

2500

Z 2000--

00

0
1000 -

500-

0
-1000 -500 0 500 1000
x direction body force, Fbx(N)

Figure 5-4: Weighted force production envelope for a front leg with forward shoulder velocity .= 2 m s-.
As colors change from blue to red the number of configurations that allow forces within the contour to be
generated increases.

It must be noted that the weighted force map provides no sense of continuity within a
stride, and is therefore a first pass analysis tool that requires relatively little computational
effort. This is to say that some of the forces that are theoretically possible may be difficult
to achieve as part of a cohesive stride because they represent leg configurations that are
separated by substantial differences in joint configuration. To remedy this deficiency and
provide a more developed but intuitive way of understanding and constructing coherent
strides the force map planning tool was developed.

5.2 Force Map Stride Planning

If the control of the body is approached by decoupling the leg from the body at the shoulder
as proposed, the task of the body level controller is to generate a desired shoulder trajectory
and a force profile that accompanies it. The job of the leg then is to match both of these
requirements dynamically, most probably by implementing a type of impedance control
[21]. Execution of this task is made more difficult by an additional decision that the leg level
controller must make. Specifically, what configuration of the leg should be selected as the
touchdown configuration (cf. Figure 5-5). This decision is difficult because the touchdown
configuration influences the subsequent evolution of the stride, and if the wrong configura-
tion is selected the leg may not be capable of producing the desired forces. The inability
of the leg to generate the desired shoulder forces may be brought about by several factors
such as configuration or velocity dependence, exceeding the limits of friction at the foot or
perhaps even lifting the foot off the ground.

Choosing an appropriate touchdown configuration also assures that adequate control

authority remains to take up disturbances that may arise throughout the stride. As the
locomotion speed increases the bounds of friction and the actuator limits move ever closer,

closing the touchdown configuration window until ultimately the window is a single point
and locomotion speed cannot be increased further (cf. Figure 5-6). If attaining high running
speeds is the design goal this point, which depends on the shoulder trajectory and force
profile, must be located with precision. However the target touchdown configuration may
change throughout the speed range and also with changes in the desired force profile or
shoulder trajectory.

2ms-1 [
5ms-1 [
10 m s-1 [

Figure 5-5: Acceptable touchdown configura- Figure 5-6: Schematic representation of how
tion range, the foot may be placed within the the range of acceptable touchdown configura-
shaded area. tions closes with increasing speed.

Variation of the ideal touchdown configurations with all of these parameters makes de-
termining target configurations both difficult and essential. A brute force approach, in which
the leg dynamics are solved directly, is difficult to implement due to the computational limits

of embedded control hardware. Even if the shoulder trajectory and force profile are known,
there is still a large range of touchdown configurations, each of which must be simulated
with the leg dynamics to develop the control torques. The friction interaction then must be
evaluated at every instant throughout each of these strides to rule out trajectories that would
cause the foot to slide. The following section addresses this problem in stride planning and
develops an intuitive graphical method for evaluating the force production capabilities of a
leg over a stride.

5.2.1 Selecting a Stride Primitive

The process of developing a stride planner begins with the selection of a representative
shoulder trajectory through space. This trajectory often depends on gait and the character-
istics and design of the body and leg. The model must capture the essence of the dynamic
interaction between the leg and the body. Specifically this means that an identical force pro-
file, if applied to the physical hardware must not produce a substantially different shoulder
trajectory. The complexity of the model required to accomplish this varies withe leg and
body design, in this case of a rotary gallop a straight line shoulder profile is a reasonable
approximation. This is due to the fact that the majority of the joint accelerations arise from
matching the shoulder forward velocity. The vertical velocity and changes in the horizontal
velocity are small in comparison and thus contribute only a minor part to accelerations of
the generalized coordinates.

One substantial advantage of selecting a straight line profile is that it makes it possible to
decouple the selection of touchdown configuration from the remainder of the stride. That is
to say that regardless of the precise touchdown configuration selected the shoulder remains
on the same line through space. If the stride primitive were not linear different touchdown
configurations would traverse different positions through space requiring a larger number
of calculations (cf. Figure 5-7). However, a disadvantage of the straight line profile simpli-
fication is that it neglects the configuration dependence of the shoulder force that will exist
with the additional compression of the leg that typically occurs in a real stride. This may
lead the results to indicate that a stride is possible, when in actuality it falls just outside of
the actuator capability.
Figure 5-7: Straight as opposed to curved shoulder profiles permit the decoupling of touchdown configuration
selection from stride torque calculation.

5.2.2 Calculating the Force Production Envelope

For a specific set of position, velocity and acceleration conditions of the shoulder it is possible
to calculate the state dependentforce production envelope or SDFE. This is the region of the Fb.-
Fby plane that contains all body forces that the leg is capable of generating under the given
conditions (cf. Figure 5-8). The limits of the forces that may be produced are defined by two
factors, the actuator torque limits and the contact interaction at the foot. In the model that
follows contact requires two conditions to be met. Fist, in order to ensure the foot remains
fixed the forces at the foot must lie within the friction cone. Second, the vertical force at the
foot must be in the positive direction, that is to say the foot may not pull on the ground.

All of these limitations are reflected in the characteristics of the SDFE and have physi-

cally significant, intuitive interpretations (cf. Figure 5-9). The lowest point on the Fy scale
represents the shoulder force required to actuate the leg while maintaining zero contact force
with the ground (F2f). For a well designed leg this force should always be near zero, mean-
ing that only a small portion of the forces that occur at the shoulder are due to accelerations
of the leg. If this body force is large, actuation of the leg will affect body dynamics, which
will likely be undesirable. A massless leg will always place this point at the origin.

The cone that originates at Fzf is the friction cone of the foot translated into shoulder

force coordinates. If the friction coefficient (p) is made larger the cone widens. For low
 1600 G
- 1400

1200 - -
.....

40. 1000 .....

800 - -.-.-. oo

-S 600

~400
200 -

0
-500 0 500
x direction body force, Fbx (N)

Figure 5-8: The state dependent force envelope Figure 5-9: The bounds of the SDFE are formed
shown here is the region of the Fb,-Fv plane that by actuator limitations and the frictional interac-
contains all possible body forces for a given leg tion at the foot.
state (position, velocity and acceleration).

friction surfaces this cone may become quite narrow, possibly trimming all but one actuator
bound. In such a situation the robot is said to be predominantly friction limited.
The final set of bounds is the parallelogram produced by symmetric actuator limits. This
parallelogram is rarely rectangular, in such a case the actuators would completely decouple
horizontal and vertical force production at the shoulder. Such a situation is the typical case
for a wheeled vehicle, with the wheel torque producing a Fb: component and the suspension
generating a FbV component. In the case of a legged system an increment in elbow torque
will generally produce a force component along the vector from the foot to the shoulder. An
increment in the shoulder torque on the other hand will produce a force at a configuration
dependent angle to the leg direction.

Obtaining the points that parameterize the SDFE is not a complex as it initially appears,
this is due largely to the fact that the conversion from joint torques to shoulder forces may
be expressed as an affine transform. That is to say that for a single position, velocity and
acceleration state the shoulder force is related to the actuator torques by a translation of
the origin and a linear coordinate transformation. The physical interpretation of the origin
translation is that it represents the actuator torque required to produce zero force at the
shoulder (rz,) or conversely the force at the shoulder that produces the desired state with
zero actuator torques (Fza).
 To complete the remapping of the shoulder and knee torques (r = [ T s ]T) to the body

forces (Fb = [Fbx Fby ]T) four points in the Fbx-Fba plane must be identified. The first of
these is Fzf, the shoulder force that produces zero contact forces at the foot. When the leg
is operating under the conditions imposed by Fzf the actuators provide just enough torque

to move the members of the leg such that the foot velocity matches precisely the ground

velocity even though no contact loads are exchanged. Once this point is identified two lines
may be drawn forming an upward opening friction cone. The slope of these lines are i--1,

where [t is the friction coefficient. Body forces within the friction cone satisfy the contact
conditions at the foot.

Next, two linearly independent actuator torque states must be selected in order to iden-

tify the coordinate transformation that, in conjunction with the origin shift given by Fza,
converts joint torques to body forces. The simplest selections are rio = 1 O]T and roi

[o 1]T these torques result in shoulder forces of Fio = [Fio. Flo ]T and Foi = [Foi' Foi, ]T
respectively. Using these values the affine transformation matrix may be defined as

- Fox + Fzax -FO + Fzax -Fzsx

Ta -F 01y + Fzay -F1oy + Fzay -Fzsy (5.1)

0 0 1

where Fza = [Fzax Fzay ]T is the shoulder force that produces the desired motion of the
shoulder with zero actuator torques. Using this transformation matrix, body force may be
calculated as
Fox Tk

F y Ta TS (5.2)

1 1

With the assistance of (5.2) construction of the actuator limit curves of the SDFE requires
only matrix multiplication. After construction of the actuator parallelogram is complete the
intersection of this area with the friction cone forms the total state dependent force envelope
for the given conditions. Thus full parameterization of a single leg condition requires the

100
calculation of thee points for torque transformation, one point for inertia compensation and
the globally applicable knowledge of friction coefficient and the actuator limits. Tabulation
of these four points over the workspace for a given forward velocity would then be suffi-
cient to calculate the torque profile for an arbitrarily defined shoulder force and shoulder
trajectory.

5.2.3 Selecting Touchdown Configuration

Calculation of the state dependent force envelope over the workspace can be used to rapidly
determine a touchdown configuration that is capable of providing the desired forces to the
body. This is done by evaluating whether or not the desired force remains within the SDFE
throughout the stride. The mathematical formulation of this stride planning approach be-

gins with the selection of a desired shoulder height by the body level controller. This leaves
the horizontal positioning of the foot with respect to the body at the discretion of the leg level
controller. For each possible location of the foot relative to the shoulder the controller calcu-
lates whether this position is capable of providing the initial force by evaluating whether the
force falls within the SDFE of that position. After this initial calculation the set of acceptable
touchdown configurations will be narrowed substantially. For each of these acceptable po-
sitions the shoulder trajectory relative displacement is added to the initial shoulder position
and the evaluation of leg capability is performed again. If the leg is incapable of provid-
ing the force demanded, that trajectory is thrown out and the next time step is evaluated
only for the remaining profiles. This approach quickly narrows the window of acceptable
touchdown configurations.
As a way to visualize the results of these stride planning calculations the capability matrix
(R) is introduced. R is a p x s matrix where p is the number of equally spaced positions on the
horizontal line selected by the body level controller for which the SDFE has been calculated
and s is the number of these positions required to match the desired stride length. Thus for
a straight line trajectory each row represents a specific leg configuration and each column
represents a given demanded force at that instant. For each case where the force falls within
the SDFE a one is added to the corresponding position of R, which initially zeroed. Thus
to identify a feasible touchdown configuration a diagonal must be identified that contains all

101
ones. If many touchdown configurations are possible one method of identifying a desirable
stride is by ranking the effort required on the part of the actuators. This can be done by
defining
2 2 2 2
+ k,max - Tk
e = 'max
e- 2'r 2 -
2±
s,mnax Tk'max

where T,max and rk,max are the shoulder and knee maximum torques respectively. Thus for
maximum effort e = 0 and for minimum effort e = 1. The diagonal of R with the largest

sum therefore represents the lowest thermal loading on the motor.

These calculations are capable of being performed in real time making their application

on modestly powered embedded systems practicable. The majority of computational time

will be spent evaluating whether or not a point in the Fx-Fy plane is within the capability
envelope for a specific configuration. Luckily an algorithm can be developed for performing
this calculation quickly. First the point must be transformed into torque coordinates, this

requires six multiplications. Next the actuator limits can be checked. Because the limits
are rectangular in torque space this is as simple as verifying that the point falls within the
actuator limits. If the point is outside of these limits then there is no need to continue, the
point is outside of the capability envelope. This eliminates the majority of points tested. If
the point is within the torque limits ground contact can be tested by evaluating whether or
not the point is above the zero foot force point. If this evaluation returns satisfactory results
the friction criterion can be checked by ensuring that the inverse of the slope from Fzf to
the point in the Fb,-Fy plane falls between -p and p. All of these calculations are relatively
simple additive or multiplicative operations which can be computed rapidly.

5.2.4 Calculating Torque Profiles

Once trajectory planning has been completed the next step, calculating the torque profiles
over the stride, has in-fact already been completed. This operation is the affine transform
from the desired shoulder forces to actuator torques. If torque values were not saved dur-
ing trajectory planning they may be recalculated for the selected touchdown configuration
without considering that the torques may fall outside of the actuator limits. If computational
resources permit, and additional accuracy is desired, the shoulder profile time step can be

102
refined and calculations performed once more to attain a smoother torque profile.
An alternative method of calculating the torques throughout the stride is available in

systems with even larger computational power. The touchdown configuration, shoulder
force and shoulder trajectory are all known, therefore the torques may be calculated from the
inverse dynamics. This calculation is somewhat time consuming for embedded systems but
provides absolute assurance that the commanded torques will provide the desired motions,
with the provision that the dynamic model used is reasonably accurate. In this scenario

the discretized profiles are used for stride planning while the inverse dynamics are used for
precise actuator torque estimation.

103
104
Chapter 6

Testing and Design Iteration

Testing of the leg design is broken down into three phases. Design proofing, single leg testing,
and multi leg testing. The first two of these phases were undertaken for the front leg design
alone and the final phase entails simultaneous testing of all four legs. The testing program
is ongoing and the control system is largely the work of Sang-ok Seok.

6.1 Design Proof

The first step in developing the leg further was testing the practicability of production meth-
ods, evaluating candidate materials and assessing the overall feasibility and performance of
the platform. The design presented in Chapter 3 is the second iteration leg design. Suf-
ficient changes were identified during the proof phase of the first iteration to warrant a
partial redesign. Two major changes were introduced in this version. First, two large di-
ameter split ring ball bearings were substituted for a Teflon impregnated bronze surface
bearing. The friction torque introduced by the surface bearings, as well as the difficulty in
achieving satisfactory alignment of the shoulder justified the additional expense of the ball
bearings. Second, the abduction axis bearing mounts were introduced to the base of the
shoulder and hip. These provisions were not considered on earlier designs. Many addi-
tional minor changes, such as reductions in motor width and the refinement of the sensor

mounting interfaces were also introduced.

The design proofing of the second iteration front leg proceeded smoothly. Fit tolerances

105
were refined from experience with the first iteration fits making assembly relatively trouble
free. Shoulder friction and alignment were improved substantially and the leg was deemed
suitable for single leg testing.

6.2 Single Leg Testing

Single leg testing was executed to quantitatively asses the performance of the leg design be-

fore committing to the fabrication of multiple units. The apparatus developed to perform

these tests is a unique variation on the standard linear slide (cf. Figure 6-1). The slide sup-
ports a body element, to which the front leg is attached. The body and leg are propelled to
a velocity of approximately 2.5 m - s-1 by a falling weight. After maximum speed is reached
the body falls from its supports until the leg impacts the ground and provides support forces.
After the contact event the body is caught by rubber wheels. CG position, inertia, angular
velocity and speed of the body are adjusted to mimic the behavior of an actual body during
a rotary gallop.

Figure 6-1: Single leg test apparatus consists of a linear slide with front leg attached to body emulator. The
body and leg enter free fall before the leg touches down for a single contact event. Following contact the body
is caught by two rubber wheels.

Initial testing of the front leg bears a remarkable resemblance to high speed footage of
a running quadruped (cf. Figure 6-2 and 6-3). A 450 ms stance phase with peak vertical

ground reaction forces of 550 N and peak propulsive force of 75 N was observed (cf. Figure

106
6-3). These results were obtained using a simple impedance controller. The front leg seemed
to operate relatively close to the actuator limits of the shoulder so a third design iteration
was proposed and is in development to add a gear reduction stage to the front leg, making
it similar in form to the rear leg.

Figure 6-2: Still images from a high-speed video of single leg testing.

Tiue (ms)
Vertical
Fore-aft

Figure 6-3: Ground reaction forces for single leg test.

6.3 Multi Leg Testing

Multi leg testing is still in it's early stages, however the robot, with a third iteration leg, has
been commanded to jump vertically (cf. Figure 6-4).

107
 7! W-

Figure 6-4: Vertical jumping of four legged robot.

108
Chapter 7

Summary/Conclusions

7.1 Design considerations

In approaching the design of a robotic system for dynamic locomotion, several character-
istics are of the utmost importance. First, the body mass must be kept as low as possible.
Every effort must be made to avoid a runaway increase in mass, as this will often lead to a
severe deterioration in performance. It is difficult or perhaps impossible to break even by in-
creasing the mass in a dynamic system due to the many nonlinearities that are encountered
along the way.

As for the leg design in particular four design principles are proposed:

- Minimize passive mechanical impedance

- Minimize mass and inertia

- Maximize actuator strength

- Find a balance between kinematics and leg use

These considerations are based on analysis of biological systems, dynamic simulation and
experimental testing. The transfer of information from biology to robotics must occur at the
lowest level, the level of fundamental principles.

109
7.2 Unique Features

Many of the features developed to enhance the performance of this particular robot may be

extended to other robotic platforms. The concentric motor topology and redistribution of
mass to the proximal extents of limbs provides a substantial increase in performance versus
an actuator at the joint topology, particularly when speeds are increased. Careful attention
must be payed to all mass placed distally in the limbs and there must be a very good reason

for placing it there and nowhere else.

Utilization of motors at the limit of their performance is something that has been largely

overlooked by many roboticists. Strict adherence to a constant voltage specification is a suit-

able way to use electromotive actuators in industrial settings but under loading conditions of
legged robots this strategy massively and unnecessarily reduces the performance envelope
of the system. This often leads to the introduction of additional performance limiting com-
promises such as an increase in gear ratio. A one or two time constant motor thermal model
is within the reach of even the simplest embedded controllers and the potential increase in
performance is enormous.

Tendons and the redistribution of bending loads, greatly enhances the overall strength
to weight ratio of the leg and avoids the need for expensive and time consuming machining
of the rather large leg links. The tendons can be either active, as the tricep of the front leg,
providing locomotive forces to the elbow, or passive as in the lower limb tendon of the front
leg.

The use of tendons enables the introduction of alternative materials and manufacturing
techniques for use in the leg links. The ideal material is highly durable, with a high yield
strength, but not necessary a very high stiffness. Polymers are a good fit for this application
and are readily available with a range of stiffness and strength characteristics. The advantage
of using thermoset polymers is that they can easily be molded into the desired leg shape,
eliminating the need to machine the numerous copies required for a complete robot. The
moulds can be made from a wide range of materials, such as wax, Delrin or aluminum. The
total mass of a plastic limb can be reduced by the in-molding of a low density polymer foam

core or by creating a link with a thin central web.

110
7.3 Design tools

The design tools proposed here are intended to guide the designer in developing an intuitive
understanding of the legged system being developed. Very few of these tools are available

to the roboticist and even simple changes, such as increasing the length of a leg link by
1 cm, may have a substantial, yet completely non-intuitive effect on the performance of
the leg during locomotion tasks. This sensitivity and complexity provides an enormous
and intimidating challenge to the designer who, without these tools, must often arbitrarily
select a geometry, typically one that is based on biological data, and forego exploration and
optimization of the leg system.
The force map is a simple graphical method of viewing the potential locomotive char-
acteristics of a leg design. Because this view is velocity dependent it can be used compare
the effects of leg geometries, masses or actuators on limit dynamic performance. Enlarging
the area of the force map, in particular the area to the right of F, = 0, is paramount to
designing a more capable leg. The effect of increasing distal leg mass for example can be
easily connected to the abrupt decrease in dynamic capability at higher speeds.
Stride planning is perhaps one of the most difficult control tasks that need to be ad-
dressed. Many aspects of the leg performance such as actuator, friction and speed limits
must be considered in making a decision about foot placement. To further complicate mat-
ters this decision must often be made quickly, with input data of questionable quality. The
method proposed in this document is an extensible yet simple approach to generating touch-
down configurations for arbitrary shoulder position and forces profiles.

111
112
Appendix A

Engineering Drawings

This appendix contains the engineering drawings for 26 parts that were custom machined
for the front and rear legs. The naming convention utilized prefixes files with either al_ for
anterior lateral components (parts used in front legs on either side), pl_ for posterior lateral
components (parts used in rear legs on either side) or oc_ for omnes crura (parts used in all
legs). The exception to this convention is the three gears which have no prefix.

113
 Px
/.-+1+
Si
1V
.
 4.1A B

AE

SECTION
A-A

- D

02A

0005

+.002 A B HI

ASMCHU3M
INTERP
BIOMIMETIC
ROBOTICS
CMBRIDGP LAB

CLM
NU ALLOY
ACCORDANCE
WIH
A A al-encoderMount
DEBR~DSNWRE~,E 1
DIELECTRIC
COATIG.
ANOD1ZE Su AT O1RV W E
o ME'AAAAASs
ARE
COATNG
AAER
C B-05 B-02
 .30

2X R.06 Q.45

05 1.594±.001

01.47

SECTION A-A .06

DRAWN DATE
INTERPRET
DRAWINGPERASMEY14.5M JAKEMCKENZIE 3/2/2D0 1
UNLESSOTHERWISESPECIFIED: CHECKED DATE BIOMIMETIC ROBOTICS LAB
DIMENSIONSAREININCHES SANGBAE KIM 4/15/2011
NOTES MASSACHUSETTS
INSTITUTE
OF TECHNOLOGY
TOLERANCES 1. MATERIAL:ABS CAMBRIDGE, MASSACHUSETTS
FRACTIONSDECIMALS ANGLES
± 1/64 XX .01 .5 TITLE
.XXX .005

MACHINED
SURFACES al_encoderWiringDivider
PROPRIETARY
ANDCONFIDENTIAL DEBURR
ANDBREAK
SHARP
EDGES
THSDOCUMENTANDTHEINFORMATIONCONTAINED
HERIN SIZE PARTNO. MDL.REV. DWG.REV.
THIRDANGLEPROJECTION
ISTHESOLE
PROPERTY
OFTHE
MIT.ANYREPRODUCTION
BIOMIMETIC
ROBOTICS
INPARTORASA WHOLE
LABOF
WITHOUT B-03 I B-01
THEWRITTEN
PERMISSION
OFTHEBRLISPROHIBITED. WT SCALE
1:1 DO NOTSCALEDWG SHEET
1 OF I
 N
lIe 141
7
lo
| I
/
N fl
K
K
7$ t-
],
2
2
2
 0.2501±.0001
.50±.01

DRAWN DATE
INTERPRET
DRAWING
PERASMEY14.5M JAKEMCKENZIE 3/1 2/2011
OTHERWISE
UNLESS SPECIFIED: CHECKED DATE BIOMIMETIC ROBOTICS LAB
AREININCHES
DIMENSIONS SANGBAEKIM 4/1 5/2011
NOTES MASSACHUSETTS OF TECHNOLOGY
INSTITUTE
TOLERANCES 1. MATERIAL: CAMBRIDGE, MASSACHUSETTS
FRACTIONSDECIMALS ANGLES
" HARDENEDSTEEL
ADND TE
1/64 XX ±01 5
.XXX± .005
2. MCMASTERPN:
MACHINEDSURFACES 97395A475 al-outputPin
DEBURR
AND BREAK
SHARPEDGES
PROPRIETARY
ANDCONFIDENTIAL
THIS
DOCUMENTANDTHEINFORMATIONCONTAINED
HERIN SIZE PARTNO. MDL REV. DWG.REV.
IS
THE SOLE OFTHEBIOMIMETIC
PROPERTY ROBOTICS
LABOF THIRDANGLEPROJECTION
B-02 B-01
MIT.ANYREPRODUCTION
INPART
ORASA WHOLE WITHOUT
THEWRITTEN
PERMISSION
OFTHEBRLISPROHIBITED. WT SCALE
5:1 DWG SHEET
DO NOTSCALE 1 OF1

2 1
F

FULLSCALE

S 001 A B

! .002 A
63 CI

-t M A B
M0OS

0005

LJ-

SECTION
A-A

menossmarsice ~~~~~~~o BIOMIMETIC

w
OANAtw
45/01
LAB
ROBOTICS
M5AHEisSTUEOTECHNOLOGY
CAMBRIDGE.M
A
5ACHEfTS
FRA~lON irF s WIBH
DEiMAL ALUMINUM
ANLES
OFb3FF F
7075-T73
*
ALLOY
TF 3E
ACCORDANCE
SolidWorks Student Edition. BEKMREGS
Z~uA
cy

COAT[NG:
2
M-Q-^MH aloutputShaft
ammuasmaoacow e v
A:-l-AFAVMmemic Use Only. C

NION0
AER
ANODIZE

ARE CB-5
COATIN4Gwi
p o

cAE- DNTCi.WG SFI

WG

-0
l

2
 10xR.08

R 11
.316 -
2XR.0

20 160

SECTION
A-A

17 -. 1.1
c

2XM3XQ.1THfU
158 .053

- 140 I-B -

4 -P43 BIOMIMETIC
ROBOTICS
4/1/2011
LAB
MAACHS
JKMCEZIAL
INERR~DAWNGERSEY43

RAON
DEIAEANG. ALUMINMALLOY
n
S
An
m t1X.7M75-Ti3
ACCORDANCEYNqH

IE~sAEGE
DEURN 2 AFT
COAIING' LE -OIN
ANODIZE
O1ELECTRIC
o ARE
MENS3ONS C B-03B02
 - 002 AI

dC B
+C.005

2X
M3X0.5THRU
MARKED"C'
|$ A B
6.001

SECTON
HB

Q .00051

SECTION
A-A

BIOMIMEIIC LAB
ROBOTICS

oscamadwHOT~n
SolidWorks Student Edition.
ltademic Use Only.
 F S
FULL
SCALE

-. 059
L002| A 4X .134 [HRUALL
248 X 90
!X 145 ON A 2.416B.C
+002|
A |B|

2 8302.00

4x 250a.0005[HRU0
+.0005 A B

ulAIEA
INEP SMEY..
MJ M CKENZIE 6/23/ O1
U TEI E BIOMIMETICROBOTICS LAB
NSN
IN
-oEANEI.
ESNG
MAEIL3334S
TECHNOLOGY-
MA5ACHUSETS
INSTUE
OF A
CAMBRIDGEMA55ACHU5E[T
M TRA:20/
DECIALS
FRAC110NS ANGLES
'l4 Ax.N2. FINLSH:CL-EAR TTLE
.xx5m PA$SIVATE
SolidWorks
PIEAYANC
Student Edition.
UNsIInEN1AL _IcNEsu-cEsg OCCarrerCap

="A' bendemic Use Only.

HARPEGEW.
DEURARDELE
ECTION
RDNGLPO C PNO, BE
B-01 B-02
 .2000±.0005

.125

.385±.001

DRAWN DATE
INTERPRET
DRAWING
PERASMEY14.5M JAKE
MCKENZIE 6/27/2010
UNLESS
OTHERWISESPECIFIED: CHECKED DATE BIOMIMETIC ROBOTICS LAB
AREIN INCHES
DIMENSIONS SANGBAEKIM 4/15/2011
NOTES MASSACHUSETTS
INSTITUTE
OF TECHNOLOGY
TOLERANCES 1. MATERIAL: 303/304 SS CAMBRIDGE, MASSACHUSETTS
FRACTIONSDECIMALS ANGLES
± 1/64 xx .01 s' 2. FINISH: CLEAR TITLE
.Xxx .005 PASSIVATE

MACHINED
SURFACES occarrierSpacer
PROPRIETARY
ANDCONFIDENTIAL DEBURR
ANDBREAK
SHARP
EDGES
THIS
DOCUMENTANDTHEINFORMATIONCONTAINEDHERIN SIZE PARTNO MDL REV. DWG.REV.
ISTHESOLE
PROPERTY
OFTHEBIOMIMETIC
ROBOTICS
LABOF THIRD
ANGLEPROJECTION
MIT.ANYREPRODUCTION
INPART
ORASA WHOLE WITHOUT A B-01 B-01
THEWRITTEN
PERMISSION
OFTHEBRLISPROHIBITED WT SCALE
5:1 DO NOTSCALE
DWG SHEET
1 OFI
 FULLSCALE 0.2000±.0005

0.125

I I

I I
I II

I .287±.001

DRAWN DATE
INTERPRET
DRAWINGPERASMEY14.5M 6/27/2010
JAKEMCKENZIE
UNLESS
OTHERWISESPECIFIED: CHECKED
SANGBAEKIM
DATE
4/15/2011
BIOMIMETIC ROBOTICS LAB
DIMENSIONS
AREININCHES
NOTES
INSTITUTE
MASSACHUSETTS OF TECHNOLOGY
CAMBRIDGE, MASSACHUSETTS
TOLERANCES 1. MATERIAL:303/304 SS
FRACTIONS DECIMALS ANGLES
1/64 .X .01 .5' TITLE
.xxx .oTs
2. FINISH:
CLEAR
PASSIVATE
MACHINEDSURFACESV" OccarrierSpacerClearance
PROPRIETARY
ANDCONFIDENTIAL DEBURR
AND BREAK
SHARPEDGES
THISDOCUMENT
ANDTHEINFORMATIONCONTAINEDHERIN SIZE PARTNO. MDL.REV. DWG.REV.
THIRDANGLEPROJECTION
ISTHESOLE
PROPERTY
OFTHEBIOMIMETIC
INPART
MIT.ANYREPRODUCTION
ROBOTICS
ORASA WHOLE
LABOF
WITHOUT
B-01 B-01
THEWRITTEN
PERMISSION
OFTHEBRLISPROHIBITED. WT SCALE
5:1 DO NOTSCALE
DWG SHEET
1 OF I
$ B00
C
A-|C
| + 0005 -B C

. .

SECTION"8

BI0O1MEIC
SECTION A-A ROBO11CS
LA

A7=7", oc-output
 1KB

T
-+7001ATB Y 0005

- - - .0020A
A-

SECTION
A-A - -2 -

c
c

'r -

0 A

DEAC
SC2:1r

OBOTIOSOIMM 0 ~ OI
RBB LAB

SECTION B- BOB5-02r~o
 4-
K)P (C '4
-- 4,

//
II
>< )~-~
/
Vr>l
127
10 0
o 0 0o
0 0
VSV
0 0
00000
 A

+1002A B

- /.0005

SECTION
A-A

BIOMIMETIC
ROBOTICS
LAB
C CCI I 151 EASCSACHETTS CI 5151511OF TCH OLO51 C

iMS13OreNENTA91M~AeORARiONCOE MD11l Caemno

noe10 p1_encoderMount
3 1 .2:.1 e. I 0
EC
C CC
C CART
A3AW'CCW
5T3 TE

DLECTRIC
WAB-ODIB-O1
t-A s M ER
CAIAC
 .002 AE EE

519 , 359t 001

FAA
3 3S " 100X 45.0'±.1
5x 60

D 0ON A .[35 B CD -

- '20500t .0005 66 1.8

.0005
9

2X R.(5 ++.0

C
A -
M1.6x0.35
THRU SECTION
A-A

U W
oluN5:NS~(EIINCE5
E

MAT
BIOMIMETIC
GBEKI
LAB
DAE ROBOTICS
MA55ACHUSETt
4/5/211 NSTITUTE
OFTECHNOLOGY
CAMBRIDGE, A
MA55ACIUS66TS
A
FR~tansorcMAL
TLERANC

a Xx&
l~ l ME
_
ALUMINUM
ANl.P
7075-173
ALLOY
-ACCORDANCI
AMS-QQ2A-200/11 plencoderMountOut
snormier~synoconneemAL 2.
nMURANDREA-SIEGE COAUNG:
DIELECTRIC gE 'A'y
ANODIZE
(BLACK)
ALL
NOMLRV.DGRV
B
B
arer
REP~ootcTI~tHNPARTORASA~toEWOCUT AFTER
ARE
sutuROEGCfl-EllOIMilRCOTCSABFDIMENSIONS
COAUNG W CL21 D oSAEW HE~
 .458

M2XO.4 THRU- .150

0 .158

0.163±.001
236±001
4 .003 A B

0.45 A
SECTION A-A

DRAWN DATE
INTERPRET
DRAWING
PERASMEY14.5M JAKEMCKENZIE 3/28/2011
OTHERWISE
UNLESS SPECIFIED: CHECKED DATE
4/15/2011
BIOMIMETIC ROBOTICS LAB
DIMENSIONS
AREININCHES SANGBAEKIM
NOTES MASSACHUSETTS OF TECHNOLOGY
INSTITUTE
TOLERANCES 1. MATERIAL: CAMBRIDGE, MASSACHUSETTS
FRACTIONS
DECIMALS
ANGLES ALUMINUM ALLOY
t 1/64 xX t .01 .5 7075-T73 TITLE
.xx S±.005 ACCORDANCE WITH

MACHINEDSURFACES AMS-Q-A-200/11 pl-encoderShaftEnd

PROPRIETARY
ANDCONFIDENTIAL DEBURR
AND BREAK
SHARP
EDGES 2. COATING:
THSDOCUMENTANDTHEINFORMATION
CONTAINED
HERIN DIELECTRIC
ANODIZE SIZE PARTNO. MDL REV DWG.REV
ISTHESOLE
PROPERTY
OFTHE ROBOTICS
BIOMIMETIC LABOF THIRDANGLEPROJECTION (BLACK) ALL
MIT.ANYREPRODUCTION
INPART
ORASA WHOLEWITHOUT DIMENSIONSARE B-03 B-01
THEWRITTEN
PERMISSION IS
OFTHEBRL PROHIBITED. AFTERCOATING WT SCALE
4:1 DO NOTSCALE
DWG SHEET
1 OF1
ripe - .
Idsiiee
rn~ LZ~

I
-7
+7
IrT
~~JK
.1
 A0 A

CBCBIC SECOCO -P
p
 8X .134THRU
24'81X90
ONA AIJOT
-C
:33 - -- - 08c |$l001|A- B C ' 45

2501± .005

(9 (9 S00

0 0
0 0 2A 23c,

.A88

002 A -B

4TP REGR - WE P : N E BIOMIMETIC

mM4/5 LAB
11ROBOTICS
MBAA
CEBB
B FF O TECHOLOGY
A
BCT - ABB N B AB

MACHINED3DREA pl-outputPin
P I BDLEPROJE-N
. O

W -S~~: OOSCALEDW - -E704

 -0.001A B

1+ .001 A B 1.0005

B-,
SECTION
A-A

24XM4X0.7THRU

002 A

DETAIL
C
SCALE4: 1

BIOMIMETIC
ROBOTICS
LAB
olu~slms~tl~css
SANB~tM
415/011
MA5ACHUSIETT51NSTITUTE
OFTECHNOLOGY A
'LANMATERIAL.
stacous EC:Mts ALUMINUM
A~gs ALLOY
A BIGAS~-JET

-"ACCORDANCE
rIH
SECTION
B-B plringGearmount
-.- DEE 2. COATING

ME~ONSAREC B-03 B-01

AFTER
COATINGwr caa cooCAE G M
CF

21
136
Appendix B

Matlab Code Listing

The code that follows is capable of producing the shape dependent force envelopes and force
maps used in the body of this document. It is broken up into six files.

- link.m -the lowest level code creates link objects for a serial chain

- chain m - arranges link objects into a chain object which is later used to derive the
equations of motion

- prof Design. m - stride profile designer, outputs force and shoulder trajectories

- strideInit.m stride initialization, calculates equations of motion and solves for

the parameters of the SDFE in each configuration

- strideComp.m stride computations, calculates the ideal foot placement locations

based on the results of stride Init .m

envPlot . m - plots the force production envelopes given the output of stride Init . m

137
 1 classdef link

5 properties

6 type = ' ';

7 1 = 0;
8 lc = 0;

9 N = eye(3);

10 ul = Inf;

111 1 = -Inf;

12 end

13

14 methods

15

16

17 function obj = link(type, 1, lc, m, I)

18 if(nargin > 0)

19 obj.type = type;

20 obj.1 = 1;

21 obj.lc = lc;

22 obj.N = diag ([m,m, I]);

23 end

24 end

25

26

27 function obj = limits(obj, 11, ul)

28 if (nargin == 3)

29 if 11 >= ul

30 temp = 11;

31 11 = ul;

32 ul = temp;

33 end

34

35 obj.ll = 11;

36 obj.ul = ul;

37 end

38 end

39

40

41 end

42 end

138
 1 classdef chain

2
3

5 properties

7 L

10 ic [];
11 g = [0;-9.81;0];

12 qs = [;
13 k = [;
14 tau = [];
15 fx = [];
16 fy = [;
17 fecc = [;
18 end

19

20 methods

21

22 function obj = chain(LN)

23

24 if(nargin > 0)

25 n = length(LN);

26

27 obj.L = LN;

28

29 obj.ic = zeros(2*n,1);

30 obj.qs = zeros(n,i);

31 obj.k = diag(zeros(i,n));

32 obj.tau = zeros(n,1);

33 obj.fx = zeros(n,1);

34 obj.fy = zeros(n,1);

35 obj.fecc = zeros(n,1);

36

37 end

38

39 end
40 end

+1 end

139
 l function profDesign

10

11

12
13 steps = 20;

14 dur = 0.6;

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30 c = sym( , [7,1]);

31 u = sym( ,[5,1]);

32 syms t ul u2 u3 u4 u5 real;

33 c([1]) = -6;

34 c([2]) = 6;

35 c(3]) 0;

36 c 4]) -12;

37 c([5]) 18;

38 c ([6]) -6;

39 c([7]) = 0;

40 fly-nom = poly2sym(c(1:3),t);

41 flxnom = poly2sym(c(4:7),t);

42 flx-asym = flxnom + 6*u(5)*t^2 6*u(5)*t;

43 fly = u(3)*compose(fly nom, u(2)*(t-u(1)), t, t);

44 flx = u(4)*compose(flxasym, u(2)*(t-u(l)), t, t);

45

140
 2 2 5 7 5
fx = matlabFunction(subs(flx, {ul,u2,u3,u4,u5}, {0,l/. , 0, , .75}));
2 2 75
fy = matlabFunction(subs(fly, {ul,u2,u3,u4,u5, {,1/. , 50, , .75}));

t steps = linspace(O,dur,steps);

fx steps = fx(tsteps);

fy-steps = fy(t-steps);

method

switch method

case 0

m = 19.75;

x ic [0
0
2.4

-. 45];

args - {m, fx, fy};

[t,x] ode45(@numint,t steps,x_ic, [],args)

xp = x(:,1);

yp = x(:,2);

141
case 1

v = 1;

ht = 0.45;

qvar = sym '

qp-i = real(sort(eval(solve(.320*sin(qvar)+.240*sin(-10*pi/180+qvar)-ht))));

xp i = .320*cos(qp_i(2))+.240*cos(-10*pi/180+qp-i(2));

xp-f = .320*cos (qp-i(1))+.240*cos(-10*pi/180+qp-i(1));

xp = linspace(xpi,xp_f,steps);

yp = ones(size(t steps))*ht;

end

save('

end

function dxdt numint(t,x,args)

m = args{1};

g = 9.81;

fx = args{2}(t);

fy = args{3}(t);

dxdt zeros(size(x));

dxdt (1:2) x(3:4);

dxdt (3) fx/m;

dxdt (4) (fy-m*g)/m;

end

142
 l function strideInit (inv, in fname)

8 close all

9 clc

10

11 initxp = 0;

12 inityp = 0;
13 initxv = in v;

14

15 is = 0.004660;
16 i k = 0.000987;

17

18 savefile = in-fname;

19

20

21

22

23

24

25 ml = 0.180;

26 11 = 0.320;
27 lcl = 11/2;
28 Ic1 = 1/12*ml*(11^2+0.02^2);

29

30 m2 = 0.200;

31 12 = 0.240;

32 lc2 = 12/2;
33 Ic2 = 1/12*m2*(12^2+0.02^2);

34

35

36

37 Li = link( ,11,lcl,ml,Icl);

38 L2 = link( , 12,lc2,m2,Ic2);

39 Li = Ll.limits(45*pi/180, 135*pi/180);

40 L2 = L2.limits(-45*pi/180, -10*pi/180);

41

42 C = chain({Ll, L2});
43 savechain(C)

44 load( )
45 load(

143
46

47

48

49

50

51

52

53 res-q = length(xp);

54 res t = 20;

55

56 args = {M fun, Qfun, Gk, Gst, C};

57

58 counter = 1;

59 cmap = colormap(jet(res-q));

60

61 qls2 = sym('

62 q2s2 = symV );

63

64 x = xp + initxp;

65 y = yp + inityp;

66

67

68 for i = 1:res-q

69 options = optimset (

70 qfunc = matlabFunction([C.L{1}.l*sin(qls2)+C.L{2}.1*sin(qls2+q2s2)
y(i);C.L{1}.1*cos(qls2)+C.L{2}.l*cos (qls2+q2s2) - x(i)], ,{[qls2;q2s2]});

71

72 qres fsolve(qfunc, [pi;-pi] ,options);

73

74 ql(i) - qres(l);

75 q2(i) = qres(2);

76 end

77

78

79

80 clear ql q2

81 qn = 20;

82 qrl = linspace(45*pi/180,135*pi/180,qn);

83 qr2 = linspace(-45*pi/180,-10*pi/180,qn);

84 for i = 1:qn

85 for j = 1:qn

86 q1 (qn* (i-1)+j) = gr1(i);

87 q2(qn*(i-1)+j) = qr2(j);

88 end

89 end

144
res-q = length(ql);

cmap = colormap(jet(length(ql)));

for j = 1:length(ql)

[nlx, nly] = pol2cart(ql(j),C.L{1}.1);

[n2x, n2y] = pol2cart(ql(j)+q2(j),C.L{2}.l);

n2x = n2x + nix;

n2y = n2y + nly;

figure(l), hold on

configplot = plot([O,nlx,n2x],[0,nly,n2y], ); axis equal

set(configplot, , cmap(j,:))

end

tau s

tau k

for i = 1:res. _q
counter 1;

for velo initxv;

C.tau(1) = taus;

C.tau(2) = tauk;

C.ic = [ql(i);q2(i);O;O];

args = {Mfun, Qfun, Gk, Gst, C);

q [ql(i); q2(i)];
vp = [velo; 0];

ap - [0; 0];

145
[q,qd,qdd] = remap(q,vpap,args);

C. ic = [ql(i);q2(i);qd(1);qd(2)];

C.tau = [0;01;
args = {Mfun, Q fun, Gk, Gst, C};

[Fpz] = fcalc (q,qd, qdd, args);

C.tau = [1;01;
args = {Mfun, Qfun, Gk, Gst, C};

[Fps] = fcalc(q,qd,qdd,args);

C.tau = [0;1];
args = {M fun, Qfun, Gk, Gst, C};

[F-pk] = fcalc(q,qd,qdd,args);

C.tau = [0;01;

args {M-fun, Q-fun, Gk, Gst, C};

[tau, F pf] = tcalc(q,qd,qdd,args);

T = [-FWpk(1)+Fpz(1), -F-ps(1)+F-pz(1), -F_pz(1);

-F-pk (2)+F-pz(2), -F_ps(2)+Fpz(2), -F-pz(2);

0, 0, 1];

tau si = sum(qdd)j*is;

tauki = qdd(2)*ik;

tau smax = 30;

tau kmax = 150;

T*[tau kmax-tauki ;tau_smax-tau si ;l];

T*[tau-kmax-tauki ;-tausmax-tau si ;1];

T*[-tau kmax-tau_ki ;tausmax-tausi ;1];

T*[-tau kmax-tau-ki ;-tausmax-tau-si ;1];

-F-pf;

[.9*300;300;1] -F-pf;

[-.9*300;300;1] -Fpf;

D(i).F f -F_pf(1:2);

D(i) .T T;
D(i) .Ti =T^-1;

146
 D(i).tau-si = tau_si;

D(i).tau ki = tauki;

figure(2), hold on, axis equal

3
plot ([pl (1) ,p2 (1) p41
(1) p3(1) pl (1)], [pl (2) ,p2 (2) ,p4 (2) p (2) pl (2)],
plot([s2(l),sl(l),s3(1)], [s2(2),sl(2),s3(2)],' ')

plot(-F pf(l),-F_pf(2),

counter = counter+l;
end
end

save ([ ,savefile], )

end

function [F_p] = fcalc(q,qd,qdd,args)

n = length(q);

Ms = argsjl}([q;qd]);

Qs args(2}([q;qd]);

C = args{5};

for k = 1:n

Gfk} = args{3}{k}([q;qd]);

Gstjk} = args{4}{k)([q;qd]);

end

147
 qs = C.qs;

K = C.k;

230 T = [0; ...

C.tau (2)];

231 FX - [0; ...

0];
232 FY = [0; . ..

0];
233 TS = C.tau(1);

234

235

236

237

238

239 SF = zeros (n, 1);

240 for k = 1:n

241 qsum = sum(q(l:k));

242 FK = C.g*C.L{k}.N(1)+[FX(k);FY (k);C.fecc(k)*(FY(k)*cos(qsum)-FX(k)*sin(qsum))];

213 SF = SF + G{k}'*(FK);

244 end

245 clear FK qsum

246

247

248

249

250

251

252 qsum = sum(q(l:n));

253 r = [(C.L{n}.l-C.L{n}.lc)*cos(qsum); (C.L{n}.l-C.L{n}.lc)*sin(qsum); 0];

254

255

256

257

258

259

260

261 TE = zeros (n,l);

148
 for k = 1:n

Gst{k} = mat2cell(Gst{k}, 3*ones(1,n), ones(l,n));

BLAH {zeros(3,1), zeros(3,1)};

for i = 1:n

for j = 1:n
BLAH{1,i} = BLAH{1,i} + qd(j).*Gst{k}{j,i};

end
end

Gj{k} = zeros(3,1);

for i = 1:n

Gj{k} = Gj{k} + BLAH{l,i}.*qd(i);

end
clear BLAH

end

for k = n:-1:1
qsum = sum(q(l:k));

Fk = C.g*C.L{k}.N(1)+[FX(k);FY(k);TE(k) ];

if k == n

Ft{k} = C.L{k}.N*(G{k}*qdd+Gj{k}) - Fk - [O;Q;T(2)] [Q;C;TS];

296

297 else

Ft{k} = C.L{k}.N*(Gf{k}*qdd+Gj{k}) - Fk + [0;O;T(2)];

end
end

Fjlx = sym(

Fjly = sym(

Fj2x = sym(

Fj2y = sym(

Fe2x = sym(

149
Fe2y = sym(

r = [(-C.L{2}.lc)*cos(sum(q(1:2))); (-C.L{2}.lc)*sin(sum(q(1:2))); 0];

Fj22 = [Fj2x; Fj2y; 0] + cross(r,[Fj2x;Fj2y;0]);

r = [(C.L{2}.1-C.L{2}.lc)*cos(sum(q(1:2))); (C.L{2}.1-C.L{2}.lc)*sin(sum(q(1:2))); 0];

Fe22 = [Fe2x; Fe2y; 0] + cross(r, [Fe2x;Fe2y;0]);

r = [(C.L{1}.l-C.L{1}.lc)*cos(q(1)); (C.L{1}.l-C.L{1}.lc)*sin(q(1)); 0];

Fj2l = [-Fj2x; -Fj2y; 01 + cross(r,[-Fj2x; -Fj2y; 0]);

r = [(-C.L{1}.lc)*cos(q(1)); (-C.L{1}.lc)*sin(q(1)); 0];

Fjll = [Fjlx; Fjly; 0] + cross(r, [Fjlx; Fjly; 0]);

forces solve(Fj22+Fe22-Ft{2},Fjll+Fj2l-Ft{1});

Fj{l} = [eval(forces.Fjlx); eval(forces.Fjly); 0];

Fj{2} = [eval(forces.Fj2x); eval(forces.Fj2y); 0];

Fe = [eval(forces.Fe2x); eval(forces.Fe2y); 0];

F-p = Fe;

r = [(C.L{n}.1-C.L{n}.lc)*cos(sum(q(1:2))); (C.L{n}.1-C.L{n}.lc)*sin(sum(q(1:2))); 0];

r22 S[ (C.L{2} .1-C.L{2}.lc)*cos(sum(q(1:2) )); (C.L{2} .1-C.L{2} .lc)*sin(sum(q(1:2))); 0];

r21 = [(C.L{1}.1-C.L{1}.lc)*cos(q(1)); (C.L{1}.l-C.L{1}.lc)*sin(q(1)); 0];

rll [(-C.L{1}.lc)*cos(q(1)); (-C.L{1}.lc)*sin(q(1)); 0];

GRF = Fj{1};

150
[nix, nly] = pol2cart (q(1) ,C.L {} .l) ;

[n2x, n2y] = pol2cart(q(l)+q(2),C.L{2}.l);

n2x = n2x + nix;

n2y = n2y + nly;

end

function [tau, F pf] = tcalc(q,qdqdd,args)

151
 n length(q);

Ms args{1}([q;qd]);

Qs = args{2}([q;qd]);

C args{5};

for k = 1:n

G5k} = args{3}{k}([q;qd]);

Gst{k} = args{4}{k}([q;qd]);

end

= C.qs;

= C.k;

430 T = [0; ...

C.tau (2)];

431 FX = [0; ...

0];
432 FY = [0; ...

0];
433 TS = C.tau(1);

SF = zeros (n,1);

for k = 1:n

qsum = sum(q(1:k));

FK = C.g*C.L{k}.N(1)+[FX(k);FY(k);C.fecc(k)* (FY(k)*cos(qsum)-FX(k)*sin(qsum))];

SF = SF + G{kl'*(FK);

152
end
clear FK qsum

qsum = sum(q(1:n));

r = [(C.L{n}.l-C.L{n}.lc)*cos(qsum); (C.L{n}.l-C.L{n}.lc)*sin(qsum); 0];

TE - zeros (n,1);

for k = 1:n

Gst{k} mat2cell(Gst{k}, 3*ones (1,n), ones(1,n));

BLAH {zeros(3,1), zeros(3,1)};
for i 1:n

for j = 1:n
BLAH{1,i} = BLAH{1,i} + qd(j).*Gst{k}{j,i};

end
end

Gj{k} = zeros(3,1);

for i 1:n

Gj{k} Gj{k} + BLAH{1,i}.*qd(i);

end
clear BLAH

end

t k = sym(;

153
ts = sym('t '

for k = n:-1:1

qsum = sum(q(1:k));

Fk = C.g*C.L{k}.N(1)+[FX(k);FY(k);TE(k)];

if k == n

Ft{k} = C.L{k}.N*(G{k}*qdd+Gj{k}) - Fk - [0;0;t_kl - [0;0;ts];

else

Ft{k} = C.L{k}.N*(G{k}*qdd+Gj{k}) - Fk + [0;0;tk];

end
end

Fjlx = sym('

Fjly = sym( 'C;

Fj2x = sym(

Fj2y = sym(

Fe2x = sym(

Fe2y = sym(

r = [(-C.L{2}.lc)*cos(sum(q(1:2))); (-C.L{2}.lc)*sin(sum(q(1:2))); 0];

Fj22 = [Fj2x; Fj2y; 0] + cross(r, [Fj2x;Fj2y;0]);

r =[ (C.L{2}.1-C.L{2}.lc)*cos(sum(q(1:2))); (C.L{2}.1-C.L{2}.lc)*sin(sum(q(1:2))); 0];

Fe22 = [Fe2x; Fe2y; 0] + cross(r, [Fe2x;Fe2y;0]);

r = [(C.L{1}.1-C.L{l}.lc)*cos(q(l)); (C.L{l}.1-C.L{1}.lc)*sin(q(1)); 0];

Fj21 = [-Fj2x; -Fj2y; 0] + cross (r, [-Fj2x; -Fj2y; 0]);

r = [(-C.L1}.lc)*cos(q(1)); (-C.L{1}.lc)*sin(q(1)); 0];

Fjll = [0; 0; 0] + cross(r, [0; 0; 0]);

forces solve(Fj22+Fe22-Ft{2},Fjll+Fj2l-Ft{1});

Fj{2} = [eval(forces.Fj2x); eval(forces.Fj2y); 0];

154
Fe = [eval(forces.Fe2x); eval(forces.Fe2y); 0];
tau [eval(forces.tk); eval(forces.ts)];

F pf = Fe;

r = [(C.L{n}.l-C.L{n}.lc)*cos(sum(q(1:2))); (C.L{n}.l-C.L{n}.lc)*sin(sum(q(1:2))); 0];

r22 = [(C.L{2}.1-C.L{2}.lc)*cos(sum(q(1:2))); (C.L{2}.1-C.L{2}.lc)*sin(sum(q(1:2))); 0];
r21 = [(C.L{1}.l-C.L{1}.lc)*cos(q(1)); (C.L{1}.l-C.L{1}.lc)*sin(q(1)); 0];
r11 = [(-C.L{i}.lc)*cos(q(i)); (-C.L{i}.lc)*sin(q(i)); 0];

[nix, nly] = pol2cart(q(i),C.L{1}.i);

[n2x, n2y] = pol2cart(q(1)+q(2),C.L{2}.1);

n2x = n2x + nix;

n2y = n2y + nly;

155
end

function [q,qd,qdd] = remap(q,vp,ap,args)

n = length(q);

C = args{5};

for k = 1:n

G{k} = args{3}{k}([q;zeros(size(q))]);

Gst{k} = args{4}{k}([q;zeros(size(q))]);

end

vqd = sym(' [n,1]);

vqdd = sym(' [n,11);

156
qsum = sum(q(1:n));

r = [(C.L{n}.1-C.L{n}.lc)*cos(qsum); (C.L{n}.l-C.L{n}.lc)*sin(qsum); 0];

w = [0;0;sum(vqd)];

vp-sym = G{n}*vqd+cross(w,r);

vp-sym = vp-sym(1:2) - vp;

qd sym = solve (vpsym);

qd = zeros (n,1);

qd(l) = eval(qdsym.vqdl);

qd(2) = eval(qd sym.vqd2);

for k = 1:n

Gst{k} = mat2cell(Gst{k}, 3*ones(1,n), ones (ln));

BLAH = {zeros(3,1), zeros(3,1)};

for i =:n

for j = 1:n
BLAH{1,i} = BLAH{l,i} + qd(j).*Gst{l}{j,i};

end
end

Gj{k} = zeros(3,1);

for i = 1:n

Gj{k} = Gj{k} + BLAH{l,i}.*qd(i);

end
clear BLAH

end

qsum = sum(q(l:n));

An = G{n}*vqdd+Gj{n};

w = [0;0;sum(qd)];

ap-sym An + cross([0;0;An(3)],r) + cross(w,cross(w,r));

157
ap-sym = ap-sym(1:2) - ap;

qdd sym = solve(ap-sym);

qdd = zeros(n,1);

qdd(1) = eval(qddsym.vqddl);

qdd(2) = eval(qddsym.vqdd2);

end

function savechain(C)

n = length([C.L{:}]);

if(true)

cfgvec [];

sym( [1,n]);

sym( [1,n]);
sym( [1,n]);
sym( '[1,n]);

sym(' '[1,n]);

for i=l:n

eval ([' ,num2str(i), ,num2str(i),

if C.L{i}.type ==
cfgvec = [cfgvec,1];

else
cfgvec = [cfgvec,O];

end

qstrn =

qstrm =

for j = 1:i
if C.L{j}.type

qstrn = [qstrn, ,num2str(j),

if j < i | i == 1

qstrm = [qstrm,' ,num2str(j),' ;

end

158
714 end
715 end
716 clear j

717

718

719

720

721 if i == 1

722 p (i) = 0;
723 x (i) = 0;

724 y (i) = 0;
725 else

726 p (i) = qstrm;

727

728 if C.L{i}.type ==

729 if C.L{i-1}.type ==
730 x(i) = x(i-l)+C.L{i-1}.l*cos(eval(qstrm));

731 y(i) = y(i-l)+C.L{i-1}.l*sin(eval(qstrm));

732 else
733 lastR find(cfgvec(l:end-1),, '- ;
734

735 dl =

736 dp = 0;

737 for j lastR+l:i-1

738 dl = [dl, ,num2str (j), ';

739 dp =dp + C.L{j}.l;

740 end
741 dl = dl(3:end-1);

742

743 x(i) = x(lastR)+(C.L{lastR}.l+eval(dl)+dp)*cos(eval(qstrm));

744 y(i) = y(lastR)+(C.L{lastR}.l+eval(dl)+dp)*sin(eval(qstrm));

745 end
746 else
747 if C.L{i-1}.type

748 x(i) = x(i-l)+C.L{i-1}.l*cos(eval(qstrm));

749 y(i) = y(i-l)+C.L{i-1}.l*sin(eval(qstrm));

750 else
751 lastR - find(cfgvec(l:end-1),l,' );

752

753 dl =

754 dp = 0;

755 for j = lastR+l:i-1

756 dl - [dl, ,num2str(j), ];

757 dp dp + C.L{j}.l;

758 end

159
 dl = dl(3:end-1);

x(i) = x(lastR)+(C.L{lastR}.l+eval(dl)+dp)*cos(eval(qstrm));

y(i) = y(lastR)+(C.L{lastR}.l+eval(dl)+dp)*sin(eval(qstrm));

end
end
end

if C.L{i}.type == 'n'

xc(i) = x(i) + C.L{i}.l c*cos(eval(qstrn));

yc(i) = y (i) + C.L{i}.1 c*sin(eval(qstrn));

else
xc(i) = x(i) + (eval(['',num2str(i)])+C.L{i}.l-C.L{i}.lc)*cos(p(i));

yc(i) = y(i) + (eval([' ',num2str(i)])+C.L{.i}.1-C.L{i}.lc)*sin(p(i));

end

end

M = sym(' ', [n,n]);

for k = 1:n

G = sym('-', [3,n]);

G(3,:) = ones(1,n);

if k < n

G(:,k+l:n) = zeros(3,n-k);

end

for i = 1:k

if C.L{i}.type

G(l,i) = y(i)-yc(k);

G(2,i) = -x(i)+xc(k);

else
G(1, i) cos (p(i) );
G(2, i) sin (p(i))

G(3, i) 0;

end
end
clear i

Gs = {};

for i = 1:n

160
 for j = l:n
Gs{i,j} = eval([' ,num2str(j), ,num2str(i),' ']);
end
end
clear i j

Gst{k} = sym(' , [3*n,n]);

if k == 1

M = eval ([ ,num2str (k),

for i = 1:n

for j = 1:n
H{i,j}

eval ,num2str(k) ,num2str(i),', ',num2str(j), 1);

Gst{k}((i-l)*3+1: (i-l)*3+3,j) = Gs{i, j};
end
end
Gk{k} = G;

else
M = evalW ,num2str (k) ,
for i = 1:n

for j = 1:n
H{i,j} eval([ ,num2str(i), num2str (j),
num2str(k), ,num2str(i), ,num2str(j), ;
Gst{k}((i-l)*3+1:(i-l)*3+3,j) Gs{i, j};
end
end
Gk{k}
end
clear i j

end
clear k

qtstr ;

varlist(l) = ql;
varlist(n+1) = qld;
for i = 2:n

varlist (i) = eval([' ,num2str(i)]);

varlist(n+i) = eval([ ,num2str(i),' ']);

qtstr = [qtstr, ',num2str(i), ':'];

end

161
 Q = qthq(H,eval([' ',qtstr, ]));

M fun = matlabFunction(M, ',{varlist'});

Q fun = matlabFunction(Q, ',{varlist'});

for k = 1:n

evalCH num2str(k),

a num2str (k), ;]
eval([ ,num2str(k),'

',num2str(k),'

end

save

else
load

end

end

function Q qthq(H, qd)

n = length(qd);

P = cell([l,n]);

for i = 1:n

for j = 1:n
if j==l

P{i} zeros(n,1);

end
P{i} = P{i} + qd(j)*H{i,j};

end
end
clear i j

Q = 0;
for i = 1:n
Q = Q + P{i}*qd(i);
end

end

162
profile clear

profile on
tic

close

clear

clc

= 20;

load(

fx = fx_ _steps;

fy = fy_ _steps;

figure

plot([0:n-1],fy, [0:n-1l],fx)

load;

qsteps length(D);
R = zeros (qsteps, n);

Tlm = 150;

T2m = 150;

mu = .9;

for i = 1:qsteps

for j = 1:n;

T = D(i).Ti*[fx(j);fy(j);l];

if abs(T(l)) <= Tlm && abs(T(2)) <= T2m

if fy(j) > D(i).F_f(2) && abs((fx(j)-D(i).F_f(l))/(fy(j)-D(i).F_f(2))) <= mu

R(i,j) = 1;

end

end

end

end

163
figure

spy (R)

testSet = find(R(:,1))

pmax = 0;

psum = 0;

for p = 1:length(testSet)

if sum(diag(R,-testSet(p)+1)) >= psum

pmax = p;

psum = sum(diag(R,-testSet (p)+l));

end
end

tc 1
Tau = zeros(3,n);

for j = 1:n

Tau(:,j) = D(j+tc-1).Ti*[fx(j); fy(j); 1];

end

Tau(l,:) = Tau(1,:)/(5.8*(1/.87));

Tau(2, :) = Tau(2, :)/ (5.8);

xlswrite( ,Tau(1:2, :))

figure

plot([0:n-1],Tau(1,:),[0:n-1],Tau(2,:))

hold on

plot([0,n-1],[Tlm/5.8,Tlm/5.8],'

plot([0,n-l],[-Tlm/5.8,-Tlm/5.8], '

t = toc

profile off

164
 t close all

2 clear all
3 clc

6 Ds={

7
8 for n 1:length(Ds);
9 load(Ds{n});

10

it qsteps = length(D);

12

13 Tkm = 23.2*5.8;

14 Tsm = 23.2;

15 mu = 0.9;

16

17 fx = linspace (-1250, 1250, 500);

18 fy = linspace(O, 2500, 500);
19 P = zeros(length(fx), length(fy));
20 Q = P;

21

22 for i = 1:qsteps
23 Tk max = Tkm-abs(D(i).tau_ ki);

24 Ts max = Tsm-abs(D(i).tau_si);
25

26 for j = 1:length(fx)

27 for k = 1:length(fy)
28 T = D (i) .Ti* [fx (j) ;fy (k) ;1]

29

30 if abs(T(l)) <= Tk max && abs(T(2)) <= Ts max

31

32 if fy(k) > D(i) .F_f(2) && abs((fx(j)-D(i) .Ff(l))/(fy(k)-D(i) .F-f(2)))

<= mu

33 P(j,k) = 1;

34 Q(j,k) = Q(j,k)+1;

35 end

36 end

37 end

38 end

39

4o end

41

42

43

-14

165
figure

contourf (fx, fy, Q', 1, , 1)

colormap([1,1,1;.8,.8,.8]);

set (gca, r 10,f ' , 'I t l )

set (gcf,

set (gcf, l n, [0, 0,2.6,2.4])

set (gcf, [500,500 ,2.6*100,2.4*100

grid on

xlabelj

ylabel

figure

contour (fx, fy,Q 10, , ,linspace (1,400,10));

set (gca, ',10

l,

set (gcf,

set (gcf, [0,0,2.6,2.4]

set(gcf, ',[500,500,2.6*100,2.4*100])

grid on

xlabel (

ylabel('

end

166
Bibliography

[1] R. M. Alexander. The gaits of bipedal and quadrupedal animals. The International
Journalof Robotics Research, 3(2):49-59, 1984.
[2] K. Aminian, B. Najafi, C. Bula, PE Leyvraz, and P Robert. Spatio-temporal param-
eters of gait measured by an ambulatory system using miniature gyroscopes. Journalof
Biomechanics, 35(5):689-699, 2002.

[3] K. Autumn, Y A. Liang, S. T Hsieh, W Zesch, W P. Chan, T W Kenny, R. Fearing,

and R.J. Full. Adhesive force of a single gecko foot-hair. Nature, 405(6787):681-684,
2000.

[4] J. E. Bertram and A. Gutmann. Motions of the running horse and cheetah revisited:
fundamental mechanics of the transverse and rotary gallop. Journalof The Royal Society
Interface, 6(35):549-559, October 2008.

[5] A. A. Biewener. Biomechanics of mammalian terrestrial locomotion. Science,

250(4984):1097, 1990.

[6] A. A. Biewener. Animal Locomotion. Oxford University Press, New York, January 2003.

[7] E. L Brancato. Estimation of lifetime expectancies of motors. ElectricalInsulationMag-

azine, IEEE, 8(3):5-13, 1992.

[8] G. Caurin and N. Tschichold-Gurman. The development of a robot terrain interaction

system for walking machines. Robotics and Automation, 1994. Proceedings., 1994 IEEE
InternationalConference on, pages 1013-1018 vol. 2, 1994.

[9] J. Charteris,
D. Leach, and C. Taves. Comparative kinematic analysis of bipedal and
quadrupedal locomotion: a cyclographic technique. Journalof Anatomy, 128(Pt 4):803,
1979.

[10] K. J. Chi and V Louise Roth. Scaling and mechanics of carnivoran footpads reveal
the principles of footpad design. Journalof The Royal Society Interface, 7(49):1145-1155,
June 2010.
[11] Cosworth. Traction Circles. Racecar Engineering, 21(2):98, February 2011.

[12] T M. Dauphinee and H. Preston-Thomas. A Copper Resistance Temperature Scale.

Review of Scientific Instruments, 25(9):884, 1954.

167
[13] M. Ehsani, Y Gao, andJ. M. Miller. Hybrid electric vehicles: Architecture and motor
drives. Proceedings of the IEEE, 95(4):719-728, 2007.

[14] M. S. Erden. Optimal Protraction of a Biologically Inspired Robot Leg. Journal of

Intelligent & Robotic Systems, 64(3-4):301-322, February 2011.

[15] E. C. Frederick. Physiological and ergonomics factors in running shoe design. Applied
Ergonomics, 15(4):281-287, 1984.

[16] M. Gienger, F Pfeiffer, and H. Cruse. Entwurf und Realisierung einer zweibeinigen
Laufmaschine. In Autonomes Laufen, page 295. VDI-Verlag, Dusseldorf, March 2005.

[17] T Hemker, M. Stelzer, 0. von Stryk, and H. Sakamoto. Efficient Walking Speed
Optimization of a Humanoid Robot. The InternationalJournal of Robotics Research,
28(2):303-314, February 2009.

[18] M. Hildebrand. Motions of the running cheetah and horse. Journalof Mammalogy,
40(4):481-495, 1959.

[19] N. Hogan. Impedance Control: An Approach to Manipulation. In American Control

Conference, 1984, pages 304-313, 1984.

[20] N. Hogan. Impedance control: an approach to manipulation: part I-theory, Part II-
implementation, Part II-Applcation. Journal of Dynamic Systems, Measurement and Control,
107(1):1-24, 1985.

[21] N. Hogan. Stable execution of contact tasks using impedance control. Robotics and
Automation. Proceedings. 1987 IEEE InternationalConference on, 4:1047-1054, 1987.

[22] P E. Hudson, S. A. Corr, R. C. Payne-Davis, S. N. Clancy, E. Lane, and A. M. Wilson.

Functional anatomy of the cheetah (Acinonyx jubatus) forelimb. Journalof Anatomy,
218(4):375-385, February 2011.

[23] I. W Hunter,J. M. Hollerbach, andJ. Ballantyne. A comparative analysis of actuator

technologies for robotics. Robotics Review, 2, 1991.

[24] S. H. Hyon. Compliant terrain adaptation for biped humanoids without measuring
ground surface and contact forces. Robotics, IEEE Transactionson, 25(1):171-178, 2009.

[25] R. K. Josephson. Contraction dynamics and power output of skeletal muscle. Annual
review of physiology, 55(l):527-546, 1993.

[26] K. Kaneko, K. Harada, F Kanehiro, G. Miyamori, and K. Akachi. Humanoid robot

HRP-3. In Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ InternationalCon-
ference on, pages 2471-2478, 2008.

[27] J. G. Karssen, M. Haberland, M. Wisse, and S. Kim. The optimal swing-leg retraction
rate for running. Robotics and Automation (ICRA), 2011 IEEE International Conference on,
pages 4000-4006, 2011.

168
[28] K. C. Kim, K. Kim, H.J. Kim, andJ. Lee. Demagnetization analysis of permanent
magnets according to rotor types of interior permanent magnet synchronous motor.
IEEE Transactionson Magnetics, 45(6):2799-2802, 2009.

[29] H. Knuesel, H. Geyer, and A. Seyfarth. Influence of swing leg movement on running
stability. Human Movement Science, 24(4):532-543, August 2005.

[30] E Lacquaniti, M. Le Taillanter, L. Lopiano, and C. Maioli. The control of limb ge-
ometry in cat posture. The Journal of Physiology, 426(1):177, 1990.

[31] D. A. Lawrence. Actuator limitations on achievable manipulator impedance. In

Robotics and Automation, 1989. Proceedings., 1989 IEEE International Conference on, pages
560-565, 1989.

[32] D. V Lee and S. G. Meek. Directionally compliant legs influence the intrinsic pitch
behaviour of a trotting quadruped. Proceedings of the Royal Society B: Biological Sciences,
272(1563):567-572, March 2005.

[33] W Lee. Designing articulatedlegsfor running machines. PhD thesis, Massachusetts Institute
of Technology, 1990.

[34] W Letko. Preliminary Wind-Tunnel Investigation at Low Speed of Stability and Con-
trol Characteristics of Swept-Back Wings. Technical report, 1946.

[35] S. Lohmeier, T Buschmann, H. Ulbrich, and F Pfeiffer. Modular joint design for
performance enhanced humanoid robot LOLA. Robotics and Automation, 2006. ICRA
2006. Proceedings 2006 IEEE InternationalConference on, pages 88-93, 2006.

[36] B. G. Lovegrove. Locomotor Mode, Maximum Running Speed, and Basal Metabolic
Rate in Placental Mammals. Physiological and Biochemical Zoology, 77(6):916-928,
November 2004.

[37] R. Lydekker and P. L. Sclater. The Royal Natural History, volume 1. E Warne, London,
January 1893.

[38] W A. Mair and D. L. Birdsall. Aircraft Performance. Cambridge Aerospace Series. Cam-
bridge University Press, March 1996.

[39] P E. Martin and R. C. Nelson. The effect of carried loads on the walking patterns of
men and women. Ergonomics, 29(10):1191-1202, October 1986.

[40] J. M. R. Martinez andJ. Duffy. An application of screw algebra to the acceleration

analysis of serial chains. Mechanism and machine theory, 31(4):445-457, 1996.

[41] T. McGeer. Dynamics and control of bipedal locomotion. Ji theor Biol, 163:277-314,
1993.

[42] T A. McMahon, G. Valiant, and E. C. Frederick. Groucho running. Journalof Applied

Physiology, 62(6):2326-2337, 1987.

169
[43] W F Milliken and D. L. Milliken. Race Car Vehicle Dynamics. Society of Automotive
Engineers, January 1997.

[44] T F Novacheck. The biomechanics of running. Gait &posture, 7(1):77-95, 1998.

[45] G.J. O'Hara. Mechanical impedance and mobility concepts. The Journal of the Acoustical
Society of America, 41:1180, 1967.

[46] E. Papadopoulos and S. Dubowsky. On the nature of control algorithms for free-
floating space manipulators. Robotics and Automation, IEEE Transactions on, 7(6):750-758,
1991.

[47] J.J. Perrine and V R. Edgerton. Muscle force-velocity and power-velocity relationships
under isokinetic loading. Medicine and science in sports, 10(3):159, 1978.

[48] S. J. Piazza. The Sprinter's Foot. Mechanical EngineeringMagazine, (October), October

2009.

[49] I. PoulakakisJ. A. Smith, and M. Buehler. On the dynamics of bounding and exten-
sions towards the half-bound and the gallop gaits. Proc. of the 2nd Int. Symp. on Adaptive
Motion of Animals and Machines, 2003.

[50] M. H. Raibert, K. Blankespoor, G. Nelson, and R. Playter. BigDog, the

Rough-Terrain Quadruped Robot. Proceedings of the 17th World Congress, pages
10823-10825, 2008.

[51] M. H. Raibert, H. B. BrownJr, M. Chepponis,J. Koechling,J. K. Hodgins, D. Dust-

man, W K. Brennan, D. S. Barrett, C. M. Thompson, andJ. D. Hebert. Dynam-
ically Stable Legged Locomotion (September 1985-Septembers1989). Technical re-
port, Massachusetts Institute of Technology, 1989.

[52] R. Rajamani. Vehicle Dynamics and Control. Mechanical Engineering Series. Springer, 1
edition, October 2005.

[53] J.P. Schmiedeler, R. Siston, and K. Waldron.

The significance of leg mass in modeling
quadrupedal running gaits. COURSESAND LECTURES-INTERNATIONAL CENTRE
FOR MECHANICAL SCIENCES, pages 481-488, 2002.

[54] J.P Schmiedeler and K.J. Waldron. The mechanics of quadrupedal galloping and the
future of legged vehicles. The InternationalJournalof Robotics Research, 18(12):1224-1234,
1999.

[55] N. C. C. Sharp. Timed running speed of a cheetah (Acinonyx jubatus). Journal of

Zoology, 241(3):493-494, 1997.

[56] R. Siegwart, P Lamon, T Estier, M. Lauria, and R. Piguet. Innovative design for
wheeled locomotion in rough terrain. Robotics and Autonomous systems, 40(2):151-162,
2002.

170
[57] J.L. Smith, S. H. Chung, and R. F Zernicke. Gait-related motor patterns and hindlimb
kinetics for the cat trot and gallop. Experimental brain research, 94(2):308-322, 1993.

[58] A. Staniforth. Competition Car Suspension: A PracticalHandbook. Haynes Publishing, 4

edition, October 2006.

[59] S. Stramigioli and H. Bruyninckx. Geometry and screw theory for robotics. Tutorial
during ICRA, 2001, 2001.

[60] Z. Vafa and S. Dubowsky. On the dynamics of manipulators in space using the vir-
tual manipulator approach. Robotics and Automation. Proceedings. 1987 IEEE International
Conference on, 4:579-585, 1987.

[61] B. Venkataraman, B. Godsey, W Premerlani, E. Shulman, M. Thaku, and R. Midence.

Fundamentals of a motor thermal model and its applications in motor protection.
Protective Relay Engineers, 2005 58th Annual Conferencefor, pages 127-144, 2005.

[62] R. M. Walter and D. R. Carrier. Ground forces applied by galloping dogs. Journalof
Experimental Biology, 210(2):208-216, January 2007.

[63] G. H. Wentink. Dynamics of the hind limb at walk in horse and dog. Anatomy and
embryology, 155(2):179-190, 1979.

[64] B. Wilcox, A. Nasif, and R. Welch. Implications of martian rock distributions on rover
scaling. Planetary Society InternationalConference on Mobile Planetary Robots and Rover Roundup,
Santa Monica CA, 1997.

[65] S. B. Williams, A. M. Wilson, L. Rhodes, J. Andrews, and R. C. Payne. Functional

anatomy and muscle moment arms of the pelvic limb of an elite sprinting athlete:
the racing greyhound (Canis familiaris). Journalof Anatomy, 213(4):361-372, October
2008.

[66] T M. Williams, G. P. Dobson, 0. Mathieu-Costello, D. Morsbach, M. B. Worley,

andJ. A. Phillips. Skeletal muscle histology and biochemistry of an elite sprinter, the
African cheetah. journalof Comparative Physiology B: Biochemical, Systemic, and Environmental
Physiology, 167(8):527-535, 1997.

[67] H. Witte, R. Hackert, K. E. Lilje, N. Schilling, D. Voges, G. Klauer, W Ilg, J. Albiez,

A. Seyfarth, and D. Germann. Transfer of biological principles into the construction
of quadruped walking machines. Robot Motion and Control, 2001 Proceedings of the Second
InternationalWorkshop on, pages 245-249, 2001.

[68] T. S. Zhao andJ. S. Dai. Dynamics and coupling actuation of elastic underactuated
manipulators. journal of Robotic Systems, 20(3):135-146, 2003.

171