• TABLE OF CONTENTS
HIDE
 Title Page
 Acknowledgement
 Table of Contents
 Abstract
 Introduction
 Sensor concept
 Description of experiments
 Experimental results
 Predictive model
 Discussion and conclusions
 Appendix
 Reference
 Biographical sketch
 Copyright














Title: Development of a dynamic touch sensor
CITATION THUMBNAILS PAGE IMAGE ZOOMABLE
Full Citation
STANDARD VIEW MARC VIEW
Permanent Link: http://ufdc.ufl.edu/UF00090208/00001
 Material Information
Title: Development of a dynamic touch sensor
Series Title: Development of a dynamic touch sensor
Physical Description: Book
Creator: Patterson, Robert Wayne,
 Record Information
Bibliographic ID: UF00090208
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
Resource Identifier: alephbibnum - 000873784
oclc - 14629091

Table of Contents
    Title Page
        Page i
    Acknowledgement
        Page ii
    Table of Contents
        Page iii
        Page iv
    Abstract
        Page v
        Page vi
    Introduction
        Page 1
        Page 2
        Page 3
        Page 4
        Page 5
        Page 6
        Page 7
        Page 8
    Sensor concept
        Page 9
        Page 10
        Page 11
        Page 12
        Page 13
        Page 14
        Page 15
        Page 16
        Page 17
    Description of experiments
        Page 18
        Page 19
        Page 20
        Page 21
        Page 22
        Page 23
        Page 24
        Page 25
        Page 26
        Page 27
        Page 28
        Page 29
        Page 30
        Page 31
        Page 32
        Page 33
        Page 34
        Page 35
        Page 36
        Page 37
        Page 38
        Page 39
        Page 40
        Page 41
        Page 42
        Page 43
        Page 44
        Page 45
        Page 46
        Page 47
        Page 48
        Page 49
        Page 50
        Page 51
        Page 52
        Page 53
        Page 54
        Page 55
        Page 56
        Page 57
        Page 58
        Page 59
        Page 60
        Page 61
        Page 62
        Page 63
        Page 64
        Page 65
        Page 66
        Page 67
        Page 68
        Page 69
        Page 70
        Page 71
        Page 72
        Page 73
        Page 74
        Page 75
        Page 76
        Page 77
        Page 78
        Page 79
    Experimental results
        Page 80
        Page 81
        Page 82
        Page 83
        Page 84
        Page 85
        Page 86
        Page 87
        Page 88
        Page 89
        Page 90
        Page 91
        Page 92
        Page 93
        Page 94
        Page 95
        Page 96
        Page 97
        Page 98
        Page 99
        Page 100
        Page 101
        Page 102
        Page 103
        Page 104
        Page 105
    Predictive model
        Page 106
        Page 107
        Page 108
        Page 109
        Page 110
        Page 111
        Page 112
        Page 113
        Page 114
        Page 115
        Page 116
        Page 117
    Discussion and conclusions
        Page 118
        Page 119
        Page 120
    Appendix
        Page 121
        Page 122
        Page 123
        Page 124
        Page 125
        Page 126
        Page 127
        Page 128
        Page 129
        Page 130
        Page 131
        Page 132
        Page 133
        Page 134
        Page 135
        Page 136
        Page 137
        Page 138
        Page 139
        Page 140
        Page 141
        Page 142
        Page 143
        Page 144
        Page 145
        Page 146
        Page 147
        Page 148
        Page 149
        Page 150
        Page 151
        Page 152
        Page 153
        Page 154
        Page 155
        Page 156
        Page 157
        Page 158
        Page 159
        Page 160
        Page 161
        Page 162
        Page 163
    Reference
        Page 164
        Page 165
        Page 166
    Biographical sketch
        Page 167
        Page 168
        Page 169
    Copyright
        Copyright
Full Text













DEVELOPMENT OF A DYNAMIC TOUCH SENSOR


By



ROBERT WAYNE PATTERSON
























A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN
PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF DOCTOR OF PHILOSOPHY


UNIVERSITY OF FLORIDA


1985
















ACKNOWLEDGEMENTS


I would like to thank the members of my graduate committee,

Professor Chen-Chi Hsu, Professor Ulrich Kurzweg, Professor Bernard

Leadon, Professor Walter Rosenbaum, and to give special thanks to

Professor Gale Nevill, Jr., who provided me the opportunity for a

graduate education.

I would also like to recognize the contributions of my mother,

Elise Patterpon-Crate, and Terri Loo in the preparation of this manu-

script.
















TABLE OF CONTENTS


PAGE

ACKNOWLEDGEMENTS ii

ABSTRACT v

CHAPTER

I INTRODUCTION . . . . . . . . ... . .1

The Human Skin . . . . . . . . ... . 2
Artificial Touch Sensing . . . . . . . . 5
Genesis of the Concept . . . . . . . . 7
Research Program Overview. . . . . . . . 7

II SENSOR CONCEPT . . . . . . . . ... .. 9

Physical Description . . . . . . . . . 9
Frequency Analysis . . . . . . . . .. 11
Spectral Signature Recognition . . . . . ... .15

III DESCRIPTION OF EXPERIMENTS . . . . . . .. .18

Experimental Set-Up. . . . . . . . . ... 18
Experiment #1--Shapes and Sizes. . . . . . ... 21
Experiment #2--Patterns of Features. . . . . ... 26
Experiment #3--Feature Orientation . . . . .. 34
Experiment #4--Relative Feature Position . . ... 43
Experiment #5--Changes in Speed. . . . . . ... 55
Experiment #6--Surface Texture/Roughness . . .. 58
Experiment #7--Repeatability . . . . . .. 73

IV EXPERIMENTAL RESULTS . . . . . . . ... 80

Experiment #1--Shapes and Sizes. . . . . . ... 81
Experiment #2--Patterns of Features. . . . . ... 87

Experiment #3--Feature Orientation . . . . .. 87
Experiment #4--Relative Feature Position . . ... 94
Experiment #5--Changes in Speed. . . . .... . 94
Experiment #6--Surface Texture/Roughness .. . . . 98
Experiment #7--Repeatability . . . . . ... 103

V PREDICTIVE MODEL . . . . . . . . ... .106

Model Description. . . . . . . . . . .107
Model Results . . . . . . . . ... . .110













VI DISCUSSION AND CONCLUSIONS . .

APPENDICES

A PATTERN VECTORS FOR EXPERIMENT #1.

B PATTERN VECTORS FOR EXPERIMENT #2.

C PATTERN VECTORS FOR EXPERIMENT #3.

D PATTERN VECTORS FOR EXPERIMENT #4.

E PATTERN VECTORS FOR EXPERIMENT #5.

F PATTERN VECTORS FOR EXPERIMENT #6.

G PATTERN VECTORS FOR EXPERIMENT #7.

REFERENCES . . . . . . . . .

BIOGRAPHICAL SKETCH . . . . . . .


PAGE

. . . . .118


. . . . . .121

. . . . . 129

S. . . . 137

. . . . . 145

. . . . . .153

. . . . . 156

. . . . . 161

S. . . . 164

. . . . . .167


. .
















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy


DEVELOPMENT OF A DYNAMIC TOUCH SENSOR

By

Robert Wayne Patterson

May 1985

Chairman: Gale E. Nevill, Jr.
Major Department: Engineering Sciences

This report describes the creation and exploration of a new

touch sensing concept for robot application. The sensing system pre-

sented here consists of an artificial "skin," a sensing element, and

means for evaluating the spectral signature. This system exploits

active, sliding motion of the skin over another surface to provide the

information required for identification. The skin is configured to

induce vibrations in the skin as a result of the sliding contact. This

vibratory motion is converted to electrical signals by the sensing

element for further processing.

The prototype sensor system, built to assess experimentally the

concept, integrates the three system components. The skin is made of

silicone rubber, and has a regular pattern of ridges extending beyond a

region of uniform thickness to serve as the contact area. The sensing

element is composed of two layers of polyvinylidene fluoride polymer

piezoelectric film. The electrical signals are examined in the fre-

quency domain, and linear discriminant analysis is used to classify the

resulting spectral patterns.









The prototype sensor was tested for a range of tactile tasks. The

results of seven experiments are presented here. The tasks explored

were to identify shapes and sizes of surface features, patterns of

features, orientation of features, and surface texture. The experiments

demonstrate the exceptional promise of the dynamic touch sensing concept

with nearly 100 percent recognition rates.

A predictive, finite element model of the prototype was also

developed. The results of this analysis are compared with one experi-

mental case, and indicate that more refined models could be used to

evaluate alternative sensor designs based on the dynamic touch sensing

concept.















CHAPTER I
INTRODUCTION


Tactile sensing, defined as "continuously variable touch sensing

over an area within which there is spatial resolution" (Harmon, 1980,

p. 1), is recognized as a principal need for next-generation robots. It

is expected that by 1990, 20 percent of all U.S. robots will have tac-

tile/touch sensing. In addition, the Japanese have embarked on a major

research effort to develop highly intelligent robots. A principal theme

of this research is the creation of very sensitive sensors for seeing

and touching (Schlussel, 1983). Promising application opportunities for

intelligent robots equipped with tactile sensitivity include exploration

and recognition of surfaces and surface features, particularly in remote

or hazardous environments, assembly of mechanical parts and electrical

components, pick and place operations, grinding, inspection, and har-

vesting. Additional application potential exists in the development of

improved prosthetic devices.

Although imitation of human functions and operations provides a

limited approach to sensor design, the expectations for ultimate sensing

capabilities are frequently shaped by analogous human, or other biologi-

cal systems. In fact, Harmon (1982) indicates that, based on question-

naire responses from researchers, administrators, students, users, and

consultants in academic, industrial, and government settings, tactile

sensors should be skin-like, composed of arrays of transducers on thin,









flexible, compliant, and durable substrates. Requirements include

continuous output, spatial resolution on the order of millimeters, and

fast response time. Researchers from around the world have pursued this

concept of tactile sensing (Bejczy, 1980; Larcombe, 1976; Stoljiljkovic

and Clot, 1977; Briot, 1979; Dario et al., 1983). Reviews of both human

touch sensing and artificial touch sensing approaches will establish the

context of this work.


The Human Skin

Classically, the human senses are vision, audition, smell, taste,

and touch. At times, however, in the history of the classification of

senses, the number of distinct senses has grown to twenty or more.

Scientists have generally used one of three approaches to categorize the

senses: (1) qualitative, based on observational similarity; (2) stimu-

lus, based on objects or energy causing a response; and (3) anatomical,

based on the sense organ responding (Geldard, 1972). The anatomical

approach has survived as the preferred approach, however, in the case of

the skin, knowledge of the anatomical mechanisms involved is somewhat

uncertain leading to a lack of classification preciseness. For conveni-

ence the skin is considered to house three types of sensory units:

mechanosensitive units, thermosensitive units, and nocipetive or high

threshold units (Valbo and Johansson, 1978; Geldard, 1972).

The skin's structure, innervation, and sensitivity vary consider-

ably over the extent of the human body. In some areas the skin is

relatively smooth, other areas are creased or wrinkled; some areas are

thick, others thin; some areas are moist, others dry; some areas readily

deform under shear while others resist deformation (Geldard, 1972).

Within the hand, the density of sensory units may vary by an order of









magnitude from the palm to the fingertip (Valbo and Johansson, 1978).

Similarly, the sensitivity of the palm is different from the fingertips;

the spatial resolution, measured by two point threshold tests, varies

from 7.7 mm in the palm to 1.6 mm on the fingertip (Valbo and Johansson,

1978).

The properties and sensory capabilities of fingertip skin most

closely resemble those desired for "artificial" skin. Fingertip skin is

composed of glabrous (nonhairy) skin, thick in comparison to other

regions of the body. The epidermis is folded forming regular patterns

of papillary ridges. The inked impressions of these patterns are the

well-known fingerprints. This corrugated structure, also projecting

into the dermis, coupled with a firm attachment of the epidermis to the

dermis enables finger tip skin to resist shifting or wrinkling due to

applied shear forces (Quilliam, 1978). Five categories of nerve endings

are present in fingertip skin: free nerve endings, Merkel type disks,

papillary nerve endings, Meissner type corpuscles, and Pacinian type

corpuscles (Quilliam, 1978). These nerve endings compose the mechano-

sensitive, thermosensitive and nocipetive sensing units.

Four types of mechanosensitive units are characterized by their

receptive field properties and their responses to sustained indentations

(Valbo and Johansson, 1978). Two types of receptive fields, one small

and distinct and the other large and indistinct, and two types of

responses, rapidly adapting and slowly adapting, are differentiable.

Rapidly adapting units respond only to changes in skin indentation.

Slowly adapting units exhibit a sustained, static discharge for steady

indentation, nevertheless, slowly adapting units with small receptive

fields (SAI units) exhibit high dynamic sensitivity. The density of the









two types of rapidly adapting units (RA units have small receptive field

and PC units have large fields) and of the SAI units increases dramat-

ically from the palm to the finger tip. Slowly adapting units with

large receptive fields (SAII units), which are highly sensitive to

tangential forces, are less dense in the fingertip than the palm. Thus,

the innervation of the fingertip seems most suited to dynamic stimula-

tion.

Touch sensing is further classified as passive or active. Passive

touch indicates the application of stimuli to the skin. Active touch,

on the other hand, is characterized by purposive exploration of an

object or surface. Activities such as reading Braille or sensing the

texture of materials are examples of active touch (Goldstein, 1980).

Active touch may also produce other stimuli, such as the sound arising

from sliding over or tapping a surface, that contribute to the entire

perception of an object. Relative motion between the fingertip and a

surface may be required to perceive certain information such as rough-

ness (Lederman, 1978).

Tactile acuity is not solely a function of innervation density, but

may also depend on the relative portion of the brain devoted to tactile

representations and on the structure of the skin itself. It has been

demonstrated that the brain region devoted to tactile processing can

change and in turn alter tactile acuity (Fox, 1984). Experiments

performed on monkeys showed that, upon loss of a finger, the brain

region devoted to the remaining fingers grew to include the region

previously devoted to the amputated finger. The tactile acuity for the

remaining fingers improved substantially. Moreover, it is believed that

the role of any individual nerve cell is unimportant, rather networks of









thousands of cells provide tactile response. In addition, the structure

of fingertip skin, particularly the papillary ridges, may contribute to

tactile perception (Quilliam, 1978). During fine movements of the

finger, the ridges create vibratory effects that propagate through the

various skin layers adding to tactile pattern recognition.

For the fingertip, it was observed that the skin is normally kept

in constant moving contact with the object surfaces to provide informa-

tion about its surface details and texture (Kirman, 1973). Moreover, it

was demonstrated that the role of movement is very different for touch

and vision; movement was determined essential to touch. Finally, Kirman

reported that a spatial arrangement shifted across the fingertip pro-

vided better discrimination than a stationary pattern of intensities.

This suggests that short-time integration of sensory signals provides

the pattern information rather than an instantaneous spatial representa-

tion.


Artificial Touch Sensing

Progress in tactile sensor development has been achieved largely

through the use of improved transduction techniques and increased spa-

tial density of sensing arrays. Early approaches using microswitches

and binary pressure sensitive pads provided little more than contact

information. More recent methods employ proportional sensing elements,

arranged in arrays or strategically located on a gripper surface, to

recognize an object's shape and orientation. These sensors are most

commonly composed of conductive elastomeric, piezoresistive, or piezo-

electric transduction materials.

Conductive elastomers have received considerable attention as

potential artificial skin materials due to their flexibility and their









promising resolution capability. However, problems with nonlinearity,

fatigue, and nonrepeatability pose formidable technical roadblocks

(Harmon, 1980, and Adjouadi, 1981). Nevertheless, tactile sensors made

from conductive elastomers have shown promising results. Experimental

sensors have been constructed with arrays of sensing cells measuring
2 2
1 mm (Allan, 1983). A 1 cm 256 element array has been used to

identify small objects such as screws, nuts, and washers (Allan, 1983).

Piezoresistive elements are sensitive, linear, reliable, and com-

patible with other semiconductor devices. One study, using the simu-

lated response of eight piezoresistive pressure sensors positioned on

multijointed fingers, demonstrated the potential of this concept for the

recognition of simple geometric shapes (Adjouadi, 1981). Practical

limitations to this approach include high unit costs and the prospect of

poor spatial distribution (Harmon, 1980).

Flexible piezoelectric polymers have also exhibited attractive

touch sensing possibilities. Polyvinylidene fluoride (PVDF or PVF2),

available in very thin films, has the particular advantages of being

rugged, light weight, broadband and having positive linearity and

hysteresis characteristics (Dario et al., 1983). Arrays of sensing

elements with unit cell dimensions on the order of 0.02 cm appear

feasible (Nevill and Davis, 1984). An intricate experimental touch

sensor composed of 16 PVDF elements with a microprocessor devoted to

each element represents one of the latest concepts (Stauffer, 1983).

PVDF also has pyroelectric properties permitting the development of

skin-like thermal sensors (Bardelli et al., 1983). The principal

drawback of piezoelectric devices is the lack of a DC response, thus

requiring the use of special signal capture techniques (Bardelli et al.,

1983; Harmon, 1982; Stauffer, 1983).









Harmon (1982) summarizes that no technology has yet provided

adequate sensing and, in view of the lack of any really powerful tech-

niques or results, there is concern that significant advances may still

be remote. Birk and Kelly (1981) also note that current touch sensors

have coarse resolution, low speed due to mass and friction, and have an

excessive number of wires. Additional related problem areas include the

trade off between resolution capability and processing speed and com-

plexity, pattern recognition, combining multi-sensory information, and

integration of sensory information with the control structure.


Genesis of the Concept

Since active, sliding motion of the fingertips is essential to

human touch sensing, a new touch sensing approach that exploits similar

motion seemed promising. In addition to employing transducer elements

and pattern recognition approaches compatible with the dynamic concept,

a skin structure, similar to the structure of fingertip skin, was used

to induce vibrations in the sensor from sliding contact. These vibra-

tions provide the information required to identify surface features.

Although this tactile sensor concept does not strictly conform to the

definition of tactile sensing given by Harmon (since continuous variable

sensing of forces in an array is not used), the sensor does exhibit

skin-like properties and could complement other technologies.


Research Program Overview

A prototype sensor based on this new concept is presented. The

performance of the prototype for a range of tactile tasks is demon-

strated primarily through experimental results. These results are

compared with the published results achieved by other tactile sensing






8


approaches. In addition, a predictive model, using the ANSYS finite

element code, is described. The model is tested for one experimental

case to give an indication of its potential usefulness. Finally, the

promise of the proposed touch sensing concept is summarized and areas of

improvement are identified.
















CHAPTER II
SENSOR CONCEPT


Based on the human skin analogy, an active touch sensor may have

three basic components affecting ultimate tactile acuity: the skin

material and configuration, the transduction element, and the signal

processing/pattern recognition techniques employed. For the prototype

touch sensor developed here, each component has been kept relatively

simple and may be far from optimal for any specific application.

Nevertheless, the prototype sensor system described below demonstrates

the promising potential this new sensing concept offers.


Physical Description

The sensor pad itself is composed of a sensing/transduction element

and a compliant, variable thickness skin. The skin serves the dual

purpose of inducing vibratory motion from sliding contact with objects

of interest and providing a durable protective covering for the sensing

element. The sensing element must capture dynamic information provided

by the induced vibratory motion.

Silicone rubber proved to be a satisfactory skin material; it is

easily molded into a variety of configurations, inexpensive, and

durable. The sensor skin was shaped by allowing the silicone rubber to

cure in an aluminum mold built for this work. The skin configuration,

shown in Figure 2-1, has seven ridges (triangular prisms) extending

below a 1 inch x 1 inch x 0.06 inch layer of silicone rubber. Each

ridge is 0.5 inches long, 0.04 inches high, and 0.065 inches wide at its



































































Figure 2-1. Prototype Silicone Rubber Skin.









base. The ridged region is the contact area of the sensor; the ridges

slide over surfaces, features, and objects of interest and induce

vibratory motion in the sensor.

The sensing element, Figure 2-2, has only two transducers. Each is

a 0.2 inch x 0.2 inch x 52 micron piece of metallized polyvinylidene

fluoride (PVDF) film. Although other researchers have noted problems

with PVDF, particularly the lack of a DC response required for quasi-

static force/pressure sensing, the film has proved suitable in a number

of applications related to this work such as accelerometers, micro-

phones, hydrophones, and ultrasonic sensors. The PVDF film has the

advantages of being flexible, light weight and having broadband re-

sponse. In addition, the PVDF film is uniaxially oriented, thus provid-

ing maximum sensitivity in one direction. Therefore, the two trans-

ducers are oriented orthogonal to each other to obtain two components of

dynamic strain. Transducer 1, Figure 2-3, provides primarily

y-direction information, orthogonal to the ridges, and Transducer 2

provides primarily x-direction information, parallel to the ridges.

The sensor (skin and sensing element) is fastened with epoxy resin

to an annular, square aluminum plate for support. This arrangement

allows skin motion and minimizes boundary interference with the central

active region. In addition, the aluminum plate can be easily attached

to an experimental apparatus.


Frequency Analysis

Frequency, or spectral, analysis is commonly used to characterize

highly irregular and nonrepeatable waveforms. Spectral analysis is

employed in a wide range of applications, including radar, sonar, speech





12


Figure 2-2. PVDF Sensing Element Attached to the Top of the Silicone
Rubber Skin.





















Electrodes


Figure 2-3. Sensing Element Cross Section.


Induced
Voltage
Channel
#2














Induced
Voltage
Channel
#1









recognition, and EEG and EKG analysis, to facilitate signal interpreta-

tion.

Random functions are typically characterized by their auto corre-

lation function 4(T) eqn 2-1, and its Fourier Transform, the power

density spectrum 0 (w) eqn 2-2

1 T
P(T) = lim f f(t)f(t+r)dt (2-1)
T-*c 2T -T





=(W) = p r(T)e-jjTdT (2-2)
-00


where f(t) is a random time function, co is frequency, and T is a time

delay parameter.

This relationship, given by the Wiener Theorem, relies on the basic

assumption that the autocorrelation function exists for every value of

the argument. The Wiener Theorem is applicable to functions that can be

described analytically or statistically provided this assumption is

satisfied (Lee, 1960). The area under the power density spectrum curve

is the total power; the power contribution by a band of frequencies is

the area of the function within the limits of the band.

For periodic, or transient functions, the analogous autocorre-

lation and its Fourier Transform are given in equations 2-3 and 2-4

respectively (Lee, 1960)



5(T) = J f(t)f(t+r)dt (2-3)
-00




(W) = f (T)eJjlTdT (2-4)
-00o









where f(t) is an periodic time function. Equation 2-4 provides the

energy density spectrum. The area under the energy density spectrum is

the total energy of the periodic function.

When significant temporal changes occur in an otherwise random

function, as in the case of speech, a representation of the function in

both the time and frequency domains is desired. This can be accom-

plished by analyzing a portion of the signal as seen through a time

window. The autocorrelation function and short-time power spectrum are

given in equations 2-5 and 2-6 respectively (Flanagan, 1972).



((T,t) = f f(A)h(t-X)f(X+T)f(t-T-X)dX (2-5)
-00



m .O
4(w,t) = f 0(r,t)e-JmdT = 2f 0(T,t)coswTdT (2-6)
-CO -00



-1
where, h(t) is a weighting function with units sec- for including past

values, A is the running integration variable, and (t T,t) is an even

function of T.

For the work presented here the power density spectrum is used to

characterize long-term random signals, and the periodic spectrum is

used for short-time phenomena.


Spectral Signature Recognition

The recognition problem is one of discrimination; the goal is to

correctly classify an observed pattern into one of several groups.

Discriminant analysis is a commonly used classification technique, and











has been employed by others to recognize spectral signatures (e.g.

Sklar et al., 1973, and Bricker et al., 1971). The simplest form of the

discriminant function is linear,



zi = ai + aix + + a 2 + ... + anx (2-7)



where zi is the classification score for group i, the x's are the n

pattern variables, and the a's are weighting coefficients selected to

maximize the separability between the groups. Each pattern variable

represents the area under the spectral density curve in a specified

frequency band.

Once the weighting coefficients have been determined, each case is

assigned membership into the group that produces the highest score. The

results of this technique will be tabulated in a classification matrix

and a jackknife classification for each experiment.

To provide a graphical depiction of the separability of the groups,

a scatter plot of the observations for each experiment is presented in

the first two canonical variables of pattern space. These variables

were found from a linear discriminant method called canonical variate

analysis (Hand, 1981; Gnanadesikan, 1977). The canonical variables are

linear combinations of the pattern variables, similar to the

classification score. The first canonical variable is the linear

combination which best discriminates between the groups, the second is

the next best linear combination orthogonal to the first, etc. The

total number of canonical variables is the smaller of n and g-1 where g

is the number of groups.











For this work, the anticipated dimensionality of the pattern vector

exceeds the sample size, this can substantially reduce the classifica-

tion accuracy (Young and Calvert, 1974). Thus, it is desirable to

reduce the dimensionality of the pattern vector by selecting a subset of

variables based on their contribution to the discrimination process.

The stepwise discriminant analysis (Dixon et al., 1983) employed enters

the variable that adds the most to the separation of the groups, and

continues to add variables that significantly contribute to the discrim-

ination. The routine also permits removal of a variable should subse-

quent entries reduce its contribution. This stepwise linear

discriminant procedure provides the classification matrix, the jackknife

classification, and the canonical variate analysis used to characterize

the experimental results.
















CHAPTER III
DESCRIPTION OF EXPERIMENTS


Seven experiments were performed with the sensor to assess its

potential to provide a variety of tactile information. An experimental

apparatus was devised and built to move surfaces and features relative

to the touch sensor. These surfaces and features were selected to test

the sensor's response to shapes and sizes, patterns of shapes, feature

orientations, changes in relative position of features, changes in

speed, and different surface textures. In addition, a brief experiment

was performed on a second touch sensor to ensure that the results were

repeatable.


Experimental Set-Up

The complete experimental set-up, shown in Figure 3-1, includes the

support apparatus for the sensor, a turntable, and a signal analyzer.

The support apparatus rigidly holds the sensor in place during contact

with moving objects. The turntable moves surfaces or features past the

sensor at a constant speed; the Hewlett Packard 5420A Digital Signal

Analyzer performs extensive measurement operations and stores

experimental results.

The prototype sensor is mounted on the end of a square prism, shown

in Figure 3-2, situated above the turntable. The sensor is positioned

vertically by moving the prism within its housing through the use of a

thumbscrew located at the top of the housing. This placement is rigidly

fixed by tightening the two thumbscrews located on the sides of the










































Figure 3-1. Experimental Set-Up.


I n o n e wI:
ounce 1
cle-w ol~





20


Figure 3-2. Mounted Prototype Sensor.









housing. The sensor is positioned laterally by sliding the entire

housing along two steel rods, and fixed in place by thumbscrews on the

front and back of the housing.

Lead wires attached to the sensing element are soldered to the BNC

connectors on the face plate of the square prism. Output from the lower

layer of PDVF, most sensitive to y-direction strains, is the input for

channel one of the signal analyzer. Similarly, the output from the

primarily x-direction sensitive upper layer of the sensing element is

the input for channel two.

The signal analyzer has a number of measurement capabilities in

both the time and frequency domains. The user selects the measurement

type and the input conditions for the measurement. A description of the

input conditions will be given for the experiments individually.

The test objects were mounted on a plastic disk which was placed on

the turntable. Each object was positioned so that its center passed

under the center of the sensor. Further, the turntable was situated so

that the diameter line passing under the center of the test object was

parallel to the ridges of the sensor when the center of the object was

directly under the center of the sensor. This arrangement served two

purposes; one, to provide a convenient reference location ensuring that

each test object traversed the same path, and two, to provide a path

close to the perpendicular direction of the ridges.



Experiment #1--Shapes and Sizes

The test objects for this experiment, shown in Figure 3-3, were

three ball bearings and three cylinders. The ball bearings have dia-

meters 5/16, 3/16, and 5/32 inches. The cylinders have diameters 5/16,















































Figure 3-3. Test Objects for Experiment #1--Shapes and Sizes.









3/16, and 1/8 inches. In addition to representing different classes of

shapes, these objects can be considered as vertexes and edges for some

global configuration. The objects were mounted so that the object

center was 2.5 inches from the center of the disk. The amount of inter-

ference shown in Figure 3-4, was also varied to ensure that the class-

ification was based on shape, not duration of contact, and to demon-

strate that, with training, results could be tolerant of inexactness in

sensor position and contact force.

Input conditions for signal measurement, shown in Figure 3-5, were

selected to calculate an energy density spectrum for the frequency band

0-800 Hz. The measurement period began when the object made contact

with the sensor and lasted 320 msec. The signal analyzer calculates the

spectrum for only one channel at a time, thus, two measurement periods

were required to provide the complete bi-directional information.

The test procedure involved mounting a test object on a plastic

disk then placing the disk on the turntable, moving the the support

housing into position, lowering the sensor to a prescribed amount of

interference and fixing it in place, and supplying power to the turn-

table permitting it to rotate at a constant angular speed of 33 1/3 rpm.

Upon contact between the the sensor and the test object, the signal

analyzer is triggered to begin its measurement period. The first

measurement was input on channel one, the spectrum was obtained and

stored. Then the input was switched to channel two and a second spec-

trum was obtained and stored. Thus, two measurement periods were

required for each observation; however, this was a restriction imposed

by the experimental equipment rather than a limitation of the sensor.

Each of the spectra was normalized by dividing by the total energy in






















Amount
of
Interference


T_
T


Figure 3-4. Side View of Sensor Skin and Test Object, Displaying Amount
of Interference.









SETUP STATE


MEASUREMENT I

AVERAGE

SIGNAL i

TRIGGER i



CENT FREQ s

BANDWIDTH s

TIME LENGTH :

dF ,


AUTO SPECTRUM

1

TRANSIENT

INTERNAL



8.0 HZ

800.000 HZ

320.000 mS

3.12500 HZ


STABLE



,CHNL 1








dT ,


312.500 uS


ADC CHNL RANGE AC/DC

1 258 mV
2 250 mV


DELAY

DC 8.0 S
DC 0.0 S


CAL (C1/C2)


1.00000
1.00800


Figure 3-5. Signal Analyzer Measurement Parameters for Experiments #1, #2, #3, #4, and #7.









the 0-800 Hz band. The two normalized spectra together represent one

experimental observation, or signature. The pattern vector is formed by

assigning the area in a specified frequency band to a pattern variable.

A 25 Hz bandwidth was selected for each variable. Thus, each spectrum

is represented by 32 variables and the signature for each observation is

described by a 64 variable pattern vector.

For each test object, observations were made for three amounts of

interference, approximately 0.02, 0.025, and 0.03 inches, then the next

object was mounted and the procedure repeated. After each object had

been tested, the entire test cycle was repeated three more times, pro-

ducing a total of 72 signatures. Average normalized signatures, pre-

sented in both real and logarithmic scales, for each object are shown in

Figures 3-6 to 3-11. The pattern vectors for all observations are given

in Appendix A.


Experiment #2--Patterns of Features

The test objects for this experiment, shown in Figure 3-12, were

made to resemble the following Braille letters: A, B, C, D, E, F, and G.

Reading Braille is an oft-cited example of active touch sensing by

humans and, thus, serves as an appropriate experimental task for this

prototype touch sensor. The individual features are 1/8 inch ball

bearings. Four possible sites, the corners of a 1/4 inch square, are

available for feature emplacement. For the purpose of providing ref-

erence positions for both the sensor and the object, the center of this

square is considered the center of the test object.

The input conditions for signal measurements are identical to those

described for Experiment #1, shown in Figure 3-5. The energy density










A SPEC


25.00 -
m


REAL

5. ..
1 i


Transducer 1 A SPEC
A SPEC


25.000
m


REAL

5.0000
M


0.0


800. 00


0.0


A SPEC
--10. 000


LGMAG
[3B


4



0.0 H8 800.00


-60.000


1


i..~..----.-~~-


0.0


Figure 3-6. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
5/16 inch Sphere, Plotted on Both Real and Logarithmic Scales.


Transducer 2


A SPEC
-18.000



LGMAS
DB



-60. 00


800.00


800. 0










Transducer 1


A SPEC
20. 00




REAL


2. 0000



A SPEC
-10. 080




LGMAG
08



-60.000


900. 00


Transducer 2




;: i ,
A
m .


A SPEC
iR. 330




REAL


2. m00



A SPEC
-10.00M


800. B


-60.000. 8


'Figure 3-7. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
3/16 inch Sphere, Plotted on Both Real and Logarithmic Scales.


0.0


0.0


""~""""~~~'-c'""~~'~ I`'--




1









Transducer 2


A SPEC

18. 880
1&08


REAL


A SPEC

208.00


REAL

5.000M


0.0 HZ


0.0


800.00


8sa.00


Figure 3-8.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
5/32 inch Sphere, Plotted on Both Real and Logarithmic Scales.


2.0000



A SPEC
-10.000



-GI4AG
D8



-0.000


0.0


A SPEC
-10.000



LGKAG
09B


800.00


.80


Transducer 1


"" .~L...-. ..~5~....~...-~.~.-~.-- - -2










Transducer 1


-p.. .-* .---..-.


(. r-


A SPEC
-10. ow




LG9AG
YBI


.1

4.....'7-- 7


Figure 3-9.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
5/16 inch Cylinder, Plotted on Both Real and Logarithmic Scales.


A SPEC





REAL


A SPEC


A SPEC
-18. 8K00




LCMAG
DB


I0.


Transducer 2


'-''~"~~-"''" ""''"'"'''""' "'~' "~' ' "I""' '-"










Transducer 1


A SPEC

12.000



REAL


2.0000
m


A SPEC
50.000



REAL


5.0000
m


A SPEC
-10.000




LGMAG
DB



-60.000


A SPEC
-10.000




LGMAG
OB



-60.000


0.0


800.00


800.00


Figure 3-10.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
3/16 inch Cylinder, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


800.00


0.0


800.00


Transducer 2










Transducer 1


A SPEC



RAL








A SPEC
-I r- "T"


A SPEC





RftAL






A SPEC
-1l. MLA!


Transducer 2





SIA
.r L t i


IsKG~t Hl



"T&~~kB fiT.. --.I..T.~r- "-1


LGRAG
DB


Figure 3-11. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for a
1/8 inch Cylinder, Plotted on Both Real and Logarithmic Scales.


~E.O %~ ES~?t~. C~'

--- -----.-. ""'"""'"" 1- ... -----~-.....-. -.,





























































Figure 3-12. Test Objects for Experiment #2--Patterns of Features.


OJ









spectrum was calculated for the 0-800 Hz frequency band from a 320 msec

record of the input signal.

The test procedure was very similar to the test procedure for

Experiment #1. The only deviation from that procedure was that the

amount of interference was held constant, at approximately 0.025 inches,

for all observations. Nine signatures were obtained for each object

from two test cycles. The first cycle consisted of five observations

per object, and the second consisted of four observations per object.

Again, the pattern vector for each signature had 64 variables rep-

resenting the fraction of the total energy contained in 25 Hz wide

frequency bands. Average signatures for the test objects are shown in

Figures 3-13 to 3-19. The pattern vectors for all observations are

given in Appendix B.


Experiment #3--Feature Orientation

One test object, the machine screw shown in Figure 3-20, was used

for this experiment. The feature of interest, the slot on the head of

the screw, was oriented at different angles (0, 15, 30, 60, 90, and 120

degrees) with respect to the direction of the ridges on the sensor skin.

While the orientation of a slot may be of interest itself, this feature

can also be thought of as an edge or a crack for part identification or

inspection applications.

The input conditions for signal measurements were the same as those

described for the previous two experiments. The energy density spectrum

was calculated, for the frequency band 0-800 Hz, from a 320 msec record

of the input signal.

The test procedure was also similar to that employed in the first

two experiments. A total of 48 observations, eight per orientation,









Transducer 2


800.00


800.00


A SPEC
25.000



REAL


5.0000
m


Transducer 1


0.0


A SPEC
16.000
m



REAL


2.0000
m


A SPEC
-10.000




LGMAG
DB



-50.000


800.00


Figure 3-13.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for
Pattern A, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


0.0


A SPEC
-10.000




LGMAG
DB


0.0


800.00










Transducer 1
A SPEC


18 000


m



REAL


?.0000


0.0


A SPEC
-10.000




LGMAG
DB



-50.000


-IL--------e--r


800.00


0.0


800.00


Transducer 2


I


0.0


Figure 3-14.


Average Normalized Spectral Signature Obtained from Transducer 1
Pattern B, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


A SPEC
10.000
m



REAL


1.0000



A SPEC
-10.000




LGMAG
DB



-50.000


800.00


0.0


800.00










Transducer 1


A SPEC
18.000
m



REAL


2. 0000
m


A SPEC
-10.000




LGMAG
DB



-50.000


A SPEC

9.0000
m



REAL


1.0000
m


A SPEC
-20.000




LGMAG
DB



-50.000


0.0


800.00


800.00


Figure 3-15.


Average Normalized Spectral Signature Obtained from Transducer 1
Pattern C, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


0.0 HZ


800.00


0.0


800. 00


Transducer 2









Transducer 2


A SPEC


A SPEC
25.000
m



REAL


5.0000
m


Figure 3-16. Average Normalized Spectral Signature Obtained from Transducer 1
Pattern D, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


10.000




REAL


1.0000
m


A SPEC
-10.000




LGMAG
DB



-50.000


8.0


800.00


A SPEC
-10.000




LGMAG
DB



-50.000


I.
0.0


HZ


800.00


0.0


800.00


0.0


880.00


u


Transducer 1










Transducer 1


A SPEC

20.000
m



REAL


2.0000
m


A SPEC
-10.000




LGMAG
DB



-50.000


800.00


800.00


A SPEC

7.0000
m


REAL


1.0000
m



A SPEC
-20.000




LGMAG
DB


0.0


Figure 3-17. Average Normalized Spectral Signature Obtained from Transducer 1
Pattern E, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


0.0


0.0


800.00


0.8


800.88


Transducer 2










Transducer 1


A SPEC
20.000
m



REAL


2.000880
m


A SPEC
-10.000



LGMAG
DB



-50.000


800.00


A SPEC

14.000
m


REAL


2.0000
m


A SPEC
-10.000



LGMAG
DB



-50.000


8.0 HZ


800.00


Figure 3-18. Average Normalized Spectral Signature Obtained from Transducer 1
Pattern F, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


0.0 HZ


0.0


800.00


800.00


Transducer 2








Transducer 1


A SPEC
20.000
m


REAL


2.0000
m


A SPEC
-10.000



LGMAG
DB


-50.000


A SPEC
18. 000
m


REAL


2.0000
m


A SPEC
-10.000



LGMAG
DB


-50.000


800.00


0.0


HZ


0.0


Figure 3-19. Average Normalized Spectral Signature Obtained from Transducer 1
Pattern G, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for


800.00


800.00


0.0


888.00


.


Transducer 2





42


Figure 3-20. Test Object for Experiment #3--Feature Orientation.












were obtained from two test cycles. The resulting average signatures

for each orientation are shown in Figures 3-21 to 3-26. The pattern

vectors for all observations are given in Appendix C.

Experiment #4--Relative Feature Position

The test objects for this experiment, shown in Figure 3-27, were

each composed of two 1/8 inch ball bearings. The first ball bearing

provided a reference position, and the second was placed in one of three

positions behind the first. The distance between the two positions was

either 0.25, 0.26, or 0.35 inches, thus, the relative change in position

was 0.01 to 0.1 inches. Three sets of objects were used to ensure that

the signature variability was caused by the change in position. The

center of the reference ball bearing also served as the center of the

test object for the purposes of positioning the sensor. This experiment

provides the closest comparison to the two-point resolution cited for

other tactile sensors.

The input conditions were the same as those for the previous

experiments. The 320 msec measuring period began when the reference

ball bearing made contact with the sensor. The time domain record was

transformed to give an energy density spectrum in the frequency domain.

The procedure for obtaining individual signatures was the same as

described for the previous experiments. However, only one test cycle,

consisting of four observations, was performed for each test object.

Thus, the total number of observations for this experiment was 36, or 12

for each position. Average signatures for the three positions are shown

in Figures 3-28 to 3-30. The pattern vectors for all observations are

given in Appendix D.









Transducer 1


800.00


A SPEC
7.0000
m


REAL


1.0000
m


800.00


A SPEC

10.000 .
m


REAL

2.0000 -



A SPEC
-10.000



LGMAG
DB


-60.000


8.0


Figure 3-21.


Average Normalized Spectral Signature Obtained from Transducer 1 and
Orientation 0 degrees, Plotted on Both Real and Logarithmic Scales.


Transducer 2 for Slot


0.0


0.0


A SPEC
-20. 000



LGMAG
DB


-60.0 0


808.88


8.0


Til


Transducer 2









Transducer 1
A SPEC

18.000
m


REAL


2.0000
m


8.0


A SPEC
-18.8000



LGMAG
DB



-50.888


800.00


0.0


Transducer 2










WlilL~iT-


A SPEC
12.000 .



REAL


2.0000
m



A SPEC
-18.00 .



LGMAG
DB



-50.0888


0.0


800.00


Figure 3-22. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for Slot
Orientation 15 degrees, Plotted on Both Real and Logarithmic Scales.


888.00


800.00









Transducer 2


S I I 0 I I
HZ 800.00


800.00


Figure 3-23.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for Slot
Orientation 30 degrees, Plotted on Both Real and Logarithmic Scales.


A SPEC

25.000
m


REAL

5.0000
m


0.0


A A^ --^^ -


A SPEC
-10.000




LGMAG
DB



-60.000


A SPEC
18. 000
m



REAL


2.0000
m


A SPEC
-10.000




LGMAG
DB



-60.000


800.00


0.0


0.0


80.000


0.0


Transducer 1










Transducer 1


A SPEC

20.000



REAL

5.0000
m


A SPEC

30.000
m


REAL


5.8000
m



A SPEC
-10.000



LGMAG
DB



-60.008


A SPEC
-10.000



LGMAG
DB



-6. 000


88800.


Transducer 2

R#,


0.0 HZ


0.0


Figure 3-24. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for Slot
Orientation 60 degrees, Plotted on Both Real and Logarithmic Scales.


800.00


0.0 HZ


0.0


800.00


80. 00









Transducer 2


A SPEC

20.000
m


REAL

5.0000
m


0.0 HZ


A SPEC
-18.000




LGMAG
DB



-60.000


0.0


888.00


m8.00.


Figure 3-25.


Average Normalized Spectral Signature Obtained from Transducer 1 and
Orientation 90 degrees, Plotted on Both Real and Logarithmic Scales.


Transducer 2 for Slot


A SPEC
30.000
m



REAL


5.0000
m


0.0


800.00


A SPEC
-10.00088




LGMAG
DB



-60.000


0.0


800.00


Transducer 1









Transducer 2


A SPEC

30. 888
m


REAL


5.0000
m



A SPEC
-10.000



LGMAG
DB



-68.088


0.0


800.00


800. 800


A SPEC


20.000


REAL

5.0000
m


0.0 HZ


A SPEC
-10.000




LGMAG
DB



-68.00888


8.0


Figure 3-26.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for Slot
Orientation 120 degrees, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


888.88


800.00


Transducer 1






























































Figure 3-27. Test Objects for Experiment #4--Relative Feature Position.


OL
0










Transducer 1


A SPEC

20.000
m



REAL


2.0000
m


A SPEC
-10.000




LGMAG
DB



-680.000


0.0


800.00


800.00


A SPEC

25.000
m


REAL

5.0000
m


0.0 HZ


A SPEC
-10.000




LGMAG
DB



-60.000


0.0


Figure 3-28. Average Normalized Spectral Signature Obtained from Transducer 1
Spacing 0.25 inches, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for Feature


0.0 HZ


800.00


800.00


Transducer 2










Transducer 1


A SPEC

28.0888



REAL

5. 000
m



A SPEC
-10.000




LGMAG
DB



-60.000


800.00


800.00


A SPEC
21 aaaI


i .JL >U .
m



REAL


2. 0000 .


m


A SPEC
-1. 000




LGMAG
DB



-60.000


0.0


HZ


0.0


Figure 3-29.


Average Normalized Spectral Signature Obtained from Transducer 1
Spacing 0.26 inches, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for Feature


0.0 HZ


8880.00


8.0


888.00


I-}- ,- -


Transducer 2


I


<
I




























Figure 3-30. Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for Feature
Spacing o.35 inches, Plotted on Both Real and Logarithmic Scales.







Trnnsducer 1
A SPEC


888. 880


A SPEC
25. 88
m

REAL

m


Transducer 2


ui


890. B


20.008
m


REAL

2. 88W


A SPEC
-10. 8


L AG


-OTW


0 .


A SPEC
-18. 8W


LGMAG
DB


. 0 HZ BM. M


. 0


8.8


mp. W









Experiment #5--Changes in Speed

Recognizing that a sensor may be required to perform at a variety

of operational speeds, an experiment was performed to demonstrate that

signatures could be shifted to account for speed,-and recognized from a

standard pattern vector determined from observations collected at

another speed. The test objects, shown in Figure 3-31, were a subset of

those tested in Experiment #4, one object for each of the three loca-

tions. For this experiment, the angular speed was increased from 33 1/3

to 45 rpm. The radial distance from the center of the disk to the

center of the reference ball bearing remained 2 1/2 inches. Therefore,

the speed of the object was increased by a factor of 1.35.

The input conditions, shown in Figure 3-32, were changed to accom-

modate the expected shift of energy to higher frequencies. The band-

width for the frequency domain was increased from 0-800 Hz to 0-1600 Hz.

Consequently, the measurement period was adjusted (automatically by the

signal analyzer) to 160 msec.

The procedure for performing the experiments remained the same,

however, the procedure for normalizing the spectral signatures was

modified. Although the spectra were calculated for a 1600 Hz bandwidth,

the bandwidth of interest was only 0-1080 Hz (1.35 times larger than the

0-800 Hz bandwidth), therefore, the spectra were normalized by dividing

each by the energy in the 0-1080 Hz band rather than the total energy

for the measurement. In addition, the 64 variable pattern vectors were

composed of the energy fraction in 33.75 Hz wide frequency bands rather

than in the 25 Hz bands used previously. This was a limited experiment

performed to assess the ease with which changes in speed could be

accounted for, thus one measurement cycle consisting of two observations































































Figure 3-31. Test Objects for Experiment #5--Changes in Speed.


Ln
0.










SETUP STATE


MEASUREMENT

AVERAGE :

SIGNAL :

TRIGGER :


CENT FREQ :

BANDWIDTH :

TIME LENGTH :

dF :


:AUTO SPECTRUM

1

TRANSIENT

INTERNAL


0.0 HZ

1.60000 KHZ

160.000 mS

6.25000 HZ


STABLE



,CHNL 1


dT :


156.250 uS


ADC CHNL RANGE AC/DC

w 1 250 mV
2 250 mV


DELAY

0.0 S
0.0 S


CAL(C1/C2)


1.00000
1.00000


Figure 3-32. Signal Analyzer Measurement Parameters for Experiment #5.









for each location was conducted. Since only six signatures were ob-

tained, signatures for each observation, rather than the averages, are

shown in Figures 3-33 to 3-38. The pattern vectors for all observations

are given in Appendix E.


Experiment #6--Surface Texture/Roughness

The test surfaces for this experiment, shown in Figure 3-39, were

six pieces of sandpaper: 36, 60, 80, 100, 180, and 220 grit. The

sandpaper was also mounted on a plastic disk and placed on the turn-

table. The sensor was centered on a position 2 1/2 inches from the

center of the disk. The angular speed of the surface was 33 1/3 rpm for

all measurements. The amount of interference could not be precisely

determined, therefore, the vertical position of the sensor was estab-

lished qualitatively by maintaining contact with the surface without

noticeably impeding the motion of the surface.

The sandpaper is a collection of small randomly distributed fea-

tures. However, instead of attempting to recognize any particular

feature, or groups of features, a characterization of the entire dis-

tribution is desired. The signal generated in the sensor is therefore

substantially different than the response induced in prior experiments;

the signal is not short-lived, or transient, rather it is the continuous

response of the sensor to the distribution of features. The measurement

set-up is shown in Figure 3-40. In addition to changing the signal

type, the 800 Hz wide bandwidth was shifted to 100-900 Hz to reduce the

effect of noise, and the measuring period was triggered randomly.

A total of 120 signatures were obtained from four test cycles, each

containing five observations per surface. Average signatures for the

six surfaces are shown in Figures 3-41 to 3-46. Since the power density










Transducer 1


A SPEC

14.000
m


REAL


2.0000



A SPEC
-10.000




LGMAG
DB



-70.000


0.0


A SPEC

20.000
m


REAL

5.0000
In


1.0000 K


A SPEC
-10.000




LGMAG
DB



-70.000


1.0000 K


0.0 HZ 1.0000 K


0.0


1.0000 K


Figure 3-33.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.25 inches, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


Transducer 2










Transducer 1


A SPEC

20.000
m


REAL

5.0000
m


A SPEC
16.000
m



REAL


2.0000



A SPEC
-10.000




LGMAG
DB



-70.000


1.0000 K


1.0000 K


0.0 HZ


0.0


1.0000 K


1.0000 K


Figure 3-34.


Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.25 inches, Plotted on Both Real and Logarithmic Scales.


A SPEC
-10.000




LGMAG
DB


0.0


0.0


Transducer 2










Transducer 1


A SPEC
20.000
m



REAL


2.0000
m


A SPEC
-10.000




LGMAG
DB



-90.000


A SPEC

14.000
m


REAL


2.0000
m



A SPEC
-10.000




LGMAG
DB



-70.000


1.0000 K


1.0000 K


Figure 3-35.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.26 inches, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


1.0000 K


0.0


1.0000 K


Transducer 2










Transducer 1


A SPEC

14.000
m


REAL


2.0000
m


1.0000 K


1.0000 K


A SPEC
16.000
m



REAL


2.0000
m



A SPEC
-10.000




LGMAG
DB



-70.000


0.0


1.0000 K


1.0000 K


Figure 3-36. Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.26 inches, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


0.0


A SPEC
-10.000




LGMAG
DB



-70.000


0.0


Transducer 2










Transducer 1


A SPEC

16.000
m



REAL


2.0000
mf


A SPEC
-10.000




LGMAG
DB



-80.000


A SPEC

14.000
m



REAL


2.0000
m


1.0000 K


A SPEC
-10.000




LGMAG
DB



-80.000


1.0000 K


0.0


0.0


Figure 3-37.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.35 inches, Plotted on Both Real and Logarithmic Scales.


0.0 HZ


1.0000 K


0.0


1.0000 K


Transducer 2








Transducer 1


A SPEC
18.000
m


REAL

2.0000


0.0


A SPEC
-10.000



LGMAG
DB


-70.000


A SPEC
14.000
m


REAL

2.0000
m


1.0000 K


0.0


0.0


A SPEC
-10.000



LGMAG
DB


-70.000


1.0000 K


1.0000 K


0.0


1.0000 K


Figure 3-38. Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 for Feature Spacing 0.35 inches, Plotted on Both Real and Logarithmic Scales.


I



AA^ , ,


Transducer 2




























































Figure 3-39. Test Surfaces for Experiment #6--Surface Texture/Roughness.


Ln









SETUP STATE


MEASUREMENT

AVERAGE :

SIGNAL :

TRIGGER :


CENT FREQ :

BANDWIDTH a

TIME LENGTH :

dF :


: AUTO SPECTRUM

1

RANDOM

FREE RUN


STABLE



,CHNL 1


500.000 HZ

800.000 HZ

320.000 mS

3.12500 HZ dT :


625.000 uS


ADC CHNL RANGE AC/DC

1 250 mV
2 250 mV


DELAY

0.0 S
0.0 S


CAL (C1/C2)


1.00000
1.00000


Figure 3-40. Signal Analyzer Measurement Parameters for Experiment #6.










Transducer 1


A SPEC


5.0000
m


REAL

1.0000
m


2008.0 HZ


A SPEC
-28.000




LGMAG
DB



-50.000


908.88


288.00


Figure 3-41.


Average Normalized
Sandpaper, Plotted


Spectral Signature Obtained from Transducer
on Both Real and Logarithmic Scales.


1 and Transducer 2 for 36 Grit


A SPEC
2W MMV


m



REAL


2.0000


200.88


900.00


m


A SPEC
-10.000




LGMAG
DB



-58.000


200.00


900.00


Transducer 2


*l


r










Transducer 2


HZ


--1--I I
900.00


900.00


Figure 3-42.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for 60 Grit
Sandpaper, Plotted on Both Real and Logarithmic Scales.


A SPEC
_1 VRR0


A SPEC

7.0000
m



REAL


1.8000
m


m



REAL


1.0080
m


A SPEC
-10.000




LGMAG
DB



-50.000


200.00


I0.00
2008.00


280.00


A SPEC
-20.000




LGMAG
DB



-50.800


980.00


988.00


200.00


1 (


Transducer 1


Q


.









Transducer 1


A SPEC

7.0000
m



REAL


1.0000
m


A SPEC
-20. 000



LGMAG
DB



-40.000


A SPEC
5.0000




REAL


500.00
U


A SPEC
-20.000



LGMAG
DB



-58.888


980.00


288.00


200.00


Figure 3-43.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for 80 Grit
Sandpaper, Plotted on Both Real and Logarithmic Scales.


980.00


280.00


200.00


900.00


Transducer 2









Transducer 1


A SPEC

6.0000



REAL


1.0000
m


A SPEC
10. 000
m



REAL


1.0000



A SPEC
-10.000



LGMAG



-48.888000


288.88


A SPEC
-28.000



LGMAG
DB



-50.000


900.00


288. 88


Figure 3-44.


Average Normalized Spectral Signature Obtained from Transducer 1 and Transducer 2 for 100 Grit
Sandpaper, Plotted on Both Real and Logarithmic Scales.


900.00


280.08


200.00


900.00


900.00


Transducer 2









Transducer 1


A SPEC
5.0000




REAL


500.00
u


A SPEC
-28.888



LGMAG
DB



-4. 000


900.08


900. 00


A SPEC
6.0000
m



REAL


1.0000
m


A SPEC
-20.088



LGMAG
DB


20. 00


Figure 3-45.


Average Normalized Spectral Signature Obtained from Transducer 1
Sandpaper, Plotted on Both Real and Logarithmic Scales.


and Transducer 2 for 180 Grit


200.00


288.00


200.00


900.00


90. 00


Transducer 2









Transducer 1


A SPEC

3.5000
m



REAL


508.00



A SPEC
-20.000




LGMAG
DB



-40.0008


A SPEC
4.0000
m



REAL


500.00
588. 88
u


A SPEC
-28.088




LGMAG
DB



-48.8 0


280.88


200.00


Figure 3-46.


Average Normalized
Sandpaper, Plotted


Spectral Signature Obtained from Transducer 1 and Transducer 2 for 220 Grit
on Both Real and Logarithmic Scales.


900.00


208.00


900.00


288.00


98. 00


Transducer 2









spectra varied slowly in comparison to the spectra encountered for

previous experiments, the pattern vector was reduced in dimension from

64 variables to 16 variables. Each of the pattern variables represented

100 Hz wide spectral bands. The pattern vectors for all observations

are given in Appendix F.


Experiment #7--Repeatability

A second prototype sensor, with the same configuration as the

first, was examined in one limited experiment to ensure that the results

presented could be reproduced. The test objects of this experiment were

the same subset of objects from Experiment #4 tested in Experiment #5.

The input conditions and test procedure were identical to those in

Experiment #4. One test cycle with two observations per object was

performed yielding the six signatures shown in Figures 3-47 to 3-52.

The pattern vectors for all observations are given in Appendix G.









Transducer 1


A SPEC
30.000



REAL


5.0000
m



A SPEC
-10.008



LGMAG
DB



-80.000


800.00


HZ 8I 0. I I
HZ 800.00


Figure 3-47.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.25 inches, Plotted on Both Real
and Logarithmic Scales.


0.0 HZ


A SPEC

18.000



REAL


2.0000



A SPEC
-10.000



LGMAG
DB



-70.000


0.0


800.00


0.0


0.0


88800. 00


Transducer 2









Transducer 1


A SPEC
30.000 J
m


REAL


5.0000




A SPEC
-10.0 00




LGMAG
Dl



-70.000


A SPEC
20.000
M


REAL


800.00


0. 0


Figure 3-48.


Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.25 inches, Plotted on Both Real
and Logarithmic Scales.


0.0


0.0


88.800


8. 800


2.0000



A SPEC
-10.000




LGMAG
DB



-68.080


_.__


Transducer 2


r -1 II I I I









Transducer 1
A SPEC

18.000
m



REAL


2.0000
_____^ ________I m


800.00


888.00


A SPEC
-18.000




LGMAG
DB


8.0


Transducer 2


80. 00


888. 0


Figure 3-49.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.26 inches, Plotted on Both Real
and Logarithmic Scales.


A SPEC
25. 08



REAL


5.0000
m


0.0


A SPEC
-108.00




LGMAG
DB



-70.000










Transducer 1


A SPEC
30.000
m



REAL


5.0000




A SPEC
-10.000



LGMAG
DB



-78.000


A SPEC
1. 000



REAL


2.0000
m


A SPEC
-10.000



LGMAG
DB



-80.080


800.00


0.8


Figure 3-50.


Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.26 inches, Plotted on Both Real
and Logarithmic Scales.


8.0 HZ


88800.00


0.0


88. 80


Transducer 2










Transducer 1


800.00


888.00


A SPEC
-10.000




LGMAG
DB



-70.000


A SPEC

14.000



REAL


2.8000
M


0.0


0.0


Figure 3-51.


Normalized Spectral Signature for Experimental Observation 1 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.35 inches, Plotted on Both Real
and Logarithmic Scales.


A SPEC

30.000
m


REAL


5.0000
a


0.0


A SPEC
-10.000



LGMAG
DB



-68.008880


8880.00


0.0


808.00


Transducer 2









Transducer 1
A SPEC
16.080
m



REAL


2. 000
m


Transducer 2


888.00


808.00


A SPEC
-18.000



LGMAG
DB



-88. 80


08.


Figure 3-52. Normalized Spectral Signature for Experimental Observation 2 Obtained from Transducer 1 and
Transducer 2 of the Second Prototype for Feature Spacing 0.35 inches, Plotted on Both Real
and Logarithmic Scales.


A SPEC
30.A00



REAL


5.80008
M


0.0


A SPEC
-18.088




LGMAG
DB



-80. 888


.LIL JI9 A1( .i1 A


0.8


808.00
















CHAPTER IV
EXPERIMENTAL RESULTS


The principal purpose of the experiments was to assess the

potential of a new touch sensor concept over a wide range of potentially

useful tactile tasks. The performance is measured by the ability to

recognize/identify a touched feature, object, or surface from a domain

of possible features, objects, or surfaces. The performance should also

be compared to the performance of alternative touch sensing approaches,

unfortunately few comparable results are available. Nevertheless, where

available, published results from other approaches will be provided.

In general, the discrimination procedure consists of a training

phase and a classification phase. The training phase establishes a

standard pattern vector, or template, from the so-called design set of

observations, for later comparison with patterns input in the classi-

fication phase. A pattern is classified into the group with the closest

fitting template.

The discrimination process was performed by the BMDP7M Stepwise

Discriminant Analysis computer program (Dixon et al., 1983). Output

from the program, including the classification matrix, the jackknifed

classification, and a scatter plot showing the location of all cases in

the first two variables of discriminant space, is provided in this

chapter.

The jackknife classification method is an estimate of the true

error rate, which is the expected error rate on future observations from









the same domain (Hand, 1981). The jackknifed classification procedure

includes pattern vectors for all observations in the design set except

one; thus, the excluded observation becomes the only case evaluated in

the classification phase. The procedure is performed for every observa-

tion, testing each independently.


Experiment #1--Shapes and Sizes

For this experiment, six objects, three spheres and three

cylinders, constituted the domain. Twelve observations, pattern

vectors, were collected for each object. These observations were

randomly divided into two groups; approximately 70% were selected for

the design set used to build the discriminant function, the remaining

observations served as the test set in the classification phase.

The results, presented in Table 4-1, show that 98% of the design

set, and 100% of the test set, of observations were correctly identi-

fied. Only one observation from the total of 72 was misclassified. The

group names refer to the ball bearings with 5/16, 3/16, and 5/32 inch

diameters, and to the cylinders with 5/16, 3/16, and 1/8 inch diameters.

The group names preceded by a "T" are the test groups not used in the

training phase.

The jackknifed classification, presented in Table 4-2, shows two

misclassifications. Both errors were misclassifications of 3/16 inch

ball bearings as 5/32 inch ball bearings. The difference in radius for

these two objects is 1/64 inch, and 22 of the 24 members of the two

groups were correctly classified. In addition, Table 4-2 shows that all

of the observations were correctly classified into the proper shape

group, sphere or cylinder. These two errors are most likely attribu-

table to experimental considerations rather than a limitation of the
















Table 4-1. Classification Matrix for Experiment #1--Shapes and Sizes.


GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP-
CORRECT
BB516 BB316 BB532 CYL516 CYL316 CYL18



BB516 100.0 10 0 0 0 0 0

BB316 100.0 0 8 0 0 0 0

BB532 87.5 0 1 7 0 0 0

CYL516 100.0 0 0 0 11 0 0

CYL316 100.0 0 0 0 0 10 0

CYL316 100.0 0 0 0 0 0 10

TBB516 100.0 2 0 0 0 0 0

TBB316 100.0 0 4 0 0 0 0

TBB532 100.0 0 0 4 0 0 0

TCYL516 100.0 0 0 0 1 0 0

TCYL316 100.0 0 0 0 0 2 0

TCYL18 100.0 0 0 0 0 0 8















Table 4-2. Jackknifed Classification for Experiment #1--Shapes and
Sizes.


GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP-
CORRECT
BB516 BB316 BB532 CYL516 CYL316 CYL18



BB516 100.0 12 0 0 0 0 0

BB316 83.3 0 10 2 0 0 0

BB532 100.0 0 0 12 0 0 0

CYL516 100.0 0 0 0 12 0 0

CYL316 100.0 0 0 0 0 12 0

CYL18 100.0 0 0 0 0 0 12



TOTAL 97.2 12 10 14 12 12 12









sensor; since the test apparatus was inadvertently misaligned during the

course of this experiment.

The scatter plot, shown in Figure 4-1, gives a feeling for the

separability of the groups. However, the figure only displays a two-

dimensional view of the five-dimensional discriminant space for this

experiment.

Many authors have described tactile sensors capable of recognizing

objects and/or surface features (Briot, 1979, Stute and Erne, 1979,

Dario et al., 1983, Snyder and St. Clair, 1978). However, specific

results, comparable to those reported here, were not presented. Other

researchers report results from experiments designed to identify objects

by grasping it with an end-effector equipped with several sensors.

Kinoshita, Aida, and Mori (1975) experimented with an artificial hand

furnished with 22 contact sensors (on-off switches). They grasped

cylinders and square pillars, five different sizes of each, 20 times to

demonstrate the recognition capability, their results are presented in

Table 4-3. Okada and Tsuchiya (1977) used both contact information and

the angular position of finger joints to identify objects in a two-step

recognition procedure. A variety of shapes were grasped 20 times to

perform the discrimination. More than 95% of the observations were

correctly classified on the basis of shape, and more than 60% of the

observations were correctly recognized as small, medium, or large. More

recently, Ozaki et al. (1982) used seven matrix-type tactile sensors,

placed on a three-fingered end-effector, for the recognition of two-

dimensional convex contours. They demonstrated the capability to regen-

erate the contours of a circular cylinder and a right prism.



















OVERLAP OF DIFFERENT GROUPS IS INDICATED BY *
+ .. . .. . . . . . . . . .. . .. . . . .. .. . . . . .. . . . .. . . . . . .. . . . . . .. . . .+ .. .





10 + B +
CC BC A
23CB AA
*B C A A
B
B A IA
C A A
A 5. + A +
N A A
0 -
N -
I -
C -
A -
L 0.
0. + +


FF
FF 6F F
F
F F


2



-10 +


E
EEE
E
EE 5
EE E


D D
DDD DD
40
D D
D


-15 +


. .... ... .... ... .. . .. .. .. .. ... +. ... +. ... +. . .. . . . . . .. ... .. ... ... .. . . *. .. . . .*. .. . .. .. ..
-22 -18 -14 -10 -6. -2. 2. 6. 10 14 1i 22
-24 -20 -16 -12 -8. -4. 0. 4. 8. 12 16 20
CANONICAL VARIABLE 1


Scatter Plot for Observations from Experiment #1.


Figure 4-1.
















Table 4-3. Comparable Results Presented by Kinoshita, Aida, and Mori
(1975).


OBJECT SIZE PERCENT CORRECT
(mm)


CYLINDER 70 (dia) 0

CYLINDER 80 90

CYLINDER 90 100

CYLINDER 100 100

CYLINDER 110 50

SQUARE PILLAR 50 (side) 25

SQUARE PILLAR 60 95

SQUARE PILLAR 70 100

SQUARE PILLAR 80 90

SQUARE PILLAR 90 100









Experiment #2--Patterns of Features

The domain of test objects for this experiment consisted of seven

different arrangements of ball bearings. Nine observations were made

for each arrangement. Approximately 70% of the observations were

randomly included in the design set, and the remainder were placed in

the test set.

The experimental results, presented in Table 4-4, show that 100% of

the observations in both the design set and the test set were correctly

identified. The group names refer to the Braille character the arrange-

ments resemble.

The jackknifed classification, presented in Table 4-5, and the

scatter plot of the observations, shown in Figure 4-2, also indicate

successful discrimination. Observations were correctly classified in

all but one case, a success rate of 98.4%.


Experiment #3--Feature Orientation

For this experiment, the feature of interest, the slot on the head

of a screw, was oriented at six different angles with respect to the

direction of the ridges on the sensor skin. As in the previous experi-

ments, the observations were randomly divided into two groups, the

design set and the test set. The groups were allotted approximately 65%

and 35% of the observations respectively.

Both the classification matrix, Table 4-6, and the jackknifed

classification, Table 4-7, show that 100% of the observations were

correctly identified. The group name indicates the slot angle with

respect to the direction of the ridges. In addition, Figure 4-3 pro-

vides a spatial depiction of the observations for the first two

canonical variables.
















Table 4-4. Classification Matrix for Experiment #2--Patterns of
Features.





GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP-
CORRECT
A B C D E F G



A 100.0 7 0 0 0 0 0 0

B 100.0 0 6 0 0 0 0 0

C 100.0 0 0 6 0 0 0 0

D 100.0 0 0 0 7 0 0 0

E 100.0 0 0 0 0 8 0 0

F 100.0 0 0 0 0 0 8 0

G 100.0 0 0 0 0 0 0 5

TA 100.0 2 0 0 0 0 0 0

TB 100.0 0 3 0 0 0 0 0

TC 100.0 0 0 3 0 0 0 0

TD 100.0 0 0 0 2 0 0 0

TE 100.0 0 0 0 0 1 0 0

TF 100.0 0 0 0 0 0 1 0

TG 100.0 0 0 0 0 0 0 4
















Table 4-5. Jackknifed Classification for Experiment #2--Patterns of
Features.


GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP-
CORRECT
A B C D E F G



A 100.0 9 0 0 0 0 0 0

B 88.9 0 8 0 0 0 1 0

C 100.0 0 0 9 0 0 0 0

D 100.0 0 0 0 9 0 0 0

E 100.0 0 0 0 0 9 0 0

F 100.0 0 0 0 0 0 9 0

G 100.0 0 0 0 0 0 0 9



TOTAL 98.4 9 8 9 9 9 10 9



















OVERLAP OF DIFFERENT GROUPS IS INDICATED BY *
.... + ... ........ .. .. .. ... ... .... ...+ ....+ .+ .. .. . + ....+ ....+ .+ . ... . . ... .


C C
C C


A A


1 A
A A A


9. +





6. +





3. +





0. +





-3. +


F
FF
F 6


D E
D I
D
D 4


G D D


B B
28
B
B


B

B -


.... ... .... + ... .. ... .. .. .. ... . +. . .+... +. + ... .. . . . .+ .. +." .... +. .. .+ . ... . +.
-9. -7. -5. -3. -1. 1. 3. 5. 7. 9. 11 13
-8. -6. -4. -2. O. 2. 4. 6. 8. 10 12
CANONICAL VARIABLE 1


Scatter Plot for Observations from Experiment #2.


0 G
G 0
G 7
D


C
A
N
a
N
I
C
A
L
V
A
R
I
A
B
L
E

2


-6. +






-10


Figure 4-2.















Table 4-6. Classification Matrix for Experiment #3--Feature
Orientation.


GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP--
CORRECT
SLOTO SLOT15 SLOT30 SLOT60 SLOT90 SLOT120



SLOTO 100.0 6 0 0 0 0 0

SLOT15 100.0 0 5 0 0 0 0

SLOT30 100.0 0 0 7 0 0 0

SLOT60 100.0 0 0 0 5 0 0

SLOT90 100.0 0 0 0 0 7 0

SLOT120 100.0 0 0 0 0 0 6

TSLOTO 100.0 2 0 0 0 0 0

TSLOT15 100.0 0 3 0 0 0 0

TSLOT30 100.0 0 0 1 0 0 0

TSLOT60 100.0 0 0 0 3 0 0

TSLOT90 100.0 0 0 0 0 1 0

TSLOT120 100.0 0 0 0 0 0 0
















Table 4-7. Jackknifed Classification for Experiment #3--Feature
Orientation.


GROUP PERCENT NUMBER OF CASES CLASSIFIED INTO GROUP-
CORRECT
SLOTO SLOT15 SLOT30 SLOT60 SLOT90 SLOT120



SLOTO 100.0 8 0 0 0 0 0

SLOT15 100.0 0 8 0 0 0 0

SLOT30 100.0 0 0 8 0 0 0

SLOT60 100.0 0 0 0 8 0 0

SLOT90 100.0 0 0 0 0 8 0

SLOT120 100.0 0 0 0 0 0 8



TOTAL 100.0 8 8 8 8 8 8




















OVERLAP OF DIFFERENT GROUPS IS INDICATED BY *
S. . ... .... +...+ ... ... .. . . ... +.... .. ... .. .+ ....+ .... .
90. +





60. +




C -
A -
N 30. + +
0 A -
N 1A
I
C DD
A D4D
L
56F
V 0.0 + EEE 3C +
A CC
R -
I
A -
B -
L
E -
-30. + +
2
B2B



-60. +





-90. +


~ ++...+. . . . ...... .... .... .... -.... .. . ... .... . ... .... ... 4" .... ... ". .... .... .... . .. .- .. .
-63. -45. -27. -9. 0 9. 0 27. 45. 63. 81. 99. 117 135
-54. -36. -18. 0 18. 36. 54. 72. 90. 108 126
CANONICAL VARIABLE 1


Figure 4-3.. Scatter Plot for Observations from Experiment #3.









Experiment #4--Relative Feature Position

Arrangements of two ball bearings were used to demonstrate the

ability to identify the relative position of one ball bearing to an-

other. The second ball bearing was placed either 0.25, 0.26, or 0.35

inches behind the reference ball bearing. Thus, the difference in

relative location was as small as 0.01. inches. Three sets of these

arrangements were tested. Unlike the previous experiments, observations

selected for the design set and the test set were not randomly chosen.

Rather, the observations from the first two sets of arrangements com-

posed the design set, and the observations from the third set constitu-

ted the test set.

Both the classification matrix, Table 4-8, and the jackknifed

classification, Table 4-9, show that 100% of the observations were

correctly classified. The group names indicate the distance between the

ball bearings in thousandths. Figure 4-4 depicts the separability

between the groups for the first two variables in discriminant space.

This experiment demonstrates the ability of the prototype sensor to

accurately discriminate differences in relative position of 0.01 inches.

These results give an indication of the dynamic resolution potential of

this touch sensing approach. The resolution of other tactile sensors is

dicated by the fineness of their transducer arrays. The demonstrated

dynamic resolution of this prototype is comparable, or superior, to the

resolution reported for other sensors and sensor concepts (Bejczy, 1980;

Briot, 1979; Nevill and Davis, 1984).


Experiment #5--Changes in Speed

For this experiment, the angular speed of the turntable was changed

from 33 1/3 to 45 rpm. One set of arrangements from Experiment #4 were




University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs