The Pennsylvania State University

The Graduate School

DESIGN OF TENSEGRITY STRUCTURES FOR

SHAPE CHANGING APPLICATIONS

A Dissertation in

Aerospace Engineering

by

Kaila M. Roffman

© 2022 Kaila M. Roffman

Submitted in Partial
Fulfillment of the
Requirements
for the Degree of

Doctor of Philosophy

December 2022
The dissertation of Kaila M. Roffman was reviewed and approved by the following:

George A. Lesieutre
Professor of Aerospace Engineering
Associate Dean for Research and Graduate Programs, College of Engineering
Dissertation Adviser
Chair of Committee

Puneet Singla
Professor of Aerospace Engineering

Xin Ning
Assistant Professor of Aerospace Engineering

Mary Frecker
Head of the Department of Mechanical Engineering

Amy Pritchett
Professor of Aerospace Engineering
Head of the Department of Aerospace Engineering

ii
 Abstract

In space applications, lightweight, deployable structures are frequently studied and

developed to meet launch vehicle constraints. This is particularly true for very large-scale

structures, where the dimensions and volume of launch vehicle fairings can severely limit the in-

space scale of structures without deployment. Tensegrity structures are very promising for such

applications, offering a high stiffness-to-mass ratio and deployability through actuation of changing

member lengths.

Tensegrity structures are self-equilibrated, pre-stressed structures. Comprised of struts and

cables, tensegrity members are modeled as uniaxial, with struts carrying compression and cables

carrying tension, meeting at nodes modeled as frictionless ball joints. A tensegrity form is

characterized by sets of nodal coordinates and member force densities (or forces), while a tensegrity

configuration refers to how the members are connected between the nodes.

Much of the literature on tensegrity structures focuses on the form-finding problem—

seeking an equilibrated combination of nodal coordinates and force densities for a particular

configuration. A wide variety of approaches to solving the form-finding problem are described in

the literature. In general, the form-finding problem is intrinsically formulated to seek only a single

equilibrated form for a particular tensegrity configuration. This is sufficient for many static

applications of tensegrity structures, where only one form would be necessary. However, tensegrity

structures can take on a wide variety of forms for a particular configuration, and little research has

been done to-date exploring the variety of shapes a particular configuration can achieve. In shape

iii
changing applications, such as deployable structures, understanding the wider variety of shapes a

structure can achieve could provide valuable insight as to how these structures can best be used.

This dissertation has two main parts. First, analytical and numerical approaches are

developed which can be used to describe the range of achievable shapes (forms) a particular

tensegrity configuration can achieve. In the analytical approach, a minimal set of design variables

describing the shapes is first developed. This minimal set is then restricted, providing bounds on

the design variables which limit the resulting shapes to only those which are also equilibrated. For

configurations too complex to solve analytically, a numerical approach is developed. This

numerical approach can work with any set of design variables and needs only a single known

equilibrated shape to begin. This shape is used as a starting point, from which parametric lines (in

terms of the design variables) are extended. These lines are used to seek new equilibrium shapes in

the span of the design variables. The end result is a representative numerical dataset which

approximates the range of achievable shapes a tensegrity structure can achieve. These approaches

demonstrate not only that a particular tensegrity configuration can achieve a wide variety of shapes,

but also that the extent of these shapes can be characterized.

The second part of this dissertation focuses on additional analysis methods which can be

used to design a tensegrity structure specifically for shape changing applications. Methods for

incorporating member length constraints are developed, including the definition of special design

variables. Additionally, a path planning approach is developed that identifies equilibrated shapes

through which a tensegrity can move, starting from and ending at designated shapes. Finally, finite

element models are developed and used to study the tensegrity’s deflections (stiffness) under

station-keeping loads, as well as its natural frequencies of vibration. These tools are used in a

demonstrative case study, investigating the potential utility of a modular tensegrity structure as a

structural backbone of a parabolic reflector on the same scale as the James Webb Space Telescope.

The resulting tensegrity structure exhibits structural characteristics comparable to if not better than

iv
those of existing parabolic reflectors in space applications. This case study not only shows the

utility of the analysis methods developed when selecting a variety of design parameters with a

particular application in mind, but also demonstrates that tensegrity structures have great potential

for use in deployable, shape changing applications.

v
 Table of Contents

List of Figures .......................................................................................................................... viii

List of Tables ........................................................................................................................... xiii

List of Symbols ........................................................................................................................ xiv

Acknowledgments.................................................................................................................... xvii

Chapter 1 Introduction ............................................................................................................. 1

1.1 Tensegrity Background .............................................................................................. 7

1.2 Equilibrium of Tensegrity Structures ......................................................................... 10
1.3 The Form-Finding Problem........................................................................................ 12
1.3.1 Iterative Form-Finding .................................................................................... 13
1.3.2 Metaheuristic Form-Finding............................................................................ 19
1.3.3 Analytical Form-Finding ................................................................................. 25
1.4 Existing Tensegrities for Shape Changing Applications............................................ 29
1.5 Organization of Dissertation ...................................................................................... 31

Chapter 2 Analytical Characterization of a Tensegrity’s Achievable Shapes ......................... 34

2.1 A 5-Step Approach to Analytically Finding the Range of Achievable Shapes .......... 35
2.1.1 Form a Geometric Matrix ................................................................................ 37
2.1.2: Form an Equilibrium Matrix .......................................................................... 40
2.1.3 Form an Augmented Matrix Using the Null Spaces of G and E ..................... 41
2.1.4: Obtain Equations Using RREF of the Augmented Matrices .......................... 43
2.1.5: Use Resulting Equations to Restrict Geometric Coefficients ........................ 44
2.2 The 2D X-tensegrity................................................................................................... 45
2.3 Triplanar Tensegrity ................................................................................................... 52

Chapter 3 Numerical Characterization of a Tensegrity’s Achievable Shapes ......................... 60

3.1 The Branching Algorithm .......................................................................................... 61

3.1.1 Exit Criteria of the Branching Algorithm........................................................ 63
3.1.2 Factors Influencing Efficiency ........................................................................ 65
3.1.3 Summary of the Branching Algorithm ............................................................ 66
3.1 Example: 2D X-tensegrity.......................................................................................... 69
3.2.1 Constrained X-Tensegrity ............................................................................... 69
3.2.2 Unconstrained X-Tensegrity ........................................................................... 72
3.2 Example: 3D Triplex .................................................................................................. 75
3.3.1 Constrained Triplex ......................................................................................... 77
3.3 A Force-Finding Method for Complex Tensegrity Structures ................................... 81
3.5 Example: 3D Triplex Module .................................................................................... 84
3.6 Branching Algorithm Conclusion .............................................................................. 87

vi
Chapter 4 Designing Tensegrities for Shape Change .............................................................. 89

4.1 Geometric Principles to Incorporate Length Constraints ........................................... 90

4.1.1 Triangle Properties .......................................................................................... 92
4.1.2 Triangular Pyramid Properties ........................................................................ 95
4.1.3 Spherical Coordinates ..................................................................................... 97
4.2 Path Planning ............................................................................................................. 100
4.2.1 Optimal Shape Fit to a Prescribed Surface ...................................................... 101
4.2.2 Using Target Start and End Shapes to Path Plan............................................. 102
4.2.3 Experimental Verification of Path Planning ................................................... 104
4.3 Finite Element Methods for Stiffness and Modal Analysis ....................................... 111

Chapter 5 Example Design Problem ........................................................................................ 115

5.1 Design Analysis of a Triplex Module ................................................................ 124

5.1.4 Effects of Pre-stress Factor.............................................................................. 126
5.1.2 Effects of Bay-Height...................................................................................... 127
5.1.3 Effects of Cable Radius ................................................................................... 128
5.1.4 Effects of Strut Radius .................................................................................... 130
5.1.5 Summary of Design Variable Effects .............................................................. 132
5.2 Design Analysis of Triplex Module Array................................................................. 133
5.2.1 Effects of Pre-stress Factor.............................................................................. 134
5.2.2 Effects of Bay Height ...................................................................................... 135
5.2.3 Effects of Cable Radius ................................................................................... 136
5.2.4 Effects of Strut Radius .................................................................................... 137
5.2.5 Summary of Design Variable Effects .............................................................. 138
5.3 Additional Design Considerations for a Triplex Module Array................................. 139
5.3.1 Fitting to a Paraboloid Surface ........................................................................ 142
5.3.2 Path Planning from Stowed to Deployed Shapes ............................................ 146
5.3.3 Summary of Triplex Module Array as a Support for a Parabolic Reflector.... 148

Chapter 6 Conclusions and Future Work ................................................................................. 151

6.1 Summary .................................................................................................................... 151

6.2 Conclusions ................................................................................................................ 156
6.3 Future Work ............................................................................................................... 157

Appendix A Using Length Constraints to Derive Design Variables ....................................... 159

A.1 Design Variable Derivation for a Constrained Triplex ............................................. 159

A.2 Design Variable Derivation for a Constrained Triplex Module ................................ 164

Appendix B Effects of PSO force-finding Parameters for a Triplex Module .......................... 169

B.1 Modification of the PSO Force-Finding for a Symmetric Shape .............................. 171

Bibliography ............................................................................................................................ 174

vii
 List of Figures

Figure 1.1: A model of the spine (left) with ligaments in red [55] adjacent to a tensegrity
model of a spine (right) with cable members in red [30]; a visual example of how
tensegrities can be biomimetic ......................................................................................... 2

Figure 1.2: Deployment stages of a scissor structure, adapted from [56]. Structure is shown
fully stowed (left), in transition (middle) and fully deployed (right). This is an example
of 1D deployment, where the structure can be deployed by moving the red node along
the dashed axis. ................................................................................................................ 5

Figure 1.3: Deployment of miura-ori origami sheet, adapted from [57]. The miura-ori fold
couples motion in two directions, such that pulling the red node in the direction of the
red arrow unfolds the sheet in two directions (the black node and y-motion on the edge
along the x-axis are held fixed). Deployment in this figure is shown from left to right.
.......................................................................................................................................... 5

Figure 1.4: Kenneth Snelson’s “needle tower” sculpture [58]; an example of a class-1
tensegrity structure. .......................................................................................................... 8

Figure 1.5: Three different tensegrity configurations, each with 6 struts and 12 nodes. Left
(a) is an icosahedron; a spherical tensegrity. Center (b) is a 6-plex cylindrical
tensegrity. Right (c) is a class-1 cylindrical tensegrity, formed by stacking two 3-plex
cylindrical tensegrities. .................................................................................................... 10

Figure 2.1: The 2D x-tensegrity, comprised of vectors representing each member, and some
cycles (a)-(d) made from the member vectors. ................................................................ 39

Figure 2.2: Geometric coefficients corresponding to equilibrated shapes of the 2D x-

tensegrity, found using a brute-force approach checking all α combinations between
0 and 2 against one another and checking each for equilibrium. ..................................... 49

Figure 2.3: Example cases of boundaries on αx2 and αx1 for the x-tensegrity .......................... 50

Figure 2.4: Combined regions for all cases of μ for the X-tensegrity ...................................... 52

Figure 2.5: A 2D triplanar tensegrity ....................................................................................... 53

Figure 2.6: Equilibrated shapes for the triplanar tensegrity, found by checking
combinations of geometric coefficient values against one another, varied between -2
and 2. ................................................................................................................................ 54

Figure 2.7: All possible cases for signs of geometric coefficient terms for triplanar
tensegrity, μ=(α1y/α3y) ...................................................................................................... 55

Figure 2.8: Example regions from case (B) of viable geometric coefficient combinations
for the triplanar tensegrity. ............................................................................................... 57

viii
Figure 2.9: Example regions from case (C) of viable geometric coefficient combinations
for the triplanar tensegrity. ............................................................................................... 58

Figure 3.1: Flow-chart of branching algorithm ........................................................................ 68

Figure 3.2: A sample iteration of the total solution set found by the numerical approach,
where color indicates Mahalanobis distance of each point. Red x’s indicate the points
that were selected as branch points for the next iteration, where the sample of points
were weighted by the Mahalanobis distance .................................................................... 70

Figure 3.3: Maximum Mahalanobis Distance for the constrained X-tensegrity. ..................... 71

Figure 3.4: Comparison of branching algorithm results to analytical results for constrained
X-tensegrity...................................................................................................................... 72

Figure 3.5: Maximum Mahalanobis distance as the algorithm iterated for the unconstrained
2D X-tensegrity ................................................................................................................ 74

Figure 3.6: The results of the numerical approach (a) compared with the results of the
analytical approach (b) shown in terms of the geometric coefficients α1x and α2x.. ......... 74

Figure 3.7: A triplex tensegrity. ............................................................................................... 75

Figure 3.8: Maximum Mahalanobis distance for the numerical approach with an
unconstrained triplex. ....................................................................................................... 77

Figure 3.9: Constrained triplex; all members fixed in length except for members 7, 8 and
9........................................................................................................................................ 78

Figure 3.10:The results from the numerical approach for a constrained triplex. The point
marked by an x is the initial equilibrated point, corresponding to the parallel triplex
shape shown in Figure 3.5................................................................................................ 79

Figure 3.11: Maximum Mahalanobis Distance over the 126 iterations it took to converge
for the constrained triplex. ............................................................................................... 80

Figure 3.12: Comparison of results from the numerical approach developed for this
research with the results of Roffman [15] for a constrained triplex. ................................ 81

Figure 3.13: A module comprised of 6 triplexes. .................................................................... 84

Figure 3.14: Top face of the triplex module, showing the inner polygon highlighted in blue.
.......................................................................................................................................... 86

Figure 3.15: An example of the spherical design variables, θ ,ϕ and λ shown in their local
frame for the first 3 nodes of the top-face cables of the triplex module. ......................... 86

Figure 3.16: Maximum Mahalanobis distance for numerical approach used with a
constrained triplex module ............................................................................................... 87

ix
Figure 4.1: A 6-sided polygon. Dashed lines show how the polygon may be broken up into
4 sub-triangles. ................................................................................................................. 92

Figure 4.2: A triangular pyramid, with vertices A, B, C, P. Point PP is the projection of P

on the x-y plane. ............................................................................................................... 95

Figure 4.3: An isolated triangle from triangular pyramid A-B-C-P in Figure 4.2 ................... 96

Figure 4.4: A triangle from the triangular pyramid A-B-C-P from Figure 4.2. ....................... 96

Figure 4.5: : Spherical representation of node k relative to node j for member i. ................... 98

Figure 4.6: Possible locations of a third node (k) as the circular intersection of two spheres,
where members of length lik and ljk must both meet at node k. Angle λ is the
parameterized angle on the circle of intersection representing geometrically possible
node k locations. .............................................................................................................. 99

Figure 4.7: A triplex module. .................................................................................................. 101

Figure 4.8: Experimental set-up. .............................................................................................. 105

Figure 4.9: Fit of triplex based on experimental structure dimensions to a paraboloid

surface. ............................................................................................................................. 108

Figure 4.10: Change in lengths between adjacent steps in the path tested. ............................. 109

Figure 4.11: Snapshots of experiment next to corresponding steps from simulation .............. 110

Figure 5.1: Dimensions from the James Webb Space Telescope used to approximate the
focal length if it were a perfect paraboloid surface. ......................................................... 116

Figure 5.2: A triplex module. ................................................................................................... 117

Figure 5.3: Array made of 19 modules, with each module shaded in different colors for
clarity. .............................................................................................................................. 117

Figure 5.4: A stowed module (left) shown adjacent to a deployed module (right).................. 124

Figure 5.5: First 6 mode shapes of the triplex module, showing only top-face members.
Mode shapes are in solid blue lines, with original members in dashed black lines. ........ 125

Figure 5.6: Effects of pre-stress factor on deformation and first natural frequency. ............... 127

Figure 5.7: Effect of bay height on the structure’s mass (MT) and force applied at each node
(FT). .................................................................................................................................. 127

Figure 5.8: Effects of bay height on deformation and first mode. ........................................... 128

Figure 5.9:Effects of cable radius on mass (MT) and resulting thrust force (FT) on each
node. ................................................................................................................................. 129

x
Figure 5.10: Effects of cable radius on deformation and first mode. ....................................... 129

Figure 5.11: Effects of strut inner and outer radii on mass (MT) and force at each node (FT).
Strut inner radii are shown as different colored lines....................................................... 130

Figure 5.12: Effects of strut outer radius on deformation and first mode. ............................... 131

Figure 5.13: Closest distance between two struts for a given triplex height (bay height). ...... 132

Figure 5.14: Effects of pre-stress factor on the average deformation and first mode. ............. 135

Figure 5.15: Effects of bay height on average deformation and first mode of triplex module
array. ................................................................................................................................ 136

Figure 5.16: Effects of cable radius on deformation and first mode in triplex module array.
.......................................................................................................................................... 137

Figure 5.17: Effect of strut outer and inner radius on average deflection and first mode for
the triplex module array. .................................................................................................. 138

Figure 5.18: Triplex Module Array with paraboloid surface overlaid in light blue................. 143

Figure 5.19: The triplex module array; the 7 modules outlined in a thick black line can be
rotated +/- 120o to achieve the full array. ......................................................................... 144

Figure 5.20: Array fit to a paraboloid surface .......................................................................... 146

Figure 5.21: Select steps from path planning between a stowed shape and a deployed shape
fit to a paraboloid surface for the triplex module array.................................................... 147

Figure A.1: A triplex tensegrity, all members are fixed in length except supporting cables,
shown in green (l7,l8,l9). ................................................................................................... 159

Figure A.2: Angles β10 and Φ with respect to triangle N1-N2-N5, used to define N5. .............. 160

Figure A.3: The triangular pyramid used to define N6............................................................. 161

Figure A.4: The triangular pyramid used to define N4............................................................. 163

Figure A.5: A triplex module. .................................................................................................. 164

Figure A.6: Top face members and nodes of the triplex module. ............................................ 165

Figure A.7: An example of spherical angles used as design variables for the first 3 nodes
in the top face of the triplex module. ............................................................................... 166

Figure A.8: Example of the angle used as a design variable when a node location is
represented as the intersection of two spheres. ................................................................ 167

xi
Figure B.1: Effects of various parameters in PSO force-finding and their effects on average
run time and convergence upon a known solution. .......................................................... 170

Figure B.2: Groupings used for a triplex module .................................................................... 172

xii
 List of Tables

Table 4.1: Comparison of supporting cable lengths found from the best-fit to a paraboloid
surface using PSO, and the corresponding measured lengths in the experimental test-
bed after performing the maneuver from stowed to the target shape. .............................. 109

Table 5.1:pre-stresses used for analysis of the triplex module in the flat shape with member
grouping. Shared or not shared refers to members shared between the same nodes of
adjacent triplexes within the module................................................................................ 119

Table 5.2: Design variable parameters upper and lower bounds to evaluate structural
properties of the triplex module and an array comprised of 19 modules. ........................ 121

Table 5.3: Effects of increasing design variables on mass, average deflection, first mode
and packaging efficiency for a single triplex module. ..................................................... 133

Table 5.4: Effects of increasing design variables on mass, average deflection, first mode
and packaging efficiency for a single triplex module. ..................................................... 139

Table 5.5: Design variables used for the case study of a triplex module array as the
backbone of a parabolic reflector ..................................................................................... 148

Table 5.6: Structural Properties of triplex module array case study. ....................................... 149

Table B.1: Sample of symmetric force densities found for a triplex module .......................... 173

xiii
 List of Symbols

C = Connectivity Matrix
q = Force density vector
qi = Individual force density of member i
M = Number of members
n = Number of nodes
d = Dimension (i.e., 2D vs 3D)
M = Mass matrix
K = Stiffness matrix
f = External force vector
⃑ 𝒊,𝒋
𝒗 = Vector of member connecting nodes i and j
G = Geometric matrix
𝜟 = Vector of member length components
𝛥 = Individual member length components
nx, ny, nz = Nodal coordinate vectors
m = Member type vector (+1 cable, -1 strut)
𝜟𝒈 = Member length components, signs from G

E = Equilibrium matrix

̃
𝒒 = Modified force density vector

𝛼 = Geometric coefficient

v = Vector of design variables

vi = Individual design variable i

NL = Number of lines

nt = Number of points checked on a line

ne = Number of equilibrated points saved

NBP = Number of branch points

dm = Mahalanobis distance

𝜮 = Covariance matrix

xiv
 𝜇̅𝑖 = Average value of design variable i
NP = Number of points
Nv = Number of variables
li = Length of member i
θ = Azimuth in spherical coordinates
𝜙 = Elevation in spherical coordinates
λ = Angle on a parametric circle
f = Focal length (of paraboloid)
xp, yp, zp = Coordinates of center of paraboloid
wxy = Weighting factor
KT = Tangent stiffness matrix
KL = Linear stiffness matrix
E = Modulus elasticity
A = Cross-sectional area
KNL = Non-linear stiffness matrix
T = Transformation matrix
pf = Pre-stress factor
X = Amplitude
𝜔 = Natural frequency
𝜌 = Material density
σy = Tensile strength
Pcr = Critical buckling load
rout = Outer radius of strut
rin = Inner radius of strut
Δv = Change in velocity
FT = Thrust force
MT = Total structure mass
rc = Cable radius
a = acceleration
Mt = tensegrity structure mass
Vstow = volume of stowed tensegrity
rstow = circumscribing radius of stowed tensegrity
hstow = height of stowed tensegrity

xv
Ntop = number of nodes on top face of tensegrity
𝛥𝑧 = average distance from node-z and paraboloid
̅̅̅̅̅
∆𝑥𝑦 = average distance nodes moved in x-y

xvi
 Acknowledgments

This work was supported via federal funding. Any opinions, findings, conclusions, or

recommendations expressed in this work are those of the author and do not necessarily reflect the

official views of the federal government.

I would first like to thank my advisor, Dr. George A. Lesieutre, for his endless support and

guidance during my Penn State academic career. His encouragement and sage counsel was

invaluable throughout my time here; without either I’m not sure I could ever have made it this far.

Dr. Lesieutre introduced me to tensegrity structures beginning with my Master’s work when I first

arrived at Penn State, and this dissertation would never have come together without his feedback

and guidance.

I would also like to thank my committee, Dr. Puneet Singla, Dr. Xin Ning and Dr. Mary

Frecker. This dissertation only came to the best version it could be thanks to their insightful

comments and suggestions.

Pursuing my doctoral studies offered challenging experiences in more ways than one, and

I could not have done it without the boundless support of my friends and family. I’d like to thank

father, whose unyielding encouragement got me through most aspects of life, my academic career

being no exception. Knowing he was always there rooting for me was grounding in even the most

trying times. I’d also like to thank my aunt; I wouldn’t be who I am today without her love and

care. I am also grateful for my uncle; it was his suggestion that first led me to engineering, a path

which has challenged me in all the right ways. Last but not least, I’d like to thank my fiancé. His

support and encouragement was there from the very beginning. Having him by my side was

invaluable in getting where I am today.

Kaila M. Roffman

xvii
 Dedication

To my father and my fiancé

xviii
 Chapter 1

Introduction

Tensegrity structures are truss-like, lightweight, pre-stressed structures. Classical

tensegrities offer a collection of struts suspended by a network of tensioned cables. A lack of strut-

to-strut connections gives the impression of floating members, an appealing visual that undoubtedly

caught the eye of the first artists and architects who constructed them. Indeed, the earliest

constructions of tensegrity structures are found almost exclusively in art and architecture

applications, with the word “tensegrity” itself being coined by Buckminster Fuller, an architect [1].

While Fuller coined the word, his student, Kenneth Snelson, was the first to construct an actual

tensegrity, making the inventor of the tensegrity structure a contested title.

While Snelson built the first reported tensegrity structure in 1948 [2], it wasn’t until 1985

that the first scientific investigation of tensegrities as biomimetic structures was presented—in that

case, as a model for the cytoskeletal system of cells presented by Ingber and Jamieson [3]. While

most biomimetic scientific endeavors were first inspired by nature and then studied for human

contrived applications, tensegrities are an interesting inverse of order of operations. It takes only a

superficial understanding of tensegrities to readily see how they mimic the musculoskeletal systems

found in most animals (e.g., see Figure 1.1). Similar to Ingber and Jamieson’s cytoskeletal model

[3], nature’s solution to mammalian locomotion combines stiff members (bones, struts) with

flexible members (muscles/tendons, cables), to offer a flexible, adaptable system naturally capable

of withstanding a wide variety of external loadings and disturbances.

Despite their biomimetic relevance, most current tensegrity applications are more akin to

their architectural origins: as lightweight alternatives to classical load-bearing structures

1
(deployable booms [4,5,6,7], supports for space structures [8,9,10], etc.). Tensegrities offer an

excellent solution for scenarios requiring high stiffness-to-mass ratios (such as space applications),

but many of these load-bearing applications fail to exploit the structure’s biomimetic ability to offer

dramatic shape change in complex environments. In this regard, tensegrities have the most

opportunity to shine in general shape-change-oriented applications, including movement. Surely

mother nature has the upper hand on developing structural efficacy for locomotive applications;

after all, evolution for efficient methods of something so critical to survival has had millions of

years to mature. We humans, comparatively, have only had a measly few thousand to consider it

(with some of the earliest rumors of moving automata dating back to the 10th century BC [16]).

Engineers at their most ingenious seek the path of least resistance to most effectively complete the

task—what better way than to borrow from the countless years of work nature has already devoted.

Figure 1.1: A model of the spine (left) with ligaments in red [55] adjacent to a tensegrity model of a spine (right) with
cable members in red [30]; a visual example of how tensegrities can be biomimetic

Having existed for over 70 years now, one might wonder why tensegrities, for all their

advantages, aren’t found more often built in the real world. Conceivably, the most direct answer is

that they simply haven’t been studied long or seriously enough to pass the tests of time and

experience. Tensegrities have faded in and out of focus within the scientific community as more

technologies became available to make their physical realization more achievable (such as 3D

printing). Even with the advent of new technologies making tensegrity models more accurate and

2
easier to build, their geometric complexity as well as the requirement of pre-stress makes their

assembly complicated, especially if viewed in the context of mass/automated production.

Furthermore, nearly all conceivable tensegrity applications have well known legacy technologies

that have the advantage of centuries of human development and improvement. Truss structures

have been produced and used throughout human history, with some primitive trusses speculated to

date back to the bronze age [17]. Robotic manipulators (as they’ve come to be known) can be dated

back to the 1960’s [18]. Even in the (relatively) new field of space exploration, deployable

structures have been used since 1963, with a gravity-gradient stabilization boom on the 1963-22A

satellite [19]. Moving rovers, similarly, have existed since the Soviet Lunokhod rovers in the 1970’s

[20]. Even the half century of development of rovers (brief when compared to the millennia of

trusses) reflects a depth of understanding and comfort that is difficult to match.

The adoption of new technology is slow (at best), when capable heritage technology

alternatives exist. Proving new technology to be a worthwhile and superior alternative to an older

counterpart is a challenge overcome, in part, by thorough scientific analysis of the new technology,

and all its inherent properties. Many structural characteristics have been considered in great detail

for tensegrities, including structural stiffness, vibration properties, equilibrium, stability, form-

finding and others [21,22,23,24,25]. Each of these aspects are critically important to designing

structures for real-world applications, and undoubtedly were considered for the few tensegrities

seen throughout the world (such as the Kurilpa bridge in Australia).

While tensegrities are lightweight alternatives to stiff, static structures like bridges, their

inclusion of cable members also lends them to more dynamic applications. For example, cable

members tend to be much more elastic than their strut alternatives, thus making tensegrities capable

of flexing to absorb impact without breaking. This has been considered in the literature for

locomotive applications when structures might be expected to move on uneven terrain, or for

structures being dropped from a high altitude [26, 27]. Similarly, the potential flexibility of the

3
overall structure has piqued interest in soft robotics applications [28, 29]. The same members which

provide flexibility can also be used to enact shape change via actuation of relatively few members,

fully taking advantage of tensegrity’s biomimetic nature [30,31]. Applications designed to use of

tensegrities for their unique cable properties in addition to the more typical characteristics

associated with stiff, truss-like structures are most advantageous. Tensegrities are extremely

promising for applications requiring morphing, flexible structures. To date, the most

technologically advanced tensegrity structure physically built is the Super Ball Bot [26]; designed

as a planetary lander which could be dropped from significant heights and then self-rolled for

locomotion to explore while protecting an inner payload.

Existing tensegrities for applications that benefit from dramatic shape change, like the

Super Ball Bot, are few and far between. The field of deployable structures, however, has seen a

burgeoning interest in tensegrities. Deployment itself is a type of shape change, though it is

simplified in that two primary end states are considered. Most often these structures are designed

around the stowed and deployed shapes, where the transition between the two requires shape

change, but is designed to be as simple as possible. “Simple” here refers to the ability to transition

with as little actuation/directions of actuated motion as possible. For example, Figure 1.2 shows a

deployable scissor structure with a single mechanical degree of freedom. For these scissor

structures, shape change is enacted by pulling one end of the structure (assuming the other end is

fixed) along a single axis. The middle image in Figure 1.2 is a transition shape, showing how the

structure moves through different shapes while transitioning from stowed to deployed, where shape

refers to a particular set of member orientations. Deployable structures can also be designed with

coupled motions, such that actuation in a single direction enacts shape change in multiple

directions. Figure 1.3 shows an example of deployment of a miura-ori origami structure. This

structure, like the scissor structure in Figure 1.2, also requires only a pulling motion in a single

direction for deployment, but enacts shape change in two directions simultaneously due to

4
mechanical coupling. These two examples visually display how deployment inherently requires

shape change, while using simple types of actuation to enact the deployment process.

Figure 1.2: Deployment stages of a scissor structure, adapted from [56]. Structure is shown fully stowed (left), in
transition (middle) and fully deployed (right). This is an example of 1D deployment, where the structure can be
deployed by moving the red node along the dashed axis.

Figure 1.3: Deployment of miura-ori origami sheet, adapted from [57]. The miura-ori fold couples motion in two
directions, such that pulling the red node in the direction of the red arrow unfolds the sheet in two directions (the black
node and y-motion on the edge along the x-axis are held fixed). Deployment in this figure is shown from left to right.

Deployable structures are often used in space applications, in which launch vehicles impose

strict mass and volume constraints. Tensegrities are particularly appealing for space applications

as deployable structures, since they may be designed to offer a superior stiffness-to-mass ratio as

compared to truss alternatives. Tensegrities described in the literature are most often deployed by

changing cable lengths, while holding strut members fixed in length [4, 11, 12]. In this way, strut

5
orientation is dictated by the controlled change of the cable lengths. Some studies also allow strut

members to change length, though this is less common [13,14].

Cable actuation can offer controlled, perhaps automated deployment with a relatively

simple motorized spool. Deployable structures are typically designed with only two main shapes

in mind. However, the same mechanisms used to actuate cables for deploying a tensegrity could

readily be re-used to enact more dramatic shape change, even beyond the shapes associated with

the deployment process. While some research exists exploring the extent of tensegrity structures’

ability to dramatically change shape via cable actuation [15], tensegrities are infrequently designed

explicitly for shape change applications beyond deployment .

With tensegrity structures’ distinguishing characteristic being their flexibility (both

literally and in terms of their adaptability to be designed for particular applications), shape changing

applications have the most room for further analysis. Shape change analysis is built fundamentally

upon the form-finding problem: determining a structural shape (form) satisfying equilibrium and

structural requirements. While form-finding problems can sometimes be modified to understand all

achievable shapes, there is very little literature suggesting such studies have been done—in part

due to the complexity required to find and describe a particular shape (let alone a range of shapes).

Existing research on the form-finding problem has shown beyond a doubt that a vast variety of

shapes are achievable with tensegrities. However, the shapes a single tensegrity can achieve, and

design of tensegrity configuration specifically with shape changing applications in mind, is as yet

unexplored. Shape change—in reference to adjusting the lengths of various members—goes hand

in hand with deployable structures, robotics, and locomotive applications. With these types of

applications likely being some of the greatest opportunities for tensegrities, understanding the

shapes a tensegrity can achieve to implement these desired actions, and how well it can achieve

these goals throughout the variety of shapes it takes on, would be of great import to the further

development and realization of tensegrities in real-world applications.

6
 In this chapter, a literature review of existing form-finding methods is conducted. This

review shows the variety of approaches that exist to find equilibrated forms of tensegrity structures.

While there are many approaches to the form-finding problem, these approaches are inherently

oriented around finding a single equilibrated shape. A development of the understanding of

tensegrity configuration and its ability to achieve shape change can greatly enhance not only an

understanding of tensegrity structures in general, but also the ability to incorporate tensegrities in

real-world applications.

1.1 Tensegrity Background

Tensegrity structures are truss-like structures comprised of struts and cables connected at

nodes. Classical definitions of tensegrity structures specify that the members are uniaxial, with

struts carrying only compression and cables carrying only tension. These structures require pre-

stress to be in equilibrium. This requirement is particularly obvious for the most traditional type of

tensegrity, a so-called class-1 tensegrity. Class-n of a tensegrity denotes n struts at most meeting at

any single node. Figure 1.4 shows a class-1 tensegrity tower, where one can readily imagine how

a lack of pre-stress in the structure would result in slack cables, and thus collapse of the tower (at

least in the absence of gravity). While a class-1 tensegrity is the most traditional tensegrity, recent

research allows higher classes to fall under the tensegrity definition, as these structures still offer

many benefits of the inclusion of pre-stress and cables, while the increased strut-to-strut

connections can offer greater stiffness.

7
 Figure 1.4: Kenneth Snelson’s “needle tower” sculpture [58]; an example of a class-1 tensegrity structure.

Tensegrity structures can vary widely in appearance. Figure 1.5 demonstrates three

drastically different types of tensegrity structure, each containing six struts and twelve nodes. The

types of tensegrities shown in Figure 1.5 fall into commonly found subcategories of tensegrity type:

spherical and cylindrical. As their names imply, spherical tensegrities have all nodes lying on the

surface of a sphere (Figure 1.5 (a)), while cylindrical tensegrities have all nodes lying on the surface

of a cylinder (Figure 1.5 (b) and (c)). The cylindrical subcategory can be further characterized by

the n-plex of the structure, where n refers to n nodes on a face of the cylinder. Figure 1.5 (b) shows

a 6-plex cylindrical tensegrity, while Figure 1.5 (c) is a tensegrity made by combining two 3-plex

tensegrities into a tower. Figure 1.5 clearly demonstrates how different the forms of tensegrity may

appear, even with common characteristics such as number of struts or nodes. Spherical and

cylindrical tensegrities are common types of tensegrities, but tensegrities can readily take on forms

which fall in neither of these categories.

8
 In this research, tensegrities will be primarily described by their configuration and their

shape. Configuration refers to a particular topology of the tensegrity, meaning the connectivity and

the member types. The connectivity, described by a connectivity matrix, specifies how nodes are

connected to one another by members in the structure. Member type refers to designation of

members in a connectivity as either a strut or a cable. A particular configuration of a tensegrity may

take on a variety of shapes by changing the location of the nodes, without changing the

configuration itself. A particular shape, therefore, refers to a particular set of nodal locations for a

given configuration. The design variables in this research at generally tied to the shape of the

structure, where the particulars of the problem (e.g. any length constraints) are used to define a

minimal set of design variables that can be used to find nodal coordinates (or else are directly the

nodal coordinates themselves) and thus define the shape of the structure. The class of a tensegrity

structure may be determined from the information provided within a specified configuration.

These definitions are important for an understanding of the general type of tensegrity

structure being described. The configuration and shape also directly determine the equilibrated

states of the structure. For a given configuration, not every shape that is geometrically possible can

be equilibrated by a set of positive cable tensions and compressive strut loads. In any practical

application, a lack of equilibrium would result in perhaps catastrophic failure due to structural

instability or collapse. Thus, it is critically important to understand which shapes, for a given

configuration, are in equilibrium in order to properly design a shape-changing structure.

9
 Figure 1.5: Three different tensegrity configurations, each with 6 struts and 12 nodes. Left (a) is an icosahedron; a
spherical tensegrity. Center (b) is a 6-plex cylindrical tensegrity. Right (c) is a class-1 cylindrical tensegrity, formed by
stacking two 3-plex cylindrical tensegrities.

1.2 Equilibrium of Tensegrity Structures

Structural analysis of tensegrities is a well-developed field. Methods of analysis developed

for classical truss structures are readily applied to tensegrities, with some modification to

accommodate the inclusion of cable members, which cannot carry compression. The modifications

which enable the use of classic analysis methods commonly include the following assumptions:

• Members may only carry axial loads

• Nodes are negligible in size and mass, and are represented as frictionless ball

joints

• Struts and cables carry compression and tension respectively and exclusively

• Materials used for all members obey Hooke’s Law

• All loads are applied (or can equivalently be represented) at nodes

• No dissipative forces act on the system

The effects of gravity and material mass are often assumed negligible in tensegrity analysis

as well, though the accuracy of this assumption depends heavily on the type of analysis being

10
conducted and the intended applications. With these assumptions, the method of joints is readily

applicable to tensegrity structures for analyzing equilibrium. The equations derived from balancing

forces at each node can be compactly collected into matrix form. In the nomenclature common to

tensegrity analysis, the matrix form of the equations use a connectivity matrix (C), and a vector of

force densities (q). The connectivity matrix is m x n, for m members and n nodes. It is populated

with values of 0, 1 or -1, with the columns containing 1 and -1 denoting the start and end nodes of

the member which that particular row represents. The equations of equilibrium are linearized by

first dividing the member forces by member lengths. The forces in each member then become a

force per-unit-length, i.e., a force density (q). The force density vector, q, is thus a vector containing

the force density of each member, constructed in the same order as the connectivity matrix. The

final resulting equations of equilibrium in the x-y-z directions (for no external loads) are then

written as:

𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒙)𝒒 = 𝟎
𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒚)𝒒 = 𝟎 (𝐸𝑞. 1.1)
𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒛)𝒒 = 𝟎

Where a superscript T denotes transpose, diag refers to a diagonal matrix with the elements

of the vector argument on the diagonal, and x y z are the nodal coordinate vectors. Equation 1 is

often equivalently represented as:

𝑨𝒒 = 𝟎
𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒙)
(𝐸𝑞. 1.2)
𝑨 = [𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒚)]
𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒛)

Where A is the equilibrium matrix. The equations of equilibrium are thus written in terms

of the element force densities. Alternatively, they may be rearranged to be written in terms of nodal

coordinates, as shown in Eq. 1.3.

𝑫[𝒙 𝒚 𝒛] = [𝟎 𝟎 𝟎]
(𝐸𝑞. 1.3)
𝑫 = 𝑪𝑇 𝑑𝑖𝑎𝑔(𝒒)𝑪

11
 The matrix D is referred to as the force density matrix. Equations 2 and 3 are equivalent,

and either may be used as is most suited to the method of analysis. Solving for an equilibrated

configuration of a tensegrity structure—be it in terms of internal member forces, nodal coordinates

or some combination thereof—is the basis of the form-finding problem. Equilibrium for tensegrity

structures requires not only that the forces be balanced, but also that cables only carry tension.

Classically, struts are also required to only carry compression (and this often results naturally from

requiring cables to be in tension) though, in reality, it is possible for struts to carry tension as well.

Understanding feasible tensegrity shapes aligns with form-finding analysis, and thus an

understanding of equilibrium and how it is integrated in the form-finding methods already existing

is essential for building up an understanding of tensegrities for morphing applications.

1.3 The Form-Finding Problem

The form-finding problem seeks a particular “form” (i.e., shape) such that the structure

maintains equilibrium. The problem statement and solution methods can vary dramatically based

on what information is assumed to be known, and the resulting solutions have a similar degree of

variety. In some cases, numerous solutions are possible given the known parameters, while in

others, the given parameters may result in no possible solutions. Tensegrity structures pose an

additional challenge over classical truss structures, since cable elements are restricted to only carry

tension. Thus, the tensegrity structure form-finding problem must not only find equilibrated

solutions, but also those which do not violate member loading types.

The form-finding problem is quintessential to understanding tensegrity structures, and has

been thoroughly explored in the literature. Tibert and Pellegrino [32] offered a review of form-

finding methods, and categorized them into two main categories: kinematic and statical methods.

The kinematic methods described work by either maximizing strut lengths or minimizing cable

12
lengths, with the other type of element fixed in length. These methods readily apply to determining

equilibrated regular, say cylindrical tensegrities, and generally struggle to deal with irregular

tensegrities, tensegrities with a large number of nodes, or tensegrities that are not known / well

defined. Statical methods are described to work with a known topology (connectivity and member

types) and to find a set of nodal coordinates and force densities that result in equilibrium. The

majority of form-finding problems throughout literature fall into this category, as it is more robust

for finding unknown, irregular, or generally complex tensegrity structures.

Since the review conducted by Tibert and Pellegrino [32], a large number and variety of

form-finding studies based on statical methods have been conducted. These can be partitioned into

3 main types of approaches: iterative, metaheuristic and analytical methods.

1.3.1 Iterative Form-Finding

There are numerous ways of using an iterative approach for the form-finding problem.

Many methods are built upon the force density method, a subcategory of the statical methods as

defined by Tibert and Pellegrino [32]. The force density method builds upon the linearization of

the equations of equilibrium, as described in the previous section. These methods typically require

the connectivity and specification of member types as inputs to work. An initial guess of either the

nodal coordinates or the force density vector is used to solve Eq. 1.2 or Eq. 1.3 respectively. In

most cases, a force density vector is used as the initial guess, since a +1 and -1 can be assigned for

tension and compression members respectively as a method of enforcing member types. With an

initial force density vector (q), the force density matrix (D in Eq. 1.3) can be formed, and then Eq.

1.3 may be solved for nodal coordinates. Since the force density vector is likely not an equilibrated

solution, the resulting coordinates will not be the final coordinates. These approximate coordinates

can be used to form the equilibrium matrix (A in Eq. 1.2), which can in turn be used via Eq. 1.2 to

13
solve for a new force density vector. The iterative methods work to update nodal coordinates and

force density vectors in this way, iterating through the two equations until both are satisfied. The

main distinction between force density methods comes from the approach used to solve Eq. 1.2 and

Eq. 1.3. Sources [33-36] all build upon the force density method directly, with varying approaches

to solving Eq. 1.2 and 1.3, or else different approaches to handle the initial guesses of nodes/force

densities.

Estrada, Bungartz and Mohrdieck [33] offer an iterative method which requires knowledge

of the connectivity and member types. Their method works by using an initial guess of the force

density vector (q), assigning members a +1 or -1 for tension and compression respectively, based

on whether the member is a cable or a strut. This guess is used to initialize the process of iterating

between Eq. 1.3 and Eq. 1.2 to find nodes and force densities satisfying both. Rank conditions on

the equilibrium and force density matrices are defined to help with the iterative process, as shown

in Eq. 1.4.

𝑟𝑎𝑛𝑘(𝑨) ≤ 𝑀 − 1
(𝐸𝑞. 1.4)
𝑟𝑎𝑛𝑘(𝑫) ≤ 𝑛 − 𝑑 − 1

Where M is the number of members, n is the number of nodes, and d is the dimension of

the structure (i.e., d=2 in 2D or d=3 in 3D). The first rank condition on A is derived from ensuring

there is at least one or more feasible state of self-stress. The second rank condition on D ensures

there are at least d relevant particular solutions. In Eq. 1.4,. Estrada Bungartz and Mohrdieck use a

Schur decomposition of the force density matrix as shown in Eq. 1.5, where U is a basis of nodal

coordinates and V is a diagonal matrix of zero eigenvalues.

𝑫 = 𝑼𝑽𝑼𝑇 (𝐸𝑞. 1.5)

Since, in most cases, the force density vector being used will not result in equilibrium, a minimal

length condition (i.e., selecting a basis for nodal coordinates from U which minimizes the lengths

14
of members) is used to select approximate nodal coordinates for the current iteration. A reduced

form of matrix U (denoted Ũ) is generated by eliminating nodal bases which require members to

have zero length (in this method, lengths are minimized but should not be zero). Then, LU

decomposition is applied to a matrix of the member lengths (generated by multiplying the

connectivity matrix by Ũ), and the pivots of the upper triangular matrix which are set to zero are

used to select the linearly independent nodal coordinates. The resulting coordinates approximately

satisfy Eq. 1.3. Using these coordinates, Eq. 1.2 is then checked via the singular value

decomposition of A (Eq. 1.6).

𝑨 = 𝑼𝑽𝑾𝑇 (𝐸𝑞. 1.6)

The signs of the last column of the resulting W should be checked against the force densities (such

that the signs are the same as those of the initial guess that designated member type). Additional

columns of W are included as needed until the least squares quantity in Eq. 1.6 is minimized while

also satisfying the same signs as the initial force density vector.
2
̃ − 𝒒0 ‖
‖[𝒘𝑗 ⋯ 𝒘𝐾 ]𝒒 (𝐸𝑞. 1.7)

In Eq. 1.7, wK is the last column of matrix W, and j is the index of the first column included from

W (starting at k-1 and subtracting additional indices from k as necessary). The methodologies

presented in this paper are efficient for finding a particular state of an equilibrated tensegrity

structure, however, it does not explore structures which may have multiple states of self-stress. The

method only requires initial information of connectivity and member type.

Tran and Lee (2010 [34]) offer an approach very similar to Estrada, Bungartz and

Mohrdieck [33], using eigenvalue and singular value decomposition of the force density and

equilibrium matrices respectively. The eigenvalues of D (or more specifically, the number of

eigenvalues less than or equal to zero) and what they imply are explored in this paper, using

different approaches in updating the next iteration of the force density vector for the different cases.

Numerous examples demonstrate the robustness of this methodology, and it is likely possible to

15
use this method to find irregular tensegrity shapes as well. As is the case with many iterative

approaches, controlling which solution the algorithm will converge to can prove challenging, and

restrict its ability to find new, unknown solutions.

Lu, Li and Shu [35] address this by building off the methods described by Estrada [33] and

Tran and Lee (2010 [34]) by specifying some of the nodal coordinates and solving for the rest. This

method requires knowledge of the structure’s connectivity, an initial force density vector, and some

known (specified) nodal coordinates. The force density vector is initialized, again, as +1 for cables

and -1 for struts. In this case, Eq. 1.2 and Eq. 1.3 are solved for force densities and nodal coordinates

using singular value decomposition and eigenvalue decomposition respectively. The eigenvalue

decomposition of the force density matrix is shown in Eq. 1.8. Any specified nodal coordinates are

incorporated with the eigenvalue decomposition, such that for n0 known nodes corresponding to

coordinates x0, the orthogonal matrix 𝜳 is partitioned into 𝜳𝟎 (n0 vectors of 𝜳) and 𝜳∗ (the

remaining vectors of 𝜳 corresponding to the remaining unknown nodes x*).

𝑫 = 𝜳𝜮𝜳𝑇
(𝐸𝑞. 1.8)
𝒙 = 𝜳∗ 𝜳−1
∗
𝟎 𝒙𝟎

The paper shows the ability of this method to find irregular tensegrity shapes successfully by

specifying some nodes. In many instances, iterative methods are hard to control in terms of the

exact solution to which they converge outside of severe restrictions within the process itself, or else

well-educated initial guesses. In this sense, it can be difficult to understand the variety of shapes a

tensegrity may achieve through iterative approaches. While this paper offers greater success in

finding irregular shapes, the requirement to provide some nodal locations makes the method less

useful for finding new, unknown shapes for a particular tensegrity configuration.

Tran and Lee (2015 [36]) also offer an approach in which singular value decomposition is

used on both the force density and equilibrium matrices. Tran and Lee (2015) identify the use of

SVD instead of eigenvalue decomposition for the force density matrix as beneficial since the most

16
appropriate singular values found in the decomposition are easily identified as they are non-

negative. The algorithm works to modify the force density vector as little as possible to encourage

the last d+1 singular values of D to become null, where d is the dimension of the structure.

Furthermore, this paper offers a method of exploring numerous shapes of a particular structure,

namely by considering affine transformations of the nodal coordinates. Affine transformations are

transformations which, when applied, preserve collinearity (e.g., points that are on a line prior to

transformation are still on a line after the transformation) and the ratios of distances between points.

Alternative nodal coordinates can be found that result in an asymmetric structure this way, though

the force density vector will remain symmetric. This implies that only the member lengths

corresponding to independent nodes may be controlled directly with this method.

Another iterative approach is based on dynamic relaxation, a subset of the kinematical

methods of form finding as described by Tibert and Pellegrino [32]. The dynamic relaxation method

is built upon the dynamic equations of equilibrium (Eq. 1.9).

𝑴𝒅̈ + 𝓓𝒅̇ + 𝑲𝒅 = 𝒇 (𝐸𝑞. 1.9)

In Eq. 1.9, M, 𝓓, K and f are the mass matrix, damping matrix, stiffness matrix and vector of

external forces respectively. These matrices are lumped matrices, with values only at the nodes.

The variables 𝒅̈, 𝒅̇ and 𝒅 are the acceleration, velocity and displacement vectors. The mass and

damping matrices are usually set as diagonal for simplicity, and the initial velocities and

displacements are also started at zero. The equation of equilibrium is used to compute out-of-

balance forces (caused by pre-stress) to determine acceleration of the structure. A kinetic damping

method can be used, tracing the undamped motion of the structure and setting the velocities to zero

when a local peak of kinetic energy is found. Zhang, Maurin and Motro [37] use an iterative

dynamic relaxation with kinetic damping approach. The accelerations, velocities, and

displacements are iteratively updated until the system reaches a steady state equilibrium. Different

approaches of handling the stiffness of the structure are explored for their ability to influence the

17
final shape of the results. This approach is advantageous for its inherent integration of stability

considerations, and its ability to handle more complex or irregular shapes. However, by nature of

iteration, it is still difficult to control the exact result to which the structure will converge. This is

addressed to some extent by careful handling of the stiffness matrix. However, guidance of the final

shape still requires some user input, implying an understanding of the desired end result. In this

sense, while a variety of shapes may be found with this method, it is likely not an effective approach

for searching for a wide variety of unknown shapes.

Zhang et al. [38] offer a stiffness-based approach. Similar to the dynamic relaxation

method, out-of-balance forces resulting from the pre-stress of the structure are used based on

iterative calculations of the nodal coordinates to determine whether the structure is in equilibrium.

This method, however, is built around analyzing the stiffness matrix directly, rather than looking

at the whole dynamic equation of equilibrium. Since negative eigenvalues of the stiffness matrix

result in an unstable structure, and zero eigenvalues result in divergence, constraints are selected

and implemented to eliminate any such eigenvalues and ensure the positive definiteness (after

excluding rigid body motions) of the stiffness matrix. This method requires user inputs for the

connectivity, rest lengths and axial stiffnesses of members, and an initial set of nodal coordinates.

The nodal coordinates do not need to be close to the final configuration. A structural stiffness matrix

is constructed from the initial nodal coordinates. Constraints are then applied based on their

stochastic selecting algorithm or restricted step algorithm, seeking constraints which make the

resulting stiffness matrix positive definite. Nodal displacements are then calculated based on the

stiffness matrix, using a line search algorithm to update each iteration’s final nodal coordinate

vector. The stiffness matrix is then updated based on these new nodal coordinates, and the process

iterates until a state of self-equilibrium is found, as determined by out-of-balance forces (resulting

from the internal stresses of the structure). This method is capable of finding multiple states of

stable configurations, and can handle irregular and highly complex (i.e., many nodes/members)

18
tensegrities. While this iterative approach is fairly robust, it is still difficult to control the final

resulting configuration to which this method converges. This is a challenge inherent to iterative

approaches.

On the whole, iterative approaches require little initial input and are often highly efficient

and quickly converging. In some cases, they are even well equipped to handle irregular or complex

tensegrities. In understanding the shape changing capabilities of a structure, the ability to quickly

realize a variety of shapes a particular topology can achieve is critical. It is possible that iterative

approaches could be used to do this, however, they would be difficult to control efficiently for

searching the space of achievable, unknown and new shapes for a particular structure. Controlling

the final shape upon which these approaches converge is challenging, and often requires

incorporating some knowledge of what that final shape is to be in the initialization parameters.

1.3.2 Metaheuristic Form-Finding

Metaheuristic algorithms are often implemented for extremely complex problems. These

methods often require very little to be known initially, and can offer a far faster approach than their

iterative counterparts for problems with many design variables. When seeking previously unknown

shapes of tensegrity structures, metaheuristic approaches offer a promising approach to solving the

form-finding problem. These approaches typically start with a randomly generated initial

population. Each set of design variables (typically nodal coordinates in the form-finding problem)

in that population is a possible solution randomly selected between limits set by the user. Each set

in the population is then evaluated for how well it fits the intended outcome. The design variables

in every set of the population are then updated. The method of updating to a new set of variables is

typically the distinguishing factor between metaheuristic approaches. In all cases, the algorithm

terminates when a solution set is found satisfying some specified criteria.

19
 Genetic algorithms are among the more frequently used metaheuristic approaches for form-

finding problems. While the exact methods in genetic algorithms may vary, in general they work

in the following way:

1. Initialize a population. Typically, each “member” of the population is a set of design

variables, often randomly selected within limits set by the user.

2. Evaluate the fitness of each member of the initial population. The members of the

population are sorted by the goodness of their fitness, and the best members are

preserved, while the worst are discarded.

3. A new population is generated. This new population includes the best members from

the previous generation, as well as cross-overs and mutations of those preserved

members. Cross-overs are members are formed by selecting design variables directly

from preserved members. Mutations are generated by randomly perturbing other

members of the population.

4. The new generation, comprised of preserved, cross-over and mutated members, is then

evaluated for fitness (assigned a numerical value based on a fitness function). The

process repeats from step 2 until a member is found with a sufficiently “good” fitness.

Within the genetic algorithm category, different research approaches use different design

variables, objective functions, constraints and filters on the population. Xu and Luo [39] offer two

form-finding approaches using a genetic algorithm. In their first approach, the aim is to find as

many 3- dimensional tensegrities as possible for a given topology and set of members. This

approach requires the connectivity, number of members/nodes, elastic moduli of members, and rest

lengths of all members as inputs, with only the initial nodal locations designated as variables. The

second approach is the same as the first, except that the elasticities of the cables are included as

20
design variables as well. In this case, the cable rest lengths are assumed to be zero, and the end goal

is still to find as many 3-dimensional tensegrities as possible. In both cases, their approach uses

dynamic relaxation to determine whether or not a structure is in equilibrium, looking at imbalances

caused by internal forces for given nodal coordinates. This method was successful in finding known

and unknown shapes of tensegrity structures, as well as finding numerous solutions (by running the

code in parallel) in equilibrium when they exist for a given topology. Similar to iterative methods,

however, this method struggles to control where the algorithm will converge. In this sense, it is

hard to control which exact shape will result from the code. Furthermore, the authors question the

efficacy of this method for larger numbers of nodes, quoted as an “inherent shortage of the dynamic

relaxation algorithm.”

Koohestani [40] offers another approach using a genetic algorithm. This method seeks to

minimize the d+1 lowest eigenvalues (for d dimensions of the structure) of the force density matrix

(D in Eq. 1.3). The fitness function is defined as the sum of the absolute values of the d+1 lowest

eigenvalues of the force density matrix, divided by the sum of the absolute values of the force

densities of the members. The force densities are described for their role in the fitness function as

having no effect on the global minimum of the problem, but in improving the behavior of the

problem such that it is more likely to converge on upper or lower bounds of the force densities. The

nodal coordinates and force densities are the design variables in this case, and the problem can be

simplified and expedited by enforcing symmetry within the calculations. Koohestani notes that this

method requires further study for viability with non-regular tensegrities.

Lee and Lee [41] offer yet another approach using a genetic algorithm. Based on

Koohestani’s [40] method, the first d+1 smallest eigenvalues divided by the absolute value of the

sum of the force densities are included in the fitness function. An additional multiplier is also

included by Lee and Lee: a parameter denoting the standard deviation of the force densities. This

additional multiplier is included to encourage the force densities of cables to be more uniform. This

21
method requires only defining a connectivity, with the algorithm itself determining which members

are best suited as struts and cables without violating the class (i.e., number of struts allowed to

touch at a node) of the tensegrity structure. Nodal coordinates are the design variables in this case,

with the algorithm set to output any combination of struts locations and nodal coordinates that can

result in equilibrium. While this method is not designed to specifically require regularity, the fitness

function designation certainly favors it, as uniform force densities are more likely to result in

uniform member lengths, and this is reflected in the selected examples shown in the paper. This

method could likely be modified (either by exclusion of the constraints mentioned in the paper, or

else by modification of the fitness function) to encourage searching for irregular tensegrities, but it

is unclear what effects this would have on efficacy of the algorithm.

As a final example, Gan et al. [42] offer a genetic algorithm approach specifically designed

to handle complex and irregular tensegrities. The only input required is the number of nodes

(required to be an even number), from which the connectivity, number of struts (determined to be

half the number of nodes to ensure no two struts must connect at a single node) and a prototype

force density (+1 and -1 for cables and struts respectively) are generated. The initial connectivity

and force densities are randomly generated for the initial population (without violating the

tensegrity/class constraints). The fitness function sums four values, the first of which is the rank

deficiency of the singular value diagonal matrix from the singular value decomposition of the

equilibrium matrix (matrix V in Eq. 1.6). The second factor is a difference between a predefined

minimum number of members and the current iteration’s number of members at any given node.

The third factor is a difference between the allowed number of struts at a node and the current

iteration’s number of struts at a node, and the final factor is a parameter discouraging intersection

of members with one another. This method proves efficient at finding irregular tensegrities with

very little user input. The authors note that increasing the number of nodes increases the required

number of generations for convergence (e.g., run time). The solutions to which this algorithm

22
converges are dependent upon the initial configurations generated—typically done randomly—and

can thus be difficult to direct if certain types of shapes are being sought.

Other metaheuristic analyses can be used with similar implementation. Another found

throughout the literature for tensegrity structures is Particle Swarm Optimization (PSO). Particle

Swarm Optimization is similar to genetic algorithm approaches, in that the initial population used

is a collection of sets of the design variables, typically generated randomly within user-set

boundaries. The general procedure of PSO is outlined below:

1. Initialize the population; typically, each “particle” of the population is a set of the

design variables randomly selected within the user set boundaries

2. Evaluate the fitness of each particle, based on a prescribed fitness function.

3. Each particle is assigned a velocity which will direct it to a new “location” (i.e., the

design variables for the next iteration). This velocity is determined by balancing the

best fitness level found in that particular particle’s history, as well as the best fit found

of all the particles in the population throughout the entire history of the algorithm run

thus far.

4. Terminate when a tolerance criterion is met. Typically, this is an acceptably small

deviation between the best fitnesses of new iterations.

Like the genetic algorithm approach, the differences in methods comes primarily from the

formation of the optimization problem itself. The choice of fitness functions and design variables

greatly influence the end results. Particle Swarm Optimization is less common than genetic

algorithms in tensegrity optimization, and is more commonly found in applications focused on

structural properties of tensegrities [43,44] than form-finding specifically. Roffman and Lesieutre

[45] used a PSO for form-finding with cylindrical tensegrities assembled as an array to cover a 2D

23
area. The objective function was chosen to align one surface of the structure with a prescribed

surface (a paraboloid in the examples shown). The algorithm was found to effectively converge on

solutions for the tensegrity structures explored. Further work could be done to improve the efficacy

of the algorithm as described in the paper, particularly if it were to be applied for more complex

objective shapes.

In another example of PSO used for form-finding, Chen et al. [46] use a hybrid approach

to efficiently find equilibrated configurations. The approach is built upon the fundamentals of

force-density methods, but integrates use of PSO to calculate member force densities. The objective

function used to find the force densities is a weighted sum of three factors. This first is the sum of

the absolute values of the pre-stress (force density) of a design member and the average pre-stress

of all members—effectively favoring uniform pre-stressing. The second factor encourages the code

to seek solutions resulting in cables in tension and struts in compression. The final factor uses rank

deficiencies previously discussed to help ensure the final solution is in equilibrium. This method

was developed to work for symmetric tensegrities specifically, and was effective in finding novel,

complex tensegrity shapes.

Another metaheuristic approach is the Monte Carlo method. Monte Carlo methods include

randomization in their initialization, but vary a bit from the genetic algorithm or PSO approaches.

The general outline of Monte Carlo methods is as follows:

1. Define limits on the design variables, and randomly initialize the sets of solutions,

typically determined based on some sort of probability distribution

2. Analyze the fitness of the solutions

3. Accept or reject each solution based on prescribed criteria

4. Perturb accepted results randomly, and repeat until termination criteria is met

24
 Li et al. [47] use Monte Carlo methods to find novel, large scale and irregular tensegrities.

Their method seeks an equilibrated configuration with nodal coordinates as design variables. Inputs

include the number of struts, cables, elements and nodes, the connectivity of the structure, resting

lengths of elements and the initial nodal coordinates (these can be randomly generated).

Equilibrium (and thus the termination criteria) is met when the resultant forces from internal force

densities are net zero at all nodes. The examples in this paper demonstrate this method’s efficacy

at handling structures with very large numbers of members/nodes, as well as handling irregular

tensegrities.

Metaheuristic analysis methods are often used to handle complex problems with unknown

solutions. The results of these methods can vary somewhat dramatically based on the formulation

of the problem, and can often be designed to direct the algorithm towards desired solutions. That

said, by the nature of randomization, metaheuristics can struggle in cases seeking numerous

solutions, as they may converge on local minima, or else offer different solutions with every run.

In some cases, running these algorithms numerous times and comparing the results from many runs

is sufficient for understanding the problem. In seeking the variety of shapes a particular tensegrity

could achieve, metaheuristics may be useful in finding some examples, but would likely lose their

efficacy when seeking all achievable shapes.

1.3.3 Analytical Form-Finding

Analytical solutions are the preferred approach whenever possible, but are often the most

difficult to achieve. These solutions are exact, and can often lead to significant insight about the

nature of the problem as a whole. In some cases, analytical solutions are only possible with

significant simplification or assumptions. Tensegrity structures are fairly well studied, and have

25
analytical results for understanding many aspects of their structural properties. Analytical solutions

for form-finding specifically are less common, but do exist.

Analytical approaches to form-finding of tensegrities are built around finding equilibrated

solutions. Williamson, Skelton and Han [48], as well as Masic, Skelton and Gill [49] offer examples

of linear algebra treatment of the equilibrium equations (e.g., Eq. 1.2 and 1.3) to solve for tensegrity

shapes (and force densities). The methods developed in both of these papers benefit from numerical

tools when the structures become asymmetric, or else have many members. That said, the

fundamentals established in these papers offer linear algebraic treatment of the equilibrium

conditions of tensegrities. Linear algebra approaches for handling of problems are often desirable

for problems that must be solved using a computer (be that completely numerically, or using a

symbolic solver), as linear algebra manipulations are often computationally efficient. Furthermore,

Masic, Skelton and Gill [49] show that tensegrity equilibrium is preserved under affine

transformations. Affine transformations can also be represented through linear algebraic

manipulations, and offer a way of using existing, known tensegrity equilibria to find new tensegrity

shapes. Affine transformations are useful for taking known equilibrated tensegrity shapes, and

extending that single solution to encompass a larger set of shapes, found by translating, rotating

and shearing the original shape. Such transformations enable discovery of shapes that may have

been previously unknown, but do not necessarily encompass all shapes a particular tensegrity may

be able to achieve.

More recently, alternative analytical approaches have been explored. Zhang et al. [50] offer

an approach not unlike the force-density numerical methods. Analyzing the force-density matrix

explicitly, Zhang et al. found that forming a symbolic force density matrix could be used to derive

allowable relationships between force densities of the members in a tensegrity. Fundamentally, this

method is built upon the inequality in Eq. 1.4 on the force density matrix (D). It can equivalently

be rewritten in terms of the null space of D for d dimensions as:

26
 𝑛𝑢𝑙𝑙(𝑫) ≥ 𝑑 + 1 (𝐸𝑞. 1.10)

The eigenvalues of the force density matrix can be calculated from the characteristic equation (Eq.

1.11, where n is the number of nodes, λ are the eigenvalues and I is an identity matrix), the first

d+1 coefficients (e.g., P0, P1 and P2 for a 2D tensegrity) of which should equal zero in order to

satisfy the Eq. 10 inequality.

𝑑ⅇ𝑡(𝑫 − 𝜆𝑰) = 𝑃0 − 𝑃1 𝜆 + 𝑃2 𝜆2 − 𝑃3 𝜆3 + ⋯ + (−1)ⅈ 𝑃ⅈ 𝜆ⅈ + ⋯ + (−1)𝑛 𝑃𝑛 𝜆𝑛 = 0 (𝐸𝑞. 1.11)

The coefficients (Pi) will result in equations in terms of the force densities of the members of the

tensegrity. By properties of the force density matrix, the first coefficient, P0, is always zero. This

leaves d coefficients set equal to zero. Grouping of elements is used throughout the examples in

the paper, both to simplify the equations as well as to encourage finding more regular results that

can be checked against known, equilibrated tensegrities. While these equations may be solved by

hand for more simple structures, these equations can quickly become cumbersome with more

members. Even the simplest 3D example shown in this paper, a triplex, simplified to have only 3

groups of force densities, resulted in complex equations for the coefficients that require the use of

commercial software to address. In the cases shown, software was used for symbolic operations,

so the solutions are still exact solutions rather than numerical approximations. This analytical

approach is promising in its ability to encompass the whole range of what a tensegrity structure can

achieve in shape, in that the results offer ratios of force densities in the members (corresponding to

different nodal locations for the same connectivity). Zhang et al. do not address, however, the

efficacy of this approach for more allowable force-density groupings of the members (e.g., more

groups of force densities each containing fewer members, or else none at all). This approach offers

d equations to be solved, and all examples studied had d or fewer groupings, meaning d (or less)

unknowns for d equations. The methods described are no less valid for larger numbers of variables

(force densities), but the ability to solve the resulting equations certainly elicits further

investigation, if not an entirely numerical approach.

27
 Koohestani and Guest [51] offer a different analytical approach, taking advantage of the

fact that both the geometric and the equilibrium constraints inherent to a tensegrity can be written

in terms of the nodal coordinates. In order for elements to meet at the same nodes as specified by

connectivity, relationships between how vector representations of the elements can be explicitly

written. These relationships are geometrical limitations, written such that certain vector

representations of elements should add and subtract to start and end at the same node. These

geometrical constraints can be written in matrix form as a cycle-member incidence matrix. If the

cycle-member incidence matrix is denoted B, then the relationship between the cycle-member

incidence matrix and the nodal coordinates is as follows:

𝑩[𝒙 𝒚 𝒛] = [𝟎 𝟎 𝟎] (𝐸𝑞. 1.12)

How the nodal coordinates are directly related to the geometric constraints as well as the force

densities can be seen by comparing Eq. 1.3 and Eq. 1.12. The two can be combined to result in a

single (set) of equations, shown in Eq. 1.13.

𝑩
𝑯[𝒙 𝒚 𝒛] = [ ] [𝒙 𝒚 𝒛] = [𝟎 𝟎 𝟎] (𝐸𝑞. 1.13)
𝑫

Koohestani and Guest suggest solving Eq. 1.13 analytically via gaussian elimination and careful

pivoting to form H into an upper triangular matrix. The paper offers examples which were solved

analytically, as well as some in which the method is modified to be solved numerically for more

complex cases. The examples clearly demonstrate the method’s efficacy for handling tensegrities

with a very large number of members/nodes, as well as irregular shapes. For more complicated

tensegrities, the numerical approach becomes necessary. The authors point out the inability to

control the shapes to which the numerical methods (iterative in nature) converge, and mention that

further investigation is worthwhile to improve that.

Analytical solutions are the most desirable approach (if available/possible) to

understanding the range of shapes a structure could achieve. Ideally, the equilibrium conditions of

28
a tensegrity might result in a relationship between member lengths (or, by extension, force

densities) that could describe the whole range of shapes possible. It is likely that inclusion of the

geometric compatibility relations (represented by B from Koohestani and Guest [51]) can help

guide a more comprehensive understanding of the morphing capabilities of tensegrity structures,

but the methods discussed require further investigation for more complex or asymmetric

tensegrities.

1.4 Existing Tensegrities for Shape Changing Applications

Tensegrity structures designed specifically for morphing applications are largely

unexplored. Deployable structures have been a common application of tensegrities [4-8,11-13],

however these types of applications represent a subset of more general shape change. This research

aims to explore the effects of tensegrity design on the structure’s ability to change shape

dramatically. While deployment would be a particular application, the methods used to explore

shape change for deployment are not comprehensive enough to extend to more dramatic morphing.

Few examples in literature can be found in which tensegrities are designed specifically for

shape changing applications where the topology itself is considered as part of the design problem.

Deployable structures aside, the Super Ball Bot [26] and the Laika tensegrity spine quadruped [31]

are both examples where the tensegrity is intended for locomotion, and thus must inherently change

its shape. In both cases, however, there is no discussion of how the tensegrity topology is chosen,

and likely the authors decided the general tensegrity design based on intuition with simplicity of

control and (at least in the case of the tensegrity spine) with biomimetics in mind.

While a comprehensive approach to designing tensegrities for morphing applications has

yet to be developed, there is some existing literature exploring a few aspects of the problem. Oh et

al. [52] explore the analysis of moving tensegrity nodes to a target location through equilibrated

29
shapes only. The methods developed by Oh et al. seek optimized force elongation of cables to

prevent slack throughout the shape change process. By specifying target node locations and

requiring movement through equilibrium, this work addresses some aspects of a tensegrities ability

to achieve a “goal” without collapsing. Selection of target nodal locations is an important aspect of

the design problem for morphing structures, and this paper addresses methods of shape change with

that in mind. Similar consideration of nodal target positions will likely dictate minimal

requirements for the morphing structures designed with the methods developed for this research.

Lai, Plummer and Cleaver [53] consider multiple practical aspects of realizing morphing

tensegrity structures. When building a physical tensegrity, the ability to construct nodes such that

all members meet at a single point is a challenge—and a challenge exacerbated when members

require integrated actuation. Lai, Plummer and Cleaver develop a form-finding method that

considers nodes having a finite dimension to address this. They further offer methods for control

that are applicable to any number of actuators within a tensegrity structure. The methods developed

in this paper are not aimed at understanding a structures ability to morph, however, they are

important to consider when attempting a design with practical, real-world requirements.

As a final note, Chen, Liu and Skelton [54] considered tensegrity structures for morphing

airfoil applications. This paper is the closest example the in literature to date considering topology

as it relates to morphing capabilities. Specific airfoil shapes are specified as the intended goal for

the structure, and different levels of “complexity” of the tensegrity are analyzed for their ability to

achieve said shapes. Complexity in this paper refers to the number of repeated tensegrity units. The

connectivity of each unit is pre-defined. In this sense, while some consideration is given to topology

of the structure, the methods in this paper are simplified by scaling member lengths to

accommodate more or fewer tensegrity units within the airfoil. The focus of their paper is aimed

more at the control law of the morphing tensegrity wing, but nonetheless offers the best example

30
found thus far of considering topological effects of a tensegrity structure on its morphing

capabilities.

The idea of using tensegrity structures for morphing applications is not novel, and has been

explored in application-specific studies throughout the literature. However, finding and describing

the range of achievable shapes a particular tensegrity topology can achieve is largely unaddressed,

and can provide significant insight into the shape changing abilities of these structures. If the variety

of shapes a structure can achieve is well understood, then the optimal shapes for a given application

can be found, and the structure can thus be designed for shape changing applications.

1.5 Organization of Dissertation

Tensegrity structures are highly adaptable structures with immense potential for shape

changing applications. However, there is incomplete understanding of how a particular topology

influences a structure’s ability to change shape. This dissertation seeks to characterize the range of

achievable equilibrated shapes for a tensegrity structure, and demonstrate how that knowledge may

be used throughout the design process to develop tensegrity structures that are well suited for shape-

changing applications. This dissertation is organized as follows:

Chapter 2 describes an analytical approach to describing the range of achievable shapes of

a given tensegrity structure. As discussed in Chapter 1, numerous methods of form-finding exist in

the literature, but these existing methods lack the ability to characterize all shapes (forms) a

particular tensegrity can achieve. The null space of a geometric matrix (i.e., the cycle-basis matrix

developed by Koohestani and Guest [51]) is used to describe the range of geometrically viable

shapes of a tensegrity structure. The linear combination of the vectors in the null space describe

any geometrically viable shape of the structure. The coefficients of these vectors are constrained to

ensure the resulting shapes are equilibrated. Example cases for a 2D x-tensegrity and a 2D triplanar

31
tensegrity are developed in detail. For more complex tensegrity structures, a numerical approach is

required.

Chapter 3 describes a numerical approach used for more complicated tensegrity structures.

This method works by using one known equilibrated shape as a starting point, from which the

design variables are adjusted to search the overall design space for additional equilibrated

tensegrities. Chapter 3 discusses the features and efficacy of this numerical approach in detail.

Example cases for the numerical approach are shown for the 2D x-tensegrity and a 3D triplex.

Chapter 4 discusses additional design considerations, and how they may be used in tandem

with understanding a tensegrity structure’s achievable shapes. Geometry-based approaches to

incorporating member length constraints into a minimal set of design variables are developed.

Additionally, a path-planning approach is developed which can be used to identify a set of discrete

equilibrated shapes through which a tensegrity structure can move between target start and end

shapes. Methods used to select optimal fits to prescribed shapes are developed in tandem with the

path-planning methods. Finally, additional structural properties are considered which may

influence the design of the overall structure, such as the structure’s ability to maintain a prescribed

shape in the presence of external loads, and the structures dynamic properties via modal analysis.

Chapter 5 examines a particular design case, namely using a tensegrity structure as the

backbone of a parabolic reflector. This structure will be evaluated for its ability to achieve a surface

paraboloid shape of specified focal length. The configuration of candidate structure was developed

with modularity in mind, where each tensegrity unit—called a module—can change shape

independently. Modules could be stowed and launched in separate vehicles, then later assembled

in space with sufficient nodal connections between modules to form a larger, stable structure. The

module is evaluated using the shape change analysis developed in Chapter 3, and with the

additional design considerations developed in Chapter 4. Altogether, this chapter provides insight

32
to the capabilities of these modules to perform as parts of a deployable, backbone of a parabolic

reflector.

By nature of being comprised of struts and cables, tensegrities are highly adaptable

structures that lend themselves well to actively controlled shape change. This dissertation seeks to

address two primary objectives. The first objective is to offer a general approach to characterizing

a wide variety of shapes a particular tensegrity structure can achieve, without significant

assumptions of symmetry or member grouping. Characterizing the range of achievable shapes for

a given connectivity provides insight as to how that configuration relates to the structure’s ability

to change shape. The second objective is to address additional relevant analysis methods that would

be necessary when designing tensegrity structures specifically for shape changing applications. The

design analysis in this dissertation shows how tensegrity structures can be efficiently designed for

specific shape changing applications, and demonstrates their potential superior utility for adaptable

structures.

33
 Chapter 2

Analytical Characterization of a Tensegrity’s Achievable Shapes

The form-finding problem is a well-researched subject for tensegrity structures. A

tensegrity’s form can be characterized by the set of its nodal coordinates and force densities. The

equations of equilibrium, as discussed in Chapter 1, can be written in a way that emphasizes either

these nodal coordinates or force densities, but a complete specification of a tensegrity form requires

a compatible combination of the two. This results in a large set of variables for a smaller set of

equations which, particularly when combined with the added axial loading constraints dictated by

member type, evince the inherent complexity of the form-finding problem.

While many approaches to the form-finding problem have been developed over the years,

most are intrinsically oriented around finding a single form which satisfies these equations. While

finding a single satisfactory form is sufficient for many applications (particularly static

applications), applications which require the tensegrity to change shape would benefit greatly from

a deeper understanding of the range of achievable shapes particular to a tensegrity’s configuration.

A description of the range of achievable shapes as a function of a tensegrity’s configuration would

provide great insight not only in the ability to achieve particular shapes, but also provide a dataset

to rely on in unforeseen circumstances, where additional, unplanned shapes may be beneficial.

Particularly for space applications, this is likely beneficial as it provides potential resilience in

unpredictable environments. For a rover, there may be unforeseen terrain. For a robotic

manipulator, the widest range of achievable shapes allows for the most types of possible

maneuvers. For an area-spanning platform, shape adjustments could be used to compensate for

deviations of shape when high precision is necessary, accommodating for the effects of

disturbances such as hysteresis or thermal loading. Understanding the shapes that a particular

tensegrity is capable of achieving not only allows for a deeper understanding for which applications

34
a configuration is most equipped to handle, but the capability of a mobile structure to adapt to

unforeseen circumstances also highlights the robust characteristics that may make tensegrities more

appealing alternatives to existing structures.

Analytical solutions to the equations of equilibrium, even for—or perhaps especially for—

simple tensegrities, offer much insight regarding the range of achievable shapes. Most existing

form-finding methods in the literature are numerical, or else require significant assumptions

(typically some sort of symmetry) that restrict the types of solutions found. Koohestani and Guest

[51] offer a promising analytical approach which uses geometric relationships between members

(characterized by the cycle basis matrix) in tandem with the equations of equilibrium to seek

solutions. The work of [51] uses member grouping (e.g., selective symmetry that defines sets of

members) to aid in simplifying the equations for seeking analytical solutions, and resorts to an

iterative numerical approach for more complex structures. While the analytical solutions could

provide insight into a range of achievable shapes, the symmetry assumptions and iterative nature

of the numerical approach restrict the viability of this approach for finding a wider range of

achievable shapes. This research proposes a new method of analytically finding the range of

achievable shapes.

2.1 A 5-Step Approach to Analytically Finding the Range of Achievable Shapes

One of the many challenges associated with finding the range of equilibrated shapes a

tensegrity can achieve comes from describing the results themselves. In this research, the geometry

of the structure is used as a starting point for considering viable shapes. The cycle-basis described

by Koohestani and Guest [51] effectively describes the geometric relationships between members,

and is used in this research to define geometric coefficients. These geometric coefficients are later

used (with limits to enforce equilibrium) to describe the range of achievable shapes. This differs

35
from most form-finding approaches in the literature, which typically use nodal coordinates or force

densities (or some combination thereof) to define a tensegrity shape. In this case, geometric

coefficients are appealing as they naturally fall from the null space of the geometric matrix, and

can subsequently be limited directly to enforce equilibrium requirements. While it is possible to

convert geometric coefficients back into another set of variables (such as nodal coordinates) to

define the shape, this would involve additional steps.

In developing this analytical method, a few assumptions are made:

• The connectivity matrix is given

• The member types are given (i.e., which members are struts/cables)

• Member 1 is restricted to lie on the x-axis (this eliminates shapes which would be

equivalent, but rotated about the z-axis)

• If the tensegrity is 3-dimensional, member 2 is restricted to lie in the x-y plane

(this eliminates shapes which would be equivalent, but rotated about the x or y

axis).

• Member 1 is of length 1 (this scales the results, eliminating shapes that would have

the same relative angles and length ratios, but are multiplied by some constant

factor)

• Member 1 has a force density of 1 (this scales the force densities, eliminating

results that are otherwise equivalent, but multiplied by some constant factor)

These assumptions are in addition to the typical assumptions made when evaluating the

equilibrium of tensegrity structures discussed in Chapter 1, and shown again here:

• Members may only carry axial loads

• Nodes are negligible in size and mass, and are represented as frictionless ball joints

36
 • Struts and cables carry compression and tension respectively and exclusively

• Materials used for all members obey Hooke’s Law

• All loads are applied (or can equivalently be represented) at nodes

• No dissipative forces act on the system

With these assumptions, a new approach to describe a tensegrity’s range of achievable

shapes is developed. This approach can be broken down into 5 main steps: (1) form a geometric

matrix (i.e., the cycle basis in [51]); (2) form the equilibrium matrix; (3) create an augmented matrix

using the null spaces of the geometric and equilibrium matrices; (4) obtain the reduced row echelon

form (RREF) of the augmented matrices to form a set of equations; and, finally, (5) use the resulting

equations to impose restrictions on possible values of the geometric coefficients. The following

subsections describe each of these steps in greater detail.

2.1.1 Form a Geometric Matrix

For a particular set of nodes and connectivity, relative member locations are fully defined.

If a single node is moved and the connectivity is unchanged, then by geometric necessity members

connected to that node must change length and/or orientation to accommodate the node movement.

These relationships can be represented with vector addition, representing each member as a vector

pointing from the lower to the higher numbered node for the two nodes defining said member. (This

is also how positive and negative 1’s are typically assigned in the connectivity matrix.) Node

ordering does not matter, so long as it is consistent. A set of vectors that start and end at the same

node comprise a cycle. These cycles represent the geometric requirements of the structure’s

connectivity (the same cycles in the cycle basis of Koohestani and Guest [51]). For illustration,

Figure 2.1 shows a 2D x-tensegrity, and some of the cycles within the structure.

37
 The geometric matrix is formed as a minimum set of cycles needed to describe the

structure. This minimum set is linearly independent, and contains a set of cycles such that each

member and node is represented at least once. For example, in Figure 2.1, the x-tensegrity contains

all four cycles shown, (a)-(d). Cycle (d), however, is a linear combination of cycles (a) and (c).

Thus, for the x-tensegrity, the minimum set of cycles can be represented by only 3 cycles, (a)-(c).

Including additional cycles, such as (d), would be mathematically viable, but results in a larger

geometric matrix than necessary. The resulting cycles can be written as vector equations. For

example, Eq. 2.1 shows the vector equations for the x-tensegrity (cycles (a)-(c)). These equations

are written in matrix form in Eq. 2.2. These equations are more compactly shown in Equation 2.3,

⃑ corresponds to the vector representation of

where the matrix G is the geometric matrix, and 𝒗

members.

𝒗1,2 + 𝒗2,3 − 𝒗1,3 = 𝟎

𝒗1,2 + 𝒗2,4 − 𝒗1,4 = 𝟎 (𝐸𝑞. 2.1)
𝒗1,3 + 𝒗3,4 − 𝒗1,4 = 𝟎

⃑⃑⃑⃑⃑⃑⃑
𝐯𝟏,𝟐
⃑⃑⃑⃑⃑⃑⃑
𝐯𝟐,𝟑
0
1 1 0 0 −1 0 ⃑⃑⃑⃑⃑⃑⃑
𝐯𝟑,𝟒
[1 0 0 −1 0 1] = { 0} (𝐸𝑞. 2.2)
⃑⃑⃑⃑⃑⃑⃑
𝐯𝟏,𝟒 0
0 0 1 −1 1 0 0
⃑⃑⃑⃑⃑⃑⃑
𝐯𝟏,𝟑
{⃑⃑⃑⃑⃑⃑⃑
𝐯𝟐,𝟒 }

38
 Figure 2.1: The 2D x-tensegrity, comprised of vectors representing each member, and some cycles (a)-(d) made from
the member vectors.

⃑ =𝟎
𝑮𝒗 (𝐸𝑞. 2.3)
Each member vector v is comprised of the sum of the vector length components in each

⃑ ⅈ = 𝛥𝑥𝑖 𝒙
dimension (e.g., in 2D for some member i: 𝒗 ̂ + 𝛥𝑦𝑖 𝒚
̂ where 𝜟 is the member length

̂ and 𝒚
component, and 𝒙 ̂ are unit vectors). Thus, the vector Eq. 2.3 can be split into 2 scalar

equations for 2D structures, as shown in Eq. 2.4.

𝑮𝜟𝒙 = 𝟎 𝑎𝑛𝑑 𝑮𝜟𝒚 = 𝟎 (𝐸𝑞. 2.4)

39
Where 𝜟𝒙 and 𝜟𝒚 are vectors of length components in x and y respectively. The length component

vectors can be found in terms of the nodal coordinates (nx and ny in x and y respectively) and

connectivity matrix (C) as:

𝜟𝒙 = −𝑪𝒏𝒙 𝑎𝑛𝑑 𝜟𝒚 = −𝑪𝒏𝒚 (𝐸𝑞. 2.5)

In a conventional connectivity matrix, the column corresponding to the lower numbered node (start

node) is a +1, while the column corresponding to the higher numbered node (end node) is -1. When

using the connectivity matrix and nodal coordinates directly to find member vectors, this results in

vectors mathematically pointing from the higher node to the lower node, where the geometric

matrix is the reverse. A negative pre-multiplication is included in Eq. 2.5 to accommodate this.

2.1.2: Form an Equilibrium Matrix

One of the most commonly used forms of the equations of equilibrium is written in terms

of the force densities (Eq. 1.1), re-written here for convenience:

𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒏)𝒒 = 𝟎 (𝐸𝑞. 2.6)

with one such equation in each dimension, and n representing x or y (or z) for each dimension. From

Eq. 2.5, recall that -Cn represents the member length components 𝜟 from the geometric matrix.

Modifying the geometric matrix components with this negative sign and denoting them as 𝜟𝒈 =

−𝜟 = 𝑪𝒏, the length components from the geometric matrix can be directly substituted into Eq.

2.6. Furthermore, the vector q contains the force density of each member. Assuming the member

types (struts or cables) are known, a vector (m) may be constructed containing the required signs

of the members’ force densities (mi), where:

+1 𝑖𝑓 𝑚ⅇ𝑚𝑏ⅇ𝑟 𝑖 𝑖𝑠 𝑎 𝑐𝑎𝑏𝑙ⅇ
𝑚ⅈ = {
−1 𝑖𝑓 𝑚ⅇ𝑚𝑏ⅇ𝑟 𝑖 𝑖𝑠 𝑎 𝑠𝑡𝑟𝑢𝑡

With this vector m, and 𝜟𝒈 , Eq. 2.6 can be rearranged as follows:

40
 𝑪𝑇 𝑑𝑖𝑎𝑔(𝒎)𝑑𝑖𝑎𝑔(𝜟𝒈 )𝒒
̃ = 𝑬𝑑𝑖𝑎𝑔(𝜟𝒈 )𝒒
̃ = 𝑬 {𝑑𝑖𝑎𝑔(𝜟𝒈 )𝒒
̃} = 𝟎 (𝐸𝑞. 2.7)

From Eq. 2.7, the matrix E is the equilibrium matrix of interest. Note that by moving the signs of

the force densities into the CT matrix via multiplication with m, the elements of the modified force

̃, are required to be positive, so that the required signs of each members’ force
density vector, 𝒒

density will not change.

2.1.3 Form an Augmented Matrix Using the Null Spaces of G and E

From Eq. 2.2 and Eq. 2.3, the null space of G defines a basis for solutions describing 𝜟.

The size of the null space will be M x (M - rG), where M is the number of members in the tensegrity,

and rG is the rank of the geometric matrix, G. Denoting the number of vectors in the null space as

k = M - rG, and each vector in the null space as g, the geometrically viable combinations of member

length components 𝜟 (with 𝜟 defined by the conventions of G, as in Eq. 2.4) can be defined as:

𝜟 = ∑ 𝛼ⅈ 𝒈ⅈ (𝐸𝑞. 2.8)
ⅈ=1

In Eq. 2.8, the coefficient 𝛼 is designated the geometric coefficient, and is the principle variable

type that will be used to characterize the range of achievable shapes in this geometric approach.

Note that the geometric matrix G is comprised of 0’s, 1’s, and -1’s, and therefore the null space of

G can also be found as vectors comprised of 0’s, 1’s, and -1’s. Additionally, while the geometric

matrix G is the same in all dimensions, the member length components 𝜟 are independent in each

dimension. Thus, 𝜟 in all dimensions can be characterized by an equation of the form of Eq. 2.8,

but the geometric coefficients 𝛼 are independent in each dimension.

In this step, some assumptions regarding scaling and orienting the design space can be

imposed. First, member 1 is assumed to lie on the x-axis. This is imposed by setting 𝛥𝑦1 = 0, and

41
is used with Eq. 2.8 to eliminate one geometric coefficient. Next, member 1 is assumed to be of

length 1. Combined with the prior assumption that member 1 is on the x-axis, this is imposed as

𝛥𝑥1 = 1, and is used with Eq. 2.8 to eliminate a second geometric coefficient. If the problem is 2D,

these are the only assumptions imposed at this step. If the problem is 3D, then member 1 lying

strictly on the x-axis also imposes 𝛥𝑧1 = 0, which is used to eliminate an additional geometric

coefficient. In 3D, a final assumption is imposed, requiring member 2 to lie on the x-y plane. This

is imposed as 𝛥𝑧2 = 0, eliminating one more geometric coefficient. In total, for a geometric matrix

having k vectors, the number of geometric coefficients used to fully define the tensegrity structure

with these assumptions is:

𝑑𝑘 − 2(𝑑 − 1) (𝐸𝑞. 2.9)

Where d is the dimension of the problem (d=2 in 2D, and d=3 in 3D).

From Eq. 2.7, the null space of the equilibrium matrix E describes the member length

components multiplied by the magnitude of their respective force densities. (The products are the

force components.) The member length components in Eq. 2.7, 𝜟𝒈 are the negative of the member

length components used in the formation of the geometric matrix, 𝜟. With this, Eq. 2.8 can be

substituted into Eq. 2.7 to result in Eq. 2.10:

̃} = 𝟎
𝑬 {−𝑑𝑖𝑎𝑔 (∑ 𝛼ⅈ 𝒈ⅈ ) 𝒒 (𝐸𝑞. 2.10)
ⅈ=1

Note that the terms within the curly brackets are a single vector, in which each element corresponds

to a member, and each member is therefore a function of the geometric coefficients and force

density magnitudes. Since the term within the curly brackets of Eq. 2.10 is a single vector, the span

of the null space of E can be used to characterize this vector. Denoting the null space vectors of E

as a matrix e, an augmented matrix is formed in Eq. 2.11 to ensure that solutions for member length

components are also in equilibrium.

42
 𝑘

̃]
[𝒆 |−𝑑𝑖𝑎𝑔 (∑ 𝛼ⅈ 𝒈ⅈ ) 𝒒 (𝐸𝑞. 2.11)
ⅈ=1

The geometric coefficients are independent in each dimension, and thus this augmented matrix

should be formed for each dimension.

2.1.4: Obtain Equations Using RREF of the Augmented Matrices

To ensure shapes characterized by the span of the null space of the geometric matrix are

also in equilibrium, these length components from Eq. 2.8 are multiplied by the magnitudes of their

respective members’ force densities; the resulting product of which must be in the null space of the

equilibrium matrix. The augmented matrix formed in Eq. 2.11 can be used to enforce this

requirement. After performing row operations to convert the left-hand side of the augmented

matrix, e, into reduced row echelon form (RREF), rows of the left-hand side which are all zeros are

used to identify relevant equations from the right-hand side. In order for the vector on the right-

hand side of the augmented matrix in Eq. 2.11 to be in the null space of the matrix on the left-hand

side of Eq. 2.11 (as is required to meet equilibrium constraints), the elements of the right-hand side

in the same rows which are all zero on the left-hand side must be equal to zero as well. This results

in a set of equations which are functions of the geometric coefficients and force density magnitudes.

These equations must be evaluated independently in each of the dimensions, since the geometric

coefficients are independent in each dimension; however, the row operations needed to obtain the

equations will be the same in all dimensions.

The number of equations to evaluate is determined by the rank of the null space of e. If the

rank of the null space is denoted re, then the number of equations is defined by Eq. 2.12, with d

being the number of dimensions and M being the number of members in the tensegrity:

𝑑(𝑀 − 𝑟𝑒 ) (𝐸𝑞. 2.12)

43
2.1.5: Use Resulting Equations to Restrict Geometric Coefficients

The equations resulting from the RREF of the augmented matrix in Eq. 2.11 are functions

of the geometric coefficients and force density magnitudes. One force density variable can be

eliminated by requiring q1=1. For force components to be balanced, the ratio of force densities

must be balanced. These ratios of balanced force densities can be multiplied uniformly by any

scalar, called the prestress coefficient, and preserve equilibrium. In practice, this would be

equivalent to tensioning the system without changing member lengths. By setting q1=1, the pre-

stresses are scaled to the first member, and therefore do not result in any loss of generality of the

results, since they could be multiplied by any pre-stress coefficient. With this assumption, there are

k geometric coefficients and M-1 force densities, all of which are variables to be found from the

d(M-re) equations.

Since the member type vector, m, was incorporated in the formation of the equilibrium

matrix, all force density magnitudes in the resulting equations must maintain their positive signs.

A negative force density variable would equate to switching the type of loading (i.e., a strut in

tension or a cable in compression), and thus violate the member requirements of tensegrity

structures. The equations can be solved for the force density magnitudes, resulting in ratios that are

functions of the geometric coefficients. Careful consideration of the numerators and denominators

can discern what types of combinations of geometric coefficients result in positive force densities.

Any set of geometric coefficients which satisfy all the equations, without allowing a force density

component to become negative, corresponds to a viable, equilibrated tensegrity shape.

44
 2.2 The 2D X-tensegrity

Pictured in Figure 2.1, the 2D X-tensegrity is among the simplest of tensegrity structures. This

tensegrity comprises 4 nodes, 4 cables and 2 struts, where the member vectors connecting node 1

to node 3 and node 2 to node 4 in Figure 2.1 are the struts. The connectivity matrix of the X-

tensegrity is:

1 −1 0 0
0 1 −1 0
0 0 1 −1 (𝐸𝑞. 2.13)
𝑪=
1 0 0 −1
1 0 −1 0
[0 1 0 −1]

As described in Section 2.1.1, the minimum set of cycles for the X-tensegrity are cycles

(a), (b) and (c) in Figure 2.1, and the resulting geometric matrix is:

1 1 0 0 −1 0
𝑮 = [1 0 0 −1 0 1] (𝐸𝑞. 2.14)
0 0 1 −1 1 0

A member type vector is then defined as m = [1, 1, 1, 1, -1, -1], since the first 4 members

correspond to cables and the last 2 members correspond to struts. This is used with the connectivity

matrix to form the equilibrium matrix according to Eq. 2.7 as:

1 0 0 1 −1 0
−1 1 0 0 0 −1 (𝐸𝑞. 2.15)
𝑬=[ ]
0 −1 1 0 1 0
0 0 −1 −1 0 1

Next, the null spaces of the matrices are found for the augmented matrix. First, the null

space of the geometric matrix is found as:

1 0 −1
−1 1 1
1 −1 0 (𝐸𝑞. 2.16)
𝑛𝑢𝑙𝑙(𝑮) =
1 0 0
0 1 0
[0 0 1]

45
In the case of the X-tensegrity, the null space has k=3 vectors in each of the 2 dimensions, resulting

in a total of 6 geometric coefficients. This null space is used to represent the member length

components with the geometric coefficients as follows:

∆𝑥1 1 0 −1
∆𝑥2 −1 1 1
∆𝑥3 1 −1
= 𝛼1𝑥 + 𝛼2𝑥 + 𝛼3𝑥 0 𝑎𝑛𝑑
∆𝑥4 1 0 0
∆𝑥5 0 1 0
{∆𝑥6 } { 0 } { 0 } { 1}
∆𝑦1 1 0 −1
∆𝑦2 −1 1 1
∆𝑦3 1 −1
= 𝛼1𝑦 + 𝛼2𝑦 + 𝛼3𝑦 0 (𝐸𝑞. 2.17)
∆𝑦4 1 0 0
∆𝑦5 0 1 0
{∆𝑦6 } { 0 } { 0 } { 1}

From the assumption that member 1 is of length 1 and lies on the x-axis, Eq. 2.17 is used to find:

𝛼1𝑥 − 𝛼3𝑥 = 1 ⇒ 𝜶𝟏𝒙 − 𝟏 = 𝜶𝟑𝒙

𝛼1𝑦 − 𝛼3𝑦 = 0 ⇒ 𝜶𝟏𝒚 = 𝜶𝟑𝒚 (𝐸𝑞. 2.18)

The results of Eq. 2.18 can then be substituted back into Eq. 2.17, and the member length

components can be substituted into Eq. 2.11 with the null space of E to yield the following

augmented matrices:

0
−1 1 0 1 −1 1 0 𝑞2 (𝛼2𝑦 )
−1 1 1| 𝑞2 (1+𝛼2𝑥 ) −1 1 1|
𝑞 (𝛼 − 𝛼2𝑦 )
−1 0 1 𝑞3 (𝛼1𝑥 − 𝛼2𝑥 ) −1 0 1 3 1𝑦 (𝐸𝑞. 2.19)
𝑎𝑛𝑑 𝑞4 (𝛼1𝑦 )
1 0 0 𝑞4 (1𝛼1𝑥 ) 1 0 0
|
0 1 0 𝑞5 (𝛼2𝑥 ) 0 1 0| 𝑞5 (𝛼2𝑦 )
(0 0 1 𝑞6 (𝛼1𝑥 − 1) ) 0 0 1
𝑞6 (𝛼1𝑦 ) )
(

Next, the RREF of these matrices must be found. This is straightforward, since the left-hand side

of these augmented matrices is comprised of 1’s and 0’s. In the case of the x-tensegrity, the results

are as follows:

46
 1 0 0 𝑞4 (𝛼1𝑥 )
0 1 0| 𝑞5 (𝛼2𝑥 )
0 0 1 𝑞 6 (𝛼1𝑥 − 1)
𝑎𝑛𝑑
0 0 0 𝑞4 (𝛼1𝑥 ) − 𝑞5 (𝛼2𝑥 ) + (1)
0 0 0|𝑞4 (𝛼1𝑥 ) − 𝑞5 (𝛼2𝑥 ) − 𝑞6 (𝛼1𝑥 − 1) + 𝑞2 (𝛼2𝑥 − 1)
(0 0 0 𝑞4 (𝛼1𝑥 ) − 𝑞6 (𝛼1𝑥 − 1) + 𝑞3 (𝛼1𝑥 + 𝛼2𝑥 ) )

𝑞4 (𝛼1𝑦 )
1 0 0 𝑞5 (𝛼2𝑦 )
0 1 0|
0 0 1 𝑞6 (𝛼1𝑦 )
(Eq. 2.20)
0 0 0 𝑞4 (𝛼1𝑦 ) − 𝑞5 (𝛼2𝑦 )
0 0 0|𝑞 (𝛼 ) − 𝑞 (𝛼 ) − 𝑞 (𝛼 ) + 𝑞 (𝛼 )
4 1𝑦 5 2𝑦 6 1𝑦 2 2𝑦
0 0 0
( 𝑞4 (𝛼1𝑦 ) − 𝑞6 (𝛼1𝑦 ) + 𝑞3 (𝛼1𝑦 + 𝛼2𝑦 ) )

In order to ensure the augmented vector is in the span of null(E), and thus in equilibrium, the final

3 elements of the augmented vectors in both x and y must be equal to zero. This results in the

following 6 equations:

𝑞4 (𝛼1𝑥 ) − 𝑞5 (𝛼2𝑥 ) + (1) = 0

𝑞4 (𝛼1𝑥 ) − 𝑞5 (𝛼2𝑥 ) − 𝑞6 (𝛼1𝑥 − 1) + 𝑞2 (𝛼2𝑥 − 1) = 0
𝑞4 (𝛼1𝑥 ) − 𝑞6 (𝛼1𝑥 − 1) + 𝑞3 (𝛼1𝑥 + 𝛼2𝑥 ) = 0
𝑞4 (𝛼1𝑦 ) − 𝑞5 (𝛼2𝑦 ) = 0 (𝐸𝑞. 2.21)
𝑞4 (𝛼1𝑦 ) − 𝑞5 (𝛼2𝑦 ) − 𝑞6 (𝛼1𝑦 ) + 𝑞2 (𝛼2𝑦 ) = 0
𝑞4 (𝛼1𝑦 ) − 𝑞6 (𝛼1𝑦 ) + 𝑞3 (𝛼1𝑦 + 𝛼2𝑦 ) = 0

Recall that use of the member type vector m incorporates information about member loading

requirements into these equations such that all of the force densities are required to be positive.

These equations are then solved for each force density, with results as follows:
−𝛼1𝑦
𝑞2 =
𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦
𝛼2𝑦 𝛼1𝑦
𝑞3 =
(𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 )(𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 )
−𝛼2𝑦
𝑞4 = (𝐸𝑞. 2.22)
𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦
−𝛼1𝑦
𝑞5 =
𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦
−𝛼2𝑦
𝑞6 =
𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦

47
Since each force density is required to be positive, the terms of the numerators and denominators

to help determine limitations on the geometric coefficients. In this case, the 4 repeated terms are

𝛼1𝑦 and 𝛼2𝑦 for the numerators, and (𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 ) and (𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 +

(𝛼1𝑥 − 1)𝛼2𝑦 ) for the denominators. Careful observation reveals the following requirements:

• 𝛼1𝑦 and 𝛼2𝑦 must have the same sign

• For 𝛼1𝑦 , 𝛼2𝑦 < 0, (𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 ) > 0 𝑎𝑛𝑑 (𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 −

1)𝛼2𝑦 ) > 0

• For 𝛼1𝑦 , 𝛼2𝑦 > 0, (𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 ) < 0 𝑎𝑛𝑑 (𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 −

1)𝛼2𝑦 ) < 0

These requirements can used to derive bounds on acceptable values of geometric

coefficients. For validation of results and to provide a point of reference, a data set using a brute-

force approach was generated first. This set was generated by incrementally varying each geometric

coefficient between 0 and 2 and testing all combinations of the coefficient values against one

another for equilibrium. This range of geometric coefficient values was chosen relative to the fixed

length member 1, which has an x and y length component of 1 and 0 respectively. Since geometric

coefficients are directly related to member length components (via Eq. 2.8), this loosely translates

to +/- the fixed component length for the unfixed components. In this case, all 𝛼1𝑦 and 𝛼2𝑦 are

positive, thus the first requirement is automatically met and any value of these coefficients in the

tested range is acceptable. The remaining geometric coefficients, 𝛼1𝑥 and 𝛼2𝑥 , have limitations on

combinations which are in equilibrium (indicated by the sign requirements on the denominator

terms) and are plotted for reference in Figure 2.2.

48
 Figure 2.2: Geometric coefficients corresponding to equilibrated shapes of the 2D x-tensegrity, found using a brute-
force approach checking all α combinations between 0 and 2 against one another and checking each for equilibrium.

Since the geometric coefficients in y are both positive, the second requirement on the denominator

terms must be enforced in this case:

(𝛼1𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 ) < 0 𝑎𝑛𝑑 (𝛼2𝑦 − 𝛼2𝑥 𝛼1𝑦 + (𝛼1𝑥 − 1)𝛼2𝑦 ) < 0 (𝐸𝑞. 2.23)

Rearranging these expressions, results in:

𝛼2𝑦
( ) (𝛼1𝑥 − 1) + 1 < 𝛼2𝑥
𝛼1𝑦
(Eq. 2.24)
𝛼2𝑦
( ) (𝛼1𝑥 ) < 𝛼2𝑥
𝛼3𝑦

Since both expressions must hold for any particular set of geometric coefficients, whichever

expression is more restrictive is used to set the lower bound on 𝛼2𝑥 . For a particular combination

of 𝛼2𝑦 and 𝛼1𝑦 , these expressions yield lines in 𝛼2𝑥 and 𝛼1𝑥 . Defining 𝜇 = 𝛼2𝑦 /𝛼1𝑦 , Eq. 2.24 can

be re-written as:

(𝐼) 𝜇𝛼1𝑥 + 1 − 𝜇 ≤ 𝛼2𝑥

(𝐸𝑞. 2.25)
(𝐼𝐼) 𝜇𝛼1𝑥 ≤ 𝛼2𝑥

Thus, the two lines that could be potential lower bounds are parallel lines of slope 𝜇, whose

intercepts are 1- 𝜇 (I) and 0 (II). Recall that 𝛼2𝑦 and 𝛼1𝑦 are required to have the same sign;

49
therefore, all possible values of 𝜇 are greater than 0. Since the more restrictive line determines the

lower bounds, the lower bound for a particular value of 𝜇 is therefore determined by whichever

𝛼2𝑥 -intercept has a lower value. Given that 𝜇 is always positive, there are 3 possible cases of 𝜇

values to consider: 𝜇>1 (𝛼2𝑦 >𝛼1𝑦 ), 𝜇=1 (𝛼2𝑦 =𝛼1𝑦 ) and 𝜇<1 (𝛼2𝑦 <𝛼1𝑦 ). Examples of these 3 cases

are shown in Figure 2.3.

Figure 2.3: Example cases of boundaries on αx2 and αx1 for the x-tensegrity

For the case where 𝜇=1, the two lines are one and the same and are shown in black in Figure 2.3.

For the case where 𝜇>1, the line (I) will always have an intercept less than zero (since the intercept

is 1-𝜇), therefore making line (II) the lower bound for these values of 𝜇. When 𝜇<1, line (I) will

always have an intercept greater than zero, thus making line (I) the more restrictive case. Thus, the

boundaries can be summarized as follows:

𝛼2𝑦 𝛼2𝑦 𝛼2𝑦

( ) 𝛼1𝑥 + 1 − ( ) ≤ 𝛼2𝑥 𝑓𝑜𝑟 ( )≤1
𝛼1𝑦 𝛼1𝑦 𝛼1𝑦
(𝐸𝑞. 2.26)
𝛼2𝑦 𝛼2𝑦
( ) 𝛼1𝑥 ≤ 𝛼2𝑥 𝑓𝑜𝑟 ( )≥1
{ 𝛼1𝑦 𝛼1𝑦

50
 The results generated using a brute-force approach (Figure 2.2) display the equilibrated

shapes purely in terms of possible 𝛼1𝑥 and 𝛼2𝑥 combinations, independent of the particular value

of 𝜇. To replicate this graph, all results encompassed within Eq. 2.26 must be found and combined.

For the case, 𝜇<1, the lowest value of 𝜇 (since 𝜇 strictly positive) would be 𝜇=0. In this case, the

boundary line (I) is the line 𝛼2𝑥 =1. As the value of 𝜇 is decreased, the slope becomes less and less

steep, and the intercept of (I) rises, resulting in regions contained within the region 𝛼2𝑥 >1. The

corresponding region is shaded in blue in Figure 2.4.

For the case where 𝜇>1, the lowest value of 𝜇 would be exactly the line 𝜇=1. The region

using 𝜇=1 in line (I) as the lower-bound is pictured in Figure 2.4 in gray. Increasing 𝜇 beyond 1

would only increase the slope of the line, thus generating regions already contained by the case of

𝜇=1 (for strictly positive 𝛼1𝑥 and 𝛼2𝑥 ). The combined cases of these two regions (i.e., combining

the blue and gray regions of Figure 2.4) result in the same results as found in the brute-force

approach, thus validating the analytical approach. Similar derivations can be applied for the case

where 𝛼1𝑦 and 𝛼2𝑦 are both negative.

51
 Figure 2.4: Combined regions for all cases of μ for the X-tensegrity

Similar derivations can be applied for the case where 𝛼1𝑦 and 𝛼2𝑦 are both negative.

2.3 Triplanar Tensegrity

For additional illustration of this analytical approach, it can be applied to a 2D triplanar

tensegrity. Pictured in Figure 2.5, the triplanar tensegrity comprises 4 nodes, 3 cables and 3 struts.

The connectivity matrix is as follows:

1 −1 0 0
0 1 −1 0
1 0 −1 0 (𝐸𝑞. 2.27)
𝑪=
1 0 0 −1
0 1 0 −1
[0 0 1 −1]

52
 \
Figure 2.5: A 2D triplanar tensegrity

The geometric matrix and equilibrium matrix are shown in Eq. 2.28 and 2.29 respectively as:

1 1 −1 0 0 0
𝑮 = [1 0 0 −1 1 0] (𝐸𝑞. 2.28)
0 0 1 −1 0 1

1 0 1 −1 0 0
−1 1 0 0 −1 0 (𝐸𝑞. 2.29)
𝑬=[ ]
0 −1 −1 0 0 −1
0 0 0 1 1 1

With these matrices, the steps described in Sections 2.1.3-2.1.5 are applied, and result in the

following equations for the force densities:

−𝛼1𝑦
𝑞2 =
𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦
−𝛼1𝑦
𝑞3 =
𝛼3𝑦 𝛼1𝑥 − 𝛼3𝑥 𝛼1𝑦
−(𝛼1𝑦 − 𝛼3𝑦 )
𝑞4 =
𝛼3𝑦 𝛼1𝑥 − 𝛼3𝑥 𝛼1𝑦 (𝐸𝑞. 2.30)
−(𝛼1𝑦 − 𝛼3𝑦 )
𝑞5 =
𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦
(𝛼1𝑦 (𝛼1𝑦 − 𝛼3𝑦 ))
𝑞6 =
(𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦 )(𝛼3𝑦 𝛼1𝑥 − 𝛼3𝑥 𝛼1𝑦 )

Similar to the x-tensegrity, the null space of the geometric matrix of the triplanar tensegrity has

k=3 vectors, resulting in a total of 6 geometric coefficients. With the same scaling assumptions as

imposed on the x-tensegrity, two geometric coefficients are eliminated, leaving Eq. 2.30 in terms

of the force densities and the 4 remaining geometric coefficients, 𝛼1𝑥 , 𝛼3𝑥 , 𝛼1𝑦 and 𝛼3𝑦 .

53
 In this case, the terms of interest are 𝛼1𝑦 and (𝛼1𝑦 − 𝛼3𝑦 ) in the numerators and (𝛼3𝑦 𝛼1𝑥 −

𝛼3𝑥 𝛼1𝑦 ) and (𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦 ) in the denominator. Once again, the force densities are all

required to be positive, so the following requirements are derived via careful observation of Eq.

2.30:

• 𝛼1𝑦 and (𝛼1𝑦 − 𝛼3𝑦 ) must have the same sign

• For 𝛼1𝑦 and (𝛼1𝑦 − 𝛼3𝑦 )>0, (𝛼3𝑦 𝛼1𝑥 − 𝛼3𝑥 𝛼1𝑦 ) and (𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦 )<0

• For 𝛼1𝑦 and (𝛼1𝑦 − 𝛼3𝑦 )<0, (𝛼3𝑦 𝛼1𝑥 − 𝛼3𝑥 𝛼1𝑦 ) and (𝛼3𝑦 − 𝛼1𝑥 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦 )>0

Numerical results using the same approach described for the x-tensegrity (checking

combinations of coefficient values against one another for shapes that result in equilibrium) were

found and are shown in Figure 2.6. Each coefficient was varied incrementally between -2 and 2,

and the results are shown in terms of 𝛼1𝑥 vs. 𝛼3𝑥 in Figure 2.6.

Figure 2.6: Equilibrated shapes for the triplanar tensegrity, found by checking combinations of geometric coefficient
values against one another, varied between -2 and 2.

54
Similar to the x-tensegrity, the denominator terms from Eq. 2.30 can be used to define boundaries

as follows:

(𝐼) 𝛼3𝑦 + 𝛼3𝑥 𝛼1𝑦 = 𝑎1𝑥 𝛼3𝑦

(𝐸𝑞. 2.31)
(𝐼𝐼) 𝛼3𝑦 𝛼1𝑥 = 𝛼3𝑥 𝛼1𝑦

Whether or not the lines represent upper or lower bounds in 𝑎1𝑥 (as a function of 𝑎3𝑥 ) will depend

on the particular case. The sign of 𝛼1𝑦 determines the sign of all other terms. For 𝛼1𝑦 >0, the other

numerator term shows 𝛼1𝑦 > 𝛼3𝑦 . This case results overall in negative-valued numerators for force

densities, and therefore indicates that both denominator terms must be negative. Letting 𝜇 =

(𝛼1𝑦 /𝛼3𝑦 ), this case also indicates that the magnitude of 𝜇 > 1. The coefficient 𝛼3𝑦 can be positive

or negative so long as it satisfies 𝛼1𝑦 > 𝛼3𝑦 ; the sign of 𝛼3𝑦 determines whether lines (I) and (II)

are upper or lower bounds. Conversely, when 𝛼1𝑦 <0, the other numerator term indicates 𝛼1𝑦 <

𝛼3𝑦 , and therefore the magnitude of 𝜇 < 1. Again the sign of 𝛼3𝑦 can be positive or negative, and

is used to determine whether or not lines (I) and (II) are upper or lower bounds. This case results

in overall positive numerators, and therefore indicates the denominator terms must also be positive.

In total, there are four possible cases for the triplanar tensegrity. These cases are labeled (A)-(D)

and summarized in Figure 2.7.

Figure 2.7: All possible cases for signs of geometric coefficient terms for triplanar tensegrity, μ=(α1y/α3y)

55
 Starting with case (A), the numerator terms are both positive, indicating the denominator

terms must be negative, and the magnitude of (𝛼1𝑦 /𝛼3𝑦 )> 1. In this case, 𝛼3𝑦 is taken as positive.

This means the denominator terms are both still less than zero, and can be rearranged and combined

to show:

(1 + 𝛼3𝑥 𝜇) < 𝛼1𝑥 < 𝛼3𝑥 𝜇 (𝐸𝑞. 2.32)

In this case, line (I) forms a lower bound while line (II) forms an upper bound. With the substitution

of 𝜇 = (𝛼1𝑦 /𝛼3𝑦 ) into lines (I) and (II), it is clear that these two lines are parallel to one another,

and offset by +1. No matter what the value of 𝜇, the line (1 + 𝛼3𝑥 𝜇) = 𝛼1𝑥 (i.e., line (I)) will

always be above the line 𝛼1𝑥 = 𝛼3𝑥 𝜇 (i.e., line (II)). This indicates that the viable regions for 𝛼1𝑥

(being greater than line (I) and less than line (II)) do not overlap, and thus there are no viable

solutions for this case.

In case (B), the numerators are both still positive, but 𝛼3𝑦 is taken as negative. The division

by the negative value of 𝛼3𝑦 flips which line is an upper and lower bound, showing:

(1 + 𝛼3𝑥 𝜇) > 𝛼1𝑥 > 𝛼3𝑥 𝜇 (𝐸𝑞. 2.33)

This case offers viable solutions, therefore, in the region bounded between lines (I) and (II). Since

𝛼1𝑦 > 0 and 𝛼3𝑦 < 0, 𝜇 is negative in this case, and will always have a magnitude greater than 1.

The largest value of 𝜇 that satisfies these requirements would be 𝜇 = −1, while the most negative

value of 𝜇 would approach negative infinity as 𝛼3𝑦 approaches zero. As 𝜇 approaches negative

infinity, the slopes of the two lines become increasingly negative. When 𝛼3𝑦 = 0, the two lines

would over lap one another at a value of 𝛼3𝑥 = 0 for any value of 𝛼1𝑥 . For any value of 𝜇 between

negative infinity (region shaded blue) and -1 (region shaded red), the region of overlap would lie

between these cases, as shown in Figure 2.8.

56
 Figure 2.8: Example regions from case (B) of viable geometric coefficient combinations for the triplanar tensegrity.

Cases (C) and (D) stem from values of 𝛼1𝑦 < 0, resulting in positive numerators, and

requiring the denominator terms to each also be positive. Since this case requires 𝛼1𝑦 < 𝛼3𝑦 , 𝜇

will always have a magnitude less than one. As before, the sign of 𝛼3𝑦 determines which line is the

upper or lower bound. Case (D) shows the same upper and lower bound designation as case (A),

and can therefore be eliminated without further inspection, since the resulting boundaries have no

region of overlap, and therefore no viable combinations of 𝛼3𝑥 and 𝛼1𝑥 . This leaves case (C), where

𝛼3𝑦 > 0, resulting once again in negative values of 𝜇. This case has lines (I) and (II) as the upper

and lower bounds respectively, just as in case (A). Since 𝜇 must be negative, the largest value 𝜇

can have would be 𝜇 = 0. Similarly, since the magnitude of 𝜇 must be less than 1, the smallest

value 𝜇 can have would be 𝜇 = −1. These two cases are pictured in Figure 2.9, with 𝜇 = −1 shaded

in blue and 𝜇 = 0 shaded in red. As the slope is increased from -1 towards 0, the region bounded

by the curves sweeps an area between these two cases.

57
 Figure 2.9: Example regions from case (C) of viable geometric coefficient combinations for the triplanar tensegrity.

The numerical results shown in Figure 2.6 from the brute-force approach encompass all

combinations of 𝛼1𝑥 and 𝛼3𝑥 without distinguishing particular pairs of 𝛼1𝑦 and 𝛼3𝑦 values. Thus

to compare the analytical results against the brute-force approach results, the regions derived from

both cases (B) and (C) would be combined. Visual comparison of these results (i.e., Figure 2.8

combined with Figure 2.9 vs the brute-force results in Figure 2.6) verifies the success of this

approach. In the case of the triplanar tensegrity, the lines defining the upper and lower bounds of

the regions (line (I) and (II) respectively) do not change, so the results can be concisely summarized

as follows:

𝛼1𝑦 𝛼1𝑦
1 + 𝛼3𝑥 ( ) > 𝛼1𝑥 > 𝛼3𝑥 ( ) 𝑓𝑜𝑟 (𝛼1𝑦 > 0 𝑎𝑛𝑑 𝛼3𝑦 < 0) 𝑂𝑅 (𝛼1𝑦 < 0 𝑎𝑛𝑑 𝛼3𝑦 > 0) (𝐸𝑞. 2.34)
𝛼3𝑦 𝛼3𝑦

The results offered by this analytical approach are a set of inequalities on a minimal set of

design variables (geometric coefficients) which succinctly describe the entire solution space of

equilibrated shapes of a tensegrity structure. These results offer insight into the variety of shapes

58
these structures can achieve, and offer them in a compact description. However, this approach

requires extensive analysis of the problem unique to the tensegrity in question. The equations that

results from requiring the force components (member length components times force densities) to

be in the range of the null space of the equilibrium matrix can quickly become too complex to solve

directly, as is the case with even the simplest 3D tensegrity, the triplex. In such cases, a numerical

approach is necessary; this is addressed in the next chapter.

59
 Chapter 3

Numerical Characterization of a Tensegrity’s Achievable Shapes

Understanding the range of shapes a particular tensegrity configuration can achieve is a

complex problem. Throughout the literature, a vast variety of form-finding approaches exist

seeking equilibrated shapes of tensegrities. These approaches often employ symmetry assumptions

to help simplify the problem, and are typically structured to seek only one form that is equilibrated.

The analytical approach developed in Chapter 2 addresses this for reasonably simple tensegrity

structures, developing a way to characterize the variety of shapes a tensegrity structure can achieve,

symmetric or otherwise. However, as the number of members and nodes in a tensegrity structure

increases, the complexity of the resulting equations also increases, rapidly resulting in equations

too unwieldy to solve directly. In these cases, a numerical approach is necessary to characterize the

range of achievable shapes.

Numerical approaches in the literature employ a variety of different methods, but are

typically developed with the goal of seeking a single shape. Shifting the goal of the algorithm to

seek many shapes instead of a single shape dramatically affects the formulation of the problem. In

some ways, seeking many equilibrated shapes is akin to employing form-finding problems seeking

a single shape many times over. However, employing an algorithm designed to converge for a

single shape repeatedly is not an efficient means of finding many different equilibrated shapes, and

would prove particularly inefficient when seeking a set of shapes that represent the variety of shapes

a particular tensegrity can achieve.

In the literature Porta and Hernández-Juan [82] and Juan and Tur [83] discuss a similar

problem, looking at collision free path planning between equilibrated tensegrity configurations.

Intrinsic to these path planning approaches is finding a set of shapes which are equilibrated between

60
the desired start and end points of this path. The approaches addressed in these papers use

probabilistic methods in tandem with optimization to seek equilibrated paths (that also avoid

collisions) between target shapes. Porta and Hernández-Juan [82] in particular define an

equilibrium manifold to aid in the path planning process. The equilibrium manifold is essentially a

method of describing shapes, where kinematic and static constraints are used in the definition of

the shapes. As noted in [82], a full description of the equilibrium manifold (e.g. the full description

of the variety of shapes a tensegrity can achieve) is difficult, and typically approached with some

form of sampling and/or optimization of a parameterized form of the manifold. The work of [82]

and [83] focuses on path planning for robotic applications, and thus incorporates consideration of

finding numerous equilibrated shapes as required by the path. These approaches, however, do not

seek to describe the variety of shapes the structure can achieve in general. The numerical approach

developed in this chapter focuses on seeking a wide variety of shapes, akin to developing a

representative dataset of shapes which the entire equilibrium manifold of the tensegrity structure

would describe. It is likely that this approach could be used in tandem with path planning

approaches like those described in [82] and [83], however, is outside the scope of this research.

This chapter develops a novel numerical approach, seeking a representative dataset of

equilibrated shapes. The result is a dataset which represents a wide variety of static, equilibrated

shapes the particular tensegrity configuration can achieve. This dataset is demonstrative of the types

of asymmetric shapes a tensegrity can achieve.

3.1 The Branching Algorithm

The numerical approach developed herein uses a known equilibrated shape as a branching

point to search for other equilibrated shapes. With one known equilibrated shape, the first step in

this numerical approach is identifying the design variables. These variables can be geometric

61
coefficients like those used in the analytical approach described in Chapter 2, nodal coordinates as

often used in form-finding problems throughout the literature, or any other set of design variables

which fully describe the structure’s shape. Once the design variables are identified, the initial

starting shape is specified in terms of the selected variables. This set of design variables,

corresponding to a known, equilibrated shape, becomes the first branch point. Upper and lower

bounds on the design variables can be set based on how they correspond to the structure’s shape,

and can help narrow the scope of the algorithm to only consider a portion of the design space

relevant to the problem. This branching algorithm results in a representative dataset which

approximates the variety of shapes the tensegrity configuration in question can achieve in

equilibrium.

In this approach, a branch point is used as a starting point from which parametrized lines

are drawn. A parameterized line in terms of design variables, v, is shown in Eq. 3.1:

𝑣1 𝑣1,BP 𝑣1,𝑘 − 𝑣1,BP

{ ⋮ } = { ⋮ }+{ ⋮ }𝑡 (Eq. 3.1)
𝑣𝑁 𝑣𝑁,BP 𝑣𝑁,𝑘 − 𝑣𝑁,BP

In this equation, N is the number of design variables, BP denotes the branch point, and k is an

arbitrary point on a parameterized line. The generalized slope of the parameterized line is

determined by the vector corresponding to the point k (vk). Choosing different values for t, the

parametric variable, generates new sets of design variables v along the parametric line, where t=0

corresponds to the branch point. Different values of t on a particular parameterized line correspond

to different shapes, each of which can be checked for equilibrium. For each line defined this way,

values of t corresponding to the upper and lower bounds set on each of the variables is found.

Values of t are increased beyond t=0 until a boundary is hit for one of the design variables, setting

the upper bound of the t value on that particular line. The lower bound of t is found in the same

way, decreasing t below 0 until a boundary is hit. Defining the range of t this way results in lines

62
that span broader ranges of the solutions space, to encourage finding a variety of shapes even if the

branch point is near a boundary of the design variables.

A number of NL parameterized lines, specified by their slopes via vk, are then defined from

this initial branch point. A number of points (nt) on each of these lines is then checked for structural

equilibrium, with shapes corresponding to the resulting variables v. Any points which do not satisfy

equilibrium are discarded, while satisfactory points are added to a solution set. To discourage very

similar shapes from any single line skewing the distribution of points in the broader solution set, a

maximum number (ne) of points is specified to add to the solution set.

After this solution set is updated during a branch iteration, a subset is selected as NBP new

branch points. From each of these NBP points, NL new lines are drawn, on each of which nt resulting

shapes are checked for equilibrium. The numerical algorithm continues in this way, updating the

solution set through each iteration to include any new equilibrated shapes found.

3.1.1 Exit Criteria of the Branching Algorithm

As the numerical solution set updates, it is important to identify proper convergence criteria

to indicate when the algorithm should terminate. The goal of this algorithm is to provide a

representative dataset of the range of achievable shapes. Since the end results are generally not

known, the variety of shapes in the overall dataset is tracked to quantify when the algorithm stops

finding shapes in new regions of the design space. This is quantified using the Mahalanobis

distance.

Mahalanobis distance is frequently used in statistical analysis as a measure of how a point

deviates from the mean of all points in a dataset [75-80]. This is calculated via Eq. 3.2:

̅ )𝜮−1 (𝒗ⅈ − 𝝁
𝑑𝑚,ⅈ = √(𝒗ⅈ − 𝝁 ̅ )′ (𝐸𝑞. 3.2)

63
Where dm,i is the Mahalanobis distance of point i, vi is the vector of the design variables for point i,

𝝁 is the vector of the average values of the design variables in the entire dataset. For a dataset with

Np points, where each point has Nv variables, the covariance matrix, 𝜮, is given by Eq. 3.3:
2
∑𝑁 𝑃
ⅈ=1(𝑣ⅈ,1 − ̅̅̅)
𝜇1 ∑𝑁𝑃
ⅈ=1(𝑣ⅈ,1 − ̅̅̅)(𝑣
𝜇1 ⅈ,𝑁𝑣 − ̅̅̅̅̅)
𝜇𝑁𝑣
⋯
𝑁𝑃 − 1 𝑁𝑃 − 1
𝜮= ⋮ ⋱ ⋮ (𝐸𝑞. 3.3)
𝑁𝑃 𝑁𝑃 2
∑ⅈ=1(𝑣ⅈ,𝑁𝑣 − ̅̅̅̅̅)(𝑣
𝜇𝑁𝑣 ⅈ,1 − ̅̅̅)
𝜇1 ∑ⅈ=1(𝑣ⅈ,𝑁𝑣 − ̅̅̅̅̅)
𝜇𝑁𝑣
⋯
[ 𝑁𝑃 − 1 𝑁𝑃 − 1 ]

Where the non-bold vi and 𝜇̅ are a design variable in point i and the average value of a design

̅ respectively). A high Mahalanobis distance indicates a strong

variable (elements of vectors vi and 𝝁

deviation away from the average of the whole dataset. In the context of this numerical algorithm, a

high Mahalanobis distance for some set of design variables vi would correlate to shape i having

design variable values significantly different than most of the rest of the dataset. This in effect gives

a measure of how different shape i is from the rest of the shapes in the data set.

To use this to help guide termination of the algorithm, the maximum value of Mahalanobis

distance in the solution set is calculated at each iteration. As the algorithm iterates and branches,

the solution space fills out with equilibrated shapes. As the dataset fills, the maximum Mahalanobis

distance will stop changing significantly. This indicates that shapes in new regions of the design

space (relative to the dataset as a whole) are no longer being found. Once the change in maximum

Mahalanobis distance relative to prior iterations effectively stops changing (within some percentage

for a specified number of iterations), the algorithm considers the exit criterion to be met, and

terminates.

64
3.1.2 Factors Influencing Efficiency

The efficacy of the numerical algorithm is influenced by a variety of factors. The number

of points Np checked in each iteration is shown in Eq. 3.3, and is affected by of the number of

branch points NBP, the number of lines off each branch point NL, and the number of points checked

on each line nt.

𝑛𝑡
𝑁𝑝 = (𝑁𝐵𝑃 )(𝑁𝐿 ) (𝐸𝑞. 3.3)

Each design point must be checked for equilibrium—a non-trivial calculation when the

number of design variables become large. Making an informed selection of each of these variables

and thus the resulting number of points in each iteration is therefore an important consideration for

the efficiency of the algorithm. The lines off the branch points are given random directions (within

the variables’ bounds)—including more lines offers a greater spread of directions from a particular

point, and can offer greater coverage across the design space. Increasing the number of points

checked on each line similarly offers a greater spread of points checked across the design space.

In general, increasing the number of lines is given greater priority over increasing the

number of points off each line, as the potentially novel “directions” of each line are more likely to

find points in new areas of the design space than increasing the density of points on a smaller

number of lines. The points on the line nt are evenly spaced in t values (from Eq. 3.1) between t

values in the positive and negative direction along the line (relative to the branch point where t =

0). The maximum and minimum value of t is selected by adjusting t in each direction until a

boundary is first hit for any of the design variables.

The number of branch points NBP also influences the number of points checked directly, as

well as the spread of the resulting points. Since each line in successive iterations is drawn from one

of the selected branch points, not only is the quantity of branch points significant in its influence

65
on algorithm behavior, but how these points are selected can also greatly influence the speed of

convergence of the algorithm.

The Mahalanobis distance is calculated to track the maximum value in the solution set to

monitor convergence. This requires the Mahalanobis distance to be calculated for each point in the

solution set. Thus, the same calculations used to assess convergence can also be used to help select

a subset of points from the overall solution set to use as branch points in the next iteration.

This approach uses the Mahalanobis distance to weight each point in the solution set. Using

MATLAB’s datasample function, the Mahalanobis distance is used as a weight for each point to

provide a weighted sample of the solution set. The sample favors points with higher Mahalanobis

distance values, and is thus more likely to select points in outlying, less-dense portions of the

solution set. By favoring points in less dense regions of the design space, new lines will be drawn

in successive iterations in these sparser regions, thus encouraging the discovery of new points.

3.1.3 Summary of the Branching Algorithm

The numerical approach is summarized as follows:

1. Identify appropriate design variables, and corresponding upper and lower bounds on

the design variables. For one shape known to be in equilibrium, find the corresponding

values of the design variables. The design variables for this shape are used as the first

branch point.

2. For each branch point, define NL lines off the branch point by randomly specifying the

vector of variables vk (and associated line directions) using Eq. 3.1.

66
 3. Select nt points evenly spaced on each line, setting the upper and lower bounds of t

values such that each is the first value above and below t = 0 at which any design

variable hits a bound.

4. Check all of the points thus generated for equilibrium. Discard any points which are

not equilibrated, then add the remaining equilibrated points to the overall solution

dataset.

5. Calculate the Mahalanobis distance for each point in the solution dataset. Calculate the

maximum Mahalanobis distance of all points in the solution set.

6. If the current iteration’s maximum Mahalanobis distance has not changed by a

designated percentage during a designated number of iterations, the exit criteria has

been satisfied, so exit the algorithm. If not, continue to step 7.

7. Select NBP points from the solution dataset using a weighted sample, where the

Mahalanobis distance is used to provide more weight to points with higher

Mahalanobis distance values. Return to step 2.

Figure 3.1 also summarizes this approach in a flow chart. This numerical approach can

work with any type of design variable and any initial shape as a branch point. The resulting solution

dataset provides a representative sample of shapes, lending insight into a significant variety of the

range of achievable shapes a tensegrity structure can achieve. Example cases for the 2D X-

tensegrity (addressed in Chapter 2), a 3D triplex, and a triplex module are shown in the following

sections.

67
Figure 3.1: Flow-chart of branching algorithm

68
 3.1 Example: 2D X-tensegrity

The 2D X-tensegrity provides a good demonstration of the efficacy of this numerical

approach. Since the X-tensegrity can be solved analytically, the solution to this problem is already

known. Furthermore, the constrained X-tensegrity case provides an example with only 2 design

variables, making it easier to visualize the weighting approach and convergence behavior.

3.2.1 Constrained X-Tensegrity

The geometric coefficients used in the analytical approach are used as design variables.

Addressing the constrained X-tensegrity problem, the strut lengths are set to √2. Combined with

member 1 being set to length 1, this leaves two geometric coefficients, α1x and α2x. The number of

branch points was set to 100 (all points are used if the total number of points in the data set is less

than 100), with 20 lines off of each point and up to 5 equilibrated points are saved from each line.

Convergence criterion was met if the percent difference in the maximum Mahalanobis distance did

not exceed 1% for 15 iterations. The Mahalanobis distance was used to weight the points in the

solution set for selection of the 100 branch points.

A sample iteration is shown in Figure 3.2. The light blue point corresponds to the initial

branch point, and other shapes in the dataset are shown as colored points, with the colors

corresponding to the Mahalanobis distance. The red crosses show those points which are selected

as branch points for the next iteration. From Figure 3.2, the Mahalanobis distance can be seen to

favor points in less-dense regions of the design space, and is thus useful in encouraging finding

new shapes that span the entire solution space.

69
 Figure 3.2: A sample iteration of the total solution set found by the numerical approach, where color indicates
Mahalanobis distance of each point. Red x’s indicate the points that were selected as branch points for the next
iteration, where the sample of points were weighted by the Mahalanobis distance

The maximum Mahalanobis distance of the solution dataset is tracked for every iteration,

and is shown in Figure 3.3. Exit criterion was met in 21 iterations. The algorithm took

approximately 70 seconds to run, and found a total of 218,415 equilibrated points. The results show

excellent agreement with the known, analytically found solution set for the constrained X-

tensegrity, as shown in Figure 3.4. To find an equilibrated dataset of similar size, 650 increments

of each geometric coefficient between the same upper and lower bounds are checked against one

another for equilibrium. This brute-force approach takes 50 seconds to run, and results in 212,737

equilibrated points. While the brute-force approach is more efficient in this particular case, the

branching algorithm was able to find similar results in a similar amount of time, and show good

agreement with the known solutions. The brute force approach checks every combination of the

tested variables against one another, and thus takes exponentially longer with the addition of

70
variables (i.e., if there are n increments tested for each variable and k variables, then nk points must

be checked). The numerical approach does not generate more or less points based on the number

of variables in the problem. While additional variables may slow down the numerical approach,

since there are thus more factors contributing to equilibrium, it will not slow down exponentially

like the brute-force approach. Despite being slower than a brute-force approach, this example is a

good demonstration of the accuracy of the branching algorithm.

Figure 3.3: Maximum Mahalanobis Distance for the constrained X-tensegrity.

71
 Figure 3.4: Comparison of branching algorithm results to analytical results for constrained X-tensegrity.

3.2.2 Unconstrained X-Tensegrity

The numerical approach shows excellent agreement with the unconstrained X-tensegrity

as well. In this case, there are 4 geometric coefficients, α1x , α2x, α1y and α2y. Like the constrained

72
case, 100 points weighted by the Mahalanobis distance were used as branch points. The number of

points off each line was maintained at 5, while the number of lines off each branch point was

increased to 50 to accommodate the larger design space. Additionally, the exit criterion was

considered satisfied when no more than a 5% difference in maximum Mahalanobis distance was

found for 75 consecutive iterations. The larger percent difference was used to accommodate the

larger design space (which results in an increased likelihood of finding new solutions with a larger

Mahalanobis distance), while the number of iterations increased to ensure the spread across the

solution space was reasonably dense.

The algorithm converged in 124 iterations, took approximately 30 minutes to run and found

3,303,533 equilibrated points. The maximum Mahalanobis distance for each iteration is shown in

Figure 3.4, and the results in α1x vs α2x are shown with the analytical results for comparison in

Figure 3.5. These results are a significant improvement upon a brute-force approach which would

check all combinations of coefficient values against one another for equilibrium. Using the same

boundaries as the geometric coefficients and dividing each coefficient into 50 incremental values

between the bounds, the brute force approach checks 6,250,000 (504) points, 2,284,781 of which

are equilibrated, and took over 2 hours to run. Thus for a similar amount of equilibrated points, the

branching approach not only converged significantly faster, but also resulted in more equilibrated

points overall.

73
 Figure 3.5: Maximum Mahalanobis distance as the algorithm iterated for the unconstrained 2D X-tensegrity

Figure 3.6: The results of the numerical approach (a) compared with the results of the analytical approach (b) shown
in terms of the geometric coefficients α1x and α2x..

74
 3.2 Example: 3D Triplex

The triplex tensegrity is the simplest configuration of a 3D tensegrity (sometimes called

the simplex). The triplex is comprised of 3 bottom face cables, 3 top face cables, 3 supporting

cables and 3 struts; in total 9 cables and 3 struts connecting 6 nodes. A well-known equilibrated

shape for the triplex tensegrity is pictured in Figure 3.6, where the top and bottom faces of the

triplex are parallel.

Figure 3.7: A triplex tensegrity.

With 6 nodes, this structure has 18 degrees of freedom. The problem can be scaled and

oriented by setting one node as the origin, setting member 1 to length 1 on the x-axis, and requiring

member 2 to lie on the x-y plane. This reduces the number of degrees of freedom to 11, and thus

11 variables must be identified to fully define the shape. Geometric coefficients can be derived for

the triplex using the same methods developed for the analytical approach, or the remaining free

nodal coordinates can be used directly as the design variables.

75
 In the case of the X-tensegrity, the design space was relatively small (4 variables at most),

and the maximum Mahalanobis distance was well-behaved, generally decreasing steadily. In the

case of the triplex, a tensegrity requiring 11 variables, the solution space is much more complex.

As the iterations progress, it is possible for the maximum Mahalanobis distance to steadily decrease

over a number of iterations but, if a new branch in the updated population finds a solution in a

significantly different area of the design space, the maximum Mahalanobis distance can increase.

The convergence criteria counts the number of iterations over which the percent difference in

maximum Mahalanobis distance has not changed significantly. While, for example, a 5% difference

might be considered small, a 5% increase is likely to indicate a solution found in a new portion of

the design space. Thus, whenever the maximum Mahalanobis distance increases from the prior

iteration by more than a specified percentage, the convergence criteria counter resets.

The numerical approach was applied to a triplex tensegrity using the geometric coefficients

as design variables, derived using the same approaches described in Chapter 2. In this case, 100

points weighted by the Mahalanobis distance were used as branch points, with 10 lines off of each

branch point and 5 points off of each line. The convergence criteria was considered satisfied after

50 consecutive iterations with a maximum Mahalanobis distance percent difference of no more

than 5% between successive iterations. The counter for consecutive iterations was reset if the

maximum Mahalanobis distance increased by 2.5% or more between successive iterations. In this

case, the algorithm converged in 274 iterations; 24,361 equilibrated shapes were found and the

algorithm took approximately 1 hour to run. The density of the final dataset can be increased by

running the algorithm with more strict convergence criteria, or a higher number of branch points,

lines off each branch point, and/or points off of each line. Nevertheless, this case found a significant

number of unique shapes for the free triplex tensegrity, demonstrating the efficacy of this algorithm

to find a wide variety of asymmetric shapes across the breadth of the solution space.

76
 Figure 3.8: Maximum Mahalanobis distance for the numerical approach with an unconstrained triplex.

3.3.1 Constrained Triplex

As another example, the numerical approach was also applied to a triplex tensegrity which

has 9 members fixed in length, allowing only the 3 supporting cables to change length. These length

constraints significantly reduce the number of degrees of freedom (bringing it down to only 3, with

the same assumptions of member 1 lying on the x-axis and member 2 on the x-y plane).

Identifying suitable design variables which fully describe the structure’s shape while

directly incorporating these length constraints, however, is much more complex than in the case of

the X-tensegrity. With these constraints, 3 out of the 4 members connected to a single node are

constrained. Using traditional nodal coordinates (or geometric coefficients) as design variables,

these constraints result in numerous, inter-related nonlinearities that quickly become too

77
cumbersome to solve. Figure 3.8 numbers the nodes and members of the triplex in question. All

members are fixed in length except for the supporting cables (members 7, 8 and 9). The known

lengths of the remaining members, in combination with the bottom-face being fixed on the x-y

plane and the first member fixed on the x-axis, can be used to derive a set of 3 variables that fully

define the shape. With careful consideration of the geometric properties incorporating fixed

member lengths, the 3 variables used are the lengths of two supporting cables (l8 and l9), and one

angle, (𝛷). This angle is also depicted in Figure 3.8, and is the angle of rotation of the triangle

formed by N1, N2 and N5 about the x-axis. Chapter 4 discusses the general geometric properties

used to derive these design variables, and Appendix A.1 details the specific derivation of these

variables for the constrained triplex in question.

Figure 3.9: Constrained triplex; all members fixed in length except for members 7, 8 and 9.

Figure 3.9 shows the results for this constrained triplex in. In this case, 100 points weighted

by the Mahalanobis distance were used as branch points, with 30 lines off of each branch point and

5 points off of each line. The convergence criterion was considered satisfied after 75 consecutive

78
iterations with a maximum Mahalanobis distance percent difference of no more than 5% between

consecutive iterations. The counter for consecutive iterations was reset if the maximum

Mahalanobis distance increased by 2.5% or more between consecutive iterations. This case

converged in 126 iterations, and the behavior of the maximum Mahalanobis distance is shown in

Figure 3.10.

Figure 3.10:The results from the numerical approach for a constrained triplex. The point marked by an x is the initial
equilibrated point, corresponding to the parallel triplex shape shown in Figure 3.5.

79
Figure 3.11: Maximum Mahalanobis Distance over the 126 iterations it took to converge for the constrained triplex.

Previous work by Roffman [15] derived the solution set for this type of constrained triplex.

The approach in [15] was developed by checking a wide variety of nodal coordinates against one

another using a spherical representations of nodes. The present numerical approach, in combination

with the geometric approach to length constraints developed in Chapter 4, results in a solution set

in excellent agreement with the results of [15]. The results in [15] are shown overlaid on the results

from this numerical approach in Figure 3.12, with the results from this approach in black and the

results from the approach of [15] in blue. Furthermore, this branching algorithm offers a significant

improvement in run-time as compared to the approach in [15], with the branching algorithm taking

approximately 1.5 hours to run and the approach in [15] taking approximately 3 hours to run.

80
 Figure 3.12: Comparison of results from the numerical approach developed for this research with the results of
Roffman [15] for a constrained triplex.

3.3 A Force-Finding Method for Complex Tensegrity Structures

This numerical approach can, in principle, work for any tensegrity structure comprised of

any number of members and nodes. As the structure becomes more complex, however, traditional

decomposition approaches to find force densities—many of which require symmetric shapes—do

not always find viable force density combinations. For complex structures in this research, such as

the triplex module example in Section 3.3, Particle Swarm Optimization (PSO) is used in

combination with decomposition of the equilibrium matrix to find viable force density sets even

for complex, asymmetric tensegrities.

For a particular shape defined by a set of nodal coordinates to be equilibrated, the tensegrity

structure must also have a corresponding set of member force densities which both satisfy the

loading requirements of the members (cables in tension, struts in compression), as well as the

general equations of equilibrium, shown again as:

81
 𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒙)
[𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒚)] 𝒒 = 𝟎 (𝐸𝑞. 3.4)
𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒛)

Where C is the connectivity matrix, x y z are the nodal coordinate vectors and q is the vector of

force densities.

Recall from the analytical approach in Chapter 2 that the member loading requirements can

be separated from the force density vector by incorporating the member loading signs in a vector

m, and multiplying this vector into the equilibrium matrix:

𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒙)𝑑𝑖𝑎𝑔(𝒎)
[𝑪𝑇 𝑑𝑖𝑎𝑔(𝑪𝒚) 𝑑𝑖𝑎𝑔(𝒎)] 𝒒
̃=𝟎 (𝐸𝑞. 3.5)
𝑇
𝑪 𝑑𝑖𝑎𝑔(𝑪𝒛) 𝑑𝑖𝑎𝑔(𝒎)

̃ now represents the magnitudes of the force densities, each of which is required to remain
Where 𝒒

positive to avoid violating member loading constraints.

The matrix in Eq. 3.5 is a modified equilibrium matrix, E*, with information on member

length components and member types built in. The null space of this equilibrium matrix, null(E*),

can therefore be used to characterize the force density magnitudes as:

𝑁𝑬∗

∑ 𝛾ⅈ 𝒆∗ⅈ = 𝒒
̃ (𝐸𝑞. 3.6)
ⅈ=1

Where 𝑁𝑬∗ is the number of vectors in the null space of E*, 𝒆∗ⅈ are corresponding null space basis

vectors, and 𝛾ⅈ are the constant-valued coefficients corresponding to the null space vectors.

Any combination of coefficient values, 𝛾ⅈ , which results in a set of positive force densities,

̃>0 is a valid combination of the null space vectors that satisfies all equilibrium requirements for
𝒒

the set of nodal coordinates used to form E*. Different sets of nodal coordinates will result in

different dimensions of the null space, and thus a different number of coefficients to find. No matter

the number of coefficients, a PSO method is implemented to seek a set of coefficients which results

82
in a set of positive force densities. To encourage finding positive force densities, the fitness function

to be minimized is defined as follows:

̃)
− min(𝒒
𝑓𝑖𝑡𝑛ⅇ𝑠𝑠 = (𝐸𝑞. 3.7)
̃𝑇 𝒒
√𝒒 ̃

Where superscript T indicates transpose.

Additionally, since the zero vector is always in the span of a null space, this vector of zeros

will also always be a solution to the problem. This would correspond to a solution where none of

the members are pre-stressed, which is undesirable for a tensegrity. To help mitigate the possibility

̃ is less than a tolerance

of finding this solution, a penalty is added if the minimum value of 𝒒

(typically set to 0.001 here).

This method was successfully used to find force densities for a wide variety of shapes for

complex structures, including all shapes found in the following section for the triplex module. In

general, the results typically converge quickly when a solution is present. However, shapes that are

not equilibrated result in solutions that either falsely converge (e.g., on a fitness value around the

assigned penalty value), or else hit limits on the number of iterations or runs set for the optimization

problem. Excessive iteration is problematic when used for something like the numerical approach

developed in this chapter, since many shapes will be tested for equilibrium that are ultimately not

equilibrated.

To minimize the effects on computation time for such cases, a known equilibrated shape

was studied in detail to determine the effects of different optimization parameters on convergence.

Appendix B shows the results of this study, used to determine appropriate values of PSO parameters

(initial population size, maximum number of iterations, and maximum number of runs) to use for

the larger form-finding numerical approach developed in this chapter.

83
 3.5 Example: 3D Triplex Module

As a more complicated example, a triplex module is pictured in Figure 3.12. This module

is comprised of 6 triplexes. The geometry was conceived with large systems in mind, where

numerous modules would be assembled together to form a larger array, while maximizing node-

to-node connectivity between adjacent triplexes. (These modules are used later in Chapter 5 for an

example study.)

Without any constraints, a triplex module has 19 nodes, and thus 57 degrees of freedom.

This is reduced to 53 by assuming member 1 is a known, fixed length on the x-axis, and member 2

lies on the x-y plane. To further reduce the number of variables (and accommodate application

scenarios discussed in Chapter 5), the bottom face cables (blue in Figure 3.12) are fixed in space

such that all cables are on the same plane and of the same length. Additionally, the top face cables

(black in Figure 3.12) are assumed to be fixed in length. In total, this reduces the number of degrees

of freedom to 18, and allows the shape of the module to be defined effectively by description of

the top-face nodes.

Figure 3.13: A module comprised of 6 triplexes.

As the number of variables increases, the span of the design space for all types of shapes

the structure can achieve increases exponentially. While this numerical approach could work in

84
principle for any number of design variables, it is limited in practice by the efficiency of the

computer(s) used to run it, as well as the time limitations of the user. Increasing the number of

variables significantly increases the time it takes to converge, and further increases the likelihood

of convergence prior to fully spanning the complete range of potential shapes a structure can

achieve (e.g., increases the likelihood of missing shapes significantly different than those found in

the solution set).

In cases with a larger number of variables, such as this triplex module, it is therefore useful

to limit the design space to a region of interest. In this case, the module will later be used to connect

multiple modules together to form a larger array. For ease of connecting modules (and to increase

the likelihood of finding equilibrated shapes), it is helpful to consider shapes that are near the shape

shown in Figure 3.11. Shapes which laterally shift the top face far from the bottom face of a

particular module are more likely to interfere with adjacent modules when arranged together.

Furthermore, shapes close to this known, equilibrated shape are more likely to also be equilibrated.

Thus, the numerical approach sets upper and lower bounds on the design variables as deviations

from their values in this known, flat shape.

The triplex module shape, thanks to the imposed constraints, is fully defined by the shape

of its top face. There are 12 top face nodes (nodes n8-n19), 6 of which form an inner polygon,

highlighted in Figure 3.13. This inner polygon is first described, specifying n8 as the first 3 design

variables (n8x, n8y, n8z), then specifying additional inner-polygon nodes n10, n16 then n12 and n18

sequentially relative to other inner polygon nodes using spherical coordinates. Figure 3.14 shows

the two design variables resulting from using spherical coordinates for n10 relative to n8 as θ10 and

𝜙10. These relative angles, θ and 𝜙, are specified for n10, n16 n12 and n18. The remaining outer

nodes, and node 14, can be specified by a single angle thanks to the length constraints of the top

face. This angle can be found as the intersection of two spheres, resulting in a parametric circle

whose normal in 3D space is along the axis of an inner polygon member (or the line drawn between

85
n12 and n16 in the case of n14). The angle, λ, is shown for n9 relative to the inner polygon member

connecting n8 and n10 in Figure 3.14. Chapter 4 discusses the use of these spherical coordinates

and spherical intersections in detail. Additionally, the detailed derivation of all the design variables

is shown in Appendix A.2. In summary, the resulting 18 design variables are: n8x, n8y, n8z, θ10 ,𝜙10,

θ12 ,𝜙12, θ16 ,𝜙16, θ18 ,𝜙18, λ14, λ9, λ11, λ13, λ15, λ17 and λ19.

Figure 3.14: Top face of the triplex module, showing the inner polygon highlighted in blue.

Figure 3.15: An example of the spherical design variables, θ ,ϕ and λ shown in their local frame for the first 3 nodes of
the top-face cables of the triplex module.

The bounds on the design variables are set as small deviations from their values in the

original, flat shape shown in Figure 3.12. The nodal coordinates of the node 8 were allowed to vary

between +/- 0.25 of their original value in x-y, and +/- 1.5 in z, and the design-variable angles were

86
allowed to deviate by +/-45o. The numerical approach was run using up to 100 branch points

between iterations, weighted by maximum Mahalanobis distance, with 10 lines off each branch

point and up to 5 points from each line. The maximum Mahalanobis distance is shown in Figure

3.16 over the 284 iterations for which the algorithm was run. The percent difference between

adjacent iterations was no more than 0.7%, and the percent difference between the final iteration

and 50 iterations prior was 15.6%. The algorithm could be run longer to satisfy more strict exit

criterion, however, this case successfully found 29,131 unique, asymmetric shapes within the

restricted bounds set on the design variables, demonstrating the general capabilities of this

algorithm with more complex structures.

Figure 3.16: Maximum Mahalanobis distance for numerical approach used with a constrained triplex module

3.6 Branching Algorithm Conclusion

This chapter details a branching algorithm approach, designed to characterize the range of

achievable shapes a particular tensegrity configuration can achieve. The algorithm uses a single

87
shape known to be in equilibrium as a branch point, then defining parametric lines in terms of the

design variables in random directions to search for new equilibrated shapes. This algorithm was

tested against known cases for the 2D X-tensegrity, and the 3D triplex tensegrity, and showed

excellent agreement with the known results. Additionally, the algorithm was used for two cases—

the unconstrained triplex and a constrained triplex module—where the results were not previously

known. All example cases in this chapter demonstrate the ability of the branching algorithm to

generate a representative dataset of the variety of shapes a tensegrity structure can achieve in

equilibrium. These datasets demonstrate the ability of tensegrity structures to achieve a wide variety

of asymmetric shapes—a property highly beneficial in shape changing applications.

88
 Chapter 4

Designing Tensegrities for Shape Change

Designing a structure for a specific application entails a variety of considerations. In shape

changing applications, the primary consideration is the ability to achieve prescribed shapes. With

a target shape (or numerous target shapes) identified, a structure can be evaluated for its ability to

achieve the shapes necessary for the particular application.

Imposing member length constraints (i.e., members which are fixed in length) is an

important consideration that significantly affects a structure’s ability to achieve target shapes.

Members which cannot change length affect the geometry of the structure, and limit the range of

achievable shapes. It is often necessary to implement such constraints, be it for purposes of

attaching to rigid components or structures, or reducing the number of actuated members, thus

saving costs in weight, complexity, and energy consumption. Since member length constraints

directly impact the range of achievable shapes, methods to deal with these constraints are

considered in detail and incorporated directly in the analysis.

In general, quasi-static shape-change involves transitions between nearby equilibrium

states. A path-planning method is developed which seeks discrete, equilibrated shapes through

which a tensegrity can move to change shape between two prescribed end shapes. This path

planning can be used for deployment, and/or for additional shape adjustments post-deployment.

Finally, conventional structural characteristics are also important when evaluating a

tensegrity structure’s ability to perform a prescribed mission. A non-linear Finite Element Model

(FEM) is used to evaluate the structure’s stiffness under external loads. Structural stiffness lends

insight to the structure’s ability to maintain prescribed shapes in the presence of external

89
disturbances. Additionally, the FEM is used to consider the structure’s natural frequencies of

vibration. This modal analysis lends additional insight into the structure’s ability to maintain its

shape under dynamic loading.

Both stiffness analysis and modal analysis can be used to help identify a good pre-stress

state for the mission requirements. Pre-stress can be proportionally adjusted by a pre-stress factor;

once the ratios of pre-stress amongst the members within a tensegrity structure are identified,

multiplying all stresses by a constant factor will still result in structural equilibrium. Increasing or

decreasing this pre-stress factor is akin to tensioning (or un-tensioning) the structure without

changing member lengths, and influences the structure’s stiffness and natural frequencies. This

analysis can be used to identify reasonable pre-stress factors within material limitations for the

mission’s needs.

4.1 Geometric Principles to Incorporate Length Constraints

In many applications, ensuring that some members do not need to change length can be

very beneficial to the overall design. Each member which changes length must be actuated,

typically by incorporating motors of some kind. Motors by necessity add weight and energy

requirements to a system, both of which must be minimized in space applications to accommodate

launch vehicle limitations and ensure longevity in space where energy systems are limited.

Additionally, if the structure is mounted to other components (e.g., where the structure is attached

to the rest of the spacecraft, or if it is supporting another structure), attempting to change member

lengths at these stiff connection points could result in unnecessary strain on the structure.

To accommodate member length constraints in the shape change analysis methods

discussed in Chapters 2 and 3, the design variables of the problem which describe the shape are

selected to incorporate length constraints more directly. In the case of the 2D x-tensegrity in

90
Chapter 2, incorporating length constraints with the traditional Euclidean distance equations is

relatively straightforward, since there are only 4 geometric coefficients (design variables) in total.

In cases where there are more design variables, the equations needed to quantify member lengths

as a distance between nodes involve more design variables. As more members are constrained, the

nonlinear equations containing more variables become inter-related, each containing shared

variables. These equations quickly become too complex to solve directly using conventional,

Euclidean distance-based methods. In these cases, exploiting the geometric characteristics of the

structure’s connectivity can lead to better selection of design variables which more directly

incorporate length constraints without highly complex equations.

There are 3 main geometric principles used throughout this research to aid in development

of improved design variables. The first involves the general properties of triangles. Since any

polygon can be broken down into a collection of triangles sharing sides, intrinsic triangle properties

(how their lengths can add, how their angles are related, etc.) can be used to set limits on design

variables and relate members within the structure to one another. A second geometric consideration

uses the properties of a triangular pyramid—essentially a 3D extension of triangle properties. Just

as with triangles, triangular pyramid properties can be used to relate member lengths and angles to

one another. Finally, spherical coordinates and properties of spheres in 3D space can be used.

Spherical coordinates are a convenient way to represent nodal coordinates when a member is fixed

in length, since spherical coordinates can be written such that the radius is equal to member length,

and the free end of a member can trace out the surface of a sphere. Such spherical surfaces can be

used to realize where a member node can reach with the length constraint imposed, and thus the

properties of the multiple spheres representing constrained members can be used together to select

the best design variables.

The following subsections address each of these geometric principles in additional detail.

91
4.1.1 Triangle Properties

In a properly tensioned tensegrity structure (without cable slack and not so tensioned that

any struts start to buckle), all members can be represented as straight lines connected between

nodes. Any set of members that connect to form a closed loop—just like the cycles used for the

geometric matrix in Chapter 2—can be viewed as the edges of a polygon. Any polygon can be

broken down into a sub-set of triangles with shared edges. A 6-sided polygon is shown in Figure

4.1, for example, where the main polygon has solid edges and the additional edges drawn to divide

it into sub-triangles are shown as dashed lines. Even if this polygon is in 3D space, such that the

planes defined by each triangle are not all the same, the triangles themselves form their own planes,

and the typical 2D triangle properties hold.

Figure 4.1: A 6-sided polygon. Dashed lines show how the polygon may be broken up into 4 sub-triangles.

The two main useful properties of a triangle are the conditions on their side lengths,

and conditions on their relative angles when all sides are known. For a triangle of 3 side lengths la

≤ lb ≤ lc, a triangle can only be formed if lc+lb ≤ lc, and lc-lb ≤ la. If a side length is considered a

variable, l, and is not known, then the side length can be bounded by this property as:

|𝑙ⅈ − 𝑙𝑗 | ≤ 𝑙 ≤ 𝑙ⅈ + 𝑙𝑗 (𝐸𝑞. 4.1)

92
where members i and j are the other two sides of the triangle. This is particularly useful for

implementing length constraints when a triangle can be formed within a tensegrity polygon where

two members (i and j) are fixed length members, and the third member (l) can be used as a design

variable, now constrained by this triangle property of Eq. 4.1.

If the variable side length is shared between two triangles, then the bounds on the length

variable can be set by whichever triangle imposes more restrictive conditions. For example,

triangles A-B-C and A-C-D share side 7. The possible lengths of member 7, l7, can be found as:

max(|𝑙1 − 𝑙2 |, |𝑙8 − 𝑙3 |) ≤ 𝑙7 ≤ min(𝑙1 + 𝑙2 , 𝑙8 + 𝑙3 ) (𝐸𝑞. 4.2)

In cases where more than one of the sides involved in the relevant triangles are not fixed

in length, then design variables can be imposed sequentially, such that as lengths are selected within

whatever bounds can be identified, and additional constraints based on these triangle properties can

be imposed based on the selected length values. For example, consider all solid members in Fig 4.1

(1-6) as having known lengths, and the dashed members (7-9) as having lengths to be determined.

Then member 7 in Eq. 4.2 is identified by Figure 4.1 as one of the unconstrained members, in

addition to member 8. Since members 1 and 2 are fixed in length, member 7 can be fixed first,

selecting a length that satisfies |l1-l2| ≤ l7 ≤ l1+l2. With length 7, triangle A-C-D now only has the

length of member 8 left to be determined, where member 8 may be assigned a length subject to |l7-

l3| ≤ l8 ≤ l7+l3. Unconstrained lengths in polygons with some fixed lengths can be sequentially

imposed this way, intrinsically incorporating members which are fixed in length by setting

boundaries on design variables that incorporate these properties.

Sequentially specifying member lengths using triangle properties incorporates length

constraints more directly, but this does not necessarily guarantee that a geometrically viable shape

will be identified, particularly from a numerical approach perspective. For example, the closed

polygon of six sides shown in Figure 4.1 comprises 4 triangles, with 3 variable lengths. The first

triangle A-B-C can be used to select a compatible l7 relative to members l1 and l2. Then, the selected

93
value of l7 can be used with member l3 to set bounds on a second length variable l8. This second

variable, l8, is constrained by two triangles: A-C-D and A-E-D. Only one of these triangles has two

known side lengths (A-C-D having l7 and l3 now specified), and thus only one of the triangles can

be used to set bounds on length 8. Since information from the triangle A-E-D is not incorporated

into this selection, it is possible that a length for member 8 results in no possible length of member

9 which satisfies equation 4.2. This would be the case when max(|l8-l4|,|l6-l5|)>min(l8+l4,l6+l5).

For structures with many polygons (or many-sided polygons), tracking all the combinations

of lengths quickly becomes complicated. Numerical approaches using these triangle properties can

use these constraints as a way to check if a variable set is viable, discarding cases which result in

length sets that cannot form a triangle (i.e., the bounds in the inequality in Eq. 4.2 are contradictory).

While these cases do occur, setting up variables in this way, one that accounts for some of the

triangle inequalities by sequentially specifying lengths, makes solutions that are viable more likely,

since at least some of the variables are selected in a way that is guaranteed to satisfy the triangle

side-length inequalities. This is far better than an approach where variables are selected

simultaneously and at random, since that does not guarantee that any of the resulting triangles

would satisfy the necessary inequality of Eq. 4.2.

Once all side lengths of a relevant triangle are known, the law of cosines can be used to

solve for the angles between members. The law of cosines is shown in Eq. 4.3:

𝑙𝑎2 + 𝑙𝑏2 − 𝑙𝑐2

cos 𝜉 = (𝐸𝑞. 4.3)
2𝑙𝑎 𝑙𝑏

Where 𝜉 is the angle between the sides of lengths la and lb (𝜉 is opposite side of length lc). The

approaches developed herein do not typically use the law of cosines to specify a design variable as

an angle in this way; however, the law of cosines often becomes useful when translating the design

variables back to global cartesian coordinates; a step necessary for checking the resulting shape for

structural equilibrium.

94
4.1.2 Triangular Pyramid Properties

Similar to 2D triangles, the properties of a triangular pyramid can be used to identify design

variables for 3D tensegrities. A triangular pyramid is shown in Figure 4.2. The pyramid in Figure

4.2 is comprised of vertices A, B, C and P. Point Pp is the projection of point P on the x-y plane. In

this research, the pyramid properties described are used for a case where the position of node P is

unknown, but all member lengths in the pyramid (lAB, lBC, lAC, lAP, lBP, lCP) are known. In this case,

the law of cosines and the Pythagorean theorem can be used to solve for the nodal coordinates of P

(u, v, w).

Figure 4.2: A triangular pyramid, with vertices A, B, C, P. Point PP is the projection of P on the x-y plane.

The triangular pyramid is oriented such that the triangle A-B-C is on the x-y plane. The x-

y-z coordinate system of Figure 4.2 can be considered a local coordinate system, with the origin at

point A and with point B on the x-axis.

To solve for the nodal coordinates, consider triangle A-B-P, as show in Figure 4.3. Since

the 3 side lengths of A-B-P (lAP, lBP, lAB) are known, the law of cosines can be used to solve for the

95
angle 𝜀. Line s is drawn as a projection of P, such that it is perpendicular to side A-B. Thus with

angle 𝜀, the x-coordinate of P (u) can be determined as:

2 2 2
𝑙𝐴𝐵 + 𝑙𝐴𝑃 − 𝑙𝐵𝑃
𝑢 = 𝑙𝐴𝑃 ( ) (𝐸𝑞. 4.4)
2𝑙𝐴𝐵 𝑙𝐴𝑃

Figure 4.3:
Figure 4.3: An isolated triangle from triangular pyramid A-B-C-P in Figure 4.2

Once the x-component of P is known, triangle C-P-PP can be used to solve for w. Figure

4.4 shows this triangle. Note that Q is the distance from PP (the projection of P to the x-y plane) to

C, and can therefore be found as the distance between C and PP. If the coordinates of C are not

known directly, they can be solved for using the law of cosines with triangle A-B-C, since all side

lengths of this triangle are known, and this triangle lies on the x-y plane.

Figure 4.4:
Figure 4.4: A triangle from the triangular pyramid A-B-C-P from Figure 4.2.

With the coordinates of C=(uc,vc,0), Q is solved for using Eq. 4.5:

𝑄 2 = (𝑢𝐶 − 𝑢)2 + (𝑣𝐶 − 𝑣)2 (𝐸𝑞. 4.5)

96
Note that, since A is the origin, 𝑙𝐴𝐶 2 = 𝑢𝐶 2 + 𝑣𝐶 2 , and that, from triangle C-P-PP, 𝑄 2 = 𝑙𝐶𝑃 2 +

𝑤 2 . With these substitutions, Eq. 4.5 can be expanded to show:

2 2
𝑙𝐶𝑃 = 𝑙𝐴𝐶 − 2𝑢𝐶 𝑢 − 2𝑣𝐶 𝑣 + 𝑢2 + 𝑣 2 + 𝑤 2 (𝐸𝑞. 4.6)

Since 𝑢2 + 𝑣 2 + 𝑤 2 = 𝑙𝐴𝑃 2 , Eq. 4.6 can be solved for v as:

2 2 2
𝑙𝐴𝐶 − 𝑙𝐶𝑃 +𝑙𝐴𝑃 − 2𝑢𝐶 𝑢
𝑣= (𝐸𝑞. 4.7)
2𝑣𝐶

Finally, w can be found using the Pythagorean theorem as:

2
𝑤 = ±√𝑙𝐴𝑃 − 𝑢2 − 𝑣 2 (Eq. 4.8)

The solutions for w are involve pyramids of the same geometry, but one having a positive

z-value and the other a negative. When this approach is used to solve for nodal coordinates, both

of these solutions are geometrically viable with respect to the isolated pyramid, but all combinations

of results must be considered and checked when these results are incorporated with those for the

rest of the tensegrity. Appendix A shows the derivation of variables for a highly constrained triplex

tensegrity, which makes extensive use of these pyramid properties. In the case of the constrained

triplex, the pyramid properties are used such that variable member lengths may be specified as

design variables, and used with these properties to translate back to cartesian coordinates.

4.1.3 Spherical Coordinates

Finally, spherical coordinates are also used herein as a method of dealing with length

constraints. Figure 4.5 shows how spherical coordinates local to member i could be used to define

node k. The axes represent a local x-y-z coordinate frame with its origin at node j.

97
 Figure 4.5: : Spherical representation of node k relative to node j for member i.

The relevant angles of member i, 𝜃ⅈ and 𝜙ⅈ , are defined relative to node j to represent node k via

Eq. 4.9:

𝑛𝑘,𝑥 𝑙ⅈ cos 𝜃ⅈ cos 𝜙ⅈ + 𝑛𝑗,𝑥

𝑛
{ 𝑘,𝑦 } = { 𝑙ⅈ sin 𝜃ⅈ cos 𝜙ⅈ + 𝑛𝑗,𝑦 } (𝐸𝑞. 4.9)
𝑛𝑘,𝑧 𝑙ⅈ sin 𝜙ⅈ + 𝑛𝑗,𝑧

Spherical coordinates are often a better coordinate system to use than cartesian coordinates when a

member is fixed in length, since spherical coordinates directly incorporate member length and

represent the remaining 2 degrees of freedom as independent angles (𝜃 and 𝜙). As their name

implies, for a fixed radius, spherical coordinates can be thought of as representing possible node

locations as the surface of a sphere. In terms of Figure 4.5, the sphere for member i would have a

center at node j and a radius equal to the length of member i. The possible locations of node k (if

member i is fixed in length) are then anywhere on the surface of that sphere, identified by 𝜃ⅈ and

𝜙ⅈ .

Representing fixed-length members as spheres in this way can also help reduce the number

of variables needed to describe a shape when relating fixed-length members to one another. Figure

4.6 shows two spheres, representing two different members (of lengths ljk and lik respectively). One

member connects nodes i and k, the other member connects nodes j and k, and node k is shared by

both members.

98
Figure 4.6: Possible locations of a third node (k) as the circular intersection of two spheres, where members of length
lik and ljk must both meet at node k. Angle λ is the parameterized angle on the circle of intersection representing
geometrically possible node k locations.

Since node k is shared by both members (and their lengths are considered fixed), the

possible locations of k must be somewhere on both of the spherical surfaces which represent the

two members. Two spherical surfaces intersect either in a circle or a point (if they intersect at all),

based on their radii. If the sum of their radii is less than the distance between the sphere centers,

then the spheres do not intersect (and there is no possible combinations of variables that satisfy the

structure’s connectivity and length constraints). If the sum of their radii is equal to the distance

between the sphere centers, the intersection is a point. If the sum of the radii is greater than the

distance between the centers, then the intersection is a circle. This circle can be parameterized, such

that a point on the circle is identified by a single angle (λ).

In this sense, the location of node k can be represented by only one variable while

incorporating length constraints of other members. This is an improvement over the direct use of

Cartesian coordinates with three variables, and even spherical coordinate with two variables. Using

these types of intersections to relate connected members of fixed lengths is often extremely useful

in more complex tensegrity geometries.

99
 4.2 Path Planning

Deployment is critical for large scale structures where a launch vehicle fairing cannot

accommodate the full-size structure. In these cases, ensuring the structure can change shape from

a stowed to a deployed state is critical to meeting mission requirements. Identifying start and end

target shapes and using known equilibrated shapes to form a path between them is a critical step in

the design of deployable, shape changing structures for space applications. In most applications

(e.g., deployable booms or reflectors), a target shape is identified as the final, deployed state of a

structure. The structure’s ability to achieve the prescribed shape is typically a driving design feature

of the structure itself, especially in cases like that of a reflector, where surface accuracy is critical

to mission success. Additionally, ensuring the path moves through equilibrated shapes helps ensure

robustness to unpredicted external disturbance and to maintain control of the structure throughout

deployment.

The details of target shapes depend on the specifics of the application for which the

structure is being designed. In some cases, it is possible to described target shapes in terms of the

same variables used to characterize the shape itself. For example, if the stowed state of a triplex

tensegrity is defined as the smallest height of the triplex, maintaining the geometry of the structure

otherwise, then the z-values of the top-face nodes can be evenly minimized directly to find the

smallest overall height to use as the stowed shape (accommodating any additional considerations

as necessary).

In other cases, the desired shape cannot necessarily be directly translated into the

structure’s nodal coordinates (or their other characteristic design variables). For example, Chapter

5 describes a process for fitting the top face of a triplex module like that shown in Figure 4.7 (also

Figure 3.11) to a paraboloid surface. In this application, where the surface would be used to support

a reflector, the top-face nodes should be selected to maximize surface accuracy. With the inclusion

100
of certain member length constraints, not every node can necessarily be directly on the paraboloid

surface. In these cases, an optimization method is used, seeking the best possible fit considering

structural constraints in terms of the selected design variables.

Figure 4.7: A triplex module.

4.2.1 Optimal Shape Fit to a Prescribed Surface

The specifics of the optimization approach for matching a target shape depend on the

particular application. In general, the target shape is first described in terms of nodal coordinates.

This could include all nodes, a sub-set of nodes, or some other parameter which can be translated

to nodal coordinates (such as triplex height interpreted as a uniform z-value of top-face nodes).

Once the target shape is identified in terms of nodes, a fitness function can be defined in terms of

the design variables chosen to characterize shapes, where the fitness function is minimized to

identify the configuration having the smallest possible deviation from the target shape.

Member length constraints can sometimes be incorporated into the selection of design

variables, and a penalty should be added to the fitness function to addresses cases that violate

member length constraints. Equilibrium can also be treated as a constraint, where violations of the

structure’s equilibrium add an additional penalty to the fitness function. Other factors can be added

101
to the fitness function as necessary to influence the final best-fit shapes, but such factors are

particular to the specific application in question.

Upper and lower bounds are determined based on the target shape and how the design

variables relate to the target shapes. For example, if the goal is to seek a stowed state of a triplex

by minimizing height without changing the x-y values of node locations, then the design variables

corresponding to the z-values of the top face nodes can be identified as vz. These variables would

then be given a smaller lower bound, such that vz can encompass the smallest possible stowed state,

as defined by a small triplex height. To minimize the change in the x-y values of nodes, the variables

corresponding to x-y nodal coordinates could be given upper and lower bounds as small deviations

from their initial values (if any deviation is allowed at all), thus encouraging the likelihood of

finding a viable stowed state with a small height and maintaining similar x-y values of nodes.

In this thesis, Particle Swarm Optimization (PSO) is used when an optimization approach

is needed, chosen for its use in tensegrity structure optimization in the literature [43, 44], including

in form-finding approaches directly [45, 46].

4.2.2 Using Target Start and End Shapes to Path Plan

Once target shapes are identified, a path-planning approach is implemented to select

equilibrated shapes between the two target end states. This may involve a stowed-to-deployed

process for deployable structures, or may be a change between two different deployed shapes to

enact shape change post-deployment. In either case, a start and end shape are identified as the start

and end of the path.

To identify the details of the path itself, the design variables in the start and end shapes are

used to define a parametrized line, shown in Eq. 4.10:

102
 v1,𝑐 v1,𝑠𝑡𝑎𝑟𝑡 v1,𝑒𝑛𝑑 − v1,𝑠𝑡𝑎𝑟𝑡
{ ⋮ }={ ⋮ }+{ ⋮ }𝑡 (Eq. 4.10)
v𝑁𝑣 ,𝑐 v𝑁𝑣 ,𝑠𝑡𝑎𝑟𝑡 v𝑁𝑣 ,𝑒𝑛𝑑 − 𝑣𝑁𝑣 ,𝑠𝑡𝑎𝑟𝑡

Where v is the design variable, the subscript c refers to the current shape in the path, Nv is the

number of design variables, the subscripts start and end refer to the start and end shapes

respectively, and t varies from 0 to 1 along the parameterized path from the start shape (t = 0) to

the end shape (t = 1).

Using Eq. 4.10, the path can be broken up into any number of desired smaller steps, using

the corresponding t values for each step to calculate the variables and thus the shape. Each

intermediate shape must be checked for equilibrium and satisfaction of the constraints. While the

design variables are generally selected to encourage shapes that satisfy the constraints, not all

combinations of design variables guarantee satisfaction of constraints. To accommodate this, a PSO

subroutine is implemented whenever equilibrium and/or length constraints are not immediately

satisfied in a particular step. This optimization subroutine is set up in a very similar way to that of

the PSO used to find the best shape of the tensegrity to achieve the target shape. In this case, the

fitness function is defined as the sum of the absolute value of the difference between the design

variables. Penalties are added for sets of design variables which do not satisfy equilibrium and/or

member length constraints. The upper and lower bounds of the variables are determined as

deviations from the design variables on the current step in the path which did not meet the

constraints (vc in Eq. 4.10). This approach acts to find an equilibrated shape as close as possible to

the initial shape (at that step) on the target path that was found not to be in equilibrium.

Additionally, movement between adjacent steps in the path are checked to ensure no

members are required to pass through or collide with one another. The methods detailed in [15] are

used to check for member interference. Members are assumed to move linearly between adjacent

steps in the path. This linear assumption allows the members to be represented as parametric lines.

The nearest distance between members as they move between adjacent steps can then be calculated.

103
If the nearest distance is less than or equal to the sum of the members’ radii, then member

interference has occurred, resulting in an invalid path.

The start and end shapes throughout this dissertation are chosen not only to fit the desired

parameters of the intended application, but also to decrease the likelihood of these types of paths

resulting in non-equilibrated shapes, or paths requiring interference. This was generally successful,

as no paths were found that required member interference. Regardless, member interference is

checked in all cases, and the same PSO sub-routine used to implement a local search for nearby

equilibrated shapes could also be used to seek similar nearby shapes not requiring member

interference, had such occurred in the designated path.

4.2.3 Experimental Verification of Path Planning

As a demonstration of the variety of shapes a tensegrity may achieve, and the utility of

using these numerical datasets to aid in path planning, a triplex tensegrity with variable geometry

was built. The triplex geometry was specified to be like that of one of the triplexes in the module

of Figure 4.7, where the top face of the triplex is inscribed in the bottom face of the triplex in a top-

down view. Both the strut lengths and the lengths of the supporting cables are controllable.

Figure 4.8 shows the experimental set up. Firgelli bullet series 36 cal. linear actuators are

used as the struts. These actuators have a 6” stroke length, and an eye-to-eye length of 16.34” when

fully retracted. The cables are made of 0.88 mm diameter Ultra High Molecular Weight

Polyethylene (UHMWPE) braided cord. The top and bottom face cables of the triplex are tied to

the linear actuator eye holes directly. The bottom face cables are tied to a length of 11.25” each,

and the top face cables are tied to a length of 6.5” each. The supporting cables are attached to

carabiners at the top face nodes, then threaded through the bottom face eye holes, where they are

re-directed through eye hooks to custom spools attached to Nema-14 stepper motors with 19:1

104
planetary gearboxes. These motors are mounted on the fixed test bed. Additionally, the bottom face

nodes are tied to eye hooks fixed to the test bed, to ensure that the bottom face nodes do not move

during the experiment.

Figure 4.8: Experimental set-up.

For this experimental demonstration, the top face of the triplex is fit to a paraboloid surface,

as if the triplex were part of a larger module from Figure 4.7. This module would be centered with

the same x-y coordinates as the center of the paraboloid, and the triplex is selected as the one from

the module with its first bottom face cable aligned on the positive x-axis, with one node at the

origin. The paraboloid surface to which it is fit has a focal length of f=10, and a z-shift of zp=17”.

This is selected as a case of dramatic curvature to ensure visible changes in orientation of the top

face. The best fit shape to the prescribed paraboloid surface is found using the PSO approach. The

paraboloid surface is defined in Eq. 4.11:

2 2
(𝑥 + 𝑥𝑝 ) + (𝑦 + 𝑦𝑝 )
𝑧= + 𝑧𝑝 (𝐸𝑞. 4.11)
4𝑓

105
where (xp, yp, zp) is the center of the paraboloid and f is the focal length.

To fit the top nodes of the triplex to the paraboloid surface, the average difference in z-

value, Δz, is defined in Eq. 4.12, calculated as the magnitude of the difference between the z value

on the paraboloid surface (Eq. 4.11) for the x-y values of the nodes, and the actual z value of the

nodes:
6 2 2
(𝑥ⅈ + 𝑥𝑝 ) + (𝑦ⅈ + 𝑦𝑝 )
∑ | + 𝑧𝑝 − 𝑧ⅈ |
4𝑓
ⅈ=4
𝛥𝑧 = (𝐸𝑞. 4.12)
3

The smaller the value of Δz, the better the fit of the top nodes to the prescribed surface.

Additionally, the average distance of the nodes in x-y from their original position is added to the

fitness function, pre-multiplied by a weight, wxy. This is included to encourage finding top face

nodes which are in similar x-y positions, since this would make node-to-node connections in the

final module easier to implement directly, and decreases the likelihood of finding shapes which

would flip the orientation of the top-face triangle relative to its starting flat shape—which is desired

to prevent member interference when path planning is later implemented. Thus, the total fitness

function is defined by Eq. 4.13, using Δz from Eq. 4.12:

2 2
∑6ⅈ=4 √(𝑥ⅈ − 𝑥ⅈ,𝑓𝑙𝑎𝑡 ) + (𝑦ⅈ − 𝑦ⅈ,𝑓𝑙𝑎𝑡 )
𝑓𝑖𝑡𝑛ⅇ𝑠𝑠 = 𝛥𝑧 + 𝑤𝑥𝑦 (𝐸𝑞. 4.13)
3

This triplex configuration has 6 nodes, and 6 length constraints (all face cables, fixed in

length). Additionally, the bottom face is assumed to have member 1 on the x-axis, and member 2

(and therefore also member 3) on the x-y plane. This results in 6 degrees of freedom, and therefore

6 variables are required to define the structure’s shape. The variables are identified as the nodal

coordinates of node 4, the spherical azimuth and elevation of node 5 relative to node 4 (𝜃5 and 𝜙5 ),

and the angle on a parameterized circle from the intersection of spheres defined by nodes 4 and 5

106
(λ6). Upper and lower bounds are set on the variables as +/- 2 for the x-y coordinate of node 4, +/-

6 for the z coordinates of node 4, and +/- 45o for the angles.

For a particular set of variables, the resulting shape can be converted back to nodal

coordinates. Then, the nodal coordinates are used to evaluate the fitness function defined by Eq.

4.13. Additionally, the resulting shape is checked for equilibrium. A penalty is added to the fitness

function if the shape is not equilibrated.

The strut lengths are calculated from the resulting nodal coordinates and the structure’s

connectivity. The linear actuators used as struts have a maximum stroke length of 6”. An additional

penalty is added to the fitness function if the resulting strut length is shorter than the fully retracted

length, or longer than the fully extended length. Note, however, that finding violations of strut

lengths is unlikely given the restricted range of design variables.

The results of this optimization approach are shown in Figure 4.9, along with the

paraboloid surface to which the top face was fit. A weight factor of wxy=0.75 was used to limit the

x-y motion of top face nodes. The final shape resulted in an overall fitness value of 0.0795 in. This

fitness value is the sum of the average distance of the top-face nodes from their projected location

of the paraboloid, as well as the average distance of the nodes from their original position (with the

second term multiplied by the weight factor). In this case, the fitness to the paraboloid surface alone

was a perfect fit (e.g., fitness=0), indicating a shape was found where the top-face nodes were

directly on the paraboloid surface.

107
 Figure 4.9: Fit of triplex based on experimental structure dimensions to a paraboloid surface.

The shape found using this optimization approach is then used as the target end shape for

path planning, with the start shape defined as the flat shape when the struts are fully retracted. The

path is then defined using a parametric equation, as shown in Eq. 4.10. The number of steps for t is

set to 35, chosen such that the smallest length change in any given member is 0.1”. (0.1” is selected

since it is the smallest increment that can be commanded of the struts.) Each step along the path is

checked for equilibrium. If the shape specified by the path is not in equilibrium, then a localized

PSO search for nearby equilibrated shapes is implemented as described in Section 4.2.

This approach was used successfully to determine a sequence of equilibrated shapes,

resulting in a path between the start and end shapes. Figure 4.10 shows the resulting changes in

length required of each supporting cable and strut to achieve the current step’s shape.

The experiment was recorded as the triplex moved through the prescribed path. The

supporting cable lengths were measured using a tape measure (with associated potential error) at

the end of the maneuver. Table 4.1 shows the measured lengths compared to the lengths found from

optimization. The measured cable lengths show good agreement with the expected member lengths.

108
 Additionally, Figure 4.11 shows snapshots of the experimental triplex juxtaposed with the

modeled triplex shape. The exact camera angle was not measured, so the angles in the simulated

and actual photos were aligned by eye. That said, the photos show good agreement in movement

throughout the shape change path.

Figure 4.10: Change in lengths between adjacent steps in the path tested.

Member Length from PSO [in] Measured Length [in]

Supporting Cable 1 17.801 17.75

Supporting Cable 2 19.737 19.75

Supporting Cable 3 18.784 18.75

Table 4.1: Comparison of supporting cable lengths found from the best-fit to a paraboloid surface using PSO, and the
corresponding measured lengths in the experimental test-bed after performing the maneuver from stowed to the target
shape.

109
Figure 4.11: Snapshots of experiment next to corresponding steps from simulation

110
 4.3 Finite Element Methods for Stiffness and Modal Analysis

In the design of any structure, understanding its behavior in the presence of external loads

is extremely important to ensure the success of the mission. Significant deformation in the presence

of external loads not only puts additional strain on the materials used in the structure, but can also

inhibit the intended use of the structure itself, particularly where shape maintenance is important,

as it is for a reflector. Additionally, understanding vibrational behavior is important to avoid

resonance (and therefore potential structural failure) and minimize dynamic response during

operation.

In this thesis, a non-linear Finite Element Model (FEM) developed by Yildiz [4] is used.

A non-linear approach is particularly important for tensegrity structures, since the inclusion of pre-

stress in the structure is critical to its ability to maintain equilibrium and carry external loads.

Additionally, non-linear approaches such as that used in [4] offer greater accuracy in cases of large

deformations; when the deformation is large, the elements in the structure change length, which in

turn affects the stiffness of the structure. In the approach in [4], the load applied to the nodes is

incrementally increased until it reaches the final applied load. In each step of incremental loading,

the resulting nodal coordinates are calculated, using corresponding member lengths to update the

stiffness matrix. In addition, if any cable member has a length at any step less than its original zero-

load length, the cable has gone slack; to account for this, the pre-stress force and the material

stiffness in that cable are set to 0 for any load increments where this remains true.

The methods of this non-linear FEM approach are described in detail in [4]. In this

approach, each member is defined as a single element. To calculate deformation under external

loads, Eq. 4.14 is used:

𝑲 𝑇 ∆𝑼 = 𝒇 − 𝑸 (𝐸𝑞. 4.14)

111
where KT is the tangent stiffness matrix, comprised of the sum of the linear (KL) and the non-linear

(KNL) stiffness matrices (Eq. 4.15 and 4.16 respectively), ∆𝑼 is the displacement vector, f is the

external load vector and Q is the internal load vector from the force densities. The internal loads in

each member are found by multiplying the force density (q) in each member by the member’s

length (l). If a pre-stress factor (pf) is used, this is also included in the product, such that a single

element’s internal load Q is found as Q=pf ql. The matrices and vectors of Eq. 4.14 are global,

obtained by transforming and summing all the local matrices and vectors for each element.

The linear stiffness matrix comes from the axial stiffness of each element as dependent

upon their material properties, and is shown in the local element frame as:

1 0 0 −1 0 0
0 0 0 0 0 0
𝐸𝐴 0 0 0 0 0 0
𝑲𝐿 = (𝐸𝑞. 4.15)
𝑙 −1 0 0 1 0 0
0 0 0 0 0 0
[0 0 0 0 0 0]

where E is the modulus elasticity, A is the cross-sectional area of the member, and l is the nominal

length of the member.

The non-linear stiffness matrix in the local element frame is found as:

1 0 0 −1 0 0
0 1 0 0 −1 0
0 0 1 0 0 −1 (𝐸𝑞. 4.16)
𝑲𝑁𝐿 =𝑞
−1 0 0 1 0 0
0 −1 0 0 1 0
[0 0 −1 0 0 1]

where q is the element’s force density. Each of these matrices can be transformed to the global

coordinate system, and then summed to find the tangent stiffness matrix, KT, used in Eq. 4.14. The

transformation is applied by pre-multiplying local matrices (and local vectors, when necessary) by

the following transformation matrix T:

112
 ∆𝑥 ∆𝑦 ∆𝑧
0 0 0
𝑙 𝑙 𝑙
0 0 0 0 0 0
𝑻= 0 0 0 0 0 0 (𝐸𝑞. 4.17)
∆𝑥 ∆𝑦 ∆𝑧
0 0 0
𝑙 𝑙 𝑙
0 0 0 0 0 0
[0 0 0 0 0 0]

This transformation matrix T is particular to the element in question. In Eq. 4.17, ∆𝑥, ∆𝑦 and ∆𝑧

are the difference in the two nodal coordinates of the member in question, and l is the element’s

length. Each non-zero element of the transformation, therefore, is a direction cosine of the member.

To update member lengths throughout the incremental application of loads, Hooke’s law

is first used to calculate the rest length (l0) of each element. Hooke’s law is applied to each element

as shown in Eq. 4.17:

𝐸𝐴
𝑝𝑓 𝑞𝑙 = (𝑙 − 𝑙0 ) (𝐸𝑞. 4.18)
𝑙

Which can be re-arranged to solve for the rest length as:

𝐸𝐴𝑙
𝑙0 = (𝐸𝑞. 4.19)
𝐸𝐴 + 𝑝𝑓 𝑞𝑙

In Eq. 4.19, to calculate the rest length of each element, the initial length l of each member is

calculated based on the nodal coordinate locations. In application of each incremental loading, Eq.

4.19 is re-arranged based on the initial calculated value of l0 and the current lengths l of each

member under the loading conditions to solve for that increment’s corresponding force density in

each member (q). This q is used in Eq. 4.16 to update the stiffness matrix.

In addition to analyzing deformations under external loads, the non-linear FEM developed

by Yildiz [4] is used to analyze the natural frequencies of the structure. The natural frequencies are

found by considering the application of harmonic motion in the form x=Xeiωt. The eigenvalue

problem is then written as:

𝑲 𝑇 𝑋 = 𝜔2 𝑴𝑋 (𝐸𝑞. 4.20)

113
where X is the amplitude, ω is the natural frequency, and M is the mass matrix. The same stiffness

matrices in Eq. 4.15 and Eq. 4.16 are used to form the global KT in the same way as used in Eq.

4.14. Additionally, the mass matrix (M) in each element can be found as:

2 0 0 1 0 0
0 2 0 0 1 0
𝜌𝐴𝑙 0 0 2 0 0 1
̅ =
𝑴 (𝐸𝑞. 4.21)
6 1 0 0 2 0 0
0 1 0 0 2 0
[0 0 1 0 0 2]

where 𝜌 is the element’s material density, and A and l are again the cross-sectional area and length

̅ can be transformed and assembled into a

of the element respectively. The local mass matrix 𝑴

global mass matrix M in the same way as the stiffness matrices. With this, eigenvalue analysis of

Eq. 4.20 can be used to find the natural frequencies of the tensegrity structure of interest.

This analysis, along with analysis of the deformation of the structure under external loads,

allows for a deeper understanding of the structure’s behavior in the working environment, enabling

better capability of structural design to ensure mission success. Using the methods detailed in this

Chapter, the next Chapter takes a detailed look at an example case. The geometric principles from

section 4.1 are used to select design variables of a complex tensegrity structure incorporating

numerous length constraints. Additionally, the FEM approaches described in this section are used

to evaluate how varying different structural parameters influences the structure’s overall

deformation and first mode frequency. Finally, the approach in sections 4.2 and 4.3 are used to

identify a target shape, and employ a path between designated shapes through equilibrium.

114
 Chapter 5

Example Design Problem

To demonstrate the efficacy of the methods developed in this thesis, a representative design

problem is considered, namely a tensegrity structure for use as the backbone of a parabolic reflector.

The tensegrity developed could be used either for a mesh reflector, or a solid reflector. The

tensegrity is evaluated for its ability to maintain a shape as close to the prescribed paraboloid

surface as possible.

The James Webb Space Telescope (JWST) is an excellent demonstration of a large

reflector recently launched in space. As such, the general dimensions of the JWST are used as a

point of reference for the geometric scaling of the tensegrity structure. The JWST is designed for a

6.5 m diameter reflector, comprised of 18 hexagonal pieces (with a 19th hexagon in the center that

does not hold a mirror), each of which has a diameter of 1.32 m [61,62]. While the JWST is not

designed to have a paraboloid surface, the dimensions of the JWST can be used to approximate a

paraboloid surface of similar scale.

Figure 5.1 details the geometry of the JWST. In addition to the diameter of the main

mirrors, the arms holding the secondary reflecting mirror are approximately 7.62 m long [63].

These arms are built in a tripod formation, but are approximated as being exactly the diameter apart

and aligned for these purposes, as pictured in Figure 5.1. Assuming that secondary mirror (which

has a diameter of 0.74 m) is placed such that the reflected rays would have been focused to cover

its full area, the focal point of the mirror can be approximated as the location at which lines running

along the arms would cross. With this information, the focal length is approximated as 7.6 m. This

focal length is used as a demonstrative case; the approaches developed in this research could be

applied to any desired focal length for fitting to a paraboloid surface.

115
Figure 5.1: Dimensions from the James Webb Space Telescope used to approximate the focal length if it were a perfect
paraboloid surface.

With these general dimensions in mind, the triplex module (pictured again in Figure 5.2)

is used to study how a tensegrity could be used for such an application. The module algins well

with the JWST example, as the bottom face of the module forms a hexagon. Each module has a

side-to-side length of 1.32 m, to align with the size of the primary mirror segments of the JWST.

This results in bottom face cable lengths of 0.762 m and top face cable lengths of 0.44 m. To cover

the same area as the JWST primary mirror, 19 triplex modules can be arranged just as the mirror

hexagons of the JWST, as shown in Figure 5.3.

116
 Figure 5.2: A triplex module.

Figure 5.3: Array made of 19 modules, with each module shaded in different colors for clarity.

117
 The design parameters to be determined are the bay height (the height from bottom to top

face of the module when in the flat shape), the member radii (cable and strut, where struts may be

hollow with wall thickness to be determined), and the pre-stress coefficient. The bay height is tested

for values between 0.7 m and 3.35 m, where 0.7 m is selected as twice the depth of the primary

mirror paraboloid (to provide room within the module to change shape for fitting to the curvature),

and 3.35 m is the approximate depth of the JWST support structure [64]. The cable radii are varied

between 0.25 and 15 mm, the strut outer radii are varied between 10 and 30 mm, and inner radii

values of 0, 5, 10, 15, 20 and 25 mm are tested for each value of outer strut radius (barring

combinations where inner strut radius would be greater than or equal to outer strut radius).

The pre-stress coefficient corresponds to pre-multiplying all pre-stresses (force densities)

in the structure by a constant value, thus effectively tensioning or loosening the entire structure.

Each of these parameters is studied for its effects on the structure’s mass, deployment

efficiency, first natural frequency, and the structure’s ability to hold shape in the presence of

station-keeping type loads. These design parameters are first studied in the single triplex module,

and then applied to the 19-module array. The analyses conducted in this chapter are only applied

to the flat shape, where the top and bottom faces are parallel, with the exception of evaluating

fitness to a prescribed surface directly, as well as the associated path-planning to reach this shape

from the flat shape.

For calculations requiring pre-stress, the force-finding approach developed in Section 3.3

is used, with a modification to group like members with the same force density. The modification

to encourage member grouping is detailed in Appendix B. The force densities are detailed in Table

5.1. The values in this table are all multiplied by the same pre-stress factor pf when considering the

effects of pre-stress.

118
 Member grouping Pre-stress [N/m]

Bottom face cables, shared 1.000

Bottom face cables, not shared 0.595

Top face cables 1.595

Supporting cables, shared 3.190

Supporting cables, not shared 1.595

Struts 1.595
Table 5.1:pre-stresses used for analysis of the triplex module in the flat shape with member grouping. Shared or not
shared refers to members shared between the same nodes of adjacent triplexes within the module.

The materials selected are Unidirectional Mitsubishi K13C2U UHN /epoxy (60% fiber

volume fraction) for the struts (E=536 GPa, ρ=1,840 kg/m3) and Kevlar 49 (E= 124 GPa, ρ= 1,440

kg/m3) for the cables, chosen for their high modulus of elasticities and low densities [4, 65, 66].

The tensile strength of the cables is σy= 3,600 MPa. Euler buckling (Pcr) for the struts is calculated,

defined in Eq. 5.1:

4
𝜋 3 𝐸(𝑟𝑜𝑢𝑡
4
− 𝑟ⅈ𝑛 )
𝑃𝑐𝑟 = (𝐸𝑞. 5.1)
4𝑙

The value of the critical buckling load (Pcr) is used to calculate the pre-stress factor corresponding

to an axial force in the struts exactly equal to the critical load based on the force densities and

lengths of the struts. The pre-stress factor in cables corresponding to loads equal to the tensile forces

corresponding to σy is also calculated. The two resulting pre-stress factor values (one for struts and

one for cables) are used to inform the selection of the maximum pre-stress factor in an unloaded

tensegrity.

The critical buckling load is dependent upon the outer and inner radii of the struts. Since

the buckling load is decreased (and thus more limiting) for a smaller outer radius, the strut limiting

load is calculated as the most extreme case tested, rout=10 mm and rin=5 mm. Additionally, the

critical buckling load is decreased for larger values of l. The longest strut length tested

119
(corresponding to the highest bay height) is l=3.42 m when the bay height is equal to the selected

1 m. This results in a critical buckling load of 11,391 N. Cases with shorter struts and larger cross-

sectional areas will have larger values of Pcr, with the upper-bound of tested cases being a solid

strut of radius router=30 mm, and of length 0.97 m (for a height of 0.7 m), resulting in a maximum

Pcr=3,468,000 N.

The limiting case for the cables is based on the tensile strength (3,600 MPa) and the cross-

sectional area of the cables. The cables have solid circular cross sections, and thus the limiting force

in the cables is smallest (and thus most restrictive) when the area is smallest. This comes from the

lower bound of tested cases, when cable radius is 0.25 mm, where the corresponding force in the

cable is then 706.9 N. The maximum limiting force in cables comes from the largest area, when the

radius is 15 mm, and the resulting limiting cable force would be 2,544,700 N.

The force in any given member is related to the pre-stress factor by Eq. 5.2:

𝐹ⅈ
𝑝𝑓 = (𝐸𝑞. 5.2)
𝑞ⅈ 𝑙ⅈ

Where pf is the pre-stress factor, Fi is the force in member i, qi is the force density in the member,

and l is the length of the member.

As an initial look at the effects of the design parameters, a middle-range value of force

based on the results of the critical buckling load in struts and the limiting load in cables is used to

calculate the maximum pre-stress factor. For an internal member force of Fi=200,000 N, the

maximum acceptable pre-stress value (based on the force densities and the lengths in the tallest

tested height) is 18,335. Rounding up, a pre-stress factor is varied from 1 to 20,000 to study the

effects on the structural properties.

Table 5.2 summaries the bounds set on the variables evaluated for their effects on the

overall structures mass, packaging efficiency, deflection under loading in z, and first natural

120
frequency. These bounds are used both for the single module, as well as for the larger array

comprised of 19 modules.

Design Variable Lower bound Upper bound

Pre-stress factor 1 20,000

Bay height [m] 0.7 3.35

Cable radius [mm] 0.25 15

Strut outer radius [mm] 10 30

Strut inner radius [mm] 0 25

Table 5.2: Design variable parameters upper and lower bounds to evaluate structural properties of the triplex module
and an array comprised of 19 modules.

When testing the effects of the different design variables, the other design variables are

held at their nominal values. These nominal values are 10,000 for the pre-stress factor, 2.025 m for

the bay height, 7.625 mm for the cable radius, 20 mm for the outer strut radius and 10 mm for the

inner strut radius; all selected to be in the middle of the range of values considered.

The calculation of the deformation and first natural frequency are detailed in Section 4.3.

In the case of deformation, loads are applied based on that magnitude of the forces the structure

may experience during station-keeping maneuvers. In general, station-keeping maneuvers are

typically characterized by the required Δv. The resulting acceleration felt by the spacecraft can then

be found as the intended Δv divided by the duration (t) of the maneuver (thruster firing). The

resulting thrust force (FT) on the spacecraft can then be found as FT=MTa, where MT is the total

spacecraft mass and a= Δv/t. The exact Δv required for the maneuver is highly specific to the

particulars of the mission, orbit and spacecraft, and the duration of the maneuver is selected

accordingly. Detailed consideration of the orbit, thrusters and over-all spacecraft specifications are

outside the scope of this research. Overall, station keeping maneuvers typically require fairly low

Δv.

121
 For example, the JWST orbits at the second Lagrange point (L2), about the sun (L2 being

chosen for its alignment with the Earth as it orbits). Petersen [67] discusses station-keeping

maneuver strategies for the JWST, seeking minimal Δv maneuvers. The required Δv varies

throughout [67] depending on the particular type of maneuver, but ranges from approximately 0.01-

2 m/s. The larger values of Δv come from non-optimal conditions, and suggestions for minimizing

Δv requirements are the focus of [67].

For a general consideration of deflection behavior, the larger value of Δv=2 m/s is used to

calculate the external load. The mass at launch of the JWST is approximately 6,200 kg [62]. The

burn duration for station-keeping maneuvers is dependent upon the particulars of the spacecraft

mass, thrusters and orbit. In the case of the JWST, burns around 1 minute long every 21 days are

required for station-keeping [67, 68]. This 1-minute maneuver duration is used to calculate the

acceleration as 0.033 m/s and a corresponding thrust force of 206.7 N with the JWST mass. The

mass of the JWST includes far more than just the backbone of the structure; the mass of the

backbone alone is approximately 970 kg [64]. To provide some consideration to the mass of the

structure (and therefore more reasonable force values to what the structure might experience during

maneuvers), an additional 275.3 kg (from dividing the 5,230 kg of additional mass by 19 modules)

is added to the tensegrity structures mass for an individual module when calculating the thrust. The

acceleration of 0.033 m/s from Δv = 2 m/s and t = 60 seconds is held constant. The mass of the

structure is updated based on the parameters as shown in Eq. 5.2:

18 42
2 2 2
𝑀𝑡 = ∑ 𝜌𝑠 𝜋(𝑟0,ⅈ𝑠 − 𝑟ⅈ,ⅈ𝑠 )𝑙ⅈ𝑠 + ∑ 𝜌𝑐 𝜋𝑟𝑐,ⅈ𝑐 𝑙ⅈ𝑐 (𝐸𝑞. 5.2)
ⅈ𝑠=1 ⅈ𝑐=1

Where ρs and ρc are the strut and cable material densities respectively, router is the outer

radius of the struts, rinner is the inner radius of the strut, rc is the cable radius, l is the member length,

and subscripts is and ic refer to strut members (18 in total) and cable members (42 in total)

respectively. The resulting thrust force is calculated accordingly as the mass of the tensegrity

122
structure (Mt) summed with 275.3 kg allotted to spacecraft components (total mass MT=Mt+275.3).

The deformations calculated apply the resulting thrust force uniformly over the top face nodes in

the negative-z direction, simulating a thrust aligned with the central axis of the tensegrity structure.

Packaging efficiency is also calculated. The packing method is designed assuming that a

module would be launched fully assembled, and that additional actuation for deployment should

be avoided. The stowed state is thus taken as the equilibrated flat shape shown in Figure 5.2, with

the shortest possible bay height (calculated as the shortest bay height when strut members first

touch as the height is decreased). The minimal bay height is calculated by decreasing bay height

from its initial height until the minimum distance between struts is equal to the sum of the radii of

two struts, indicating the members have just started touching one another. Decreasing the height

any lower would require attempting strut members to pass through one another, causing undesirable

stresses (or even failure) in the structure. From the geometry of the structure, since only the z-value

of the top face nodes are changing, the first member collision will always occur between the struts,

and it will occur at the same instant for all triplexes in the module thanks to symmetry. Thus,

packaging efficiency, for this method of stowing, is only affected by the outer radius of the strut.

The stowed volume is calculated as the smallest cylinder in which all nodes of the module

fit. Figure 5.4 shows an example stowed module adjacent to an example deployed module. The

radius of the cylinder circumscribing the outer nodes of the module is the same in the stowed and

the deployed states (the circumscribing radius of the bottom-face nodes). The only variable in

stowed volume subject to change, therefore, is the stowed height. The stowed volume is simply the

volume of the cylinder (Vstow=πr2stowhstow), and thus packaging efficiency is characterized by the

stowed height of the module, to be minimized for better packaging efficiency.

123
 Figure 5.4: A stowed module (left) shown adjacent to a deployed module (right).

5.1 Design Analysis of a Triplex Module

A single triplex module is comprised of 6 sub-triplexes, with a total of 19 nodes

and 60 members, 18 of which are struts. For the single module, all 7 of the bottom face nodes are

considered fixed in space for this analysis. The pre-stress factor, bay height, cable radius, and strut

inner and outer radius are all evaluated for their effects on the module’s first natural frequency,

deflection under loading in z, mass, and packaging efficiency. For deformation and modal analysis,

all bottom faced nodes are fixed in x-y-z as the boundary conditions. The deformation is

characterized by the average magnitude of the deviation of the nodes from their underformed

position in z. While some deformation may occur in the x and y directions as well, these

deformations are negligible when compared to the z deformations, since the loading is uniformly

applied in the negative z direction.

124
 The effects of the design parameters on the mode shapes themselves are not studied in

detail for every case of design variables tested. The mode shapes are calculated alongside the

natural frequencies. The first 6 mode shapes of the module with the nominal design-variable values

are shown in Figure 5.5, showing only the top face members for visual clarity, with the original

members in dashed black lines, and the mode shape of members shown in solid blue.

Figure 5.5: First 6 mode shapes of the triplex module, showing only top-face members. Mode shapes are in solid blue
lines, with original members in dashed black lines.

125
 For analysis of the structure’s deflection under external loads, the same loading conditions

are used for all sets of tested design parameters. With the bottom nodes fixed in space, all top-face

nodes are loaded uniformly in the negative-z direction with a load given by Eq. 5.3:

(𝑀𝑡 + 275.3) ∗ (0.033)

𝐹𝑇 = (𝐸𝑞. 5.3)
𝑁𝑡𝑜𝑝

where FT is the force at each node in Newtons, Mt is the tensegrity mass calculated from Eq. 5.2 in

kg, 275.3 is additional mass allotted to the total spacecraft in kg, 0.033 is the acceleration in m/s,

and Ntop is the number of nodes on the top-face of the module (Ntop=12 for the single module).

Deformation of the structure is evaluated as the average deformation of all free nodes in z.

5.1.4 Effects of Pre-stress Factor

The pre-stress factor for this module is varied between 1 and 20,000. The pre-stress factor

determines the magnitude of force in the members in the absence of external loads, and therefore

affects the structures’ deformation behavior and first natural frequency, but does not affect the

structures’ mass nor its stowed height. With the nominal set of design variables, the structures mass

is m=352.8 kg (including the 275.3 kg allotted to extra structural components) and a thrust force of

FT=0.97 N at each node (11.64 N total).

Figure 5.6 shows the effects of the pre-stress factor on the deflection and the first natural

frequency. In general, increasing the pre-stress factor is found to decrease deflection and increase

the first natural frequency. Thus it is desirable in both regards to design this tensegrity with higher

pre-stress values for lower deflections and higher first mode frequencies.

126
 Figure 5.6: Effects of pre-stress factor on deformation and first natural frequency.

5.1.2 Effects of Bay-Height

The bay height of the structure is varied between 0.7 and 3.35 m. Changing the bay height

of the structure changes the angles and lengths of the supporting cables and struts, and thus affects

the deformation, first natural frequency and mass of the structure. The changing mass also affects

the resulting force applied at each of the nodes. The mass (MT) and resulting thrust force at each

node FT as the bay height is varied is shown in Figure 5.7. As expected, increasing bay height

increases the structure’s mass and, proportionally, the resulting force on the nodes.

Figure 5.7: Effect of bay height on the structure’s mass (MT) and force applied at each node (FT).

127
 The results of changing the bay height of the structure on deformation and the first natural

frequency are shown in Figure 5.8. In general, increasing the bay height is found to increase the

overall structural deformation, and decrease the first natural frequency. The bay height of the

structure is found to have particularly significant effect on the first natural frequency. The natural

frequency is in fact reduced from 364 Hz to 70.8 Hz between the lower and upper bounds of bay

height respectively. Therefore, both deflection and first mode results indicate shorter bay heights

to be preferable, with the effects on the first mode being particularly significant.

Figure 5.8: Effects of bay height on deformation and first mode.

5.1.3 Effects of Cable Radius

The triplex module contains 42 cable members, each of which is modeled as having the

same material and the same radius. The radius of the cables is varied between 0.25 mm and 15 mm.

This changes the cross-sectional area of the cables, thus increasing structure’s mass and

corresponding force on each node. The effects of cable radius on mass and force are shown in

Figure 5.9.

128
 Figure 5.9:Effects of cable radius on mass (MT) and resulting thrust force (FT) on each node.

Figure 5.10 shows the effects of cable radius on the average deformation and the first mode.

In general, increasing the cable radius decreases the structure’s deformation and increases the first

natural frequency. This is to be expected, as the additional material increases the structure’s ability

to resist deformation. The effects of increasing cable radius become less significant for larger cable

radius values with respect to the average deformation.

Figure 5.10: Effects of cable radius on deformation and first mode.

129
5.1.4 Effects of Strut Radius

The effects of changing the inner and outer strut radii are considered simultaneously, as

both affect the cross-sectional area of the struts directly. The outer strut radius is varied between

10 and 30 mm, while the inner strut radius is tested for values of 0, 5, 10, 15, 20 and 25 mm. Each

inner strut radius is tested against each value of outer strut radius (excluding inner strut radii greater

than or equal to the outer strut radius). Similar to the cable radii, changing the cross-sectional area

of struts changes the structure’s mass and the resulting force at each node. The effects of increasing

the strut material used (i.e., higher outer radius and smaller inner radius) on mass and resulting

force are more dramatic than increasing cable radius despite there being only 18 strut members

compared to the 42 cable members. The struts are the longest members in the structure, and also

have higher density than the cables. This shows that reducing strut cross-sectional areas is a more

efficient means of reducing structural mass than reducing cable radius. The results of the tested

strut radii are shown in Figure 5.11.

Figure 5.11: Effects of strut inner and outer radii on mass (MT) and force at each node (FT). Strut inner radii are
shown as different colored lines.

130
 Figure 5.12 shows the effects of the various strut radii values on average deformation and

first mode. For a particular inner radius value, increasing the outer radius generally increases the

average deformation and decreases the natural frequency, with the exception of very thin-walled

struts, where increasing the outer radius for a small range of radius values that are not much larger

than the inner radius has the inverse effect. More solid members generally result in larger

deformations for the same outer radii values than the hollow counter parts, whereas more hollow

struts tend to offer higher natural frequencies for the same outer radii.

Figure 5.12: Effects of strut outer radius on deformation and first mode.

Additionally, the stow height is directly related to the outer strut radius. For a particular

triplex height (bay height), the closest distance between two struts is calculated. When this nearest

distance is equal to the sum of the radii of two struts (i.e., 2rout), the struts have just touched,

indicating the triplex height can be decreased no further. Figure 5.13 shows the closest distance

between two struts as a function of the triplex height. This plot can be used to determine the smallest

triplex height based on the strut’s outer radius. For the maximum strut radius tested (rout=30 mm),

the closest allowable distance between struts would be 0.06 m, resulting in a minimum stowed

height of 0.15 m. For the minimum strut radius tested (rout=10 mm), the nearest distance would be

131
0.01 m, resulting in a minimum stowed height of 0.047 m. Based on the dimensions of a module,

the smallest cylinder in which a stowed module could fit would have a radius of rstow=0.76 m,

resulting in stowed volumes of 0.00042 m3 and 0.000015 m3 for the larger and smaller strut radii

respectively.

Figure 5.13: Closest distance between two struts for a given triplex height (bay height).

5.1.5 Summary of Design Variable Effects

The results of varying pre-stress factor, bay height, cable radius and strut radius are

summarized in Table 5.3, where “increase” or “decrease” indicates increasing or decreasing the

design variable respectively improves the corresponding structural property. In space applications,

it is desirable to minimize mass and deflection, while maximizing the first mode and packaging

efficiency. In the case of a single module, the desired outcomes do not conflict for any given design

variable. Thus for a single bay in space applications, a triplex module with high pre-stress, low bay

height, small cable radius, and small outer strut radius with relatively larger inner radii is desired.

132
As noted, very thin walled struts have the inverse effects desired on average deflection and first

mode, and thus the exact inner radius would need to be evaluated in greater detail to optimize the

desired structural properties.

Average Packaging
Design variable Mass First mode
deflection efficiency

Pre-stress factor — Increase Increase —

Bay height Decrease Decrease Decrease —

Cable radius Decrease Decrease Decrease —

Decrease Decrease
Strut radius (fixed inner
Decrease (Increase for (Increase for Decrease
radius)
very thin walls) very thin walls)
Strut inner radius (fixed
Increase Increase Increase —
outer radius)
Table 5.3: Effects of increasing design variables on mass, average deflection, first mode and packaging efficiency for a
single triplex module.

5.2 Design Analysis of Triplex Module Array

A larger array is formed by connecting 19 modules at adjacent nodes, as shown in Figure

5.3. The geometry and scale of the array is designed to be analogous to the JWST. Each module is

considered independent of the others, such that independent modules could be stowed and launched

on separate vehicles, then later assembled in space, where each module would be capable of

maintaining (and changing) shape independently. These modules are assumed to connect at

adjacent nodes once assembled, and the larger array considers these shared nodes as a single node

between adjacent modules. This results in 259 total nodes in the array, 186 of which are top-face

nodes.

The same design parameters studied for the single module are also evaluated for the larger

grid, using the same set of variable ranges. The boundary conditions used for this array fix the

133
bottom face nodes of the central-most module in x-y-z, selected such that the larger array might be

affixed to an arm that could orient in different directions relative to the rest of the spacecraft (or

relative to a sun-shield, like the JWST). The average deformation in z of the nodes is again used to

characterize the deformation of the structure. In this case, nodes which are further from the central

module (where the bottom nodes are fixed), will have larger deformations in general. The over-all

magnitude of the deformations in the array for these station keeping maneuvers are found to be

small, however, and the average deformation in z is used as a sufficient parameter to characterize

the overall effects of the design variables on the structure’s ability to resist this type of uniform

external loading in the negative z direction. The same nominal values of the design variables are

used for each module in the larger array as in the cases evaluated for the single module. This results

in a nominal array mass of 6,703.8 kg (19 times the nominal mass of a single module), and a

corresponding nominal thrust load of 1.19 N at each node (221.23 N total).

5.2.1 Effects of Pre-stress Factor

Just as with the single module, the pre-stress factor is varied between 1 and 20,000. The

nominal mass and resulting nominal force on each node is used, since the pre-stress factor does not

affect the structure’s mass. Figure 5.14 shows the effects of pre-stress factor on the resulting

average deformation and first mode. Note that the deformation is plotted in log-scale on the y-axis.

The deformation is found to drop-off significantly with the increase in pre-stress factor, where the

largest deformation (found with the smallest pre-stress factor, pf=1) being 0.87 m, as compared to

the smallest deformation (from pf=20,000) of 0.044 mm. Deformations decrease to the mm scale

around a pre-stress factor value of 800. Increasing the pre-stress factor is also found to increase the

first natural frequency, thus showing that increased pre-stress factors are beneficial both in terms

134
of minimizing deformation and maximizing the first mode frequency, though the effects are more

significant for the deformation.

Figure 5.14: Effects of pre-stress factor on the average deformation and first mode.

5.2.2 Effects of Bay Height

The bay height is varied between 0.7 and 3.35 m. The resulting mass and forces are not

pictured for the array, as they will show the same trends as that pictured for the single module in

Figure 5.7, with the values multiplied by 19 for the 19 total bays. The effects of bay height on

deformation and first mode frequency are shown in Figure 5.15. In the case of the array, taller bay

heights result in lower deformations and lower first mode frequencies, indicating that shorter arrays

are better in terms of first mode frequency, but worse in terms of deformation.

135
 Figure 5.15: Effects of bay height on average deformation and first mode of triplex module array.

5.2.3 Effects of Cable Radius

The cable radius is varied from 0.25 mm to 15 mm. Higher cable radii result in higher

masses, and higher loads on each node in turn. The effects of cable radius on the resulting

deformation and first mode are shown in Figure 5.16. Increasing cable radius is found to increase

the average deformation and decrease the first mode frequency. Thus, it is generally better for the

array to have cables with smaller radii.

136
 Figure 5.16: Effects of cable radius on deformation and first mode in triplex module array.

5.2.4 Effects of Strut Radius

The strut outer radius is varied between 10 mm and 30 mm, with inner radii values of 0, 5,

10, 15, 20 and 25 mm tested. The effects of the various combinations of strut inner and outer radii

are shown in Figure 5.17. For a particular inner radius, increasing the outer strut radius generally

increases the deflection and decreases the first mode frequency. For a particular outer radius, struts

with larger inner radii have lower average deflections and higher first modes, suggesting hollow

struts with smaller outer radii generally result in better structural properties for the array.

137
 Figure 5.17: Effect of strut outer and inner radius on average deflection and first mode for the triplex module array.

Just as with the single module, strut radius also affects the packaging efficiency for the

larger array. However, since the bay height (and thus the stowed height) is not affected by the

number of modules in the array, the same trends as shown in Figure 5.13 hold for the array as well.

The stowed radius (i.e., the circumscribing radius of the whole array) is 3.32 m for the 19-module

array, resulting in an overall stowed volume of 5.19 m3 for the upper bound of outer strut radii (30

mm) and 1.63 m3 for the lower bound (10 mm).

5.2.5 Summary of Design Variable Effects

The results of varying pre-stress factor, bay height, cable radius and strut radius are

summarized in Table 5.5 for the 19-module array. In the case of a single module, the desired

outcomes do not conflict for any given design variable. The same general trends are seen for the

array as the single module, with the exception of bay height. In the case of the array, increasing bay

height is found to decrease the deflection as well as the first mode frequency. Thus, when designing

138
a larger array, the bay height would need to be carefully selected considering the maximum

acceptable deflection and minimum acceptable first mode frequency. The other variables show the

same trends as the single module, where a higher pre-stress, smaller cable radius, and smaller strut

with a larger inner radius offer more favorable structural characteristics.

Average Packaging
Design variable Mass First mode
deflection efficiency

Pre-stress factor — Increase Increase —

Bay height Decrease Increase Decrease —

Cable radius Decrease Decrease Decrease —

Strut radius (fixed inner

Decrease Decrease Decrease Decrease
radius)
Strut inner radius (fixed
Increase Increase Increase —
outer radius)
Table 5.4: Effects of increasing design variables on mass, average deflection, first mode and packaging efficiency for a
single triplex module.

5.3 Additional Design Considerations for a Triplex Module Array

The general trends found in section 5.2 can be used to inform the design of a triplex module

array intended for use as the backbone of a parabolic reflector. While the exact values of design

variables are not optimized in this research, values are selected based on the trends seen in section

5.2, balanced with practical consideration of general space structures applications.

First, a pre-stress factor of 15,000 is selected. While maximizing the pre-stress factor has

the best results in terms of average deflection and first mode, higher values of pre-stress also make

the odds of reaching the material limits of members more likely, since members are already pre-

tensioned to high values. The loads evaluated in this dissertation are selected based on typical

station-keeping maneuvers, which result in relatively small overall forces. If an unexpected impact,

139
due to docking for example, were to hit the structure, the loading conditions could be much more

dramatic. While the exact effects of such higher-scale forces are not considered, a pre-stress factor

is selected below the maximum tested to leave a greater factor of safety in these types of scenarios.

The effects of increasing bay height result in lower deflections and first mode frequency.

While a smaller deflection is more desirable, a lower first mode is not. In general, the deflections

in all studied cases of section 5.2 are relatively small, typically on the mm scale or less. Conversely,

the first mode frequencies were all relatively small for the cases testing bay height, with the

maximum resulting first mode still under 1 Hz. Additionally, taller bay heights require more

material, which not only take up more space, but also add mass to the overall structure. With this

in mind, 1 m is selected as the bay height, choosing a value closer to the lower bound to prioritize

increasing the first mode frequency and decreasing mass, without completely neglecting deflection

minimization.

A smaller cable radius was found beneficial for all tested structural parameters. It is worth

noting that cable failure under loads will happen more quickly for cables of a smaller radius. Thus,

for the same reasons pre-stress factor was not completely maximized, a value of 1 mm is selected

as the cable radius; a small value but not quite the minimum value tested.

In general, more hollow strut members showed favorable effects on the structural

characteristics. Figure 5.17 shows that struts of a similar wall thickness offer similar deflections

and first mode frequencies independent of outer strut diameter (i.e., the lowest outer diameters for

inner diameters of 10-25 mm). With this in mind, a strut with an inner radius of 10 mm and a wall

thickness of 2 mm (i.e., outer radius of 12 mm) is selected.

For the selected strut radius (12 mm), the smallest possible stowed height of the array

would be 0.024 m, based on the results shown in Figure 5.13. This is likely an impractical stow

height, however, when considering how struts would be actuated to accommodate length changes.

Most likely, some sort of telescoping strut would be required to enact shape change. To change

140
length from 0.024 m to the 1 m bay height in evenly length segments, 41 segments would be

required. This is likely an impractical number of when considering actuation. If the struts are made

using the same type of linear actuator concept as the linear actuators used for the experiment in

Chapter 4, the struts more practically would be broken into 2 segments. With this in mind, the

stowed state is considered to be the height resulting from halving the strut lengths, such that the

struts could be comprised of 2 even-length members, where one would store inside the other. This

results in a stowed height of 0.45 m, and a corresponding stowed volume of 15.6 m3 (as compared

to a deployed volume of 34.7 m3). Investigation of more segments to further improve the packaging

efficiency towards the limits found based on strut diameter could certainly be employed, but are

outside the scope of this thesis.

With the design parameters selected, the triplex module array can be studied in greater

detail to understand its structural behavior. These design parameters result in a structure mass of

5,420 kg, including the 275.3 kg per-module for additional spacecraft weight (the mass of the

tensegrity itself thus being only 189.3 kg). The resulting first natural frequency in the flat shape

with the designated bay height of 1 m is found to be 0.82 Hz. Design parameters could be modified

to increase this first mode; however, many reflector space structures are designed with first mode

frequencies less than or equal to 0.82 Hz once deployed in orbit [69-71], with Hyde et al. [72]

modeling the JWST first mode as 0.4 Hz. These natural frequencies typically involve incorporation

of more spacecraft elements than the supporting structure alone (e.g., the JWST model by Hyde

includes representation of the optical telescope, science instrument module, spacecraft and

sunshield), and incorporation of additional spacecraft elements is outside of the scope of this

research. However, the first natural frequency calculated for this structure is deemed reasonably

large to proceed with this set of design variables, as 0.8 Hz is similar to the first mode frequencies

of reflector spacecraft described in the literature [69-73].

141
 The average deformation in z as defined in sections 5.1 and 5.2 is calculated for this set of

design variables as well, resulting in an average deformation of 0.16 mm. This is a relatively small

deflection unlikely to cause any issues regarding material failure in the structure (the maximum

resulting stress in any member was found to be 1.96 MPa).

Finally, the structure’s ability to achieve the desired shape in the deployed state must be

considered. To that end, section 5.3.1 describes the methods used to find the best-fit shape to a

paraboloid surface in terms of surface accuracy. Additionally, a path from a flat, stowed shape to

the resulting shape fit to the paraboloid surface is also considered, demonstrating the feasibility of

a deployment path through equilibrated shapes.

5.3.1 Fitting to a Paraboloid Surface

Using the JWST for scale, a paraboloid surface with a focal length of f = 7.6 is used. The

center of the paraboloid is set to (0,0,zp), where the x and y components are set to 0 to align the

paraboloid with the center for the array. The z-shift of the paraboloid, zp, is selected such that the

outermost nodes of the array are on the surface of the paraboloid when the module is in the flat

shape with the prescribed bay height of 1 m (zp = 0.64 m). This is chosen such that the maximum

required strut length would not exceed the strut length when the array is in the flat shape with a bay

height of 1 m. Figure 5.18 shows the paraboloid surface with the flat shape of the array.

142
 Figure 5.18: Triplex Module Array with paraboloid surface overlaid in light blue.

Finding shapes for this array is no trivial matter. There are a total of 259 nodes, 186 of

which are top-face nodes that are to be fit to the paraboloid surface. In addition, possible locations

of those 186 are constrained by the top-face cable lengths—342 length constraints to impose in

total. Determining the design variables to use in optimization that incorporate all 342 constraints

by hand quickly becomes impractical. Instead, the structure’s symmetry, along with the weighting

factor wxy are used to reduce the scope of the optimization problem.

Figure 5.19 shows the array, where 7 of the modules are outlined. These are the only 7

modules which need to be fit to the paraboloid surface. Rotating these 7 modules by 120 o in the

positive and negative direction about the z-axis forms the larger array. Since the paraboloid surface

center is aligned with the module center, a good fit of these 7 modules will result in an equally good

fit of the whole array to the overall surface. The fitness function defined in Chapter 4 (Eq. 4.13) for

the experimental tensegrity is used for fitting this larger array to a paraboloid surface as well. The

fitness function as defined in Eq. 4.13 is shown again in Eq. 5.4:

𝑓𝑖𝑡𝑛ⅇ𝑠𝑠 = 𝛥𝑧 + 𝑤𝑥𝑦 ̅̅̅̅̅

∆𝑥𝑦 (𝐸𝑞. 5.3)

143
where 𝛥𝑧 is the average distance between the z-values of the top face nodes, and the z-value on the

̅̅̅̅̅ is the average distance between the

paraboloid surface for those nodes’ x-y coordinates, and ∆𝑥𝑦

x-y coordinates of their nodes from their original x-y coordinates in the flat shape. The inclusion of

this additional x-y distance factor is particularly important for this larger array.

Each of the 7 modules outlined in Figure 5.19 are individually fit to the paraboloid surface,

incorporating the length constraints of the individual modules into their individual optimization

problems. After all modules are fit, the distance between nodes which should be shared between

modules is calculated. If these shared nodes are sufficiently close to one another, then the

optimization is considered a success, and the resulting shape is used as the array shape fit to the

paraboloid surface. Increasing the weighting factor wxy improves the odds of finding module shapes

with negligible distance between shared nodes, but may also impede the odds of finding a shape

with a good fit to the paraboloid surface. For this particular example, a value of wxy=0.75 was used.

Figure 5.19: The triplex module array; the 7 modules outlined in a thick black line can be rotated +/- 120o to achieve
the full array.

144
 The optimization routine was run for each of the 7 modules shown in Figure 5.19. The

resulting maximum distance between nodes which should be shared between adjacent modules was

8.5 mm. While the node design of modules is not considered in detail in this dissertation, some

general considerations indicate this distance is reasonably small. Recall that in this example, cables

were selected to have a radius of 1 mm, and struts an outer diameter of 10 mm. In a single module,

outer nodes of the top face (i.e., the nodes which would be shared between adjacent modules) must

accommodate the connection of 3 cables and 1 strut. Whatever mechanism is used to accommodate

connection between adjacent modules would need to, at a minimum, accommodate the connection

of two struts. Nodes in these simulations are modeled as discrete points, not as a physical node with

a finite volume. The maximum distance between shared nodes is less than the radius of a single

strut, and thus it is reasonable to assume that the body of the node, however it is designed, can

accommodate a discrepancy of this size between adjacent nodes.

Once the fit is found for the 7 modules outlined in Figure 5.19, these modules are rotated

+/- 120o to form the larger array (except for the central-most module, which only needs to be

represented once). The larger array, with nodes shared between adjacent modules, is then used to

calculate the overall fit of the array to the paraboloid surface, based on the deviation of the top-face

nodes from the surface in z. The fit in this case was found to be 0.12 mm.

Figure 5.20 shows the resulting fit shape with the paraboloid surface. This tensegrity

structure is meant to be the backbone of the reflector, and thus the exact surface accuracy of the

reflector itself would vary depending on how the reflective surface is implemented. If this structure

were used as the backbone to a mesh reflector, and the mesh were connected directly at nodes, then

the overall surface accuracy would be closely tied to the accuracy calculated here. If the structure

is instead used as the backbone of a solid reflector, like the JWST, then the surface accuracy would

be more closely tied to the accuracy of the mirrors. While the components attaching the reflectors

to the backbone could accommodate some discrepancies in surface accuracy, it is still desirable to

145
have good surface accuracy in the supporting structure directly, as this would reduce the need for

and demands on the ancillary structures.

Of these two cases, a mesh reflector would require more strict accuracies of the tensegrity

structure. Even in this stricter case, the surface accuracy of 0.12 mm offers a strong starting point

for the reflective surface. Commercially available mesh reflectors offer surface accuracies on the

mm scale, such as the AstroMesh reflector (accuracy listed as “better than 3 mm” [73]) and the

unfurlable mesh reflector from L3 Harris (accuracy listed as 4 mm [69]). While the end-state

accuracy of the reflector would depend upon the mesh material and the methods used to tension the

mesh, the tensegrity itself shows surface accuracy on the same scale as these commercially

available reflectors, and would prove a promising structure for such an application.

Figure 5.20: Array fit to a paraboloid surface

5.3.2 Path Planning from Stowed to Deployed Shapes

With a target shape identified as that fit to the paraboloid surface, the path planning

approach detailed in Chapter 4 is used to identify a path between the start shape (the stowed shape)

and the target end shape. The stowed shape is taken as the flat shape with a height of 0.45 m. The

path from this stowed shape to the shape fit to the paraboloid surface is evaluated for the 7 modules,

146
then rotated +/- 120o to detail the shape of the array as a whole. In this case, the path was broken

into 50 steps. A few steps are shown in Figure 5.21, showing the successful deployment from the

stowed shape to the target shape through statically equilibrated shapes.

Figure 5.21: Select steps from path planning between a stowed shape and a deployed shape fit to a paraboloid surface
for the triplex module array.

147
 Any number of steps between he stowed and deployed shape could be evaluated based on

the specifications of the system. Generally speaking, the number of steps should be sufficiently

small such that assuming linear movement of nodes between adjacent steps is an acceptable

assumption. This assumption is used when checking for member interference between adjacent

steps. Additionally, the steps can be used for commanding actuators during the shape-changing

process (as they were in the experiment in Chapter 4), and thus the smallest increment the actuators

can be commanded to move can be used to inform the number of steps as well. Each individual

step on the path is checked for structural equilibrium, member length constraints, and potential

member interference when moving from the prior step. As mentioned in Chapter 4, a subroutine

can be used to search for new shapes whenever one of these conditions is not met. In this case,

however, no steps along the path required additional searching.

5.3.3 Summary of Triplex Module Array as a Support for a Parabolic Reflector

This chapter investigated the structural properties of a triplex module array with the

intended use as the structural backbone for a parabolic reflector. The design variables selected are

summarized in Table 5.5. These variables were chosen based on the findings of section 5.2, seeking

a combination of low mass, high stiffness (characterized in terms of both deflection and first mode

frequency), small stowed volume, and high surface accuracy. The resulting structural

characteristics are summarized in Table 5.6.

Pre-stress factor 15,000

Bay height 1m
Cable radius 1 mm

Strut outer radius 12 mm

Strut inner radius 10 mm
Table 5.5: Design variables used for the case study of a triplex module array as the backbone of a parabolic reflector

148
 Mass 5,420 kg

Stowed volume 15.6 m3

Average deflection in z 0.16 mm

First mode 0.82 Hz

Surface accuracy 0.12 mm

Table 5.6: Structural Properties of triplex module array case study.

As the backbone of a parabolic reflector, surface accuracy is of critical importance. A best

fit shape to the paraboloid surface in question (having a focal length of f = 7.6 and a z-shift of zp =

0.64 m) was found that resulted in a surface accuracy of 0.12 mm (where surface accuracy is

determined by the average deviation in z from the prescribed surface). The design of the structure

itself does not consider whether the tensegrity would be used to support a mesh reflector or a solid

reflector.

The James Webb Space Telescope is a solid reflector which offers mirror alignment

accuracies on the nanometer scale [74]. The JWST is designed with such fine control because its

intended use in the red to mid-infrared wavelength range (0.6-28.8 micro meters) [61]. If this

tensegrity array were used as a support for a solid reflector, the end result surface accuracy would

also be dependent upon the mirrors themselves, as well as whatever mechanism was used to mount

the mirrors to the tensegrity. If the mirrors are mounted directly at the nodes, then some amount of

discrepancy in surface accuracy can be accounted for in the mounting structures themselves, such

that the final position of the mirrors is more closely aligned with the prescribed surface. However,

the JWST is designed for additional alignment after deployment; these mechanisms provide

alignment precision with 10 nm accuracy. The tensegrity structure offered a surface accuracy of

0.12 mm; as such, it is unlikely that the tensegrity array would offer the stringent accuracies

required for wavelength observations on the micrometer scale.

149
 As the backbone of a mesh reflector, the tensegrity array may be used with the mesh tied

directly to nodes, adding additional tension elements throughout the structure as needed. While the

specifics of additional tensioning are outside the scope of this thesis, the surface accuracy of the

nodes that are directly tied to the mesh would, in this case, be the calculated 0.12 mm. This level

of surface accuracy is comparable to, if not better than, that of existing mesh reflectors in space.

The AstroMesh by Northrop Grumman offers a surface accuracy “better than 0.3 mm”, for use up

to Ka band (4 - 7.5 mm) [73], and L3 Harris offers an unfurlable mesh reflector with 4 mm surface

accuracy, intended for S band (75 - 150 mm) [69]. The smaller the wavelength, the more stringent

the surface accuracy requirements. Based on the surface accuracy of 0.12 mm, this tensegrity array

would prove more suitable for applications up to the Ka band.

150
 Chapter 6

Conclusions and Future Work

This dissertation investigates the design of tensegrity structures for shape-changing space

applications. With origins in art and architecture, tensegrity structures are most often found built

for static applications, and most research exploring tensegrity structure capabilities is focused

around such. The true potential of tensegrity structures, however, is not met by static applications;

by actuating the lengths of members within the structure, tensegrity structures can be designed for

shape change, whereby their inclusion of cable members can be fully exploited. This dissertation

explores methods to characterize the range of equilibrated shapes a particular tensegrity structure

can achieve, as well as explores a variety of methods to analyze and design these structures with a

particular shape changing application in mind.

6.1 Summary

First, a literature review of existing form-finding methods for finding equilibrated shapes

of tensegrity structures was conducted, categorizing these methods as iterative, metaheuristic, and

analytical. Most methods in the literature seek only a single equilibrated shape. This dissertation

addressed this gap by developing methods to explore the broader range of equilibrated shapes a

tensegrity may achieve, and by exploring the use of these structures in shape changing applications.

An analytical approach was developed which characterizes the entire range of equilibrated

shapes a tensegrity structure can achieve. The geometry of the structure was described via a

geometric matrix, formed by representing members as vectors and detailing the independent cycles
these vectors form within the structure. The span of the null space of this geometric matrix was

used to describe all geometrically viable shapes the structure can achieve. This geometric

description of such shapes was characterized by the geometric coefficients, which are then

substituted into the equations of equilibrium. The matrix form of the equations of equilibrium were

then solved for the force densities. The resulting equations were used to impose limitations on the

values of geometric coefficients which result in shapes that are strictly equilibrated. This analytical

approach was used to derive such limitations for a 2D x-tensegrity and a triplanar tensegrity,

successfully describing all possible equilibrated shapes the tensegrities can achieve in terms of their

geometric coefficients.

For cases where the resulting equations become too complex to develop and solve

analytically, a numerical approach was developed. This approach starts with a single shape known

to be in equilibrium, with the geometry described in terms of appropriate variables. Parameterized

lines are defined from this starting point, using the variables selected to define the shape of the

tensegrity. A number of points on each line, corresponding to different tensegrity shapes, are

checked for equilibrium. Resulting shapes which are equilibrated are added to a numerical dataset,

while non-equilibrated shapes are discarded.

In successive iterations, multiple shapes from the data set are selected as new shapes from

which lines will branch; points chosen from the data set are selected using a sampling approach

which favors points with a higher Mahalanobis distance value. As the data set grows, the maximum

Mahalanobis distance is monitored; once the Mahalanobis distance stops changing significantly for

a specified number of iterations, the numerical approach is considered converged, with the resulting

set of shapes forming a representative data set.

This approach was checked against the analytical results of the 2D x-tensegrity,

successfully finding a representative dataset which agreed with the analytical results. Additionally,

the approach was used to find representative datasets for the 3D triplex. Triplex cases where

152
members were unconstrained and where all but 3 members were constrained were both evaluated,

each converging on a representative dataset. Finally, the approach was used to find a local dataset

of representative shapes for a tensegrity module comprised of 6 triplexes. The bottom face of the

triplex was held fixed, and the lengths of top face cables were also fixed. Even with these

constraints, this tensegrity has 18 of degrees of freedom. This case demonstrates how the numerical

approach can be used for more complex tensegrity structures.

Additional analysis methods were developed to aid in the understanding of how tensegrity

structures can be used in shape changing applications. Geometric principles were developed to

guide the incorporation of length constraints into the representation of tensegrity shapes. Their

application results in a minimal set of design variables that describe the shapes while incorporating

fixed lengths as needed.

Additionally, an approach to path planning was described. This approach uses the design

variables selected to characterize the tensegrity’s shape to define a parametric line between

specified start and end shapes. This line is split into any desired number of steps, where shapes at

each step are checked for satisfaction of equilibrium and any length constraints. If a step is found

to not satisfy either, a Particle Swarm Optimization routine is implemented to search for a set of

design variables which does satisfy these conditions, and is also as close to the set of design

variables found from the parametric line as possible. This approach can be used to identify paths

that connect equilibrated shapes between the desired start and end shapes of the tensegrity.

This path planning approach was verified in the laboratory using a tensegrity structure built

using linear actuators for struts and supporting cables with lengths adjusted using spools attached

to stepper motors.

Furthermore, a nonlinear finite element model was used to study deformations of tensegrity

structures under external loads as well as to determine the structure’s natural frequencies of

153
vibration. These analyses can be used to understand a tensegrity structure’s ability to achieve

desired shapes, as well as to provide insight into a tensegrity’s structural performance.

Finally, a case study evaluating the use of a tensegrity structure as the backbone of a

parabolic reflector was conducted. A tensegrity module comprised of 6 triplexes was defined and

selected for its geometric compatibility to connect many modules together at outer nodes. This

module is constrained with the bottom face fixed in space and the top face cables fixed in length,

using the geometric principles described in previous chapters to characterize the structure’s shape

in a way that incorporates the length constraints.

A single module was first evaluated for how the pre-stress factor, bay height, cable radius,

and strut inner and outer radius affect the tensegrity’s mass, deflection under uniform loading in

the z-axis, first mode natural frequency of vibration, and packaging efficiency. In summary,

increasing pre-stress decreases the deflection and increases the first mode frequency. Increasing

bay height increases the mass, increases the deflection and decreases the first mode frequency.

Increasing cable radius increases the module’s mass, decreases the deflection and increases the first

mode frequency. Increasing the outer strut radius for a particular inner radius value increases the

mass, while increasing the inner radius for a particular outer radius value decreases the mass. In

general, increasing the outer strut radius for a particular inner strut radius increases the deflection

and decreases the first mode frequency, while increasing the inner radius for a particular outer

radius decreases the deflection and increases the first mode frequency. For very thin-walled struts,

the inverse was true; increasing the outer radius while the walls were thin decreased the deformation

and increased the first mode. Finally, the outer strut radius is the only parameter that influences the

packaging efficiency. A stowed state is defined by decreasing the structure’s height until strut

members just touch. Packaging efficiency is improved with smaller outer strut radius values.

The same analysis was then applied to an array comprised of 19 triplex modules, sized in

a general way by reference to the JWST. The pre-stress factor, cable radius and strut inner and outer

154
radius were found to follow the same trends in their effects on mass, deformation, first mode and

packaging efficiency as the single module. The bay height, however, has the opposite effect on

deformation. Increasing bay height for the array decreased deformation under station-keeping

maneuvering loads. Increasing bay height still reduced the natural frequency of the first mode of

vibration, indicating that the bay height has opposing effects on deformation and first mode

frequency for the larger array, and posing a trade-off in design.

These findings were then used to guide the selection of a set of design variables and

evaluate the resulting structure in greater detail. A pre-stress factor of 15,000, bay height of 1 m,

cable radius of 1 mm, and strut inner and outer radius of 10 mm and 12 mm respectively were

selected. The resulting deformation for uniform loading in z was found to be 0.16 mm, and the first

mode frequency was 0.82 Hz; these values are in general agreement with existing reflectors in

space applications.

Particle swarm optimization was used to find a best-fit array shape to a prescribed

paraboloid surface. Defining surface accuracy by the deviation of nodes in the z-direction from

their projected values on the paraboloid surface, the array was found to have a surface accuracy of

0.12 mm for a paraboloid surface with focal length f=7.6 and z-shift zp=0.64 m. This level of surface

accuracy is comparable to that of existing mesh reflectors used for Ka band applications, indicating

this tensegrity array would be a good candidate as the backbone of reflectors for similar

applications.

Additionally, a path from a stowed shape—specified as a flat shape in which struts are half

their final height when the deployed bay height is 1 m—to the best-fit deployed paraboloid shape

was identified. This path was verified as passing through only equilibrated shapes for each module,

and did not result in any member interference between the start and end shapes with the selected

member radii.

155
 6.2 Conclusions

The objectives of this dissertation were posed to explore the capabilities of tensegrity

structures in shape changing applications. The first objective was addressed by developing a novel

method of characterizing the broad range of equilibrated shapes a tensegrity structure can achieve.

The analytical and numerical approaches developed in this thesis demonstrate not only that these

structures can achieve symmetric and asymmetric shapes, but also that the wide variety of shapes

can be systematically characterized. This characterization addresses a gap in the literature, which

typically formulates the general, well-studied form-finding problem around finding a single (and

often symmetric) shape for a particular tensegrity configuration.

The second objective of this dissertation seeks to understand how these structures can be

designed for a particular shape changing application. To address this, a variety of analysis methods

were developed which handle inclusion of member length constraints, targeted shape selection, and

quasi-static path planning. Additionally, the pre-stress factor, bay height, and radii of the struts and

cables were studied in detail for their influence on overall structural mass, packaging efficiency,

deflection under external loads and first mode frequency. Using these methods, a detailed modular

tensegrity structure was developed and analyzed for its potential as the backbone of a parabolic

reflector. This case study demonstrates the utility of understanding the effects of different tensegrity

properties on its ability to perform in a shape changing application. While the particular example

focused on a parabolic reflector, the methods developed could be applied to any general tensegrity

configuration and with any shape changing application in mind. These methods demonstrate the

immense potential of tensegrity structures in shape changing applications, especially in space

applications where minimal mass and volume are dictated by launch vehicle constraints.

156
 6.3 Future Work

This dissertation describes methods that can be used to understand the range of achievable

shapes for different types of tensegrity structures, as well as analysis tools that can be used to design

such structures specifically for shape changing applications. Nevertheless, additional research

could be conducted to further improve the methods proposed in this thesis and enhance the ability

to design tensegrities for shape change.

First, in both the analytical and numerical approaches characterizing achievable shapes for

tensegrity structures, stability analysis was not considered. Understanding which shapes are stable

and which are not is very important in any practical application. During active operation, any highly

unstable should would be avoided. Additionally, some forms of unstable shapes may be of interest

when designing the structure, to exploit passive deployment methods or bistable shapes. Whatever

the application, and additional step of stability analysis within the characterization of equilibrated

shapes would be a highly insightful tool and improvement upon the approaches developed in this

thesis.

The numerical approach for characterizing the range of achievable shapes was successfully

used for a variety of tensegrity structures. However, as the structures became more complex, the

time it took this approach to converge increased drastically. Even using parallel cluster computing,

it would be beneficial to consider alternative approaches when determining the directions of

branching lines. In this dissertation, line directions are selected randomly, ensuring that trial design

points remain within the design variable boundaries. While some amount of randomness will likely

always be required for initiating the approach, where only one shape is known, it would be

worthwhile to investigate selection of line directions which favor areas of the data set known to be

in equilibrium as the solution dataset grows. Alternatively, seeking line directions influenced by

more sparse regions of the data set may also encourage growth in new directions. This is included

157
in the current approach to some extent by favoring points with higher Mahalanobis distance values

for use as branch points, but factoring this into the line directions as well could potentially improve

the efficiency of this approach.

Incorporating length constraints into the characterization of a structure’s range of shapes is

highly important for the efficacy of a numerical approach, as well as for exploring realistic

structures that might actually be used for space applications. The geometric properties explored in

this thesis could, in theory, be used for any tensegrity structure of any complexity. However, these

approaches described need to be manually implemented with careful consideration of the

structure’s geometry. While this was accomplished for a fairly complex tensegrity, as the structures

become more and more complex with more and more variables to keep track of, avoiding human

error would be difficult. A more generalized approach to incorporating member length constraints

in a tensegrity’s shape characterization would be highly beneficial, particularly if it could be

implemented to work autonomously for any connectivity.

Finally, there is much room for advancement in the physical implementation of shape

changing tensegrity structures. The experimental set up in this dissertation used linear actuators as

struts, and stepper motors to actuate cables. These actuation methods were implemented open-loop

based on calculated sensitivities, with no feedback to verify that the members were moving exactly

as desired. In applications where precision shapes are necessary, implementation of feedback would

be critical to improving performance. Both global metrology systems (such as a motion capture

system) and local member sensors would surely be useful for providing feedback in real-world

shape changing applications.

158
 Appendix A
Using Length Constraints to Derive Design Variables

A.1 Design Variable Derivation for a Constrained Triplex

Chapter 3 addresses a constrained triplex with 9 members fixed in length, leaving only the 3

supporting cables to change length, pictured below in Figure A.1 with supporting cables in green.

In addition, the bottom face is fixed on the x-y plane, and the member 1 is on the x-axis, With these

constraints, this triplex has only 3 degrees of freedom. The design variables defined to incorporate

member length constraints are l8, l9, and Φ.

Figure A.1: A triplex tensegrity, all members are fixed in length except supporting cables, shown in green (l7,l8,l9).

First, the length of supporting cable 8 is specified as l8. With this cable length, the

triangle N1-N2-N5 has all fixed side lengths, and the law of cosines can be used to define the angle

between member 1 and member 10 (both of which are fixed in length) as:

𝑙12 + 𝑙10
2
− 𝑙82
𝛽10 = 𝑐𝑜𝑠 −1 ( ) (𝐸𝑞. 𝐴. 1)
2𝑙1 𝑙10

159
With this angle, the location of N5, if members 10 and 8 were on the x-y plane, would be defined

as:

𝑁5 = {𝑙10 cos(𝛽10 ) , 𝑙10 sin(𝛽10 ), 0} (𝐸𝑞. 𝐴. 2)

Then, an angle Φ can be defined as the second design variable, designating the angle of

rotation of member 10 about the x-axis (relative to the x-y plane). Angles 𝛽10 and Φ are depicted

in Figure A.2.

Figure A.2: Angles β10 and Φ with respect to triangle N1-N2-N5, used to define N5.

With the angle Φ, a rotation matrix R10 can be defined to rotate member 10 about the x-axis, as

shown in Eq. A.3. Then, this rotation matrix is applied to N5 as defined by Eq. A.2 to fully fix N5

in space for the current set of design variables (Φ and l8), as shown in Eq. A.4.

1 0 0
𝑅10 = [0 𝑐𝑜𝑠 𝛷 − 𝑠𝑖𝑛 𝛷] (𝐸𝑞. 𝐴. 3)
0 𝑠𝑖𝑛 𝛷 𝑐𝑜𝑠 𝛷

𝑙10 cos(𝛽10 )
𝑁5 = 𝑅10 { 𝑙10 sin(𝛽10 ) } (𝐸𝑞. 𝐴. 4)
0

With N5, defined, only N4 and N6 remain. Both these nodes can be fixed by specifying one

additional length, l9, and using it with two triangular pyramids.

160
 This first triangular pyramid is N2-N3-N5-N6, shown in Figure A.3. Nodes are moved
relative to their original locations in Figure A.1 for visual clarity of the pyramid in question.

Figure A.3: The triangular pyramid used to define N6.

From Eq. A.4, node 5 is fully defined, and thus the length l3,5 can be calculated as the distance

between N3 and N5. Node 6 is first found relative to the plane defined by N2-N3-N5. The angle

between members 2 and 8 (𝛽8 ) can be found using the law of cosines
2
𝑙22 + 𝑙82 − 𝑙3,5
𝛽8 = 𝑐𝑜𝑠 −1 ( ) (𝐸𝑞. 𝐴. 5)
2𝑙2 𝑙8

Then, 𝛽8 can be used to define node 5 relative to the plane defined by N2-N3-N5 as N5={u5,v5,w5}

with Eq. A.6:

𝑢5 𝑙8 cos(𝛽8 )
𝑣
{ 5 } = { 𝑙8 sin(𝛽8 ) } (𝐸𝑞. 𝐴. 6)
𝑤5 0

161
Finally, using Eq. 4.4, Eq. 4.7 and Eq. 4.8, Node 6 can then be defined as N6={u6,v6,w6} as shown

in Eq. A.7:

𝑙22 + 𝑙211 − 𝑙29

𝑢6 𝑙11 ( )
2𝑙2 𝑙11

𝑣6 𝑙28 − 𝑙25 +𝑙211 − 2𝑢5 𝑢6

= (𝐸𝑞. 𝐴. 7)
2𝑣5
{𝑤6 }
±√𝑙211 − 𝑢6 2 − 𝑣6 2
{ }

Node 6 as defined by Eq. A.7 is in a local coordinate system, defined with triangle N2-N3-

N5 as the local x-y plane and side N2-N3 as the local x-axis and N2 as the local origin. This can be

translated back to the global x-y-z system using Eq. A.8:

𝑵𝟔 = 𝑵𝟐 + (𝑢6 𝒖
̂ ) + (𝑣6 𝒗
̂) + (𝑤6 𝒘
̂) (𝐸𝑞. 𝐴. 8)

̂, 𝒗
Where N6 and N2 are the global coordinates of nodes 6 and 2 respectively, and 𝒖 ̂ and 𝒘
̂ are the

local x-y-z unit vectors as defined by Eq. A.9:

̂
𝒖 (𝑵𝟑 − 𝑵𝟐 )
𝑛𝑜𝑟𝑚(𝑵𝟑 − 𝑵𝟐 )
̂ =
𝒘 (𝑵𝟓 − 𝑵𝟐 ) × (𝑵𝟑 − 𝑵𝟐 ) (𝐸𝑞. 𝐴. 9)
𝑛𝑜𝑟𝑚((𝑵𝟓 − 𝑵𝟐 ) × (𝑵𝟑 − 𝑵𝟐 ))
̂
𝒗 ̂ ×𝒖
𝒘 ̂

The final node, N4, can be defined in much the same way as N6, using the triangular pyramid N5-

N3-N6-N4 as pictured in Figure A.4. Nodes are re-arranged in Figure A.4 relative to their locations

in Figure A.1 for clarity.

162
 Figure A.4: The triangular pyramid used to define N4.

Just as with the prior pyramid, all side lengths are now known, as are all node locations

besides N4. The location of N4 can be derived relative to the plane defined by triangle N5-N3-N6 just

as N6 was derived relative to the plane defined by triangle N2-N3-N5. Thus, N4={u4,v4,w4} is defined

by Eq. A.10:

𝑙23,5 + 𝑙24 − 𝑙212

𝑢4 𝑙4 ( )
2𝑙3,5 𝑙4

𝑣4 𝑙25 − 𝑙26 +𝑙24 − 2𝑢̃6 𝑢4

= (𝐸𝑞. 𝐴. 10)
2𝑣̃6
{𝑤4 }
±√𝑙24 − 𝑢4 2 − 𝑣4 2
{ }

Note that 𝑢
̃6 and 𝑣
̃6 are the local x-y components of N6 relative to the plane defined by N5-N3-N6,

and not the u6 and v6 defined in Eq. A.7. These local coordinates of N4 are translated back to the

global x-y-z coordinates as defined by Eq. A.11:

163
 𝑵𝟒 = 𝑵𝟓 + (𝑢4 𝒖
̂ ) + (𝑣4 𝒗
̂) + (𝑤4 𝒘
̂) (𝐸𝑞. 𝐴. 11)

̂, 𝒗
Where in this case 𝒖 ̂ and 𝒘
̂ are defined by Eq. A.12:

̂
𝒖 (𝑵𝟑 − 𝑵𝟓 )
𝑛𝑜𝑟𝑚(𝑵𝟑 − 𝑵𝟓 )
̂ =
𝒘 (𝑵𝟑 − 𝑵𝟓 ) × (𝑵𝟔 − 𝑵𝟓 ) (𝐸𝑞. 𝐴. 12)
𝑛𝑜𝑟𝑚((𝑵𝟑 − 𝑵𝟓 ) × (𝑵𝟔 − 𝑵𝟓 ))
̂
𝒗 ̂ ×𝒖
𝒘 ̂

Thus, all node locations can be defined by using l8, l9 and Φ as design variables for this

type of constrained triplex.

A.2 Design Variable Derivation for a Constrained Triplex Module

A constrained triplex module is used throughout this thesis for a variety of considerations.

The bottom face of the module (pictured below in Figure A.5) is held fixed in space, and the lengths

of all top-face cables are considered fixed. Supporting cables and struts are free to change length.

This leaves the nodes of the top-face of the structure to be defined, based on their connected

member length constraints.

Figure A.5: A triplex module.

164
 The top face of the triplex module is pictured in Figure A.6, and all individual members

have the same length. The outer-most nodes of the top face (n9, n11, n13, n15, n17 and n19) can each

be defined by one angle, as a rotation about their respective member on the inner polygon

(highlighted in blue in Figure A.6). The inner polygon is defined using the spherical coordinates

and intersection of spherical surface properties as defined in Chapter 4. In total, the shape of the

module top face (and thus the whole module) is defined by one set of nodal coordinates (node 8),

and 15 angles.

Figure A.6: Top face members and nodes of the triplex module.

First, 3 variables are specified as one node’s coordinates directly. This translates the top

face relative to the bottom face. Node 8 is specified as n8={x8, y8, z8}. Then, nodes 10 and 18 can

be defined using spherical coordinates relative to node 8 as:

𝑙8,10 cos 𝜃10 cos 𝜙10

𝒏𝟏𝟎 = 𝒏𝟖 + { 𝑙8,10 sin 𝜃10 cos 𝜙10 } (𝐸𝑞. 𝐴. 13𝑎)
𝑙8,10 sin 𝜙10

165
 𝑙8,18 cos 𝜃18 cos 𝜙18
𝒏𝟏𝟖 = 𝒏𝟖 + { 𝑙8,18 sin 𝜃18 cos 𝜙18 } (𝐸𝑞. 𝐴. 13𝑏)
𝑙8,18 sin 𝜙18

where bold-face indicates vector notation of the node, and li,j is the length of the member connected

between nodes i and j. Figure A.7 shows the spherical angles for node 10 relative to node 8 as an

example.

Figure A.7: An example of spherical angles used as design variables for the first 3 nodes in the top face of the triplex
module.

With nodes 10 and 18 defined, nodes 12 and 16 can be defined relative to nodes 10 and 18

respectively in the same way:

𝑙10,12 cos 𝜃12 cos 𝜙12

𝒏𝟏𝟐 = 𝒏𝟏𝟎 + { 𝑙10,12 sin 𝜃12 cos 𝜙12 } (𝐸𝑞. 𝐴. 14𝑎)
𝑙10,12 sin 𝜙12

𝑙18,16 cos 𝜃16 cos 𝜙16

𝒏𝟏𝟔 = 𝒏𝟏𝟖 + { 𝑙18,16 sin 𝜃16 cos 𝜙16 } (𝐸𝑞. 𝐴. 14𝑏)
𝑙18,16 sin 𝜙16

With nodes 16 and 12 defined, the final node of the inner polygon, n14, must be within the

fixed member lengthsl14,16 and l14,12 from nodes 16 and 12 respectively. To be at the required

distance from each node (n12 and n16), n14 must lie on the surface of two spheres. The first sphere

is centered at n12 with a radius equal to l14,12, and the second sphere is centered at n16 with radius

166
equal to l14,16. The intersection of two spheres (if it exists), is either a circle in 3D space (if the sum

of the spheres’ radii is less than the distance between the centers) or a single point (if the sum of

the spheres’ radii is equal to the distance between the centers). In the case of the circular

intersection, the circle’s parameterized equation can be written as:

̂ 𝑐ⅈ𝑟𝑐𝑙𝑒 + 𝑟𝑐ⅈ𝑟𝑐𝑙𝑒 sin 𝜆𝒗

𝒏𝑐ⅈ𝑟𝑐𝑙𝑒 = 𝑪𝑐ⅈ𝑟𝑐𝑙𝑒 + 𝑟𝑐ⅈ𝑟𝑐𝑙𝑒 cos 𝜆𝒖 ̂𝑐ⅈ𝑟𝑐𝑙𝑒 (𝐸𝑞. 𝐴. 15)

where ncircle is a coordinate on the circle, Ccircle is the center of the circle, rcircle is the radius of the

̂ 𝑐ⅈ𝑟𝑐𝑙𝑒 and 𝒗
circle, 𝜆 is the parametric variable corresponding to the angle on the circle, and 𝒖 ̂𝑐ⅈ𝑟𝑐𝑙𝑒

are local x-y unit vectors in the plane of the circle. The center, radius and unit vectors of the circle

can be found based on the properties of the two spheres [60]. With Eq. A.15, the location of n14 can

thus be determined by specifying a single angle 𝜆. Figure A.8 shows an example of this angle for

nodes i and j that each have members of fixed length meeting mutually at node k.

Figure A.8: Example of the angle used as a design variable when a node location is represented as the intersection of
two spheres.

In the same way that n14 can be specified by a single angle because it is connected by known

lengths to two known nodes, n9, n11, n13, n15, n17 and n19 can all be defined relative to their respective

inner-polygon node pairs (n8-n10, n10-n12, n12-n14, n14-n16 and n16-n18 respectively). This corresponds

167
to specifying a additional 5 𝜆 angles, the final 5 design variables required to fully define the shape

of the top-face of the triplex module, and thus the shape of the whole module itself.

In summary, the triplex module shape is specified by assuming that the bottom nodes are

fixed in space, then specifying n8 as nodal coordinates directly. Next, nodes 10 and 12 are specified

with two angles each (𝜃10 , 𝜙10, and 𝜃12 , 𝜙12 respectively) using spherical coordinates. Relative to

these nodes, nodes 16 and 18 are then specified with two angles each in the same way (𝜃16 , 𝜙16,

and 𝜃18 , 𝜙18 respectively). Finally, the remaining 7 nodes are specified with 𝜆 angles (𝜆14,

𝜆9 , 𝜆11 , 𝜆13 , 𝜆15 , 𝜆17 , and 𝜆19). This is a total of 18 variables needed to specify the shape, 3 of which

are coordinates, and the remaining 15 of which are angles.

168
 Appendix B
Effects of PSO force-finding Parameters for a Triplex Module

To ensure that the numerical approach described in Chapter 3 is as computationally

efficient as possible, the parameters of the Particle Swarm Optimization used for force-finding were

studied to help minimize run-time, especially when a particular shape was unlikely to converge on

a solution.

The parameters considered were the size of the initial population (in this case, the

population is comprised of different sets of coefficient values 𝛾ⅈ ), the number of iterations it took

to converge, and the number of runs it took to find a solution. Since the numerical approach was

developed simply to find shapes that could be equilibrated, the exact values of the force densities

resulting in equilibrium were not evaluated for their magnitudes, but simply for whether or not they

satisfied equilibrium. The optimization seeks valid combinations of force densities, and therefore

would terminate early if a viable set of coefficients were found.

Figure B.1 summarizes the findings of this approach. Each combination of parameters was

evaluated 10,000 times. Different color bars correspond to different initial population values.

Clustered bars represent cases described in terms of ratios of the maximum number of iterations to

the maximum number of runs. (E.g., the left-most cluster reads 100:1, and therefore tested 100

maximum iterations with 1 run.)

Darkened bars with a red “x” indicate that some evaluation cases missed a known

equilibrated shape. The details of the missed cases are highlighted in boxes of colors corresponding

to the initial population size, where the first number is the number of times the known solution was

missed and the second number is the average time it took to run in a case where the solution was

not found. (E.g., the left-most case for an initial population of 1,000, being blue, missed 4,813 out

of 10,000 cases and took an average of 0.277 seconds to run in a missed case.)

169
Figure B.1: Effects of various parameters in PSO force-finding and their effects on average run time and convergence upon a known solution.
 Based on the results of Figure B.1, an initial population of 1,000 with 100 maximum

iterations and 50 maximum runs was subsequently used for this force finding method. In all cases,

no more than 14 runs were required to converge; however, 50 was selected as the maximum number

of runs, as a higher number of runs for this lower population value does not significantly increase

run time, and may be necessary for more complex shapes resulting in more coefficients as variables.

It is worth noting that this force finding approach, as implied by the missed cases of Figure

B.1 is not exhaustive. It is possible for equilibrated shapes to be missed regardless of the parameters

selected. However, the results showed sufficient success in finding a broad set of shapes for all the

examples tested in Chapter 3. The resulting solution sets are meant to be representative samples,

not exhaustive sets of solutions.

In this sense, this approach is more than sufficient to find a representative sample of

solutions. Furthermore, if a particular shape is highly desirable, the parameters of the PSO method

can be modified to increase the likelihood of finding a solution (e.g., larger initial population, more

runs etc.) if one exists.

B.1 Modification of the PSO Force-Finding for a Symmetric Shape

In some cases, it is possible to modify the fitness function in Eq. 3.7 (shown again in Eq.

B.1) to encourage specific types of solution sets. In particular, symmetric shapes are studied

throughout the literature, as their symmetry simplifies the form finding process, and further can

simplify the manufacturing process by reducing the number of member types. Symmetric force

densities, to go with the symmetric shapes, can also be desirable so that members experience similar

loading conditions, rather than a-symmetric cases where particular members may experience

significantly more stress or strain than other members in the tensegrity.

 172
̃)
− min(𝒒
𝑓𝑖𝑡𝑛ⅇ𝑠𝑠 = (𝐵. 1)
̃𝑇 𝒒
√𝒒 ̃

This approach is used for the triplex module in the flat shape (where the top and bottom

faces are parallel). Based on which members are shared between adjacent triplexes, 6 member

groups are formed: (1) shared cables on the bottom face, (2) outer, non-shared cables on the bottom

face, (3) top face cables, (4) shared supporting cables, (5) non-shared supporting cables, and (6)

struts. These groups are shown in Figure B.2.

Figure B.2: Groupings used for a triplex module

To encourage finding solutions which group these force densities together, the sum shown in Eq.

B.2 is added to the fitness function defined by Eq. B.1.

𝑁1 𝑁2 𝑁3 𝑁4 𝑁5

∑ (𝑞ⅈ1 − ̅̅̅)
𝑞1 + ∑ (𝑞ⅈ2 − ̅̅̅)
𝑞2 + ∑ (𝑞ⅈ3 − ̅̅̅)
𝑞3 + ∑ (𝑞ⅈ4 − ̅̅̅)
𝑞4 + ∑ (𝑞ⅈ5 − ̅̅̅)
𝑞5
ⅈ1 =1 ⅈ2 =1 ⅈ3 =1 ⅈ4 =1 ⅈ5 =1
𝑁6

+ ∑ (𝑞ⅈ6 − ̅̅̅)
𝑞6 (𝐸𝑞. 𝐵. 2)
ⅈ6 =1

172
 173
where number subscripts indicate which group the member i is in, N is the number of members in

the group, and 𝑞̅ indicates the average value of q in that group. Sets of force densities where

individual members deviate from the average of their group result in higher values of Eq. B.2. Thus,

by adding this to the fitness function to be minimized in B.1, the final results are encouraged

towards this symmetric grouping. Eq. B.2 could be modified for any number of groupings as

required by the structure.

In the case of the triplex module, this approach found a variety of groupings, a small sample

of which are shown in Table B.1. The results are scaled so that the average force density of group

1 is always equal to 1. Interestingly, the ratios of force densities for groups 3, 4, 5 and 6 relative to

each other are always constant, while those of group 2 (corresponding to the outer-most members

of the bottom face of the triplex module) can have a variety of values. By nature of the randomized

initialization and perturbations in PSO, the final results of any two cases seeking force densities

can vary. Using something like Eq. B.2, however, adds some control over the types of solutions

found.

Group 1 1 1 1 1 1 1 1 1 1
Group 2 0.5950 0.8334 1.4918 0.2728 0.1086 0.0233 0.1185 0.0624 0.1354
Group 3 1.5950 1.8334 2.4918 1.2728 1.1086 1.0233 1.1185 1.0624 1.1354
Group 4 3.1899 3.6669 4.9836 2.5456 2.2171 2.0465 2.2370 2.1248 2.2709
Group 5 1.5950 1.8334 2.4918 1.2728 1.1086 1.0233 1.1185 1.0624 1.1354
Group 6 1.5950 1.8334 2.4918 1.2728 1.1086 1.0233 1.1185 1.0624 1.1354
Table B.1: Sample of symmetric force densities found for a triplex module

173
 Bibliography

[1] Gómez-Jáuregui, V. (2009), “Controversial Origins of Tensegrity,” Proceedings of the IASS Symposium
2009.

[2] C. Sultan, Tensegrity: 60 Years of Art, Science, and Engineering. In Hassan Aref and Erik van
der Giessen, editors: Advances in Applied Mechanics, Vol. 43, Burlington: Academic Press, 2009,
pp. 69-145. ISBN: 978-0-12-374813-3

[3] Ingber, D.E., Jameison, J.D.: Cells as tensegrity structures: architectural regulation of
histodifferentiation by physical forces transduced over basement membrane. In: Gene Expression during
Normal and Malignant Differentiation. (eds. Anderson LC, Gahmberg GC, Ekblom P), Academic Press,
Orlando, FL, pp. 13-32, 1985.

[4] Yildiz, K., “Cable Actuated Tensegrity Structures for Deployable Space Booms with Enhanced
Stiffness,” Ph.D. Dissertation, Aerospace Engineering Dept., Pennsylvania State Univ., State College, PA,
2018

[5] Chen, M., Goyal, R., Majji, M. and R. E. Skelton (2020) “Deployable Tensegrity Lunar Tower,”
arXiv:2009.12958. https://arxiv.org/pdf/2009.12958.pdf

[6] Tibert, G. and S. Pellegrino (2003) “Deployable tensegrity masts,” in 44th AIAA/ASME/ASCE/AHS/
ASC Structures, Structural Dynamics, and Materials Conference, p. 1978.

[7] Zawadzki, A. and A. Al Sabouni-Zawadzka (2020) “In Search of Lightweight Deployable Tensegrity
Columns,” Applied Sciences, 10(23), 8676.

[8] Knight, B., Duffy, J., Crane, C., and J. Rooney (2001) “Geometry and Mechanics of Deployable
Tensegrity Structures,” in 19th AIAA Applied Aerodynamics Conference.

[9] Yang, S. and C. Sultan (2016) “Modeling of tensegrity-membrane systems”, International Journal of
Solids and Structures, 82, pp 125-143.

[10] Chen, M., Goyal, R., Majji, M. and R. E. Skelton (2020), “Design and analysis of a growable artificial
gravity space habitat” Aerospace Science and Technology, 106, 106147.

[11] Vueve, N., Sychterz, A. C., and I. F. C. Smith (2017), “Adaptive control of a deployable tensegrity
structure” Engineering Structures, 152, pp 14-23.

[12] Ganga P.L., Micheletti A., Podio-Guidugli P., Scolamiero L., Tibert G., Zolesi V. (2016) “Tensegrity
Rings for Deployable Space Antennas: Concept, Design, Analysis, and Prototype Testing”. In: Frediani A.,
Mohammadi B., Pironneau O., Cipolla V. (eds) Variational Analysis and Aerospace Engineering. Springer
Optimization and Its Applications, vol 116. Springer, Cham. https://doi.org/10.1007/978-3-319-45680-
5_11

[13] Tibert, A. G., and S. Pellegrino (2002), “Deployable Tensegrity Reflector for Small Satellites,”
Journal of Spacecraft and Rockets, 39(5), pp 701-709.
 175
[14] Fest, E., Shea, K., Domer, B. and Smith, I. (2003) “Adjustable Tensegrity Structures,” Journal of
Structural Engineering, 129(4), pp. 515-526.

[15] Roffman, K. M., “Shape Change and Structural Performance of Cable-Actuated Cylindrical
Tensegrities” Master thesis, Aerospace Engineering Dept., Pennsylvania State Univ., State College , PA,
2019

[16] Hemal, A. K. and M. Menon. Robotics in Genitourinary Surgery. Springer International publishing,
2018.

[17] “Truss.” Encyclopædia Britannica, Encyclopædia Britannica, Inc., 20 Oct. 2014,

www.britannica.com/technology/truss-building.

[18] “Unimate - The First Industrial Robot.” Robotics Online, www.robotics.org/joseph-

engelberger/unimate.cfm.

[19] Kershner, R. B. and R.E. Fischell (2017) “Gravity-Gradient Stabilization of Earth Satellites”, in IFAC
Proceedings, 2(1): pp 249-266.

[20] “Rovers.” Rovers - Exploring the Planets | National Air and Space Museum,
airandspace.si.edu/exhibitions/exploring-the-planets/online/tools/rovers.cfm.

[21] S. D. Guest (2011), “The stiffness of tensegrity structures”, IMA Journal of Applied Mathematics, 76,
pp 57-66.

[22] Li, X. et al (2020), “Modal analysis method for tensegrity structures via stiffness transformation from
node space to task space,” Engineering Structures, 203, 109881.

[23] Ashwear, N. and A. Eriksson (2017) “Vibration health monitoring for tensegrity structures,”
Mechanical Systems and Signal Processing, 85 pp. 625-637.

[24] Zhang, J. Y., and M. Ohsaki (2012) “Self-equilibrium and stability of regular truncated tetrahedral
tensegrity structures,” Journal of Mechanics and Physics of Solids, 60(10) pp. 1757-1770.

[25] Al Sabouni-Zawadzka, A., and W. Gilewski (2018) “Inherent properties of smart tensegrity structures”
Applied Sciences, 8(5). DOI:10.3390/app8050787

[26] SunSpiral V., Agogino A., and D. Atkinson (2015), “Super Ball Bot – Structures for Planetary
Landing and Exploration” from Final Report Phase II, NASA Innovative Advanced Concepts (NIAC)
Program.

[27] Gebara, C.A., Carpenter, K.C. and A. Woodmansee, (2019) “Tensegrity Ocean World Landers”, in
AIAA SciTech 2019 Conference, https://doi.org/10.2514/6.2019-0868.

[28] Vega, J. C. et al (2020), “Influence of Elastomeric Tensioned Members on the Characteristics of

Compliant Tensegrity Structures in Soft Robotic Applications” Procedia Manufacturing, 52, pp 289-294.

[29] Liu, K., et al (2017), “Programmable Deployment of Tensegrity Structures by Stimulus-Responsive

Polymers”, Scientific Reports, 7, 3511.

[30] A. P. Sabelhaus, H. Zhao, E. L. Zhu, A. K. Agogino and A. M. Agogino, (2021) "Model-Predictive

Control With Inverse Statics Optimization for Tensegrity Spine Robots," IEEE Transactions on Control
Systems Technology, 29(1), pp. 263-277.

175
 176
[31] Sabelhaus, A. P. et al (2018), “Design, Simulation and Testing of a Flexible Actuated Spine for
Quadruped Robots,” arXiv:1804.06527. https://arxiv.org/abs/1804.06527

[32] Tibert, A. and S. Pellegrino (2011) “Review of form-finding methods for tensegrity structures,”
International Journal of Space Structures, 26(3), pp. 241–255.

[33] Estrada, G. G., Bungartz, H. J. and C. Mohrdieck (2006), “Numerical form-finding of tensegrity
structures,” International Journal of Solids and Structures, 43(22-23): pp 6855-6868.

[34] Tran, H. C. and J. Lee (2010) “Advanced form-finding of tensegrity structures” Computers &
Structures, 88(3-4): pp 237-246.

[35] Lu, J. Y., Li, N. and G. Shu (2015) “Form-finding analysis of irregular tensegrity structures by matrix
iteration” Advanced Steel Construction, 11(4): pp 507-516.

[36] Tran, H. C. and J. Lee (2013) “Form-finding of tensegrity structures using double singular value
decomposition” Engineering with Computers, 29: pp 71-86.

[37] Zhang, L., Bernard, M. and R. Motro (2006) “Form-finding of nonregular tensegrity systems” Journal
of Structural Engineering, 132(9): pp 1435-1440.

[38] Zhang, L. et al. (2014) “Stiffness matrix based form-finding method of tensegrity structures”
Engineering Structures, 58: pp 36-48.

[39] Xu, X. and Y. Luo (2010) “Form-finding of nonregular tensegrities using a genetic algorithm”
Mechanics Research Communications, 37: pp 85-91.

[40] Koohestani, K. (2012) “Form-finding of tensegrity structures via genetic algorithm”, International
Journal of Solids and Structures, 49: pp 739-747.

[41] Lee, S. and J. Lee (2014) “Form-finding of tensegrity structures with arbitrary strut and cable
members” International Journal of Mechanical Sciences, 85: pp 55-62.

[42] Gan, B. S., et al. (2015) “Node-based genetic form-finding of irregular tensegrity structures”
Computers and Structures, 159: pp 61-73.

[43] Yildiz, Y. and G. A. Lesieutre (2020) “Sizing and prestress optimization of class-2 tensegrity
structures for space boom aplications,” Engineering with Computers. https://doi.org/10.1007/s00366-020-
01111-x

[44] Chen, Y. et al. (2020) “Feasible prestress nodes for cable-strut structures with multiple self-stress
states using particle swarm optimization” Journal of Computing in Civil Engineering, 34(3): 04020003.

[45] Roffman, K. M. and G. A. Lesieutre (2021) “Morphing Tensegrity Space Platforms”, in AIAA SciTech
2021 Forum. https://doi.org/10.2514/6.2021-0428

[46] Chen, Y. et al. (2020) “A hybrid symmetry-PSO approach to finding the self-equilibrium
configurations of prestressable pin-jointed assemblies” Acta Mechanica, 231: pp 1485-1501.

[47] Li, X. et al. (2010) “A Monte Carlo form-finding metod for large scale regular and irregular tensegrity
structures” International Journal of Solids and Structures, 47(14-15): pp 1888-1898.

[48] Williamson, D., Skelton, R. E. and J. Han (2003) “Equilibrium conditions of a tensegrity structure”
International Journal of Solids and Structures, 40: pp 6347-6367.

176
 177

[49] Masic, M., Skelton, R. E. and P. E. Gill (2005) “Algebraic tensegrity form-finding” International
Journal of Solids and Structures, 42: pp 4833-4858.

[50] Zhang, L. et al. (2018) “Analytical form-finding of tensegrities using determinant of force-density
matrix” Composite Structures, 189: pp. 87-98.

[51] Koohestani, K. and S. D. Guest (2013) “A new approach to the analytical and numerical form-finding
of tensegrity structures” International journal of Solids and Structures, 50: pp. 2995-3007.

[52] Oh, C. L. et al. (2018) “Shape change analysis of tensegrity models” Latin American Journal of Solids
and Structures, 16(7): e221. https://doi.org/10.1590/1679-78255407

[53] Lai, G., Plummer, A. and D. Cleaver (2020) “Distributed actuation and control of a morphing
tensegrity structure” Journal of Dynamic Systems, Measurement and Control, 142: 071006.

[54] Chen, M., Liu, J. and R. E. Skelton (2020) “Design and control of tensegrity morphing airfoils”
Mechanics Research Communications, 103: 103480

[55] Putzer, M. et al. (2016) “A numerical study to determine the effect of ligament stiffness on kinematics
of the lumbar spine during flexion”, BMC Musculoskeletal Disorders, 16. DOI: 10.1186/s12891-016-0942-
x

[56] Atarer, F., Korkmaz, K. and G. Kiper (2017) “Design alternatives of altmann linkages,” International
Journal of Computational Methods and Experimental Measurements, 5(4): pp 495-503. DOI:
10.2495/CMEM-V5-N4-495-503

[57] T. Tachi (2011) “Freeform Rigid-Foldable Structure using Bidirectionally Flat-Foldable Planar
Quadrilateral Mesh”, in Advances in Architectural Geometry 2010, pp. 87-102. DOI: 10.1007/978-3-7091-
0309-8_6

[58] Feng, X. and S. H. Guo (2016), “Geometrical nonlinear elasto-plastic analysis of tensegrity systems
via the co-rotational method” Mechanics Research Communications, 79: pp 32-42. DOI:
10.1016/j.mechrescom.2016.12.003

[59] Osikowicz, N.S. “Shape Control of Tendon-Actuated Tensegrity Structures”, Master thesis, Aerospace
Engineering Dept., Pennsylvania State Univ., State College , PA, 2021

[60] Weisstein, Eric W. "Sphere-Sphere Intersection." From MathWorld--A Wolfram Web Resource.
https://mathworld.wolfram.com/Sphere-SphereIntersection.html

[61] “Telescope Overview.” [Online; accessed 28-September-2022]. URL

https://webbtelescope.org/news/webb-science-writers-guide/telescope-overview.

[62] “Webb Key Facts.” [Online; accessed 29-September-2022]. URL

https://webb.nasa.gov/content/about/faqs/facts.html

[63] Betz, Laura, and Rob Gutro. “NASA's James Webb Space Telescope Secondary Mirror Installed.”
Edited by Lynn Jenner, NASA, NASA, 7 Mar. 2016. [Online; accessed 29-September-2022]. URL
https://www.nasa.gov/feature/goddard/2016/nasas-james-webb-space-telescope-secondary-mirror-installed.

[64] “NASA’s Webb Telescope’s Last Backbone Component Completed.” [Online; accessed 29-
September-2022]. URL

177
 178
https://www.nasa.gov/mission_pages/webb/news/backplane.html#:~:text=The%20backplane%20support%
20frame%20also,primary%20mirror%20backplane%20support%20structure.

[65] Dalilsafaei, S., A. Eriksson, and G. Tibert (2012) “Improving bending stiffness of tensegrity booms,”
International Journal of Space Structures, 27(2-3), pp. 117–129.

[66] Murphey, T., Booms, and Trusses (2006) “Recent advances in gossamer spacecraft (Progress in
Astronautics and Aeronautics),” AIAA, 212.

[67] Peterson, J. (2019) “L2 Station Keeping Maneuver Strategy for the James Webb Space Telescope” in
AAS/AIAA Astrodynamics Specialist Conference 2019.

[68] “JWST in L2”. [Online; accessed 1-October-2022]. URL

https://www.ucolick.org/~gdi/early_jwst/now.html

[69] “Unfurlable Mesh Reflector Antennas”. [Online; accessed 30-September-2022]. URL

https://www.l3harris.com/all-capabilities/unfurlable-mesh-reflector-antennas

[70] Morterolle, S. et al. (2015), “Modal behavior of a new large reflector conceptual design.” Aerospace
Science and Technology, Elsevier, 2015, 42, pp. 74-79.

[71] Huang, H. et al. (2020), “Structural Dynamic Improvement for Petal-Type Deployable Solid-Surface
Reflector Based on Numerical Parameter Study”, Applied Sciences, 10(18), 6560.

[72] Hyde, T. et al (2004), “Integrated modeling activities for the James Webb Space Telescope: optical
jitter analysis”, in Proceedings of SPIE- the Inernational Society for Optical Engineering, 5487, pp 588-
599.

[73] “AstroMesh Unfurlable Mesh Antenna”. [Online; accessed 30-September-2022]. URL

https://www.northropgrumman.com/wp-content/uploads/AstroMesh-DataSheet.pdf

[74] Alise Fisher, “Mirror, Mirror… On Its Way!” 12 January 2022. [Online; accessed 1-October-2022].
URL https://blogs.nasa.gov/webb/2022/01/13/mirror-mirroron-its-way/

[75] Guerrero-Gonzalez, J. M., et al. (2022) “Mahalanobis distance tractometry (MaD-Tract)—a framework
for personalized white matter anomaly detection applied to TBI”, NueroImage, 260: 119475

[76] G. Verdier and A. Ferreira, (2011) "Adaptive Mahalanobis Distance and k -Nearest Neighbor Rule for
Fault Detection in Semiconductor Manufacturing," in IEEE Transactions on Semiconductor Manufacturing,
24(1): pp. 59-68, doi: 10.1109/TSM.2010.2065531.

[77] Wen, S. et al (2022) "Rotated Object Detection via Scale-Invariant Mahalanobis Distance in Aerial
Images," in IEEE Geoscience and Remote Sensing Letters, 19: pp. 1-5, 2022, Art no. 6514505, doi:
10.1109/LGRS.2022.3197617.

[78] Zhou, Y. et al. (2015) “Damage detection in structures using a transmissibility-based Mahalanobis
distance”, Structural Control Health Monitoring, 22: 1209– 1222. doi: 10.1002/stc.1743.

[79] Sarmadi, H. and A. Karamodin (2020) “A novel anomaly detection method based on adaptive
Mahalanobis-squared distance and one-class kNN rule for structural health monitoring under environmental
effects” Mechanical Systems and Signal Processing, 140: 106495.
[80] Gallego, G. et al (2013) "On the Mahalanobis Distance Classification Criterion for Multidimensional
Normal Distributions," in IEEE Transactions on Signal Processing, 61(17): pp. 4387-4396, doi:
10.1109/TSP.2013.2269047.

178
 179

[81] Modenini D., et al (2022) “Relationship between collision probability, Mahalanobis distance, and
Confidence Intervals for Conjunction Assessment” Journal of Spacecraft and Rockets, 59(4): pp 1125-1134

[82] Porta, J. M. and S. Hernández-Juan (2016) “Path planning for active tensegrity structures”, International
Journal of Solids and Structures, 78-79: pp 47-56.

[83] Juan S. H. and J. M. Mirats Tur (2008) "A method to generate stable, collision free configurations for
tensegrity based robots," IEEE/RSJ International Conference on Intelligent Robots and Systems, 2008, pp.
3769-3774, doi: 10.1109/IROS.2008.4650881.

179
 Vita

Kaila Roffman was born in 1995 in New York State, USA. She received her Bachelor’s degree in

Mechanical and Aerospace Engineering in 2017 from Rutgers University. She continued her studies

to receive a Master’s Degree in Aerospace Engineering from Pennsylvania State University in

2019. In 2020, she continued at Penn State University, starting the pursuit of her Ph.D. in Aerospace

Engineering under supervision of Dr. George A. Lesieutre. She completed her Ph.D. in 2022. Her

research areas of interest are tensegrity structures, deployable structures, and morphing spacecraft

structures.