!

"# $ % & $ ## # $ "

"' $% # #' " (
! " #

$ % &' (
" !) * +" $ , ' !

$$' $' $ - *' -' " !!) . . '

* / " .! " " 0 ! . '

+ + '

# $

.+ '

) $ $$ 1 % 1 * / " .! " " 0 ! . '

* $$ ) $# 1 $ " $ $ 1

!( & ! ++ %$ # +,-.

*/$' 0

1 $ ## 1 % 1 2 * /" 0 ! . '
" $/ 2 1 3'" & $ '1 * 4 % .,
$ 1 % 1 ' % 1 .'
) $ $$ 1 % 1 2 * /" 0 ! . '
% '1 $ 1 0 " ' 5
* $$ ) $# 1 $ " $ $ 1 ', !
 i

Abstract

In the past two decades the work of a growing number of researchers in

robotics focused on a particular group of machines, belonging to the family of
parallel manipulators: the cable robots. Although these robots share several
theoretical elements with the better known parallel robots, they still present
completely (or partly) unsolved issues. In particular, the study of their
kinematics, already a difficult subject for conventional parallel manipulators,
is further complicated by the non-linear nature of cables, which can exert
only efforts of pure traction. The work presented in this thesis therefore
focuses on the study of the kinematics of these robots and on the development
of numerical techniques able to address some of the problems related to
it. Most of the work is focused on the development of an interval-analysis
based procedure for the solution of the direct geometric problem of a generic
cable manipulator. This technique, as well as allowing for a rapid solution
of the problem, also guarantees the results obtained against rounding and
elimination errors and can take into account any uncertainties in the model
of the problem. The developed code has been tested with the help of a small
manipulator whose realization is described in this dissertation together with
the auxiliary work done during its design and simulation phases.
ii

Sommario

Negli ultimi decenni il lavoro di una parte sempre maggiore di ricercatori che
si occupano di robotica si è concentrato su un particolare gruppo di robot
appartenenti alla famiglia dei manipolatori paralleli: i robot a cavi. No-
nostante i numerosi studi al riguardo, questi robot presentano ancora oggi
numerose problematiche del tutto (o in parte) irrisolte. Lo studio della loro
cinematica nello specifico, già complesso per i manipolatori paralleli tradi-
zionali, è ulteriormente complicato dalla natura non lineare dei cavi, i quali
possono esercitare sforzi di sola trazione. Il lavoro presentato in questa tesi si
concentra dunque sullo studio della cinematica dei robot a cavi e sulla messa
a punto di tecniche numeriche in grado di affrontare parte delle problemati-
che ad essa legate. La maggior parte del lavoro è incentrata sullo sviluppo
di una procedura per la soluzione del problema geometrico diretto di un ge-
nerico manipolatore a cavi basata sull’analisi per intervalli. Questa tecnica
di analisi numeirica, oltre a consentire una rapida soluzione del problema,
permette di garantire i risultati ottenuti in caso di errori di cancellazione e
arrotondamento e consente di considerare eventuali incertezze presenti nel
modello del problema. Il codice sviluppato è stato testato attraverso un pic-
colo prototipo di manipolatore a cavi la cui realizzazione, avvenuta durante
il percorso di dottrato, è descritta all’interno dell’elaborato unitamente al
lavoro collaterale svolto durante la fase di progettazione e simulazione.
 iii

Résumé

Pendant les dernières décennies, le travail d’une partie toujours croissante

de chercheurs qui s’occupent de robotique s’est focalisé sur un groupe spéci-
fique de robots qui fait partie de la famille des manipulateurs parallèles : les
robots à câbles. Malgré les nombreux études que l’on a consacré à ce sujet,
ces robots présentent encore aujourd’hui plusieurs problématiques complète-
ment ou partiellement irrésolues. En particulier l’étude de leur cinématique,
qui se révèle déjà complexe pour les manipulateurs parallèles traditionnels,
est rendu encore plus compliqué par la nature non linéaire des câbles qui
peuvent seulement exercer des efforts de traction. Le travail présenté dans
ma thèse concentre donc son attention sur l’étude de la cinématique des
robots à câbles et sur la mise au point de techniques numériques capables
d’aborder une partie des problématiques liées à cela. La plupart du travail
se concentre sur l’élaboration d’un algorithme pour la résolution du pro-
blème géométrique direct d’un manipulateur à câbles général qui se fonde
sur l’analyse par intervalles. Cette technique d’analyse permet non seulement
de résoudre rapidement le problème mais également de garantir les résultats
obtenus en cas d’erreur de cancellation et d’arrondi et de prendre en considé-
ration les incertitudes éventuellement presentes dans le modèle du problème.
Le code développé a été testé grâce à un petit prototype de manipulateur
à câbles dont la réalisation, qui a eu lieu pendant le parcours de doctorat,
est décrite à l’intérieur du devoir en accord avec la phase de conception du
projet et de simulation.
Contents

Prologue in English xiii

Prologue in Italian xvii

Prologue in French xxi

1 Introduction 1
1.1 Serial and parallel manipulators . . . . . . . . . . . . . . . . . 2
1.2 Cable robots . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 History and main applications . . . . . . . . . . . . . . 7
1.2.2 Main components . . . . . . . . . . . . . . . . . . . . . 12
1.3 Interval analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.1 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.2 Evaluation function . . . . . . . . . . . . . . . . . . . 19
1.3.3 Problem-solving algorithm . . . . . . . . . . . . . . . . 21
1.3.4 Application in robotics . . . . . . . . . . . . . . . . . . 24
1.3.5 Advantages and disadvantages . . . . . . . . . . . . . . 27

2 Geometrico-static model for CDPRs 29

2.1 General definitions . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Working assumptions . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Geometrico-static model . . . . . . . . . . . . . . . . . . . . . 33
2.3.1 Fundamental geometric and static equations . . . . . . 34
2.3.2 Parameterization . . . . . . . . . . . . . . . . . . . . . 35
2.3.3 DGP for robots with one cable in tension . . . . . . . 39
2.3.4 DGP for robots with two cables in tension . . . . . . . 40
2.3.5 DGP for robots with three cables in tension . . . . . . 43
2.3.6 DGP for robots with four or five cables in tension . . . 44
2.3.7 DGP for robots with six cables in tension . . . . . . . 46

v
vi CONTENTS

3 DGP solving procedure 47

3.1 Code structure . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.1.1 Initial search domain . . . . . . . . . . . . . . . . . . . 48
3.1.2 Evaluation operator . . . . . . . . . . . . . . . . . . . 52
3.1.3 Filter operator . . . . . . . . . . . . . . . . . . . . . . 52
3.1.4 Influence on the bisection process . . . . . . . . . . . . 56
3.2 Code improvements . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2.1 Uniqueness of the solution . . . . . . . . . . . . . . . . 57
3.2.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.3 Parallel implementation . . . . . . . . . . . . . . . . . 64
3.3 Code structure: beta version . . . . . . . . . . . . . . . . . . . 65
3.3.1 Initialization . . . . . . . . . . . . . . . . . . . . . . . 65
3.3.2 Description of the algorithm . . . . . . . . . . . . . . . 66
3.3.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4 Experiments and Results 71

4.1 Description of the teaching prototype . . . . . . . . . . . . . . 71
4.1.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.1.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Test of the DGP solving procedure . . . . . . . . . . . . . . . 84
4.2.1 Test on a CDPR with 3 cables in tension . . . . . . . 85
4.2.2 Test on the Cogiro prototype geometry . . . . . . . . . 85
4.2.3 Test on the Marionet-VR prototype . . . . . . . . . . 86
4.2.4 Test on the teaching prototype . . . . . . . . . . . . . 86
4.2.5 Discussion of results . . . . . . . . . . . . . . . . . . . 87

5 Conclusions 97
5.1 Future perspectives . . . . . . . . . . . . . . . . . . . . . . . . 98
List of Figures

1.1 Open and closed kinematic chains. . . . . . . . . . . . . . . . 3

1.2 The Kuka R 7-dof robotic arm. . . . . . . . . . . . . . . . . . 3
1.3 The tire testing machine developed by Gough on the left, and
a modern flight simulator on the right [Airbus 2010]. . . . . . 4
1.4 The Delta IRB 360 FlexPicker by ABB R . . . . . . . . . . . . 5
1.5 An illustration of a Greek crane during building operations. . 7
1.6 On the left an image from the “de Architectura” of Marcus
Vitruvius Pollio and on the right an ancient block-and-tackle
system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.7 Bowden cables controlling an automobile throttle [Ehardt 2015]. 8
1.8 On the left the first prototype of the NIST RoboCrane and
on the right its version for aircraft painting and maintenance
([NIST 2006]). . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.9 On the left the RoboCrane as a controllable welding platform
and on the right its version for ship building and maintenance
operations [Bostelman et al. 1999]. . . . . . . . . . . . . . . . 10
1.10 On the left a drawing representing the functioning of the
AACTS ([Lee 1992]) and on the right the Large Vessel In-
terface Lift On/Lift Off (LVI Lo/Lo) Crane for ship-to-ship
container transfer at sea (source: ONR, USA) . . . . . . . . . 10
1.11 The CoGiRo robot on the left [Alexandre Dit Sandretto 2013]
and the ReelAx8 on the right: the two prototypes available at
the LIRMM [Lamaury et al. 2012]. . . . . . . . . . . . . . . . 11
1.12 (a) The Fraunhofer Institute prototype for power solar plant
construction [Pott et al. 2010], (b) the portable rescue crane
Marionet Rescue by Merlet [Merlet & Daney 2010], (c) the
cable-driven platform for automated warehouses from the Duis-
burg University [Bruckmann et al. 2013]. . . . . . . . . . . . . 12

vii
viii LIST OF FIGURES

1.13 On the left, the Athlete vehicle during a low gravity test
[Wilcox 2012] and on the right a rendering of the FAST tele-
scope [Nan 2006]. . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.14 (a) The haptic interface Spidar G-G [Murayama et al. 2004]
and (b) the Marionet Rehab by Merlet [Merlet 2010]. . . . . . 14
1.15 The SkyCam [SkyCam 2015]. . . . . . . . . . . . . . . . . . . 14
1.16 (a) The latest prototype of winches developed by the Fraun-
hofer Institute IPA [Schmidt et al. 2015] and the cable ac-
tuators with linear motors mounted on the Marionet Rehab
[Merlet 2010]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.17 A detail of the winches developed by the Fraunhofer Institute. 17

2.1 The 8-cable haptic device developed by the Fraunhofer Insti-

tute [Yang Ho et al. 2015]. . . . . . . . . . . . . . . . . . . . . 30
2.2 Some wire-end terminals. . . . . . . . . . . . . . . . . . . . . 32
2.3 Assumed position of the anchor points on the platform. . . . 33
2.4 Displacements of exit points when orientable roller guidances
are employed [Bruckmann et al. 2008a]. . . . . . . . . . . . . 34
2.5 The geometric model of a cable driven parallel robot with n
cables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Geometric model of the DGP with two cables in tension. . . . 42

3.1 The four different cases for the possible location of Bi accord-
ing to the location of Aj , the leg length ρj and the distance
between Bi and Bj . The admissible region zone for Bi is
represented by the dashed line. . . . . . . . . . . . . . . . . . 49
3.2 The solution box after the inflation process. . . . . . . . . . . 63

4.1 The control user interface of the teaching prototype. . . . . . 75

4.2 The large wrapping angle of the cable on the right on the
fishing rod top eye cause a non uniform tension along the cable. 76
4.3 The moving platform before (a) and after (b) the manual re-
distribution of cable tensions . . . . . . . . . . . . . . . . . . 76
4.4 Wrench feasible workspace determination for assigned orien-
tation and zO′ of the moving platform. . . . . . . . . . . . . . 82
4.5 Detail of the two lines bisecting the angles formed by the two
tangents of γi and γj in λij . . . . . . . . . . . . . . . . . . . . 83
4.6 Validation of Test 3 on the large scale robot MARIONET-
VR: (a) configuration No. 1 with 6 cables in tension; (b)
configuration No. 3 with 6 cables in tension. . . . . . . . . . . 86
LIST OF FIGURES ix

4.7 From Table 4.6: (a) configuration No. 1, (b) Configuration

No. 3, (c) configuration No. 4, (d) configuration No. 5. . . . . 87
List of Tables

4.1 Properties of the Phidgets R 3260E_0 gear motors. . . . . . . 73

4.2 Properties of the Phidgets R 1065_0 single motor driver. . . . 74
4.3 Solution set for a CDPR suspended by 3 cables. . . . . . . . . 90
4.4 Complete solution set for the the 8 cable robot Cogiro. Plat-
form anchor points are not coplanar. . . . . . . . . . . . . . . 92
4.5 Solution set for the CDPR suspended by 6 cables MARIONET-
VR: cable anchor points on the platform are not coplanar. . . 93
4.6 Solution set for a CDPR suspended by 6 cables: all cable
anchor points on the platform are coplanar. . . . . . . . . . . 95

xi
Prologue in English

In the late 80s some researchers began to explore the possibility of replacing
parallel manipulators’ rigid-body actuators with cables, in order to overcome
the main weaknesses of this family of robots, such as the limited workspace,
a considerably complex manufacturing process and maintenance, and high
costs. These robots, defined cable-driven parallel robots (CDPRs), or simply
cable robots, in addition to the advantages deriving from parallel architecture,
thanks to cable flexibility can also have extremely large workspaces, higher
load capacity, better kinematic performance and enhanced ease of design
and assembly. A more detailed presentation of cable robots, the literature
on them and the description of the most famous prototypes is reported in
the first part of Chapter 1 of this thesis.
However, the already difficult control and kinematic analysis of parallel
robots (that took more than 30 years to be completely solved) is for these
manipulators further complicated by the inability of cables to withstand
compressive loadings. This implies that certain poses (i.e. positions and
orientations) and/or loading conditions cannot be achieved because some
cables can suddenly become slack, thus modifying the robot configuration,
its static and/or dynamic equilibrium and compromising its control. The
work done during the three years of the Ph.D. program primarily dealt with
the study of these problems and especially on direct kinematics issues.
Direct or forward kinematics refers to compute the end-effector pose from
specified values of the joint variables, in this case cable lengths. If a robot
with n ≥ 6 cables and at least m = 6 of them are in tension, the 6 end-
effector degrees of freedom can be determined by the equations deriving
from the geometric constraints imposed by taut cables. When m < 6 these
geometric conditions are not sufficient to completely define robot pose, which
will be determined by the platform static equilibrium. In order to fulfill
equilibrium conditions, the 6 equations of the statics and the m unknown
cable tensions have to be added to the m geometric equations. Moreover only
solutions having cable tensions greater than 0 have to be taken into account.

xiii
xiv PROLOGUE IN ENGLISH

Consequently the kinematic analysis of these manipulators kinematics has

to be conducted together with their statics and taking into account the
nonlinear constraints imposed by cables. Indeed, in order to control a cable
robot in a completely safe manner it is indeed necessary to ensure that at
all times the manipulator configuration does not change due to a loss of
tension in one or more cables. So, all the possible configurations that can
be obtained if one or more cables become slack must be taken into account.
These amount, for a generic n-cable robot, to
n
X n!
N=
m!(n − m)!
m=1

These are the main reasons that make the already challenging kinematic
analysis of parallel robots even more complex for CDPRs. These issues,
together with the description of the geometric models of cable robots adopted
in this manuscript, are discussed in Chapter 2 along with an overview of the
literature devoted to direct kinematics problems of cable robots and related
methods currently adopted to deal with it.
Chapter 3 presents the work that led to the development of a problem-
solving routine that numerically solves the direct kinematic problem of a
generic cable driven parallel robot. This method can completely solve the
problem, and its results are guaranteed against numerical errors thanks to
interval analysis computational techniques. Interval analysis is a branch
of numerical analysis that includes a variety of tools that solve numeric
problems without the risk of errors due to elimination or rounding effects.
While a basic description of interval analysis and the methods used in this
thesis is reported in the second half of Chapter 1, Chapter 3 provides an
in-depth analysis of the structure of the implemented method, its different
versions and proposes possible strategies for further developments and real-
time applications.
Chapter 4 describes the results obtained during the test phase of the dif-
ferent versions of the procedure and discusses them. The method developed
during this Ph.D. program showed very good performances. It allows for
a complete analysis of direct geometrico-static problem for a generic robot
with n cables with very low computation times. Thanks to interval analysis,
it also directly searches for real solutions with positive cable tensions and its
results are guaranteed against numerical errors. This chapter also describes
the realization of a small prototype that was built in order to test the im-
plementation of the on-line version of the algorithm within a true control
scheme.
 xv

Chapter 5 summarizes the main goals achieved during this work, and
presents the principal lines of development for future research.
Prologue in Italian

Verso la fine degli anni ’80 alcuni ricercatori in ambito robotico cominciarono
ad esplorare la possibilità di sostituire i tradizionali attuatori a corpo rigido
dei manipolatori paralleli con cavi, in modo tale da superare i principali
svantaggi di questa famiglia di robot, come ad esempio il limitato spazio di
lavoro, la complessità dei processi di fabbricazione, montaggio e manuten-
zione e i costi elevati. Questi robot, definiti in inglese cable-driven parallel
robots (CDPRs), o semplicemente robot a cavi, oltre ai vantaggi derivanti
dall’architettura parallela, grazie alla flessibilità dei cavi sono in grado di op-
erare in spazi di lavoro molto vasti, posseggono maggiori capacità di carico,
migliori prestazioni cinematiche e sono di più facile progettazione e assem-
blaggio. Una dettagliata presentazione dei robot a cavi, della letteratura a
loro dedicata e la descrizione dei prototipi più famosi è riportata nella prima
parte del del Capitolo 1 di questa tesi di dottorato.
L’analisi cinematica e il controllo di questi robot, però, oltre alle prob-
lematiche già complesse proprie dei manipolatori paralleli (risolte solo dopo
più di 30 anni di ricerche), sono rese ancora più difficoltose dalla incapacità
dei cavi di sopportare sforzi di compressione. Ciò comporta che alcune pose
(ovvero posizioni ed orientamenti) e/o condizioni di carico non possono essere
raggiunte perché alcuni cavi diventano improvvisamente laschi, modificando
quindi la configurazione del robot, il suo equilibrio statico e/o dinamico e
compromettendo seriamente il controllo. Il lavoro svolto durante i tre anno
del percorso di dottorato è stato incentrato principalmente sullo studio di
queste problematiche e specialmente sugli aspetti della cinematica diretta.
Per cinematica diretta si intende il calcolare la posa del membro termi-
nale del robot per assegnati valori delle variabili di giunto, in questo caso
quindi le lunghezze dei cavi. Se si considera un robot con n ≥ 6 cavi di
cui almeno m = 6 in tensione, i 6 gradi di libertà dell’end-effector possono
essere determinati dalle equazioni derivanti dai vincoli geometrici imposti
dai cavi tesi. Quando m < 6 invece, queste condizioni geometriche non sono
sufficienti in numero a definire completamente la posa del robot, la quale

xvii
xviii PROLOGUE IN ITALIAN

sarà determinata solamente considerando anche l’equilibrio statico della pi-

attaforma. Per poter rispettare questa condizione, le 6 equazioni della statica
nelle m incognite rappresentate dalle tensioni nei cavi devono essere aggiunte
alle m equazioni derivanti dai vincoli geometrici. Inoltre solo le soluzioni in
cui le tensioni nei cavi sono positive sono da prendere in considerazione. Di
conseguenza l’analisi cinematica di questi manipolatori deve essere risolta
unitamente alla statica e nel rispetto dei vincoli non lineari dei cavi. Infatti,
per poter controllare un robot a cavi in maniera del tutto sicura è necessario
garantire che in ogni momento la configurazione del manipolatore non cambi
a causa di una perdita di tensione di uno o più cavi. Quindi, tutte le possi-
bili configurazioni che è possibile ottenere se uno o più cavi diventano laschi
vanno prese in considerazione il che equivale, per un generico manipolatore
a n cavi ad analizzare un numero di configurazioni pari a
n
X n!
N=
m!(n − m)!
m=1

Queste sono le ragioni principali che rendono la già complessa analisi

cinematica dei manipolatori paralleli ancora più ardua per i CDPR. Queste
problematiche, unite alla descrizione della modellazione geometrica dei robot
a cavi adottata in questo lavoro, sono discusse nel Capitolo 2 insieme ad una
panoramica della letteratura dedicata ai problemi di cinematica diretta e ai
relativi metodi impiegati per la sua risoluzione.
Il Capitolo 3 presenta il lavoro che ha portato allo sviluppo di una routine
per trovare per via numerica le soluzioni della cinematica diretto di un gener-
ico manipolatore a cavi. Questo metodo permette di risolvere il problema
in maniera completa e i suoi risultati sono garantiti contro eventuali errori
numerici grazie all’uso della tecnica di analisi numerica chiamata analisi per
intervalli. L’analisi per intervalli è una branca della analisi numerica che rac-
chiude una varietà di strumenti che consentono di risolvere problemi senza
il rischio di errori dovuti a cancellazione o arrotondamento. Mentre una de-
scrizione dei concetti di base dell’analisi per intervalli e i metodi richiamati
all’interno del manoscritto è riportata nella seconda metà del Capitolo 1, il
Capitolo 3 descrive nel dettaglio la struttura del metodo di calcolo imple-
mentato, le sue differenti versioni e propone strategie per eventuali futuri
sviluppi e applicazioni in real-time.
Il Capitolo 4 mostra i risultati ottenuti durante la fase di test per le
diverse versioni dell’algoritmo e presenta alcune considerazioni al riguardo.
Il metodo sviluppato durante il percorso di questo dottorato mostra un ot-
timo comportamento dal punto di vista prestazionale. Consente di condurre
 xix

una analisi completa del problema geometrico-statico diretto per un robot a

n cavi in tempi molto ridotti. Grazie all’analisi per intervalli, esso consente
anche di ottenere direttamente solo le soluzioni reali del problema aventi ten-
sioni positive nei cavi e i suoi risultati sono sicuri dal punto di vista numerico.
In questo capitolo è inoltre descritta la realizzazione di un piccolo prototipo
utilizzato a scopi didattici per testare l’implementazione delle versioni on-line
dell’algoritmo all’interno di un vero e proprio schema di controllo.
Il Capitolo 5 riassume i principali risultati contenuti in questo lavoro e
presenta le linee guida per futuri sviluppi e attività di ricerca.
Prologue in French

Vers la fin des années 80, certains chercheurs dans le domaine de la robotique

ont commencé à explorer la possibilité de remplacer les traditionnels action-
neurs à corps rigide de manipulateurs parallèles avec des câbles, afin de sur-
monter les principaux inconvénients de cette famille de robots, comme le lim-
itée espace de travail, la complexité des processus de fabrication, l’assemblage
et l’entretien et les coûts élevés. Ces robots, appelés en anglais cable-driven
parallel robots (CDPRs), ou tout simplement robot à câbles en plus des avan-
tages propres de l’architecture parallèle, grâce à la flexibilité des câbles sont
capables de fonctionner dans des espaces de travail très large, de posséder
une plus grande capacité de charge, meilleures performances cinématiques
et sont de plus faciles conception et assemblage. La première section du
Chapitre 1 de cette thèse de doctorat présent en détail les robot à câbles, la
littérature consacrée à ce sujet et la description des les prototypes les plus
célèbres.
Toutefois, l’analyse cinématique et le contrôle de ces robots, en plus des
problèmes déjà complexes propres de manipulateurs parallèles (résolu seule-
ment après plus de 30 années de recherche), sont rendues encore plus dif-
ficiles par l’incapacité des câbles de résister à des efforts de compression.
Cela implique que certaines poses (positions et orientations) et/ou condi-
tions de charge ne peuvent être atteints parce que certains câbles deviennent
soudainement lâche, modifiant ainsi la configuration du robot, son équilibre
statique ou dynamique et compromettant sérieusement le contrôle. Le tra-
vail effectué au cours du programme de doctorat a été axée principalement
sur l’étude de ces questions, et notamment sur les aspects de la cinématique
directe.
La cinématique directe est le calcul de la pose de l’élément terminal
du robot pour valeurs attribuées des variables du joints, dans ce cas, les
longueurs des câbles. Si l’on considère un robot avec n ≥ 6 câbles dont au
moins m = 6 en tension, les 6 degrés de liberté de l’end-effector peuvent
être déterminée par les équations dérivées les contraintes géométriques im-

xxi
xxii PROLOGUE IN FRENCH

posées par les câbles tendus. Lorsque m < 6, ces conditions géométriques ne
sont pas en nombre suffisant pour définir complètement la pose du robot, qui
peut être déterminée seulement en considérant l’équilibre statique de la plate-
forme. Afin de respecter cette condition, les 6 équations de la statique dans
les m inconnues représentés par les tensions des câbles doivent être ajoutés
aux m équations des contraintes géométriques. En outre, les seules solutions
dans lesquelles les tensions dans les câbles sont positives doivent être pris en
considération. En conséquence, l’analyse cinématique de ces manipulateurs
doit être résolu au même temps que la statique et en considérant les con-
traintes non linéaires dans le câbles. En fait, afin de contrôler un robot à
câbles dans un façon totalement sûre est nécessaire de garantir qu’à tout mo-
ment la configuration du manipulateur ne change pas en raison d’une perte
de tension d’un ou plusieurs câbles. Donc, toutes les configurations possi-
bles qui peuvent être obtenues si un ou plusieurs câbles deviennent lâches
devraient être prises en considération ce qui équivaut, pour un générique
manipulateur à n câbles, à l’analyser un nombre de configurations égal à
n
X n!
N=
m!(n − m)!
m=1

Telles sont les principales raisons qui font que l’analyse cinématique
déjà complexe de manipulateurs parallèles soit encore plus difficile pour les
CDPR. Ces questions, ainsi que la description de la modélisation géométrique
des robot à câbles utilisée dans ce travail, sont examinées dans le Chapitre
2 avec une revue de la littérature consacrée aux problèmes de cinématique
directe et les méthodes utilisées dans sa résolution .
Le Chapitre 3 présente le travail qui a conduit à l’élaboration d’une
routine pour trouver numériquement les solutions de la cinématique direct
d’un générique manipulateur à câbles. Cette méthode permet de résoudre
le problème complètement et ses résultats sont garantis contre les erreurs
numériques grâce à l’utilisation d’une technique d’analyse numérique appelé
analysepar intervalles. L’analyse par intervalles est une branche de l’analyse
numérique qui comprend une variété d’outils pour résoudre problèmes sans
le risque d’erreurs en cas d’élimination ou arrondissement. Bien que la de-
scription des concepts de base et des méthodes d’analyse par intervalles pré-
conisés dans le manuscrit est affichée dans la seconde moitié du Chapitre
1, le Chapitre 3 décrit en détail la structure de la méthode de calcul mise
en œuvre, ses différentes versions et propose des stratégies pour les futurs
développements et applications en temps réel.
Le Chapitre 4 montre les résultats obtenus au sein de la phase de test
 xxiii

pour les différentes versions de l’algorithme et présente quelques considéra-

tions à ce sujet. La méthode développée au cours de ce doctorat montre
un excellent comportement du point de vue des performances. Il permet
d’effectuer une analyse complète de le problème geometrico-statique d’un
robot ayant n câbles en un temps très court. En plus, l’analyse par inter-
valles permet d’obtenir directement les seules solutions réelles du problème
avec tensions positives dans les câbles et d’avoir résultats garantis du point
de vue numérique. Dans ce chapitre, il est également décrit la réalisation
d’un petit prototype utilisé à des fins éducatives pour tester la mise en œuvre
des versions en ligne de l’algorithme dans un vrai système de contrôle.
Le Chapitre 5 résume les principaux résultats de ce travail et présente les
lignes directrices pour les futures activités de développement et de recherche.
Chapter 1

Introduction

Throughout history, mankind has kept developing more and more sophisti-
cated mechanisms and machines for a great variety of uses and tasks, many
of them being represented by mechanisms capable of moving an object with
respect to a fixed base. The position and the orientation of a rigid body
in three dimensional space can be described through a maximum of 6 in-
dependent coordinates called degrees of freedom (dofs). The term robot or
manipulator includes every mechanical system that can move an object with
respect to a fixed base and control 2 or more of its degrees of freedom.
During the second half of the last century, the use of robots in industries
grew dramatically together with the interest of the scientific community in
their challenging issues. Over the last decades of the 1900s, indeed, a fruitful
combination of scientific and technological developments in both theoretical
and practical scientific fields allowed robotic devices to greatly expand their
application domains, even in quite unexpected sectors.
From a technological point of view, for example, in the last 50 years
electric motors have greatly increased their power density to the extent that
it is possible to install relatively small and lightweight motors for heavy
duties. Moreover, thanks to the discoveries in metallurgical and material
engineering, great improvements were made from a mechanical point of view
to gearboxes, that are coupled to the motors and have become more and
more compact and reliable, giving designers more freedom during the de-
velopment of new robot architectures and geometries. At the same time,
the progress made in electronic components combined with the consecutive
advent of computer science, made it possible to build and expand what can
be defined as the intelligence of robotic systems, which definitely shows the
differences with respect to traditional mechanisms and redefines the meaning

1
2 CHAPTER 1. INTRODUCTION

of automation.
During the same period, the scientific community started to show a grow-
ing interest to the more and more challenging issues emerging from robot
development. Thanks to more powerful processors and more efficient pro-
grams provided by progress in computer science, researchers ceaselessly kept
producing innovative solutions and contributions to the growth of robotics
and its applications.
Indeed, the first manipulators replaced humans for most of the easier
and repetitive operations in assembly lines, or for all those tasks that must
be carried out in potentially harmful or unhealthy environments such as
painting, welding, waste management, etc. Initially the most common robot
model had a structure very similar to that of a human arm and all the
manipulators sharing this architecture are defined serial robots.

1.1 Serial and parallel manipulators

Nowadays robots are also used in other fields besides the industrial one: there
are indeed manipulators for rehabilitation purposes, for entertainment appli-
cations, military operations, or domestic duties. Moreover, in the past two
decades, a new family of robots consisting of different kinds of autonomous
vehicles open up the range of application to an incredible extent. There
are actually unmanned vehicles for every environment and condition: fly-
ing vehicles, climbing, rolling and walking robots, sailing and underwater
autonomous boats. However these robots are not in the scope of this disser-
tation.
The majority of the robots currently in use have a structure recalling that
of a human arm, obtained by consecutively connecting different rigid bodies
with joints allowing one degree of freedom between two links. These cou-
plers are generally translational or revolute joints and this sequence of linked
bodies is called open kinematic chain (Fig. 1.1a). Manipulators having an
open kinematic chain architecture are called serial robots.
The main advantage of serial robots is a large workspace compared to
their fixed base dimension and Fig. 1.2 shows one of the most well known
models of these family of manipulators. However, open-chain robots are
affected by several weaknesses, accurately described in [Merlet 2006, Angeles
2002], and they can be all attributed to the fact that each member has to
support the weight and the inertial forces of the following ones. The most
relevant are:

• a small payload to robot weight ratio,

1.1. SERIAL AND PARALLEL MANIPULATORS 3

(a) (b)

Figure 1.1 – Open and closed kinematic chains.

• poor positioning accuracy,

• low speed and acceleration.

The drawbacks of the serial structure can be partly solved by connecting

the end-effector to the base with multiple open kinematic chains in order
to distribute the external load. The resulting structure will be thus formed
by several closed kinematic chains, like the one depicted in Fig. 1.1b. Ac-
cording to [Merlet 2006]:“a closed-loop kinematic chain mechanism whose
end-effector is linked to the base by several independent kinematic chains is
a parallel manipulator ”.

Figure 1.2 – The Kuka R 7-dof robotic arm.

The first advantage of parallel robots that immediately stands out is the
redistribution of the load applied to the end-effector on the links forming
the closed-loop chains. This implies that bodies and joints constituting the
links, also commonly called legs, can be lighter provided that the end-effector
stiffness remains constant. This fact, together with a lower positioning sensi-
tivity to joint measurement errors due to the fact that actuators are directly
4 CHAPTER 1. INTRODUCTION

connected to the end-effector, makes parallel manipulators far more accurate

than serial robots.
This reduction of weight in robot components has multiple effects on
parallel machine performances. First of all, reduced moving mass implies
that higher velocities and accelerations are possible. Moreover, the actuator
size can be significantly decreased together with, consequently, the machine
power consumption.
Another great advantage consists in the modularity of parallel manip-
ulators. Indeed all the legs are generally identical, which means that the
components can be manufactured in series. Besides the economic conve-
nience, high modularity positively affects also assembly and maintenance
operations.
A fully detailed and comprehensive introduction on parallel robots, on
their history and their applications can be found in the first chapter of [Merlet
2006]. In this thesis, only the two most widely used for industrial applications
are briefly described: the Gough platform and the Delta robot.

(a) (b)

Figure 1.3 – The tire testing machine developed by Gough on the left, and
a modern flight simulator on the right [Airbus 2010].

Although manipulators with parallel architectures were already known

before its appearance, the Gough platform is universally recognized as the
first parallel robot with a successful industrial application and it is still one of
the most famous. Its name derives from its inventor Vernon Eric Gough, who
in 1955 developed a tire testing machine [Gough 1956] (shown in Fig. 1.3a)
consisting of a hexagonal platform that moves with respect to a fixed frame
by 6 links, each one equipped with a linear actuator and connected to the
moving platform with a spherical joint and to the fixed base with a universal
1.1. SERIAL AND PARALLEL MANIPULATORS 5

joint. Thanks to this platform, tire position and orientation were precisely
controlled like the forces acting upon it. High versatility and stiffness led the
Gough platform to be successfully employed in many other domains: flight
simulators (like the one depicted in Fig. 1.3b), drilling and milling machines,
micro-positioning devices, haptic interfaces, etc.
Another successful model of a parallel manipulator is the Delta robot. It
has a limited number of degrees of freedom and it was invented by Clavel in
1988 ([Clavel 1988]). It consists of three links, each one with a parallelogram
linkage connected, by means of two revolute joints on its short sides, to the
end-effector and to a motorized lever. The Delta is generally equipped with
3 motors and consequently has 3 dofs that sometimes rise to 4 when the end
effector is equipped with a rotating head (like the one in Fig. 1.4).

Figure 1.4 – The Delta IRB 360 FlexPicker by ABB R .

The Delta became very popular for pick-and-place operations thanks to

its high end-effector accelerations (more than 100 m/s2 ). Its geometry and
architecture allows the Delta to have pure translational motion in the carte-
sian space, and also the solution of its inverse and forward kinematics is
eased by this decoupling between positioning and rotating motion of the
moving platform. It is the most common parallel machine for pick-and-place
operations.
Parallel robots nowadays have a great variety of architectures and they
are successfully employed in several domains: in medicine as surgery devices,
as simulator platforms, in manufacturing as milling CNC machines, as mirror
positioning systems in telescopes, etc.
As far as the disadvantages of parallel manipulators are concerned, the
main one is a reduced workspace compared to their dimensions. The causes
are multiple:
6 CHAPTER 1. INTRODUCTION

• most kinematic chains constituting robot legs have links connected

with universal or spherical joints which generally have restricted mo-
tion ranges;
• actuators usually have poor strokes and consequently they can provide
only limited displacements to the end-effector;
• robots legs can be quite bulky, especially when high power and stiffness
are required, leading to a high probability of interference between legs
or between legs and platform
An alternative architecture that can overcome this issue has attracted re-
searchers’ interest in the last two decades: cable robots. These manipulators
can control the end-effector using cables as actuators and the next section
describes this family of robots in more detail.

1.2 Cable robots

If rigid body legs are replaced by cables in the kinematic chains connecting
the base with the moving platform of a parallel manipulator, then it is de-
noted as a cable-driven parallel robot (CDPR). Thus the end effector pose
can be controlled by varying the cable lengths.
There are several advantages provided by this change in the actuation
scheme. First of all, since cables are flexible they can be easily coiled and
uncoiled so that actuator strokes in cable robots are way larger than in tra-
ditional rigid leg parallel machines. Even though they are flexible, cables
can have also a high tensile strength compared to their diameter and their
linear density. Thus, adopting cables instead of traditional actuators leads
to a significant reduction of moving masses and, as a consequence, to an im-
provement in terms of kinematic performances. Moreover, they can be more
easily connected with the moving platform and thanks to their flexibility
many problems caused by interferences or limited joint motion space can be
overcome.
Cables also gives CDPRs a lightweight structure which makes them more
attractive than traditional parallel robots in different domains and some of
them are presented later in this section. Another advantage provided by
cables consists in the possibility to easily connect and disconnect them to
the end effector, which makes CDPRs highly deployable and reconfigurable.
These properties, together with the cable flexibility and reduced mass of
the system, make it possible to transport even very large cable-robots with
minor effort.
1.2. CABLE ROBOTS 7

Figure 1.5 – An illustration of a Greek crane during building operations.

1.2.1 History and main applications

Cables started to be used to provide motion and forces to mechanical sys-
tems a very long time ago. Traditional cranes and sophisticated block-and-
tackle-like arrays of pulleys were already known and extensively employed
in construction sites or shipyards (Fig. 1.5 and 1.6) by the Greeks first and
the Romans later.

(a) (b)

Figure 1.6 – On the left an image from the “de Architectura” of Marcus
Vitruvius Pollio and on the right an ancient block-and-tackle system.

Nowadays, cables and flexible elements in general are currently used in

different mechanical domains. Besides the traditional lifting cranes, they
are regularly employed also in the automotive industries and for automation
purposes (like for example the bowden cable transmission shown in Fig. 1.7).
Cable driven chains are employed also in robotics and sometimes the def-
inition of cable-driven robots may be misinterpreted. Indeed, cable driving
8 CHAPTER 1. INTRODUCTION

Figure 1.7 – Bowden cables controlling an automobile throttle [Ehardt 2015].

systems are successfully used in prototypes of robotic hands like the one pro-
posed in [Grebenstein et al. 2012] and depicted in Fig. 1.8a, where they act
as human tendons to drive finger movements, or in the so-called tensegrity
robots where cables and rods are interlocked together forming rigid struc-
tures. These robots can change their shapes and stiffness or even movements
[Paul et al. 2006] by controlling cables lengths.
Another example of an interesting application of cables in robotics is
presented in [Dekker et al. 2006, Behzadipour & Khajepour 2006], where
cables replace rigid links in a Delta parallel robot (Fig. 1.8b). This choice
further reduces the mass of moving elements of the Delta architecture, but
necessarily needs an additional extensible passive or active member generally
called “spine” between the base and the end-effector that guarantees positive
tensions in cables.

(a) The DLR robotic hand presented in [Grebenstein (b) The DeltaBotTM prototype
et al. 2012]. [Dekker et al. 2006].
1.2. CABLE ROBOTS 9

However, even if all these manipulators comprise cables in their structures

or actuation schemes they do not belong to the same of family of robots
discussed in this thesis. In order to avoid possible misunderstandings, for
the remaining part of this manuscript the following definition holds:
Definition. A “cable-driven parallel robot” or simply “cable robot” is a par-
allel manipulator having the mobile platform connected to the fixed base
only by cables and its pose being controlled only by changing the lengths of
the cables.
The first manipulator that complies with this definition was developed by
the NIST and presented in [Albus et al. 1993]. It is a 6-dof crane (Fig. 1.8)
called RoboCrane used for lifting and positioning operations or to control
position, velocity and force of tools and machinery for cutting or excavating
tasks. Among the numerous application proposed for this manipulator in
[Bostelman et al. 1994], the one shown in Fig. 1.8b for airplane painting is
particularly interesting: the operator works inside a maneuverable platform
that uses computer-controlled cables to “float” around the aircraft. This
application of the NIST RoboCrane “promises to drastically reduce paint-
stripping time per airplane, cut maintenance costs and lessen incidents of
operator stress and injury” ([NIST 2006]).

(a) (b)

Figure 1.8 – On the left the first prototype of the NIST RoboCrane and on
the right its version for aircraft painting and maintenance ([NIST 2006]).

Another interesting field of application of the NIST RoboCrane is the

shipbuilding industry [Bostelman et al. 1999], where it can be used as an
appendix of traditional gantry cranes and to provide more precise control of
the welding platform (see Fig. 1.9a). Another project, again developed by
10 CHAPTER 1. INTRODUCTION

the NIST and still devoted to shipping yards, is the Flying Carpet project
[Bostelman et al. 2002], where robot cranes are used to move suspended
scaffolding and provide a safer worker access to ships, as shown by the model
in Fig. 1.9b.

(a) (b)

Figure 1.9 – On the left the RoboCrane as a controllable welding platform

and on the right its version for ship building and maintenance operations
[Bostelman et al. 1999].

Figs. 1.10a and 1.10b show how cable robots can be also installed on
ships as proposed by the AACTS project (Automated All Weather Cargo
Transfer System) [Lee 1992] or by the U.S. Navy [Quadvlieg et al. 2011] to
perform ship-to-dock or ship-to-ship operations.

(a) (b)

Figure 1.10 – On the left a drawing representing the functioning of the

AACTS ([Lee 1992]) and on the right the Large Vessel Interface Lift On/Lift
Off (LVI Lo/Lo) Crane for ship-to-ship container transfer at sea (source:
ONR, USA)

Performing freight management and pick and place tasks with cable
robots is a potentially winning strategy, especially when operations over
1.2. CABLE ROBOTS 11

large workspaces have to be performed. Several prototypes are now avail-

able and operational that can be applied in these fields and the most im-
portant are: the CoGiRo and the ReelAx8 (Fig. 1.11a and 1.11b) both
developed by the team Dexter at LIRMM [Dallej et al. 2011, Lamaury &
Gouttefarde 2013], the IPAnema family of manipulators from the Fraun-
hofer Institute of Stuttgart [Pott et al. 2010, 2013] (Fig. 1.12a shows the
prototype for solar power-plants installation), the portable and deployable
crane realized by Merlet (Fig. 1.12b) and described in [Merlet & Daney
2010] or the cable-robot-based high-bay warehouse from the University of
Duisburg-Essen [Bruckmann et al. 2013] (Fig. 1.12c).

(a) (b)

Figure 1.11 – The CoGiRo robot on the left [Alexandre Dit Sandretto 2013]
and the ReelAx8 on the right: the two prototypes available at the LIRMM
[Lamaury et al. 2012].

Cable robots provide excellent performances also if they are used as large
positioning systems. A huge 6-dof cable crane is currently in service at the
NASA to perform simulated low-gravity operations (Fig. 1.13a), and the
Chinese government started the construction of a giant telescope (FAST,
[Wilcox 2012, Nan 2006]) that can drive the light focus cabin by means of
6 cables (Fig. 1.13b). Cables’ intrinsic flexibility and compliance joined
with the ease of assembly and reconfiguration make CDPRs particularly
well suited also as haptic interfaces [Williams II 1998, Gallina et al. 2001,
Murayama et al. 2004] and rehabilitation devices [Merlet 2010, Rosati et al.
2007, Surdilovic et al. 2007] (Fig. 1.14 shows two examples of these robots,
one for each application).
Here only the most relevant prototypes and applications for cable robots
are reported. Indeed, in the last 10 years CDPRs have become very popular
among researchers and the number of teams and publications on this subject
keeps growing. But although this technology is very promising for a large
set of industrial domains, only a few are already operative and employed
12 CHAPTER 1. INTRODUCTION

(a) (b)

(c)

Figure 1.12 – (a) The Fraunhofer Institute prototype for power solar plant
construction [Pott et al. 2010], (b) the portable rescue crane Marionet Res-
cue by Merlet [Merlet & Daney 2010], (c) the cable-driven platform for au-
tomated warehouses from the Duisburg University [Bruckmann et al. 2013].

and currently there are no companies selling them except those producing
cable-driven film and television cameras like the Skycam (Fig. 1.15) or the
Spidercam.

1.2.2 Main components

In order to introduce the main subject of this thesis, a brief description of
the structure of a cable robot and of the main elements which make it up it
is presented hereafter.
A generic CDPR hardware, schematically represented in Fig. 1.16a, can
be divided, like traditional parallel robots, in three main blocks: an end-
1.2. CABLE ROBOTS 13

(a) (b)

Figure 1.13 – On the left, the Athlete vehicle during a low gravity test [Wilcox
2012] and on the right a rendering of the FAST telescope [Nan 2006].

effector or moving platform, n > 1 actuators and a supporting structure or

framework.
The end-effector does not usually have a precise shape or geometry, in-
deed it depends on the particular application. It is important to remember
that since cables can be easily attached and detached this is quite a differ-
ence with respect to traditional parallel and serial robots. The end-effector
is, therefore, generally equipped with eyebolts, rings or has other special
features to which cables will be connected (if high precision is required, uni-
versal and spherical joints are employed). For every kind of linkage it is
possible to find a point, called anchor point or attachment point, that will
define one end of each actuator. These points are part of the robot’s geo-
metric model and will be implemented in the control software. Thus it is
very important to design and build them so that their position can be easily
identified by measurements, and at the same time the reconfigurability and
assembly simplicity of these robots are preserved.
Each anchor point is connected to one (or more) cable, whose length can
be controlled by a motorized device. The most common cable actuator is
obtained by coupling an electric motor to a simple drum that, according to
its rotating direction, coils or uncoils the wire. This is probably the easiest
actuation scheme, but it also has several drawbacks. First, cable coils are
directly in contact and often rolled one above the other, and this fact may
cause premature wire failures due to wear. Then, if wires are coiled untidily,
their deployment speed may have sudden variations due to overlapping coils
and slack cable loops. Moreover, overlapping strongly affects cable measure-
ments made by reading the motor angular positions given by the encoders,
14 CHAPTER 1. INTRODUCTION

(a) (b)

Figure 1.14 – (a) The haptic interface Spidar G-G [Murayama et al. 2004]
and (b) the Marionet Rehab by Merlet [Merlet 2010].

Figure 1.15 – The SkyCam [SkyCam 2015].

an essential element for accurate robot control. More refined actuators can
be realized by equipping simple reels with a level-wind wire guide system
(1.16b) in order to have more precise and organized cable coiling. Further-
more, the most sophisticated prototypes have added a level-wind wire guide
system to specially grooved reel drums (see Fig. 1.16b). Since cable overlap-
ping is mechanically prevented, this architecture is more suitable in all those
application where higher precision (and cost) is preferred compared to minor
stroke lengths. An alternative to cable winches as actuators for these robots
is represented by the solution proposed by Merlet in [Merlet 2008], where ca-
ble lengths are controlled by pulley systems mounted on linear motors (Fig.
1.16).
From the actuators to the platform anchor points, cables are generally
routed through pulleys, rings or more complex systems in order to define
1.2. CABLE ROBOTS 15

(a) Scheme of a generic cable-driven (b) A SkyCam motorized reel [SkyCam 2015].
parallel robot [Bruckmann et al.
2008a].

possibly unique and fixed points, called exit points, that together with the
anchor points are fundamental geometric parameters for robot control. The
design of these components, indeed, may have a great influence on robot
performances, since exit points must allow cables to freely pass through
them and at the same time their position must remain possibly the same.
They are usually realized with orientable pulleys, like for example the ones
shown in Fig. 1.17, or with small ceramic eyes, although this latter solution
is practically feasible only in the case of small prototypes because it causes
wear and abrasion issues.
Then, a few words must be spent on wires, i.e. the components that
physically provide motion to the end-effector. Cable properties and appli-
cation domains are well known both in mechanics and civil engineering.
Cable-robot prototypes can be equipped with any sort of wires, ropes and
lines but the most common choices are essentially two: traditional steel-wire
ropes or aramid-fiber lines. Steel-wire ropes are generally preferred when
high load capacities are required, mainly because of the better knowledge
of their physical properties which makes it possible to easily pick the right
diameter for given safety coefficients. However, steel cables need mainte-
nance and constant lubrication, winches must be designed in order to be
extremely resistant to wear and pitting, coil overlapping is totally undesired
because it severely damages cable surfaces. Steel cables also have a high
linear density, so that mass effects like sagging must be taken into account
when dealing with large workspaces, and they have high linear thermal ex-
pansion coefficients which implies that cable lengths vary considerably with
temperature. Aramid-fiber lines, like Kevlar or Dyneema ropes and cords,
16 CHAPTER 1. INTRODUCTION

(a)

(b)

Figure 1.16 – (a) The latest prototype of winches developed by the Fraun-
hofer Institute IPA [Schmidt et al. 2015] and the cable actuators with linear
motors mounted on the Marionet Rehab [Merlet 2010].

have lower linear densities and practically inexistent thermal and humidity
effects. They can still provide high tensile strength and stiffness within a
lower weight than steel wire ropes and they guarantee lower wear issues.
They can also be fabricated so as to have signal cables woven into them
(like two of the four cables in the SkyCam system). The main drawback of
aramid-fiber lines is the highly non-linear behavior and the lack of experience
regarding their mechanical fatigue life. This latter aspect is very important
in the cable-robot design process. Indeed, as previously described, for this
application it is necessary to coil and uncoil cables on drums and make them
pass through pulley systems. This involves a great number of bending cycles
on cable fibers and demands on the one hand a careful choice of pulley and
drum diameters. On the other hand, since exit points must be realized so as
to be possibly modeled with a fixed point, the knowledge of the minimum
bending radius allowed for a given cable in relation of the maximum number
1.3. INTERVAL ANALYSIS 17

Figure 1.17 – A detail of the winches developed by the Fraunhofer Institute.

of bending cycles is very important.

1.3 Interval analysis

Since methods described in this thesis are based on interval analysis, this
section provides some elements of this technique and a general overview of
its possible applications in robotics. More details on this technique and its
mathematical framework can be found in [Moore 1966, Jaulin 2001, Hansen
& Walster 2003].

1.3.1 Basic Notions

The real interval X = [x, x] is defined as the set of real numbers y such
that x ≤ y ≤ x. When this definition is implemented on computers and
numbers must be represented
 generally in double precision format, a real
interval Xr = xr , xr is defined as a set containing all the double precision
numbers satisfying these two conditions:

• xr is the nearest double precision number such that xr ≤ x

• xr is the nearest double precision number such that xr ≥ x.

As an example, the interval representation of π on a personal computer is

Π = [3.1415926535897931, 3.1415926535897932]] (1.1)

The extrema of an interval xr and xr are returned by the functions Inf(X)

and Sup(X) respectively. The width of an interval is the quantity given by

w(X) = x − x, (1.2)
18 CHAPTER 1. INTRODUCTION

while the mean value of its extrema, i.e.

mid(X) = (x + x)/2 (1.3)

is called mid-point of an interval.

An interval vector B = {X1 , . . . , Xh }, also called a box, is a list of inter-
vals. The mid-point and the width of a box are vectors whose components
are respectively the mid-points and the widths of their interval components:

w(B) = {w(X1 ), . . . , w(Xn )} (1.4)

mid(B) = {mid(X1 ), . . . , mid(Xn )} (1.5)

Interval arithmetic
The sum among intervals is defined as
 
X + Y = x + y, x + y (1.6)

while the product between a real number and an interval is

aX = [ax, ax] (1.7)

Consequently, the opposite of an interval is

−X = [−x, −x] (1.8)

which yields to the definition of the subtraction between intervals

 
X − Y = x − y, x + y (1.9)

If an interval X does not contain 0, the opposite of an interval can be defined

as
1
= [1/x, 1/x] (1.10)
X
when x > 0, while the product of two intervals can be computed as follows
 
XY = min(xy, xy, xy, xy), max(xy, xy, xy, xy) (1.11)

It is worth noting that the interval sum and product operations are commu-
tative and associative since the following properties hold:

0 + X = X + 0 = 0, 1X = X1 = X, 0X = X0 = 0. (1.12)
1.3. INTERVAL ANALYSIS 19

On the contrary, distributive property is not generally applicable. For exam-

ple, let the following expression be considered

[1, 2] (1 − 1) = 0 (1.13)

which, expanding it, leads to:

[1, 2] 1 − [1, 2] 1 = [−1, 1] 6= 0. (1.14)

So, the equality X (Y + Z) = XY + XZ is not always true. However, the

distributive property holds in the following cases:

a (Y + Z) = aY + aZ, ∀a ∈ R (1.15)
X (Y + Z) = XY + XZ if XY > 0 (1.16)

From this introduction to interval arithmetic it is possible to make some

interesting considerations about this numerical technique. First of all, an
interval is defined by a couple of real numbers so its representation in a
computer system requires twice the memory of a floating point number.
Moreover, as is clear from eqs. (1.6), (1.7) and (1.9) to (1.11), even basic
arithmetical operations are more expensive from a computational point of
view than traditional ones, and the product in particular requires a careful
implementation in order to be effective. Also the form of expressions involv-
ing intervals has to be chosen carefully: for example, it is important to avoid,
if possible, divisions between intervals since they can easily lead to overflow
errors.

1.3.2 Evaluation function

Let f (x), with x = [x1 , x2 , . . . , xh ], be a function in h unknowns and B =
[X1 , X2 , . . . , Xh ] a box comprising an interval for each
 unknown. The interval
evaluation F (B) of f over B is an interval F , F such that, for any x ∈ B,
F ≤ f (x) ≤ F , i.e. F and F are lower and upper bounds for f when the
unknowns lie within B. An evaluation function is defined

• convergent if, for any sequence of boxes Bk , if lim w(Bk ) = 0 then

k→∞
lim w(F (Bk )) = 0
k→∞

• minimal if for any B, F (B) is the smallest interval containing every

f (x), ∀x ∈ B.

• inclusion monotonic if when B1 ∈ B2 then F (B1 ) ∈ F (B2 )

20 CHAPTER 1. INTRODUCTION

• the evaluation of a list of function F = [F1 , . . . , Fp ] is convergent only

if all the Fi are ∀ i = 1, . . . , p.
Interval analysis provides several tools to implement an interval evalua-
tion of a function [Moore 1979], but the simplest one is the natural evaluation,
in which basic arithmetic operations and elementary mathematical functions
are substituted by interval equivalents. For example, if f (x) = x2 − 2x + 1
and X = [4, 5], the natural evaluation of f over X is:

f ([4, 5]) = [4, 5]2 − 2 [4, 5] + 1 =

[16, 25] − [8, 10] + [1, 1] = [6, 17] + [1, 1] = [7, 18] (1.17)

A natural evaluation function is:

• inclusion monotonic

• usually not minimal

Indeed, the bounds provided by the natural evaluation of f are generally
not exact: the upper (lower) bound may be larger (lower) than the actual
maximum (minimum) of the function image, namely f (B) = {f (x)|x ∈
B} ⊆ F (B). For example, the image of the function f in eq. 1.17 over the
interval X = [4, 5] is [9, 16] ∈ [7, 18].
Such an overestimation dramatically affects the performances of interval-
analysis methods. This is the reason why methods that provide the tightest
interval evaluation of functions are of fundamental importance. As an exam-
ple, as shown in [Merlet 2007], using the centered form and the mean-value
theorem sharper bounds may be obtained:

f (X) ⊆ f (mid(X)) + f ′ (X)(mid(X) − X) = [8.25, 16.25] (1.18)

Anyway this improvement came at the price of computing the derivative

of the function, thus making the evaluation slightly more demanding from
the computational point of view, and this is a general trend for the strategies
of interval evaluation of functions. An interesting property is that the bounds
of the interval evaluation F are exactly the minimum and the maximum of
f (B) when f may be expressed so as to contain a single occurrence of each
unknown xi (i = 1 . . . n) [Jaulin 2001]. On the base of this property, the
expression in eq. 1.17 may be more effectively evaluated as

f (X) = X 2 − 2X + 1 = (X − 1)2 = [9, 16] (1.19)

which are the sharpest bounds for f over X = [4, 5].

1.3. INTERVAL ANALYSIS 21

This property implies that when dealing with interval analysis methods,
the choice of the parameterization is extremely important. Indeed, if the
problem is modeled by a set of parameters that lead to expressions having a
large number of multiple occurrences, the interval evaluation of each function
will be heavily overestimated, affecting algorithm performances.
In summary, it is possible to extend every mathematical function to inter-
val analysis but its implementation has to be done carefully. As mentioned
before, the distributive property does not generally hold so it is better to sim-
plify expressions manually. Moreover, function interval extensions are more
computationally expensive with respect to traditional ones implemented in
computer systems since they are built so as to provide the tightest, but at
the same time guaranteed, bounds. As a consequence, natural function eval-
uation is the best choice in order to limit the computational burden, but at
the same time functions have to be formulated so as to contain the minimum
number of multiple occurrences, which is not always possible. However, when
dealing with interval analysis based methods, if multiple parameterizations
are available, it can be better to consider simpler expressions, even at a price
of a larger number of unknowns, and avoid overestimation on the function
evaluation.

1.3.3 Problem-solving algorithm

Interval analysis may be applied for different problems in robotics, and sev-
eral procedures based on interval analysis were successfully developed in this
domain (for a detailed discussion, see [Merlet 2009]). The main part of the
work presented in this thesis is aimed at solving systems of nonlinear equa-
tions, so that only the algorithm for this particular application is discussed
hereafter.
The advantages provided by interval methods are many. One of the most
important is the capability to obtain results guaranteed against numerical
errors. Indeed, there are libraries that allow basic interval analysis tools
to be easily implemented on computer systems. Moreover, interval-analysis
algorithms have the advantage of directly searching for real solutions, which
are the only ones of practical interest. In addition to that, interval analysis
searches for solutions only within a predetermined domain. This property
turns out to be extremely useful in robotic applications, and in particular
when dealing with kinematics problems. Indeed, for any kind of robot, the
solutions of the forward kinematics have to lie within the specific workspace
of the robot or, when inverse kinematics is considered, within the ranges of
motion of its actuators. As will be explained below, in the particular case of
22 CHAPTER 1. INTRODUCTION

CDPRs, taking advantage of this peculiar feature of interval-analysis-based

algorithms, only those solutions with positive cable tensions can be obtained.
The simplest problem-solving algorithm that can be derived from inter-
val arithmetic relies on a simple branch-and-bound (or branch-and-prune)
scheme, but this approach generally leads to unsatisfying performances. In-
deed, a great number of tools and techniques must be integrated in order to
develop an effective algorithm. This fact represents one of the main draw-
backs of this technique, since choosing the right heuristics and the best way
to implement them has, on the one hand, a dramatic impact on the perfor-
mances, but on the other it requires a lot of experience, and the literature
devoted to the applications of interval analysis in robotics is limited.
The structure of the algorithm used in the code to solve a system of n
equations in n unknowns is as follows. Let B1 = [X1 , X2 , . . . , Xn ] be a box
and f = [f1 (x) , f2 (x) , . . . , fn (x)] = 0 a vector equation to be solved within
B1 . L is a list of boxes, initially set as L = {B1 }. An index i, initialized to
1, indicates which box Bi in L is currently being processed, while N denotes
the number of boxes in L. S is another list, initially empty, storing the
solutions. The interval evaluation of fj over Bi is denoted as Fj (Bi ), with
j = 1, . . . , n.
A key component of the algorithm is the evaluation operator E, which
takes a box Bi as an input and it returns:
• 1, if for any j, both w(Fj (Bi )) is smaller than a given threshold ε and
Fj (Bi ) includes 0; in this case, Bi is deemed a solution and it is stored
in S;
• −1, if Fj (Bi ) does not include 0 for at least one j;
• 0, otherwise.
A poorly chosen parameterization and a consequent large overestimation
of interval functions in the system of equations will lead thus to a large
number of bisections in order to assess whether a box may contain a solution
or not. On the other hand, using a redundant parameterization to avoid
multiple occurrences means a greater number of functions to be evaluated at
each step and a larger number of bisections to be made to meet the stopping
criterion of the algorithm for a possible solution box. Another key component
is the filter operator F, which takes a box as an input and it returns:
• −1, if there is no solution in the input box;
• a box smaller than the input one, if the filter determines that the
removed part of the input box cannot contain a solution;
1.3. INTERVAL ANALYSIS 23

Algorithm 1 Problem-solving algorithm scheme

1: i = 1, N = 1, L = {B1 }, S = {}
2: if i > N then return S;
3: end if
4: if F (Bi ) = −1 then
5: i=i+1
6: go to 2
7: else
8: Bi = F (Bi )
9: end if
10: compute E (Bi )
11: if E (Bi ) = −1 then
12: i=i+1
13: go to 2
14: end if
15: if E (Bi ) = 1 then
16: add Bi to S
17: i=i+1
18: go to 2
19: end if
20: if E (Bi ) = 0 then
21: pick xk such that w(Xk ) = max(w(Bi ))
22: bisect Xk in the middle point
23: create two new boxes B′i and B′′i from Bi
24: replace Bi with {B′i , B′′i } in L
25: N = N + 1, i = i + 1
26: go to 2
27: end if
24 CHAPTER 1. INTRODUCTION

• the input box, otherwise.

This is the most common implementation for the evaluation operator. How-
ever, a more sophisticated version of this operator is presented in Section
3.3.2. A box which is neither discarded nor deemed to be a solution is bi-
sected. The strategy adopted in the code to select which variable has to be
bisected consists in picking the variable contained in the box Bi having the
largest width. The complete scheme of the algorithm is described in Algo-
rithm 1. A strength of an algorithm of this kind is that it always terminates,
since the size of a box always decreases after a bisection. Furthermore, pro-
vided that the new boxes emerging from a bisection are put at the top of the
list, there is usually no problem of memory storage.

1.3.4 Application in robotics

Interval analysis represents a powerful tool for several problems in robotics.
One of the main strengths is that it allows for guaranteed computations
with computer systems. Indeed, when a problem is solved by a computer
algorithm, its solutions are necessarily affected by errors that come from the
fact that machines represent numbers only up to a certain precision.
Interval implementation in software libraries can take round-off and elim-
ination errors into account and thanks to its arithmetic and function exten-
sions it is possible to guarantee results coming from calculations.
The relevance of this feature can seem secondary since rounding or elim-
ination error magnitudes are usually very small quantities. However, the
literature devoted to numerical analysis is full of examples of catastrophic
calculations due to floating point representation errors (some real cases are
available at [Douglas 1998]). Some of these issues can be solved with minor
modifications of the code, others require much more expertise, others still are
impossible to predict. Interval analysis cannot obviously fix ill-conditioned
algorithms but can absolutely be very useful to identify strange behaviors in
solving procedures (for example returning large-width intervals as results).

Kinematics and control

Solving robot kinematics problems means dealing with systems formed by
both linear and nonlinear equations. In addition, certainty of obtaining
safe results, interval analysis offers several peculiar advantages with respect
to other techniques when robotic systems are considered. First, interval-
analysis-based methods generally do not work with complex numbers, so
only real solutions are returned by its algorithms which are the only ones of
1.3. INTERVAL ANALYSIS 25

practical interest. Moreover, as will be more clearly explained in the next

chapters, since interval-analysis-based algorithms require the initial search
domain of the solution to be specified, it is possible to take into account
geometrical and physical constraint directly in the initialization phase of
algorithms. In addition to that such constraints can be implemented into
the filter operator F in order to look directly for feasible solutions.
In control routines where solutions of forward kinematics have to be
computed in very short times (a few milliseconds) the most common choice is
to rely on Newton-based algorithms that are fast but not guaranteed against
numerical errors nor to provide the right solutions. On the contrary, there are
interval-based methods that can find, guarantee and assess the uniqueness
of forward-kinematics solutions in real-time.
Another great advantage that this technique may have in robotic kine-
matics problems consists in considering uncertainties affecting some robot
parameters directly in the solution algorithms. The pose of a robot end-
effector is in practice known up to a certain precision, mainly because of
clearances and/or tolerances in the assembly and manufacturing process of
mechanical components. These uncertainties can be implemented in an in-
terval analysis algorithm and from its solutions robot precision bounds may
be determined.

Workspace analysis and trajectory planning

Calculating workspaces is important both during the design phase and in

order to measure robot performances. However, it is generally impossible to
obtain their analytical expression and they must be computed through nu-
merical techniques. The majority of numerical procedures computing robot
workspaces basically obtain it by checking the feasibility of a finite number of
end-effector poses. In order to have a good approximation of the workspace
volume this number has to be very large and thus long computation times
may be required. Moreover, it is very difficult with this approach to de-
tect whether the workspace contains small unfeasible regions or singularities.
Through interval analysis, instead, it is possible to obtain workspace approx-
imation as a list of boxes in which all the end-effector poses contained within
are feasible. With the same approach the complete feasibility of a given tra-
jectory can be assessed and it is also possible to include information about
physical interferences between the robot structure and the environment or
self collisions or uncertainties in robot geometric parameters.
26 CHAPTER 1. INTRODUCTION

Calibration

Once the trajectory planning is completed and a given task is assigned to

the robot, the controller starts transmitting to the motors the appropriate
motion law obtained by means of the inverse kinematic model. Thus, the
kinematic model has a great influence on robot performances. However, a
manipulator is a mechanical system whose behavior can be affected by a large
number of phenomena and uncertainties that are extremely hard or even
impossible to include in the model. The main origins of these uncertainties
are errors in the assembling or manufacturing of robot components, their
deformation due to mechanical stress or temperature variations, and joint
backlash. The influence of these factors on robot precision and accuracy can
be considerably reduced by means of the calibration process. It basically
consists in determining a more accurate value of the geometrical parameters
of the robot, the input being external measurements of the end-effector pose
at various locations. Interval-analysis-based techniques, as shown in [Daney
et al. 2004, 2006, Wu & Rao 2007], gave very interesting results on robot
calibration routines, and recently some effective algorithms and methods
were developed for CDPRs [Alexandre Dit Sandretto 2013].

Robot design

Besides the analysis problems, i.e. determining for a given robot its perfor-
mances and/or verifying if they fulfill the requirements even in the presence
of uncertainties, interval analysis can be helpful during the design stage, also
called synthesis, of a robot. Design algorithms usually employ optimization
procedures to numerically determine a single set of geometric parameters
that minimize (or maximize) a user defined cost function. However, the
choice of the best cost function and its implementation are difficult tasks.
Through the interval-analysis based approach it is possible to obtain certified
continuous sets of solutions that satisfy all requirements and also allow the
management of uncertainties in physical realization. The structure of one
very simple algorithm for these purposes proceeds as follows. Let B be a box
of intervals in which geometric parameters can vary. Then B is processed by
considering one system of inequalities representing the first requirement and
at the end a set S1 of feasible geometries is obtained. After this process has
been repeated for every requirement, the intersection between all the feasi-
ble sets, if it exists, is the set of configurations satisfying design constraints.
More information and details about this kind of applications can be found
in [Merlet 2009].
1.3. INTERVAL ANALYSIS 27

1.3.5 Advantages and disadvantages

Interval-analysis strengths and critical issues are summarized here. The first
advantage is that interval operations are guaranteed against rounding and
elimination errors and so are the procedures based on this technique. Since
all the interval-based algorithms rely on branch-and-bound-like techniques
and they stop when certain criteria on box widths are attained, they always
terminate. Another advantage is the possibility to specify the solution search
domain and include constraints within the solving procedure. It can manage
problems affected by uncertainties and its algorithms are particularly well
suited for parallel computing implementations.
The main drawback of this technique is the lack of knowledge and litera-
ture about its methods and its possible applications. Another disadvantage
of interval-analysis algorithms is that their effectiveness cannot be a-priori
determined. This fact, combined with the aforementioned lack of expertise
about this method in certain domains (e.g. robotics), often leads to very poor
performances. Moreover the number of parameters, methods, and heuristics
that can strongly affect its effectiveness is very high. As a consequence, sat-
isfying implementations of interval-analysis-based methods are often difficult
to achieve.
Chapter 2

Geometrico-static model for

CDPRs

2.1 General definitions

This chapter presents how a generic cable-driven robot was modeled in or-
der to develop the methods described in this thesis and describes the main
assumptions made. To do so, a few definitions about CDPRs are now intro-
duced.
Cable robots can be grouped into two families: fully-constrained and
underconstrained. This distinction depends on how many degrees of freedom
of the end-effector can be controlled by robot actuators. The end-effector
of a generic robot that operates in a three-dimensional space has, like any
other rigid body, 6 degrees of freedom. As a consequence, in order to control
all of them at least 6 actuators are needed.
However, if cable flexibility provides on the one hand all the advantages
characterizing cable robots, on the other hand it introduces strong non-
linearity into CDPR analysis problems. Indeed, cables can transmit only
tensile forces and this implies that 6 cables will not generally be sufficient to
completely control the end-effector.
Due to this fact, fully-constrained (or completely-restrained ) cable robots
are equipped with additional actuators in order to prevent cables from be-
coming slack [Bosscher et al. 2007, Hiller et al. 2005, Kawamura et al. 2000,
Kossowski & Notash 2002, Ming & Higuchi 1994, Tadokoro et al. 2002].
Some CDPRs may not need redundant cables to be fully-constrained and
this happens when a constant external force with convenient magnitude and
direction acts on the moving platform, like an additional wire. This is the

29
30 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

case for example of cable-suspended parallel robots (CSPRs), i.e. CDPRs

where all forces exerted by cables have no components pointing in the same
direction of gravity and acting like multi-degree-of-freedom cranes (e.g. the
Robocrane Fig. 1.8b, the CoGiRo Fig. 1.11a, the SkyCam 1.15. etc.). Fig.
2.1 shows, instead, a CDPR prototype with a fully-constrained architecture:

Figure 2.1 – The 8-cable haptic device developed by the Fraunhofer Institute
[Yang Ho et al. 2015].

Several studies were made on fully-constrained CDPRs and CSPRs, and

most of the developed prototypes have these kinds of architectures [Albus
et al. 1993, Behzadipour & Khajepour 2006, Bouchard et al. 2010, Gout-
tefarde et al. 2011, Lamaury & Gouttefarde 2013, Kawamura et al. 2000,
Merlet 2008, 2010, Pott et al. 2010, 2013, Pusey et al. 2004, Bruckmann
et al. 2006, 2008a,b].
Otherwise, if the moving platform preserves some freedoms even if all
cable actuators are locked, the robot is referred to as underconstrained. This
happens, for example, when a CDPR has a smaller number of cables than
the end-effector dofs. Although these robots are quite overlooked in the
literature, they presents some interesting features. First, a limited number
of cables implies reduced costs and easier assembly and setup operations.
Second, having fewer cables means fewer interferences between cables or
collisions between cables and robot framework or platform. Obviously, these
advantages come at a price of reduced maneuverability, but underconstrained
CDPRs may be successfully employed for all those applications where a
limited dexterity is accepted in exchange for a reduced complexity of the
robot.
Let, for example, a CDPR with n = 3 cables be considered, with the
2.2. WORKING ASSUMPTIONS 31

anchor points on the platform being distinct. Thus, only 3 dofs of the end-
effector can be controlled, e.g. the position of a point of the platform, while
the remaining 3 (the orientation) can not. It is, however, possible to design
such a CDPR and plan its trajectories in order to move the platform along
the three directions while orientation deviations remain inside certain safety
ranges.
Another important reason that motivates the study of this family of
cable-robots is that also fully-constrained CDPRs may act like undercon-
strained ones if the external wrench applied to the end-effector requires a
negative tension in more than n − 6 cables, where 6 is the number of robot
dofs. Conversely, this fact may be also considered as an advantage and used
to extend robot workspaces to all those regions where only a subset of cables
are in tension, thus improving CDPRs reconfigurability attitudes.
Although all these reasons should encourage a careful study of under-
constrained CDPRs, little attention has been dedicated to them [Collard &
Cardou 2013, Fattah & Agrawal 2005, Fink et al. 2011, Gao et al. 2012, Hey-
den & Woernle 2006, Jiang & Kumar 2010b, Yamamoto et al. 2004, Jiang
& Kumar 2010a]. This may be caused by the complexity of their kinematic
study, which makes their control very difficult. As previously stated, since
only n dofs may be controlled (with n < 6), the platform is still movable
even if cable lengths are assigned. Thus the actual pose of the end-effector
will be ultimately determined by the wrenches acting upon it. As a conse-
quence, the set of equations deriving from the geometric constraints imposed
by cable lengths must be simultaneously solved with the relations emerging
from static equilibrium. Hereinafter, due to this coupling between kinematic
and static equations, the displacement analysis of a generic CDPR will be
defined as a geometrico-static problem. In particular the problem of finding
all the possible equilibrium poses with a given set of lengths is called a direct
geometrico-static problem (DGP).

2.2 Working assumptions

This section describes the main assumptions made to develop the model of a
generic CDPR and the procedures are then presented in the next chapters.
As described in the previous chapter, cable-robots can be built with rather
simple mechanical components. However, including a realistic model of eye-
bolts, pulleys, winch drums and cables in particular makes geometrico-static
analyses extremely complex.
CDPRs actuators are usually modeled from exit points to platform an-
32 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

Figure 2.2 – Some wire-end terminals.

chor points like rigid body extensible legs in traditional parallel manipulators,
i.e as SP S or SP U kinematic chains, where S denotes a spherical joint, P a
prismatic joint and U an universal joint. Thus, anchor points on the platform
are usually assumed to behave like spherical or universal joints. Indeed, in
some prototypes, they are made with real universal or ball-and-socket joints.
However, in order to preserve cable robot reconfigurability and ease of as-
sembly they are more often realized with hooks or carabiners connected to
eyebolts or rings, or with specific wire terminal fittings (see Fig. 2.2) or
even with simple knots. As usually assumed, hereinafter anchor points are
considered to be fixed to the end effector and corresponding, for example, to
the center of connection rings or the end of threaded sleeves (Fig. 2.3).
The methods developed and presented in this thesis are based also on
the assumption of dealing with massless and perfectly stiff cables. The main
reason for this simplification is that the primary purpose of this work is
to develop an algorithm that can rapidly solve the full direct-geometrico
static problem of a generic CDPR, considering all the possible combinations
of slack cables. This is an already complex task and the effects that both
sagging and elasticity may have on it deserve more in-depth investigations.
Some studies on cable elasticity are available in the literature (see [Pott
2010, Merlet 2008, 2010]) and also on kinematic analysis with real weighting
cable models (see [Kozak et al. 2006, Riehl et al. 2009]). However, a general
approach for the solution of the full DGP of cable robots has only recently
been presented[Merlet & Alexandre-dit Sandretto 2015]. Moreover, if a real
cable model is adopted, the part of cable that is comprised between each
winch and the corresponding exit point should also be considered.
2.3. GEOMETRICO-STATIC MODEL 33

Figure 2.3 – Assumed position of the anchor points on the platform.

As previously mentioned, cable exit points are usually modeled like spher-
ical joints, so they can be assumed as fixed points attached to the robot
framework. However, this assumption is stronger than the one made for the
anchor points on the moving platform. Indeed, cable exit points must allow
cable length changes. Generally they are made by steering roller guidance
(like the ones in Fig. 1.17) but this solution conflicts with the assumption of
considering them fixed. Indeed, as shown in Fig. 2.4, the actual exit points
vary according to end-effector pose [Gouttefarde et al. 2014, Bruckmann
et al. 2008a]. A possible solution consists in routing cables through fixed
rings with slightly bigger diameters than the cables. However, this solution
can be adopted only for small prototypes and very light loads. Otherwise,
friction effects severely modify cable behavior, thus making the analysis even
more complex with respect to the one with roller guidances.

2.3 Geometrico-static model

In CDPRs, the unilateral nature of the constraints imposed by cables does

not allow one to guarantee that in every configuration all cables remain taut.
As a consequence, in order to obtain the complete set of possible solutions
of the problem, all configurations with one or more slack cables must be
considered. Accordingly, for a generic n-n CDPR, a complete solution of
its DGP comprises not only the equilibrium configurations with n cables in
tension, but also those with 1, 2, . . . , n − 1 cables being slack. Hence, the
34 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

Figure 2.4 – Displacements of exit points when orientable roller guidances

are employed [Bruckmann et al. 2008a].

number of DGPs that must be solved is

n
X n!
NDGP = (2.1)
k! (n − k)!
k=1

This section describes how the DGP is modeled for robots with m cables
in tension, m ≤ n, and which parameterizations and equation sets are used
for each value of m. Hereinafter, an equilibrium configuration is defined
admissible if all cable tensions are positive or zero, and feasible if it is also
stable.

2.3.1 Fundamental geometric and static equations

A mobile platform is connected to a fixed base by n cables. The ith cable, i =
1, . . . , n, exits from the base at point Ai , and it is connected to the platform
at point Bi (Fig. 2.5). The platform is acted upon by a force of constant
magnitude Q applied at point G, e.g. the platform weight acting through
its center of mass. This force is described as a 0-pitch wrench QLe , where
Le is the normalized Plücker vector of its line of action. Oxyz is a Cartesian
coordinate frame fixed to the base, and O′ x′ y ′ z ′ is a Cartesian frame attached
to the end-effector (vector components in O′ x′ y ′ z ′ are denoted with the prime
mark). Without loss of generality, the fixed reference frame is chosen in such
a way that O ≡ A1 , A2 lies on the xz plane and z is directed as Le . Moreover
ai = Ai −O = [aix , aiy , aiz ], bi = Bi −O = [xi , yi , zi ], g = G−O = [gx , gy , gz ],
2.3. GEOMETRICO-STATIC MODEL 35

Figure 2.5 – The geometric model of a cable driven parallel robot with n
cables.

b′i = Bi − O′ = [x′i , yi′ , zi′ ], g′ = G − O′ = [gx′ , gy′ , gz′ ], diG = kb′i − g′ k, and
dij = kb′i − b′j k.
If ρi is the assigned length of the ith cable (taken as strictly positive),
when m cables are active (i.e. in tension) the set of geometrical constraints
imposed on the platform is formed by the following relations:
kbi − ai k = ρi , i = 1, . . . , m (2.2)
Since the platform has 6 dofs, if m < 6 its pose is ultimately deter-
mined by mechanical equilibrium. The normalized Plücker vector of the
line associated with the ith cable is Li /ρi , where, in axis coordinates, Li =
(ai − bi ) ; pi × (ai − bi ) and pi is any vector from an arbitrarily-chosen
reference point P (called, for brevity, moment pole) to the cable line. Ac-
cordingly, the wrench exerted by the ith cable on the platform is (τi /ρi )Li ,
with τi being a positive scalar representing the intensity of the cable tensile
force. Static equilibrium may then be expressed as
 
τ1 /ρ1
 . 
  . 
L1 . . . Lm Le  .  = 0 , (2.3)
| {z } τm /ρm 
M(P )
Q
with τi ≥ 0, i = 1, . . . , m.

2.3.2 Parameterization
The choice of the parameters has a great influence on the effectiveness of
interval-analysis-based methods. Indeed, if on the one hand it is better to
36 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

have a small number of unknowns in order to limit the number of boxes gen-
erated by the bisection process that have to be processed, on the other hand
a greater number of variables generally means simpler expressions which lead
to lower overestimation of the system equations.
Before introducing the adopted parameterization for the geometrico-
static model of cable robots, it is interesting to review the parameterizations
tested during the development of the algorithms and list their main issues.
Although the set of parameters currently adopted was derived from the one
presented in [Merlet 2004], the first part of the research was focused on
reviewing other parameterizations and measuring their performances when
dealing with DGPs of cable robots.

Representation of the robot pose

Let X be the list of variables representing the end-effector pose. The first
parameterization tested consists in 3 parameters that are the components
of the position vector p of a point of the platform with respect to the fixed
frame Oxyz, and 3 angles Θ = [φx , φy , φz ] (e.g. the Euler angles for three
consecutive rotations about x′ , y ′ and z ′ axis) describing
 the end-effector
orientation. So, by this parameterization X = pT , ΘT , equations (2.2) for
a generic cable i becomes

kp + R(Θ)b′i − ai k = ρi (2.4)

and, expanded, has the following expression:


2
px + cos φy cos φz b′ix − sin φz b′iy + sin φy b′iz − aix

+ py + (sin φx sin φy cos φz + cos φx sin φz ) b′ix
2
+ (− sin φx sin φy sin φz + cos φx cos φz ) b′iy − sin φx cos φy b′iz − aiy

+ pz + (− cos φx sin φy cos φz + sin φx sin φz ) b′ix
2
+ (cos φx sin φy sin φz + sin φx cos φz ) b′iy + cos φx cos φy b′iz − aiz = ρ2i
(2.5)

However, this parameterization, which is quite common in robotics, presents

two main disadvantages when dealing with interval analysis methods. The
first one consists in the great number of multiple variable occurrences in the
equation system. Indeed, in (2.5) there are 32 total occurrences of the six X
parameters, i.e.:

• 1 for px , 1 for py , 1 for pz ,

2.3. GEOMETRICO-STATIC MODEL 37

• 10 for φx , 9 for φy , 10 for φz .

This means, as explained in Sec. 1.3, that interval evaluations of these

expressions are affected by large overestimations and, thus, operators F and
E take more time to assess whether a box can can be a solution or not.
Moreover, equations coming from (2.3) are even more involved and have
higher numbers of variable occurrences than (2.5).
The second issue is that determining the tightest bounds for the initial
search domain of Θ, which have a great influence on algorithm performances,
is a challenging task. By this parameterization, the DGP with a CDPR with
m cables in tension
 is formed by a system of m + 6 relations in the 6 + m
unknowns X, τT .
As a first attempt to deal with these issues, the following substitutions
were made

cos(φx ) = cx ; sin(φx ) = sx ;
cos(φy ) = cy ; sin(φy ) = sy ;
cos(φz ) = cz ; sin(φz ) = sz ;

and consequently the three trigonometric identities were added to the system
of equations

c2x + s2x = 1; (2.6)

c2y + s2y = 1; (2.7)
c2z + s2z = 1; (2.8)

The system now has m + 6 + 3 equations in the 9 + m unknowns Y =

[pT , cx , sx , cy , sy , cz , sz , tauT ]. By this parameterization, (2.4) has a simpler
expression
 2
b′ix cy cz − b′iy cy sz + b′iz sy − aix + px
 2
+ py + (cz sx sy + cx sz )b′ix + (−sx sy sz + cx cz )b′iy − sx cy b′iz − aiy
 2
+ pz + (−cx cz sy + sx sz )b′ix + (cx sy sz + cz sx )b′iy + cx cy b′iz − aiz − rho2i
(2.9)

but there are still 32 occurrences of variables. However, while px , py and

pz still appear only once, the number of multiple occurrences for orientation
parameters drops to

• 5 for cx , 4 for cy , 5 for cz ,

38 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

• 5 for sx , 5 for sy , 5 for sz .

This not only reduces overestimation in function evaluation, but makes it

possible to easily set bounds on the orientation variables. Indeed, each c[x,y,z]
and s[x,y,z] can be initialized with the interval [−1, 1]. However, also for these
it is difficult to find the tightest initial search domain and their width still
has a great influence on computation times.
Better results were obtained by describing platform orientation with 4-
parameter representations. Quaternions, Euler-Rodrigues, and axis-angle
parameterization were tested and they all shown similar performances.
However, none of the listed parameterizations gave satisfying results (the
solution of a single DGP with these parameterizations takes more than 25
minutes), even if multiple occurrences for all the described parameterizations
may be reduced by rewriting and manipulating the equations.
All these expression, indeed, were not implemented exactly as they are
usually written in the literature because this does not assure the equation
expressions to have the lowest number of multiple occurrences. Thus, a
great deal of expertise and tests are required in order to choose the right
set of parameters for an interval-analysis-based method. Moreover, all these
parameterizations were all repeatedly tested during the development of the
codes because it may happen that improvements on filter and evaluation
operators F and E make older parameterizations more effective than the
current one.

Formulation of the static constraints

The DGP for CDPRs requires simultaneously solving the relations emerging
from both the geometrical and the static constraints. The m unknowns of
cable tensions (where m is the number of taut cables) may be eliminated
from the system of equations by observing that the system in (2.3) is linear
in τ. Thus, m linearly independent relationships may be selected with the
system and solved for the tensions. The expressions calculated this way may
be substituted back into the remaining 6 − m equation in (2.3), and, added
to the (2.2), form a system in X only.
An alternative, more elaborate, strategy to eliminate cable tensions, pre-
sented in [Carricato 2013], may be designed by observing that (2.3) admits
a solution only if
rank[M(P )] ≤ m (2.10)
Hence, by setting all (m + 1) × (m + 1) minors of M(P ) equal to zero and
by conveniently changing the moment pole, a set of linearly independent
2.3. GEOMETRICO-STATIC MODEL 39

relations only comprising the platform-pose variables may be derived, i.e.

pk (X) = 0, k = 1, . . . , Np (2.11)

where Np is an integer significantly larger than the number NX of variables

contained in X. For the DGP to admit a solution, the above equations must
be dependent, though in a non-linear way. When added to the equations
deriving from cable length constraints (2.2), (2.11) allows the pose X to be
directly solved.
However, eliminating tension with any of the two approaches presented,
as described in [Berti et al. 2013], is not convenient. The first reason is that,
by doing this, the algorithm will search for all the equilibrium configurations,
including the ones having negative cable tensions. It is still possible to
filter them by adding specific procedures to the operator F but this will
increase the computational burden. Moreover, the equations deriving from
both the approaches are remarkably more complex than the one in (2.3), and
in particular they have a higher degree, more terms and a larger number of
variable multiple occurrences. Again, this causes large interval evaluations
due to overestimation and poor performances of the algorithm.

2.3.3 DGP for robots with one cable in tension

When the platform is suspended by a single cable, under the effect of a force
QLe , it has 4 equilibrium configurations (Fig. 2.6a), but only in two of them,
i.e. Ga and Gb , is cable tension positive. Obviously, the only stable solution
is the one having the minimum potential energy, i.e. Ga .

(a) The 4 possible solu- (b) Geometric model of the DGP with
tions of the DGP with one cable in tension.
one active cable.
40 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

When only cable i is in tension, at equilibrium the end-effector can still

rotate along the z axis. However, for n ≥ 2, this rotation is bounded by
the other cables attached to the platform (Fig. 2.6b). The distance Dj
between points Bj and Aj , j 6= i, depends on the rotation angle ϑ and has
the following expression:

Dj (ϑ) = (aix + rij cos ϑ − ajx )2 + (aiy + rij sin ϑ − ajy )2 + (aiz + ρi + zij − ajz )2
(2.12)
where Oi is the projection of Bj on line Bi G, and quantities

d2ij − d2jG + d2iG

zij = kBi − Oi k =
2diG (2.13)
q
rij = kBj − Oi k = d2ij − 2
zij

only depend on the platform geometry. The minimum and maximum values
of Dj (ϑ) are obtained by setting ∂Dj /∂ϑ = 0, which yields
 
ajy − aiy
ϑDj min , ϑDj max = arctan + kπ, k ∈ {0, 1}, (2.14)
ajx − aix

where Dj min = Dj (ϑDj min ), Dj max = Dj (ϑDj max ) and Dj min and Dj max
are distinguished by evaluating the corresponding sign of the second deriva-
tive
∂ 2 Dj
= 2rij [(ajy − aiy ) sin ϑ + (ajx − aix ) cos ϑ] (2.15)
∂ϑ2
The routine outlined in Algorithm 2 computes Dj min and verifies whether
Dj min < ρj . If the latter condition is satisfied for every j = 1, . . . , n, j 6= i,
then the equilibrium configuration with the single ith cable in tension is
deemed feasible.

2.3.4 DGP for robots with two cables in tension

This problem was analytically investigated by Carricato and Merlet [Carri-
cato & Merlet 2013], and was proven to admit at most 24 real solutions.
If the taut cables are labeled with 1 and 2, then, at equilibrium, A1 , B1 ,
A2 , B2 and G must necessarily rest in a plane parallel to the z axis. Two
planar operation modes are possible, characterized by opposite orientation
of the vector (b2 − g) × (b1 − g) with respect to k × (a2 − a1 ) (Fig. 2.6).
Out-of-the-plane movements, which are very likely to occur in practice, can
cause operation-mode changes. The platform pose is described by the 9
components of the position vectors b1 , b2 and g. By conveniently rotating
2.3. GEOMETRICO-STATIC MODEL 41

Algorithm 2 DGP 1-1 for the ith cable in tension

1: for (j = 1 . . . n) do
2: if (j 6= i) then
3: if (ajx = aix ) and (ajy = aiy ) then
4: Dj min = rij 2 + (a + ρ + z − a )2
iz i ij jz
5: else
6: if (ajx = aix ) then
7: if (ajy > aiy ) then
8: ϑmin = π/2
9: else
10: ϑmin = 3π/2
11: end if
12: else
13: if (ajx > aix ) then 
a −a
jy iy
14: ϑmin = arctan ajx −aix
15: else  
ajy −aiy
16: ϑmin = arctan ajx −aix + π
17: end if
18: end if
19: Dj min = Dj (ϑmin )
20: end if
21: if (Dj min > ρj ) then
22: The configuration is unfeasible
23: Exit
24: end if
25: end if
26: end for
27: The configuration is feasible
28: Exit
42 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

Figure 2.6 – Geometric model of the DGP with two cables in tension.

the fixed reference frame along the z axis in order to have the exit points A1
and A2 lying on the xz plane, the number of variables can be reduced to 6,
since, at the equilibrium, the y component of b1 , b2 and g must necessarily
vanish. These 6 parameters are not independent, since they have to satisfy
the distance relations imposed by the geometry of the platform, namely:

kb2 − b1 k = d12
kb1 − g k = d1G (2.16)
kb2 − g k = d2G

By the aforementioned parameterization, (2.2) assumes the form

(xi − aix )2 + (zi − aiz )2 = ρ2i , i = 1, 2 (2.17)

and, by choosing O as the moment pole, matrix M in (2.3) may be written

as
 
x1 x2 − a2x 0
M = z1
 z2 − a2z −1 (2.18)
0 a2z x2 − a2x z2 x

Substituting (2.18) into (2.3) and considering relations (2.16) and (2.17),
a system of 8 equations in the 8 variables comprised in the array Y =
 T T T
X ,τ is finally obtained, where X = [x1 , z1 , x2 , z2 , xg , zg ]T and τ =
T
[τ1 , τ2 ] . This parameterization allows one to simultaneously search for so-
lutions in both operation modes.
The number of unknowns (and consequently the number of equations)
could be reduced by observing that the system in (2.3) admits a solution only
if rank(M) ≤ 2. Thus, cable tensions τ could be eliminated from equilibrium
equations by imposing det(M) = 0. This strategy, applied in [Carricato &
2.3. GEOMETRICO-STATIC MODEL 43

Merlet 2013], is not adopted here, because it does not allow an initial search
domain for the cable tensions to be specified. In this case, negative-tension
solutions should be filtered by a subsequent procedure that would increase
computation time.

2.3.5 DGP for robots with three cables in tension

The geometrico-static analysis of this family of robots was investigated in
[Carricato & Merlet 2011, 2013, Abbasnejad & Carricato 2012], where the
DGP of a robot suspended by 3 active cables was proven to admit 156
solutions in the complex field. Preliminary results obtained by solving the
same problem by interval analysis were presented in [Berti et al. 2013].
If the taut cables are labeled with 1, 2 and 3, the parameters X adopted
to describe the platform pose are the components of the position vectors b1 ,
b2 and b3 . They are not independent, since the following relations must
hold:

kbi − bj k = dij , i, j ∈ [1, 2, 3], i 6= j (2.19)

The distance equations in (2.2) assume the form

(xi − aix )2 + (yi − aiy )2 + (zi − aiz )2 = ρ2i , i = 1, 2, 3 (2.20)

and the position vector of (G − B1 ) in Oxyz may be expressed as

g − b1 = α2 b12 + α3 b13 + α4 (b12 × b13 ) (2.21)

where bij = bj − bi , and αk , k = 2, . . . , 4, are known constants obtained by

solving the system

α2 b′12 + α3 b′13 + α4 (b′12 × b′13 ) = g′ − b′1 (2.22)

From (2.21), one infers

3
X
g= αk bk + α4 (b12 × b13 ) (2.23)
k=1

where α1 = 1 − α2 − α3 . If G lies on the same plane identified by points

B1 , B2 and B3 , α4 could be set equal to zero and a simpler expression of g
would be achieved.
The relations in (2.19), (2.20) and (2.3) form a square system of 12
 T  
equations in 12 unknowns, namely Y = XT , τT = xT T T T
1 , x2 , x3 , τ . All
equations are algebraic and of degree at most 2 in the unknowns.
44 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

The number of unknowns may be reduced by eliminating cable tensions.

Different strategies can be used, and these were all tested and compared in
[Berti et al. 2013]. However, the best parameterization proves to be the one
described above, including cable tensions. This is mainly due to the reduced
overestimation provided by the simpler expression of the equations, to the
possibility of excluding negative-tension solutions from the search domain,
and to more effective filtering procedures.

2.3.6 DGP for robots with four or five cables in tension

The DGPs for CDPRs with 4 and 5 cables in tension were analytically studied
in [Abbasnejad & Carricato 2015], and they were proven to admit 216 and
140 solutions in the complex field, respectively.
Let the taut cables be labeled with 1, . . . , m (m = 4 or 5). In the following
problem-solving procedure, three different parameterizations are adopted,
depending on whether:

1. the platform anchor points Bi , i = 1, . . . , m, and G lie on the same

plane;

2. the platform anchor points Bi , i = 1, . . . , m, lie on the same plane, but

G does not;

3. The platform anchor points Bi , i = 1, . . . , m, do not lie on the same

plane.

The variables X used to describe the platform pose in case 1 are the
position vectors of 3 non-aligned points of the platform, say b1 , b2 and b3 .
These points are called reference points. The remaining m − 3 anchor points
and G are denoted as secondary points, and they are:
3
X
bi = βik bk i = 4, . . . , m
k=1
(2.24)
X3
g = αk bk .
k=1

where βi1 = 1 − βi2 − βi3 and α1 = 1 − α2 − α3 . The same approach is also

adopted in case 2, except for the position vector of G, which is expressed
using the relation in (2.23).
The parameters X in case 3 are the position vectors of 4 points of the
platform (not lying on the same plane). Similarly to case 1, the remaining
2.3. GEOMETRICO-STATIC MODEL 45

secondary points can be expressed as:

4
X
b5 = βik bk
k=1
(2.25)
4
X
g = αk bk .
k=1

where βi1 = 1 − βi2 − βi3 − βi4 and α1 = 1 − α2 − α3 − α4 . The latter

parameterization is over-redundant, since only three reference points would
be sufficient to define the platform pose. However, using only 3 points leads
to a more complex expression of the secondary points. The multiple oc-
currences of variables introduced by the cross product in (2.23) cause an
overestimation in the evaluation of the secondary points, and consequently
of the geometric-constraint equations. Indeed, the main reason for having
three distinct parameterizations, specialized according to the geometry of
the moving platform, is to be able to express each geometric constraint so as
not to contain multiple occurrences of the unknown variables. Provided that
l denotes the number of reference points, with l ∈ {3, 4}, the expressions of
the distance relations imposed by cable lengths are [Merlet 2004]:

(xi − aix )2 + (yi − aiy )2 + (zi − aiz )2 = ρ2i i = 1, . . . , l (2.26)

l
!2 l
!2 l
!2
X X X
βik xk − aix + βik yk − aiy + βik zk − aiz = ρ2i (2.27)
k=1 k=1 k=1

with i = l + 1, . . . , m.
Similarly to (2.19), another set of C2l relations (namely, 3 for l = 3 and
6 for l = 4) emerges from the distance constraints between the reference
points, i.e.
(xi − xj )2 + (yi − yj )2 + (zi − zj )2 = d2ij (2.28)

for i, j ∈ {1, . . . , l} and i 6= j.

For CDPRs with 4 or 5 active cables, equations (2.26), (2.27) and (2.28)
are not sufficient to determine the platform pose, but they must be solved
together with relations (2.3) emerging from the mechanical equilibrium, thus
adding cable tensions to the variable array, i.e. Y = [XT , τT ]T , leading to
a system of m + C2l + 6 equations in 3l + m unknowns (for l = 3 or 4,
C2l + 6 = 3l = 9 or 12).
46 CHAPTER 2. GEOMETRICO-STATIC MODEL FOR CDPRS

2.3.7 DGP for robots with six cables in tension

The DGP for this family of robots can be solved by means of the techniques
developed for the Gough-Stewart Platform. Indeed, the 6 constraints in
(2.2) imposed by the cable lengths in this case are sufficient to completely
determine the platform pose. In the literature, there are plenty of studies
about the direct kinematics of the Gough-Stewart Platform. In our code,
the DGP of a CDPR with 6 active cables is solved by the interval method
presented in [Merlet 2004], and solutions with positive cable tensions are
obtained by a filtering procedure (see Section 3.1.3).
Chapter 3

DGP solving procedure

The methods presented in this chapter solve the direct geometrico-static prob-
lem (DGP) of a generic n-n CDPR (n ≤ 6), namely a robot in which the n
exit points of cables on the base and the n cable anchor points on the plat-
form are distinct. The DGP consists in assigning the cable lengths as input
variables and in finding all possible equilibrium configurations, taking into
account the fact that one or more cables may be slack. While the DGP of a
CDPR having 6 taut cables can be dealt with using the same tools employed
for the Gough-Stewart Platform, the DGP for robots with less than 6 cables
in tension, is remarkably more complex. A successful implementation of the
methodology proposed by Carricato and Merlet [Carricato & Merlet 2010,
2013], based on exact-arithmetic elimination procedures, allowed the DGP of
2-2, 3-3, 4-4 and 5-5 CDPRs to be solved [Carricato 2013, Carricato & Merlet
2013, 2011, Abbasnejad & Carricato 2015]. In particular, least-degree uni-
variate polynomials in the ideals generated by the equations governing the
problems were found. However, the approach used in the aforementioned
contributions has the following drawbacks.

• Obtaining univariate polynomials is a challenging and not automatic

process. Moreover, they have high orders (12, 156, 216 and 140, for the
robots suspended by 2, 3, 4 and 5 taut cables, respectively), so that
their reliable solution is very difficult, as the calculation of coefficients
is very sensitive to numerical errors.

• It is not possible to incorporate constraints on the unknowns. As a

consequence, all roots (both complex and real, regardless of the tension
sign and stability) must be calculated and then post-processed in order
to discard unfeasible ones.

47
48 CHAPTER 3. DGP SOLVING PROCEDURE

Effective alternatives are provided by approaches based on floating-point

arithmetic, such as homotopy continuation. Continuation methods were
shown to be able to robustly provide all solutions for the DGP of a gen-
eral n-n CDPR [Abbasnejad & Carricato 2012, 2015]. However, this may be
quite slow and it is impossible to incorporate constraints on the unknowns.
Interval-analysis-based methods may be very effective in robotic appli-
cations [Merlet 2009, 2004] and a specific procedure for the DGP of a 3-3
CDPR was recently developed with promising results [Berti et al. 2013].
Based on these results, the described method improves the algorithm pre-
sented in [Berti et al. 2013], thus making it suitable to the DGP of a general
CDPR suspended by n cables, with n = 1 . . . 6, under the assumption that
cables are massless and perfectly stiff. As will be shown in the second text
in Chap. 4, the same algorithm can also be used to solve the DGP of robots
suspended by more than 6 cables, when cable elasticity is disregarded [Merlet
2012a].

3.1 Code structure

In this section, the most important parts of the interval-analysis code are
discussed. The code solves any geometry of cable-robots with n ≤ 6 ca-
bles. In the current version, the geometric parameters are assumed to be
known, thus not affected by uncertainties. The ALIAS C++ library [Merlet
2007] is used. ALIAS C++ is large set of interval methods mainly devoted
to the solution of robotic and kinematic problems. ALIAS C++ relies on
the BIAS/Profil library for basic interval arithmetics operations in double
precision.
The main routine of the code reads the geometry data of the n-n CDPR
from a user-defined file, thus generating a list containing NDGP sub-problems,
each solving the DGP for a subset of m cables in tension, m ≤ n. Sub-
problems are processed one at a time.

3.1.1 Initial search domain

A prerogative of any interval analysis method is that it requires an initial do-

main where solutions are searched for. For kinematics problems, this feature
is an advantage, since it allows the physical constraints of the manipulator to
be taken into account, and the code performances to be drastically improved.
3.1. CODE STRUCTURE 49

Figure 3.1 – The four different cases for the possible location of Bi according
to the location of Aj , the leg length ρj and the distance between Bi and Bj .
The admissible region zone for Bi is represented by the dashed line.

Initial bounds for the pose variables

A simple method to obtain the initial bounds of the geometrical unknowns
is the following. Let dij be the known distance between two reference points
Bi and Bj , i, j ∈ {1, . . . , l} and let the following quantities be defined:

Xij = aix + ρj + dij

Yij = aiy + ρj + dij (3.1)
Zij = aiz + ρj + dij

where dij = 0 when i = j. For each j, the coordinates of Bi must be

(component-wise) greater than [−Xij , −Yij , −Zij ]T and lower than [Xij , Yij , Zij ]T .
Thus, the interval box containing the initial search space for reference point
Bi is
 
max (−Xij ) , min (Xij ) 
 j∈{1,...,l}
 j∈{1,...,l}
 

 
Bi =  max ( −Yij ) , min ( Yij )  (3.2)
 j∈{1,...,l} j∈{1,...,l} 
 
 
max (−Zij ) , min (Zij )
j∈{1,...,l} j∈{1,...,l}

where j ∈ {1, . . . , h}.

The bounds expressed in (3.2) may be further improved. Consider a pair
of cables i, j, and denote the distance between the exit points Ai and Aj with
Dij . If the i-th cable is taut, Bi lies on the surface of a sphere Si centered in
Ai and having radius ρi . Furthermore, Bi has to lie both inside the sphere
+ −
Sij centered in Aj with radius ρj + dij and outside the sphere Sij centered
in Aj but having radius ρj − dij . Let, for the sake of simplicity, Ai and Aj
50 CHAPTER 3. DGP SOLVING PROCEDURE

+ −
be aligned on the x axis, and let Ci , Cij and Cij be the projections of Si ,
+ −
Sij and Sij in plane xz, with M1 , M2 and N1 , N2 being the intersection
+ −
points between Ci , Cij and Ci , Cij , respectively. Figure 3.1 represents the
following 4 cases:
1. if ρj + dij > Dij + ρi and max {0, ρj − dij } ≤ Dij − ρi , Bi lies on Si ;

2. if ρj + dij < Dij + ρi and max {0, ρj − dij } ≤ Dij − ρi , Bi lies on the
+
part of Si contained within Sij ;

3. if ρj + dij > Dij + ρi and max {0, ρj − dij } ≥ Dij − ρi , Bi lies on the
−
part of Si not contained within Sij ;

4. if ρj + dij < Dij + ρi and max {0, ρj − dij } ≥ Dij − ρi , Bi lies on the
+ −
part of Si contained within Sij but not contained within Sij ;
In the first case, no improvement on the bounds of Bi can be obtained
with respect to those determined by Eq. (3.2). In the second case, xM ,
which is the x coordinate of points M1 and M2 , is a better lower bound
for the first component of Bi than −Xij . In addition to that, if xM −
aix ≥ 0, also the third component of Bi can be updated to the interval
[−zM , zM ]. Similarly, in the third case, xN is a better upper bound for the
first component of Bi than Xij , and if xN −aix ≤ 0 the third component of Bi
can be narrowed to [−zN , zN ]. In the fourth case, both xM and xN improve
the initial bounds for Bi . Furthermore, thanks to the spherical symmetry of
the problem, [−zM , zM ] and [−zN , zN ] can be used to improve the bounds
for the component yi of Bi as well. If Ai and Aj are not aligned with the x
axis, a rotation matrix always exists that allows the bounds determined by
this procedure to be conveniently rotated, thus improving the actual box Bi .
For the DGP of a n-n CDPR with m cables in tension, the above procedure
has to be repeated for all pairs of taut cables i, j with i, j ∈ {1, . . . , m}. It
is possible to further improve the initial bounds for Bi , i ∈ {1, . . . , m}, by
applying the procedure for each slack cable k, with k ∈ {m + 1, . . . , n − m}
−
provided that the radius of Sik is set equal to zero. Though extremely
simple, this latter refinement may largely speed up the computation, since it
sometimes determines an empty initial search space and thus prevents any
further calculation.

Initial bounds for cable tensions

The tension variables in τ require an initial search domain, except when
n = 6. On the one hand, this can be an advantage since the lower bound
3.1. CODE STRUCTURE 51

for each element of τ can be set equal to zero and only solutions having
nonnegative cable tensions are consequently looked for. On the other hand,
choosing the upper bound of cable tensions a priori is not an easy task. For
the particular case of suspended cable robots, an upper bound for tensions
can be obtained by observing that each element of the third line of matrix
M in (2.3) is negative, i.e. M3j < 0, for j = 1, . . . , m. Thus, by considering
the third equation of system (2.3), i.e.
m
X τj
Q=− M3j (3.3)
ρj
j=1

the following inequality holds:

Qρj
τj ≤ j = 1, . . . , m (3.4)
−M3j

However, for all CDPRs which are not suspended, an upper bound for
tensions cannot be determined. Setting a too small value can cause the
algorithm to exclude feasible solutions (with high values of cable tensions),
that may actually exist within the geometrical workspace. These solutions
are the most dangerous from the mechanical point of view and it is important
to determine them, since they may possibly occur. Setting a too high value
(compared with the intensity Q of the applied load) infers a large interval
evaluation of static equations (2.3), and thus a large number of bisections
to assess whether mechanical equilibrium is attained. This leads to high
computation times.
Section 3.1.3 describes a specific filtering procedure is described, which
can drastically reduce the influence of the width of cable tension search
domain on the code performances. Thanks to this improvement, an initial
search domain for cable tension is no longer required and, when not provided,
the algorithm automatically finds all feasible solutions.

Cable numbering
For a generic n-n CDPR, with n l,the l reference points on the platform
may be chosen arbitrarily. However, this choice has an influence on both the
size of the geometric search space and the value of coefficients αk and βik
of the secondary points (cf. equations (2.23) through (2.25)), consequently
increasing or decreasing their interval evaluation. As shown in [Merlet 2004]
for the Gough-Stewart platform, these two factors, have a great influence on
the computation speed.
52 CHAPTER 3. DGP SOLVING PROCEDURE

In order to asses the performance of a given combination of reference

points, the following index is defined (for each sub-DGP) on an empirical
base:
Φ = 5Iα + 3Imin + 2IV l + IV m (3.5)
where Iα , Imin , IV l and IV m are coefficients, comprised between 0 and 1.
Iα is the (normalized) mean value of the absolute value of all coefficients αk
and βik , Imin is the (normalized) diameter of the interval having the mini-
mal width in the geometrical search domain, IV l and IV m are, respectively,
the (normalized) mean widths of the components forming the initial search
domain for the l reference points and for the m active cables. On the basis of
(3.5), the algorithm ranks all combinations of reference points and picks the
one having the lower Φ. The weights 5, 3, 2 and 1 assigned to, respectively,
Iα , Imin , IV l and IV m were chosen, after several tests, in order to provide the
combination of reference points generally leading to the best computation
time.

3.1.2 Evaluation operator

The evaluation operator E is implemented by means of the ALIAS procedure
Solve_General_Gradient_Interval (SGGI). Since the functions involved in
Eqs. (2.3,2.16,2.17,2.19,2.20,2.26,2.27,2.28) are at least of class C 2 , the Ja-
cobian matrix of the system can be analytically computed. Accordingly,
the SGGI improves the interval evaluation of the aforementioned functions
by conveniently using gradients, and by taking advantage of possible mono-
tonicities [Merlet 2007]. Moreover, SGGI exploits derivatives by applying,
during the evaluation process, a global filtering method based on the classic
interval Newton scheme, as described in [Ratschek & Rokne 1995].

3.1.3 Filter operator

The filter operator F comprises different procedures. Each of them takes an
interval box as an input and tries either to improve its bounds or to eliminate
it by using constraint propagation techniques [Jaulin 2001]. Most filtering
functions included in F rely on the 2B approach. This method consists
in rewriting an equation as the equality of two terms, determining if the
interval evaluations of both terms are consistent and, if not, using consistency
to improve the width of the interval for one or more unknowns. Let, for
instance, the equation x2 − 2x + 1 = 0 be considered. By introducing the
new variable X = x2 , the original equation may be re-written as X = 2x − 1.
Now, let [u, u] be the interval evaluation of 2x − 1. If u > 0, then the
3.1. CODE STRUCTURE 53

 √ √ 
inverse function of X indicates that x should lie in − u, u and, from
this information, the current interval of x may be updated. If u > 0, the
√ √
inverse function of X shows that x should lie outside − u, u : if the range
of x is included in this interval, then there is no solution of the equation in the
current box. The 2B approach can also be applied, with some modifications,
to constraints expressed by inequalities.

Filters on geometrical constraints

Thanks to the adopted parameterization, the 2B method applied to the
geometrical constraints expressed in (2.26) is very effective. Indeed, the 2B
method reduces the box Bi containing the intervals of coordinates xi , yi and
zi to the smallest box containing the intersection between Bi itself and the
surface of the sphere centered in Ai and with radius ρi . The same scheme is
applied for the constraints expressed by the point-to-point distance relations
(2.28).
When dealing with equations involving secondary points, such as (2.27),
first the interval components Xj , Yj and Zj of the j-th secondary point are
computed by interval evaluating relations (2.24) and (2.25); then the 2B
method is applied to all geometrical relations where secondary points are
involved, namely (2.27) and the distance relations between the secondary
points and the reference ones. If any improvement is achieved for Xi , Yi and
Zi , the 2B method is applied back to (2.24) and (2.25).
Since the 2B approach can be applied to inequalities too, the following
constraints are also considered and processed:

|bi − aj |2 ≤ (ρj + dij )2

(3.6)
|bi − aj |2 ≥ (ρj − dij )2

for any i 6= j.
When a sub-DGP is analyzed, further inequalities are added to the fil-
tering procedure, in order to take the constraints imposed by slack cables
into account, namely:
|bi − ai |2 ≤ρ2i
(3.7)
|bj − ai |2 ≤(ρi + dij )2
where i denotes one of the n − m cables considered to be slack and j denotes
one of the m active cables.
Another procedure that helps to improve the bounds on the pose variables
consists in applying the 2B method to all relations like (2.24) and (2.25)
that emerge from all possible choices of reference and secondary points (not
54 CHAPTER 3. DGP SOLVING PROCEDURE

necessarily the choice that minimizes function I in (3.5)). Finally, the 2B

approach can be applied to the geometric constraint:

(bj − bi ) · (bk − bi ) = ϕijk (3.8)

where i, j, k ∈ 1, . . . , n, i 6= j 6= k 6= i, and ϕijk = cos(B\

j Bi Bk ). The latter
is a known constant deriving from the geometry of the moving platform.

Filters on static constraints

For CDPRs having m ≤ 5 active cables, the 2B method may be applied to
the equilibrium equations expressed by (2.3). Moreover, rewriting matrix M
by choosing another exit point on the base as the moment pole, three new
equations may be obtained from the last three rows of M. Repeating this
process for all base exit points, 3(m − 1) additional relations can be used
in the filter, thus being processed with the 2B approach. However, even
if this filtering technique is quite effective, it does not reduce the negative
influence that large initial bounds on cable tension may have on the code
performances. So, another routine is implemented that almost eliminates
this dependence.
The main idea consists in trying to solve the interval linear system ob-
tained by rewriting (2.3) as:
 
L1 /ρ1 . . . Lm /ρm τ = −QLe (3.9)
| {z } | {z }
J Q

where J is a 6 × m interval matrix and Q an interval vector. When m ≤ 5,

system (3.9) is not square and admits a solution only when rank(J) ≤ m.
An approximate solution of this overdetermined system of linear interval
equations may be obtained by left-multiplying both members of (3.9) by an
arbitrary m × 6 matrix K [Rohn 1996], namely

ZJ τ = ZQ
|{z} (3.10)
C
By doing this, a square system is obtained and the interval version of the
Gauss elimination scheme provided by ALIAS [Merlet 2007] can be applied.
This algorithm, when successful (i.e. when C contains only regular matri-
ces), returns an interval vector that is the enclosure of the infinite possible
solutions of the linear interval system. According to [Rohn 1996], in or-
der to improve the algorithm performances, a good candidate for K is the
3.1. CODE STRUCTURE 55

Moore-Penrose pseudoinverse of the mid-point evaluation of J, i.e.

K = (mid(J)T mid(J))−1 mid(J)T (3.11)

However, this choice is questionable in this case, mainly due to the product
mid(J)T mid(J). Indeed, while the first three elements of each column of J
are pure numbers, the remaining three are the result of the cross product
between a dimensionless vector and a position vector whose components are
lengths. Accordingly, the elements of the matrix mid(J)T mid(J) contains the
sum between pure numbers and lengths, which is physically meaningless.
We solve this issue by considering a generalized inverse defined as

K = (WT J)−1 WT (3.12)

where W is a 6 × m matrix having the first (last) 3 rows equal to the

last (first) 3 rows of J. K is a true generalized inverse matrix, since it
satisfies the following Penrose conditions: JKJ = J and KJK = K [Penrose
1955]. By adopting W, all operations contained in (3.12) are consistent
in terms of units. Moreover, this choice still implies that, as the box size
decreases, Z tends to be the identity matrix, which is the best input for the
Gauss elimination scheme. It must be remarked that in order to reduce the
overestimation introduced by the matrix product, each element of Z can be
rewritten trying to obtain the lowest number of multiple occurrences of the
pose parameters.
A further refinement comes from considering that each element of the
interval matrix Z has the following form
6
X
Zij = Kik Jkj (3.13)
k=1

and that each column of J is a Plücker vector, so that

2 2 2
J1j + J2j + J3j −1=0 (3.14)
J1j J4j + J2j J5j + J3j J6j = 0 (3.15)

for j = 1, . . . , 6. Thus, the bounds for each element of Z may be improved

by searching the minimum and the maximum of the function
6
X
′
Zij (q) = Kik qk (3.16)
k=1
56 CHAPTER 3. DGP SOLVING PROCEDURE

where q must satisfy the following constraints:

q12 + q22 + q32 − 1 = 0

q 1 q4 + q 2 q 5 + q 3 q 6 = 0 (3.17)
Mkj ≤ qk ≤ Mkj k = 1, . . . , 6

The optimization problem emerging from relations (3.16) and (3.17) can be
analytically solved with the generalized Lagrange multiplier method and,
thus, implemented without using numeric optimization algorithms.
By virtue of the optimization of Z, improvements on the bounds of cable
tension variables can be obtained even if cable tensions are not involved in
the optimization formulation. When CDPRs having m ≤ 5 are considered,
this strategy allows the filtering procedure to partially reduce the coupling
between geometric and tension variables and, as a consequence, the influence
of the width of the tension search domain on computation time. When
dealing with the DGP of a CDPR with 6 taut cables, cable tensions are
not included in the parameterization, and the above procedure is applied to
discard boxes having negative cable tensions. In this case Z = mid(J)−1 ,
since the linear system (3.9) is square.

3.1.4 Influence on the bisection process

The effectiveness of the filtering procedures on the static constraints, espe-
cially the one relying on the Gauss elimination scheme, allows one to slightly
modify the bisection process. Indeed, rather than bisecting the largest vari-
 T
able contained in Y = XT , τT , it is more convenient to bisect the largest
variable contained in X until max(w(X)) ≥ ε (where ε is the threshold used
by operator E to stop the bisection process, see Sec. 1.3), and let F filter the
variables contained in τ. For boxes having max(w(X)) < ε, the bisection
process picks the variable having the largest width all over Y.
The advantages coming from this change on the bisection strategy are:

• a lower number of bisections

• a minimal influence of the width of the initial search domain of cable

tension variables on code performances

• small differences between the computation times for the DGP of robots
with 3, 4 or 5 cables in tension, even though they involve a different
number of variables.
3.2. CODE IMPROVEMENTS 57

3.2 Code improvements

The code proposed in the previous section, thanks to the adopted param-
eterization and the attempt to find the best combination of reference and
secondary points on the platform, can achieve really good performances.
However it is possible to further improve it by adding some features aim-
ing at obtaining better computation times as well as more reliable solutions.
Indeed, as it is presented up to now, the code still presents some issues;
these are explained hereafter together with the procedures implemented to
deal with them.

3.2.1 Uniqueness of the solution

The main issue of the algorithm presented in Sec. 3.1 is its behavior in the
case of two solutions that are close but yet distinct. Although it may be
considered only a fact of minor importance, when dealing with robot kine-
matics or, in this case, geometrico-static equations the two (or more) nearly
coincident solutions may cause an operation-mode switch due for example
to an external disturbance, which is quite likely to occur in CDPRs.
A second issue is that nothing prevents a box containing singular poses
from being considered a solution. Indeed, if the interval evaluation of a box
B∗ , with w(B∗ ) < ε, over the analyzed sub-DGP system of equations f = 0
(any of the ones presented in Sec. 2.3) contains 0, then the algorithm will
store it in the solution list, without any warning about possible singularities
contained within. Two procedures are now presented that allow these issues
to be correctly managed.

Kantorovitch operator
The numerical analysis provide a theorem that can guarantee the conver-
gence of the Newton method[Tapia 1971]. Newton-based methods, indeed,
allow the solution of a nonlinear system of equations to be found provided
that a convenient solution guess is given as an input. Though very fast, these
methods are not guaranteed to converge, and if they succeed they may con-
verge not on the solution closest to the guessed one given as input. Through
the Kantorovitch theorem both these issues can be addressed.
Let a system of h equations in h unknowns of any DGP presented in
Section 2.3 be considered. Let Y0 be a robot configuration, r0 , s0 , p0 be
positive real constants,
P U = {Y : kY − Y0 k∞ ≤ 2r0 } a ball centered in Y0 ,
and kAk∞ = maxi j |aij | be the infinity row norm of a generic matrix A.
If Y0 is such that
58 CHAPTER 3. DGP SOLVING PROCEDURE

1. the Jacobian of the system of equations has an inverse ΓY0 in Y0 such

that kΓY0 k∞ ≤ s0 ,

2. kΓY0 f (Y0 )k∞ ≤ r0,

h
X ∂ 2 fi (Y)
3. | | ≤ p0 for i, j = 1, . . . , h and Y ∈ U,
∂Yj ∂Yk
k=1

4. the constants s0 , r0 , p0 satisfy 2hs0 r0 p0 ≤ 1,

then there is a unique solution Y∗ of the system of equation f = 0 in U

and the Newton method used with Y0 as the estimate of the solution will
converge toward Y∗ .

Example 1: Let the following system be considered:

(
x2 + y 2 − 1 = 0
f (x) = (3.18)
(x − 1)2 + y 2 − 1 = 0
√
be a system
√ of equation, with solutions x∗0 = [1/2, 3/2]T and x∗1 =
T
[1/2, − 3/2] , and be
 
2x 2y
J(x) = (3.19)
2(x − 1) 2y

its Jacobian. Let x0 = [2/5, 4/5]T be a point, so matrix Γx0 is

 
−1 1/2 −1/2
Γx0 = (J(x0 )) = (3.20)
3/8 1/4

and the constant s0 may be determined as

s0 = kΓx0 k∞ = 1 (3.21)

By multiplying Γx0 by f (x0 ), the constant r0 is

1
r0 = kΓx0 f (x0 )k∞ = (3.22)
10
The third constant p0 needed by the theorem is computed by taking the max
between the infinity norms of the Hessian matrices from the two functions in
f . Neither of the Hessian matrices depends on x, since functions contained
3.2. CODE IMPROVEMENTS 59

in f are of degree 2 in the unknowns, and have infinity norm equal to 2,

which yields:
p0 = 2 (3.23)
Since the last condition required by the theorem 2hr0 s0 p0 = 4/5 ≤ 1 is sat-
isfied, the theorem guarantees that in the ball U = {x : kx − x0 k∞ ≤ 1/5}
there is a unique solution. Indeed, x∗0 ∈ [1/5, 3/5], [3/5, 1].
Then the evaluation operator E may be upgraded as follows:

• the Kantorovitch theorem is applied to mid(Bi ).

• if it succeeds

– through the Newton method, the point solution B∗i is computed

and then added to the solution list,
– from now on, all the boxes processed by F are checked for possi-
ble interferences with the box B+ i = {mid(X1 ) + [−2r0 , 2r0 ], . . . ,
mid(Xh ) + [−2r0 , 2r0 ]}, and eventually filtered;

• if it fails the box is returned to the evaluation operator unchanged.

Implementing this method in the algorithm has three main advantages:

• By this algorithm, the code can guarantee that the solution is unique
in the given box and it can calculate it within an arbitrary accuracy
and it fails only when, with the current computer precision, the value
of at least one equation cannot be determined over a box reduced to a
point: in this case, it may be shown that no other algorithm is capable
of solving the system;

• in practice, boxes no longer need to be bisected until their width reaches

the threshold ε, since Kantorovitch usually assesses the uniqueness of
a solution earlier.

• filtering the remaining boxes with B+

i may be quite effective.

As explained in [Merlet 2004], if all the equations of the solving system

are at most of degree 2 in the unknowns, implementation of the Kantorovitch
procedure can be simplified. By the adopted parameterization, the equations
involved in all DGPs fulfill this requirement, so that second-order derivatives
are constants and p0 can be precomputed at the beginning of the problem-
solving algorithm and thus, the computational burden reduced.
60 CHAPTER 3. DGP SOLVING PROCEDURE

Inflation procedure
The ball containing a single solution determined by Kantorovitch may be
widened using the inflation procedure described in [Neumaier 2001]. Let Y0∗
be a solution of the system of equations f (Y) = 0 and that Y1∗ is another
solution close to Y0∗ , so that f (Y0∗ ) = f (Y1∗ ) = 0. The mean value theorem
yields
f (Y1∗ ) = f (Y0∗ ) + J(Y)(Y1∗ − Y0∗ ) (3.24)

where J is the Jacobian matrix of the system and Y ∈ {Y0∗ , Y1∗ }. Since
(3.24) is true only when J(Y)(Y1∗ − Y0∗ ) = 0 which admits a solution only if
J(Y) is singular. Therefore, to be certain that there is no other solution in
a box centered in Y0∗ it must be checked that J(Y) is regular for any point
of the box.
This box can be obtained by means
 of the H-matrix theory described in
[Neumaier 2001, 1990]. Let B0 = [Y10∗ − δ, Y10∗ + δ], . . . , [Yh0∗ − δ, Yh0∗ + δ]
be a box centered in Y0∗ and δ a small real number, and S be an interval
matrix, obtained as follows

S = J(Y0∗ )−1 J(B0 ) (3.25)

where J(B0 ) is the interval evaluation of the Jacobian matrix over the box
B0 . Let the following interval operators also be defined

high(X) = max(|X|, |X|) (3.26)

low(X) = min(|X|, |X|) (3.27)

and the following quantities be introduced

mS = min (low(Sii )) (3.28)

i∈{1,...,h}
 
Xh
MS = max  high(Sij ) withj 6= i (3.29)
i∈{1,...,h}
j=1

If mS > MS , then J is regular over B0 and the whole process is repeated for
the new box B1 = [Y10∗ − 2δ, Y10∗ + 2δ], . . . , [Yh0∗ − 2δ, Yh0∗ + 2δ] . When
for the box Bk = [Y10∗ − 2k δ, Y10∗ + 2k δ], . . . , [Yh0∗ − 2k δ, Yh0∗ + 2k δ] this
regularity test is not satisfied, the process stops and sets Bk−1 as a new
single-solution box. This box can be used, again, to filter the other boxes
through the F operator taking advantage of possible interferences.
3.2. CODE IMPROVEMENTS 61

Also the inflation process can be simplified in the case of system of equa-
tions formed by second-degree polynomials. In this case, indeed, each com-
ponent of the Jacobian matrix  is linear in the unknowns. Let Yi0∗ be the
∗ ∗
elements of Y0 and Y∆ = Y1 + ∆, . . . , Yh0∗ + ∆ , where ∆ = [−δ, δ].
0∗

Each element Jij of the Jacobian matrix at Y∆ ∗ can be rewritten as

∗
Jij (Y∆ ) = µij + ηij ∆ (3.30)
where µij and ηij depends only on Y0∗ . Left-multiplying J by J(Y0∗ )−1 yields
J(Y0∗ )−1 J = Ih + Z (3.31)
| {z }
T
where Ih is the identity matrix of dimension h and Z an interval matrix
whose elements have the form Zij = ξij ∆ with ξij functions of η coefficients
and of the elements of J(Y0∗ )−1 . Quantities mT and MT are introduced and
computed as
mT = min (low(Tii )) = min (1 − |ξii |δ) (3.32)
i∈{1,...,h} i∈{1,...,h}
   
Xh h
X
MT = max  high(Tij ) = max δ |ξij | (3.33)
i∈{1,...,h} i∈{1,...,h}
j=1,j6=i j=1,j6=i
∗
and the following relations must be satisfied for T to be regular over Y∆
mT ≥ M T (3.34)
which yields
 
h
X
min (1 − |ξii |δ) ≥ max δ |ξij |
i∈{1,...,h} i∈{1,...,h}
j=1,j6=i
 
h
X
1− max (|ξii |δ) ≥ max δ |ξij |
i∈{1,...,h} i∈{1,...,h}
j=1,j6=i
 
h
X
1−δ max (|ξii |) ≥ δ max  |ξij |
i∈{1,...,h} i∈{1,...,h}
j=1,j6=i
1
δmin = δ ≤   (3.35)
h
X
max  |ξij | + max (|ξii |)
i∈{1,...,h} i∈{1,...,h}
j=1,j6=i

where 2δmin is the width of the box B∆ min guaranteed to contain only
solution Y0∗ .
62 CHAPTER 3. DGP SOLVING PROCEDURE

Example 2: This example shows the practical application of this latter

version of the inflation process, i.e. the one for the system of equations
consisting of functions that are at most of degree 2.
Let the same system of equation proposed in Example 3.2.1 be considered.
Assume that the solutions of the system are not known: at a certain step
of the algorithm, the Kantorovitch-theorem-based procedure finds a unique
solution in the box Bx0√ = {[1/5, 3/5] , [3/5, 1]} and then Newton returns
the solution x∗0 = [1/2, 3/2]T . Let B∗∆ = {x∗0 + ∆, y0∗ + ∆}, the Jacobian
calculated over B∗∆ is
 
∗ 2x∗0 + 2∆ 2y0∗ + 2∆
J(B∆ ) = = (3.36)
2(x∗0 + ∆ − 1) 2y0∗ + 2∆
 
∗ 2∆ 2∆
=J(x0 ) + (3.37)
2∆ 2∆

Left multiplying both sides of (3.37) by J(x∗0 )−1 brings

 
∗ −1 ∗ ∗ −1 2∆ 2∆
J(x0 ) J(B∆ ) = I2 + J(x0 ) (3.38)
| {z } 2∆ 2∆
T | {z }
Z
where, computing the product
    
1/2 1/2 2∆ 2∆ 0 0
Z= √ √ = √ √ (3.39)
3/6 3/6 2∆ 2∆ (2 3/3)∆ (2 3/3)∆

Thus, δmin can now be computed applying (3.35), which yields

√
1 3
δmin = √ √ = ≃ 0.4330 (3.40)
2 3/3 + 2 3/3 4

This result
 √ (as shown √ also in
√Fig. 3.2)
√ guarantees
 that in the box B∗∆ =
(2 − 3)/4, (2 + 3)/4 , 3/4, 3 3/4 there is a unique solution and
provides a much wider uniqueness domain. From this example it is possible
to show that if the equations in f = 0 are at most of degree 2 in the unknowns,
once the inverse of the Jacobian matrix J(Y0∗ ) is computed it is possible to
immediately obtain the largest bound of the box where the solution is guar-
anteed to be unique, since all the µ and η coefficients can be precomputed.
All the described parameterizations in Sec. 2.3 fulfill this requirement so
this simplified version of the inflation process can be implemented and the
evaluation operator E further upgraded.
3.2. CODE IMPROVEMENTS 63

Figure 3.2 – The solution box after the inflation process.

• the Kantorovitch theorem is applied to mid(Bi ).

• if it succeeds
– through the Newton method, the point solution B∗i is computed
and then added to the solution list,
– the inflation process is applied in B∗i and a larger box Bi∆ is
computed where the solution is unique
– from now on, all the boxes processed by F are checked for possible
interferences with the box Bi∆ and eventually filtered;
• if it fails the box is returned to the evaluation operator unchanged.

3.2.2 Stability
A third important refinement consists in checking the solution for stability.
All the configurations that are deemed solutions by the algorithm are indeed
equilibrium equations. However, only stable ones are feasible in practice.
Thus each solution identified by E can be processed by a procedure that
determines whether the corresponding equilibrium configuration is stable or
not. A sufficient condition for the equilibrium to be stable consists in having
the (reduced) Hessian matrix Hr being positive definite. If I3 denotes the
3 × 3 identity matrix, and ñ denotes, for a generic vector n, the skew-
symmetric matrix associated with the operator n×, Hr may be computed
as (Carricato & Merlet [2013])
Hr = NT HN, (3.41)
64 CHAPTER 3. DGP SOLVING PROCEDURE

where
m
" #
X τi I3 h −(b̃i − g̃) i
H= 1 ,
ρi (b̃i − g̃) 2 (b̃i − g̃) (g̃ − ãi ) + (g̃ − ãi ) (b̃i − g̃)
i=1
(3.42)
and N is a 6 × (6 − m) matrix whose columns generate the null space of the
Jacobian matrix JT :
 
(b1 − a1 )T /ρ1 (b1 − g) × (b1 − a1 )T /ρ1
JT = 
 .. .. 
(3.43)
. . .
(bm − am )T /ρm (bm − g) × (bm − am )T /ρm

in this case computed assuming G as the moment pole. Finally, the final
upgraded version of the evaluation operator E is

• the Kantorovitch theorem is applied to mid(Bi ).

• if it succeeds

– through the Newton method, the point solution B∗i is computed

and then added to the solution list,
– then B∗i is checked for stability, and if the test succeeds it is
marked as feasible
– the inflation process is applied in B∗i and the a larger box Bi∆ is
computed where the solution is unique
– from now on, all the boxes processed by F are checked for possible
interferences with the box Bi∆ and filtered if necessary;

• if it fails the box is returned to the evaluation operator unchanged.

3.2.3 Parallel implementation

Most interval-analysis-based algorithms are appropriate for a distributed im-
plementation, since processing a given box does not generally depend on pro-
cessing of the other boxes in the list. A master computer manages the list
and it sends a box to a slave computer. The slave executes the algorithm,
by performing a few bisections. Then, it returns the remaining boxes to the
master, and it waits for a new box to process. Such a scheme may be easily
implemented in a network of workstations. The decrease in computation
time will be, in general, less than proportional to the number of slaves, due
to the overhead of data transmission between the master and the slaves.
3.3. CODE STRUCTURE: BETA VERSION 65

This approach may also take advantage of modern multi-core CPU ar-
chitectures. By following this scheme and by using POSIX thread libraries,
a distributed implementation of the DGP code was set up on a single work-
station with a multi-core CPU. In the first step, an instance of E generates
a few boxes and it stores them in the list L. Then, a number of threads
equal to the number of CPUs is created, with each one taking a box from
L. A local instance of E performs an assigned number of bisections and it
appends the generated boxes to L. The solutions that are found, if any, are
appended to the solution list S.

3.3 Code structure: beta version

The code described in Sec. 3.1, together with the improvements presented
in Sec. 3.2, was extensively tested and shown very good performances. How-
ever, its implementation can be further enhanced and a first beta version of
the current algorithm was developed in order to address the following issues:

• since the filter operator F of the actual code was developed incremen-
tally, some modules repeat computations made by previous parts of
the code, so an in-depth review of the methods implemented within
may lead to huge simplifications;

• the actual version of the code processes each single DGP problem sep-
arately and the solution of the full DGP is made up of the sum of all
sub-DGPs, then this new version presents an approach that considers
all DGPs at the same time.

The main idea for this new version of the code consists in trying to con-
sider the DGP of a generic CDPR as a single problem and whose solution
can be computed within a unique interval algorithm. During the implemen-
tation of this approach, as will be explained below, several improvements to
the original code were made.

3.3.1 Initialization
Let a CDPR with n cables be considered. Its platform is parameterized as
described in Sec. 2.3 and a set of l reference and n − l secondary points is
conveniently chosen.
For each of the NDGP configurations with different sets of taut cables,
using the procedure described in Sec. 3.1.1, the code computes the exis-
tence bounds for the platform points bi and g and the cable tensions τ and
66 CHAPTER 3. DGP SOLVING PROCEDURE

generates an initial box Ω1 as

N[
DGP

Ω1 = BT T T
1 , . . . , B n , Bg i
(3.44)
i=1

where Bg is the box containing the initial search domain for g.

All the configurations with m < n cables in tension are preprocessed and,
for each one, the code computes the set of reference points and stores the
values of coefficients α and β for the corresponding geometry of the platform.
Finally, a constant vector w, with NDGP components is generated. Each
element of w identifies a sub-DGP configuration and it is initialized with 1 if
the corresponding search domain BT T T
1 , . . . , Bn , Bg i , with i = 1, . . . , NDGP ,
is not empty, and with 0 otherwise.
Provided that Ψ1 is the box containing the search domain for the n cable
tension, at the end of the initialization process the algorithm returns the first
input box for the algorithm, i.e.

V1 = Ω 1 , ΨT 1 , w1 (3.45)

During this initialization process, all the constants needed by the Kan-
torovitch procedure and the inflation process are computed.

3.3.2 Description of the algorithm

The algorithm structure, that was modified with respect to the one of al-
gorithm 1, taking into account the improvements made on the evaluation
operator E, proceeds as in algorithm 3. Its main components are described
hereafter.

The filter operator Fplatf orm

This operator filters the intervals of the n + 1 platform anchor points and
g. It applies the 2B method, explained in Sec. 3.1.3, to all the constraints
concerning the moving platform geometry, namely
• the distance relations between platform anchor points bi , with i =
1, . . . , n and g

kbi − bj k =d2ij (3.46)

kbi − g k =d2iG (3.47)

with i, j = 1, . . . , n, i 6= j.
3.3. CODE STRUCTURE: BETA VERSION 67

• all the relations (2.21) and (2.25) for any possible combination of ref-
erence and secondary points

• all the constraints deriving from (3.8)

If the 2B procedure finds any of these constraint to be inconsistent, Fplatf orm

returns −1. Otherwise it returns 1 if Ωi bounds are improved or 0 if the box
remains unchanged.

The filter operator Fadmissible

This filter checks if the actual pose Ωi contained in the analyzed box Vi is
compatible with the following n cable constraints

kbi − ai k ≤ρi (3.48)

kbj − ai k ≤ρi + dij (3.49)

The filter returns −1 if at least one of the two conditions is not fulfilled, i.e.
when for at least for one i or for a couple i, j, kbi − ai k > ρi or kbi − aj k >
ρj + dij .
Moreover, from (3.48), if kbi − ai k < ρi then cable i is certainly slack.
As a consequence, the filter sets to zero all configurations in wi containing
cable i, and the corresponding cable tension search domain Ψ(i). Then the
filter tries to improve bounds for Ωi by applying 2B method to inequalities
(3.48) and (3.49), and if it succeeds it returns 1. If the box Vi remains
unchanged it returns 0.

The procedure Psolve

When there is only one admissible configuration in the wi vector contained

in the box Vi , this procedure is called by the main algorithm and determines
whether Vi contains solutions for this single sub-DGP and eventually up-
dates the solution list S. As previously explained, during the initialization
procedure the code has already stored the parameterizations and equation
sets for each sub-DGP, with the corresponding set of reference and secondary
points and relative coefficients α and β. The procedure proceeds like Algo-
rithm 1, except for the evaluation operator E, which behaves as follows.

• It computes few iterations of the Newton-Raphson scheme, taking the

mid({Ωi , Ψi }) as guess point.
68 CHAPTER 3. DGP SOLVING PROCEDURE

• If the approximate solution is still in the box, the Kantorovitch-theorem-

based procedure is applied in order to check whether there is a single
solution of the system in a box centered in the approximate solution
with a known width.

– if it succeeds, then the Newton-Raphson scheme is continued and,

thanks to the Kantorovitch theorem, it is now guaranteed to con-
verge toward a unique solution. Once the solution is found and it
is stored in S, the inflation procedure is applied in order to obtain
a box where the solution is guaranteed to be unique. Finally, this
box is stored in the memory and used by the filter operator F
to reduce the size of the adjoining boxes and E returns 1. The
solution box is processed by the stability procedure.
– if the Kantorovitch test fails, then no guarantees can be provided
for the approximate solution so the E returns 0.

• It returns 0 if after a few iterations the approximate solution is outside

the considered box,

The first main difference in this version of operator E is that the evalua-
tion of the system of equation f = 0 is avoided, because it is already included
in F. Second, the system of equations for the statics of each sub-DGP was
slightly updated in order to take into account a generic external wrench act-
ing upon the platform, and not only a single force. Third, a module that
filters boxes by intersecting them with the enlarged solution boxes coming
from the inflation process is added to F.

3.3.3 Remarks
The beta version of the code structure can handle the DGP of CDPR with
n cables as a unique problem and gives interesting, but preliminary, results.
However, it is in an early development stage, its procedures are not defini-
tive since relevant changes on the algorithm structure were made and other
specific and more efficient procedures may be added to it. Moreover, it has
not completed the test phase yet and still presents minor bugs.
3.3. CODE STRUCTURE: BETA VERSION 69

Algorithm 3 Problem-solving algorithm scheme: beta version

1: i = 1, N = 1, L = {V1 }, S = {}
2: if i > N then return S;
3: end if
4: Vold = Vi
5: if Fplatf orm (Vi ) = −1 then
6: i=i+1
7: go to 2
8: else
9: Vi = Fplatf orm (Vi )
10: end if
11: if Fadmissible (Vi ) = −1 then
12: i=i+1
13: go to 2
14: else
15: Vi = Fadmissible (Vi )
16: end if
17: if wiT wi = 1 then
18: Vi has a single admissible configuration
19: Psolve (Vi , S)
20: else
21: if Vi = Vold then
22: pick k such that w[k] = 1
23: create two new boxes Vi′ = Vi′′ = Vi
24: wi′ = {. . . , wk−1 , 0, wk+1 , . . . }
25: wi′′ = wi − wi′
26: replace Vi with {Vi′ , Vi′′ } in L
27: N = N + 1, i = i + 1
28: go to 2
29: else
30: pick xk such that w(Xk ) = maxk∈{1,...,3h} (w(Ωi [k]))
31: bisect Xk in the middle point
32: create two new boxes Vi′ and Vi′′ from Vi
33: replace Vi with {Vi′ , Vi′′ } in L
34: N = N + 1, i = i + 1
35: go to 2
36: end if
37: end if
Chapter 4

Experiments and Results

In this section the results obtained during the test phase on the developed
algorithms are presented. As previously stated, the beta version of the code
is currently under development so computation times are not suited to make
comparisons between methods.
In order to investigate the feasibility of solutions found by the algorithm
and to test the real likelihood of operation mode changes, a small prototype
for teaching and experimental purposes was built during the last year of
the Ph.D. program. On this prototype, some preliminary tests were done
in order to implement a real-time compliant version of the developed DGP
solving procedures.
The first part of this chapter describes the teaching prototype and gives
an overview of the results obtained by the real-time procedures developed
for its control. In the second part some results from tests of the two versions
of the DGP solving algorithm are reported and discussed.

4.1 Description of the teaching prototype

4.1.1 Hardware
The prototype consists of a wooden framework (Fig. 4.1a) built on a square
base with sides of 800mm, and height 700mm. The exit points are realized by
means of fishing rod top guides (Fig. 4.1b, that have inner rings in ceramic
material in order to have a low coefficient of friction.
Cables are common single braided Nylon fiber lines and their diameter
is 1mm. The end-effector is a 70mm diameter steel flange and cables are
connected to it through knots that are held firm by means of nut and bolt
(Fig. 4.1a).

71
72 CHAPTER 4. EXPERIMENTS AND RESULTS

(a) An overview of the teaching (b) A detail of the exit points.

prototype.

(a) The end-effector. (b) A winch.

The winches are realized by coupling timing belt pulleys (Fig. 4.1b) to
the 6 Phidget R DC gear motors whose characteristics are reported in Table
4.1. Each motor is actuated by a Phidget R driver, whose characteristics are
listed in Table 4.2. All the drivers are powered by an AC/DC adapter from
an old PC and connected through a USB hub to a personal computer that
runs the control algorithm.

The winches are located and oriented with respect to the exit points in
order to avoid cable overlapping on pulleys and, as shown by experiments,
if this positioning is accurate, there are generally no cable coiling issues.
However, this badly affects the reconfigurability of cable exit-points.
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 73

Motor properties
Motor type DC Motor with encoder
Mechanical output power 7W
Maximum speed at rated voltage 73RPM
Rated torque 0.61Nm
Stall torque 7.04Nm
Electrical properties
Rated voltage 12V DC
Rated current 1.1A
Stall current 5.0A
Gearbox specifications
Gearbox type Spur
Gear ratio 50 : 1
Number of gear trains 4
Maximum strength of gears 1.2Nm
Shaft diameter 6mm
Shaft maximum axial load 35N
Shaft maximum radial load 25N
Encoder specifications
Encoder resolution 360CPR
Connector type E4P

Table 4.1 – Properties of the Phidgets R 3260E_0 gear motors.

4.1.2 Software

The control routine of the manipulator is managed by a personal computer

connected to the motor drivers through a USB cable. A user software inter-
face was developed in order to simplify testing operations and allows several
parameters to be controlled also during robot movements. Fig. 4.1 shows a
screenshot of the online monitoring module of the graphical user interface.
Among the other features, it makes it possible to slightly adjust coefficients
of the PID controller and switch between the different routines that compute
the forward kinematics.
74 CHAPTER 4. EXPERIMENTS AND RESULTS

Controller properties
Motor type DC Motor
Velocity resolution 0.39% duty cycle
Acceleration resolution 0.39% duty cycle/s
Acceleration min 24, 5% duty cycle/s
Acceleration max 6250% duty cycle/s
Acceleration time min 32ms
Acceleration time max 8.2s
Encoder interface
Number of encoder inputs 1
Count rate max 500000 cycles/s
Encoder interface resolution ×1
Update rate 125 samples/s
Time resolution 0.33ms
Electrical properties
Supply voltage min 9V DC
Supply voltage max 28V DC
Continuous motor current max 5A
Over-current trigger 8A
Current consumption min 20mA
Current consumption max 100mA

Table 4.2 – Properties of the Phidgets R 1065_0 single motor driver.

Inverse kinematics

The control routine has 4 different modules computing the inverse kinemat-
ics, one for each underconstrained robot with m = 3, . . . , 6 cables in tension.
When the robot has m < 6 cables in tension, 6 − m degrees of freedom of
the platform cannot be controlled.
Let the 6 degrees of freedoms of the moving platform be described by 3
parameters xO′ , yO′ and zO′ for the position of the origin O′ of the mobile
frame in Oxyz and by 3 angles ϕx , ϕy and ϕz , representing the orientation
of the platform by three consecutive rotations along the x, y and z fixed
axes. When the robot has m = 3 cables in tension, the inverse kinemat-
ics guarantees the control over the 3 position parameters, when m = 4
the user may choose to control also one of the two rotations ϕx or ϕy ,
and if m = 5 both ϕx and ϕy are managed. By means of this parame-
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 75

Figure 4.1 – The control user interface of the teaching prototype.

terization, the inverse kinematics is dealt with by solving the m distance

relations (2.2) together with the equation from static equilibrium (2.3).
This system of equation has m + 6 equations in the m + 6 + m variables
[ρ1 , . . . , ρm , xO′ , yO′ , zO′ , ϕx , ϕy , ϕz , τ1 , . . . , τm ]. Once m platform pose pa-
rameters are assigned, the system becomes square and by solving it, the m
cable lengths are obtained.
However, this scheme for the inverse kinematics of underconstrained
robot was developed only for testing purposes of the feasibility in real-time
of the algorithms for the forward kinematics problems. A new and more re-
fined module for the inverse kinematics is now currently under development,
always based on interval analysis.
In the inverse kinematics routine, the working assumptions are the same
presented in Sec. 2.2. They are proved to be reasonably correct in most of
the robot workspace, considering also the overall (moderate) manufacturing
precision of the hardware. However, for heavy payloads and/or in the upper
regions of the workspace, where cable tensions are considerably higher, there
are relevant deviations on the desired platform pose due to cable elasticity.
Moreover, friction between cables and the exit points, during cable coiling,
causes the cable segment between the pulley and the exit point to have a
higher tension than the segment between the exit point and the correspond-
ing anchor point. Indeed from Fig. 4.2 it is possible to observe that cables
have different wrapping angles on the surface of the fishing rod top eyes.
This phenomenon was observed performing a particular lifting operation
76 CHAPTER 4. EXPERIMENTS AND RESULTS

Figure 4.2 – The large wrapping angle of the cable on the right on the fishing
rod top eye cause a non uniform tension along the cable.

where both cable tensions and lengths were supposed to be all equal. As
the platform height gets closer to one of the exit points, and consequently
cable tension become higher, the error on platform orientation increases. As
shown in figures 4.3a and 4.3b, this error vanishes if, when the platform is
still, the cables are momentarily released (for example by manually tilting
the platform).

(a) (b)

Figure 4.3 – The moving platform before (a) and after (b) the manual redis-
tribution of cable tensions

A few attempts were made to model this behavior by introducing the

capstan theory, expressed by the Eytelwein formula in (4.1), in the inverse
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 77

kinematics. However, the uncertainties on the coefficient of friction ζ be-

tween the fiber lines and the rod top guide eyes and on the wrapping angle
γ lead to exponential error on the estimate of tension distribution in the two
cable sections.
τP A = τAB eζγ (4.1)
This problem can be significantly reduced or completely eliminated by
employing cables with a higher stiffness and/or by introducing rolling ele-
ments on the exit points.

Forward kinematics
The procedure that solves the direct geometrico-static problem is imple-
mented in the control routine as follows. Basically, it relies on the Psolve
procedure of the beta version of the code presented in Sec. 3.3.2. Assume
that at time t0 of the control routine the robot has the platform pose Y0 for a
given set of cable lengths ρ0 . Then, in order to follow the desired trajectory,
the controller computes the inverse kinematics and the motors move to a new
set of cable lengths ρ1 d. After a delay of few milliseconds ∆t , that is the
system interrupt time, the controller computes the cable lengths by reading
the actuator encoders. Without additional sensors, this estimate of cable
lengths is the only measurement available for feedback position control and
thus can be considered the real set of cable lengths ρ1 at time t1 = t0 + ∆t .
In order to measure the error of the platform pose with respect to the desired
trajectory, the actual pose Y1 satisfying the new cable constraints ρ1 must
be computed.
Let a CDPR with 6 cables in tension be considered. According to the
parameterization used for the inverse kinematics, assume that the upper
bounds on platform velocities vmax and ωmax are known. Thus, it is possible
to determine a bounding box for the reference points bi , with i = 1, . . . , h.
Each bounding box is centered in b0i , i.e. the coordinates of the ith reference
point at Y0 , and its edges are 2∆t (vmax + ωmax kb′i k), with i = 1, . . . , h.
The steps of the procedure are described hereafter.

• Kantorovitch is applied to Y0 to see if there is a box containing the

unique solution Y1 .

• If it succeeds, then the Newton-Raphson scheme is guaranteed to con-

verge on the unique solution and then it is computed with an arbitrary
precision.

• If it fails, the box B = {B1 , . . . , Bh } is processed by filter F.

78 CHAPTER 4. EXPERIMENTS AND RESULTS

– If the filter eliminates the box, a solution cannot be found in B.

The robot has to be immediately stopped.
– If the filter improves B bounds, the algorithm is repeated from
the beginning but Kantorovitch is applied to the mid(B).
– If the filter leaves the box unchanged, then the algorithm creates
a list L and bisects the box. Then the algorithm starts again from
the beginning and applies Kantorovitch to the mid-point of each
box in the list.

If at the end of the algorithm more than one solution has been found in the
box, the robot is stopped.
This approach, from the tests done so far on the teaching prototype,
allows the DGP for a CDPR with m cables in tension, with m = 3, . . . , 6,
to be solved with performances equivalent to standard Newton methods.
Indeed, as explained in Sec. 3.2.1, since the matrix containing the second
order derivatives can be precomputed because it is constant, the overhead
required by this test is due to the numerical inversion of the system Jacobian
matrix. From the tests done on the teaching prototype, Kantorovitch was
successful in all tests, in every point of different trajectories.
Then the test were repeated by disabling the Kantorovitch module in the
algorithm and with the only action of operator F and after a few bisections,
a solution for the DGP (with an accuracy threshold of ε = 1.e − 4) is found
usually within 4 ms, and no more than 7ms. However this is a worst case
scenario, since even if Kantorovitch fails at the first step of the algorithm, it
usually succeeds after a few bisections and the solutions in the box can still
be computed with arbitrary precision.

Remarks on the real-time procedure

Although very effective, this real-time procedure for solving the DGP of
cable robots still presents many open issues, described below in order of
importance.
The first one is that the above described procedure is not able to deal
with the sub-DGPs. Indeed it considers only the case where all the m cables
are taut. However, a module was added to check if the robot platform is
getting near to a sub-DGP solution. Let a CDPR with m cables in tension
be considered and assume that in a certain point of the followed trajectory,
the tension of one cable becomes nearly 0. When the tension of a set of
k cables, with k = 1, . . . , m − 2 is smaller than a user-defined threshold,
e.g. 0.1, the controller starts the real-time solving algorithm for a CDPR
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 79

with m − k taut cables on a parallel thread. If, in this parallel instance,

Kantorovitch succeeds in finding another unique solution in a bounded box,
then the user interface sends out a warning. Moreover, if the main algorithm
fails in finding an admissible solution because the k cable tension are 0, the
controller automatically switches to the parallel instance of the DGP solver.
The preliminary tests are promising, and by exploiting multi-core CPUs there
are no measurable differences on algorithm performances. However, this
approach still does not offer a comprehensive solution of the DGP problem
for cable robots that is real-time compliant.
The second issue concerns the behavior of the real-time module when the
Kantorovitch test fails. As described previously, when m = 6 and maximum
velocities of the platform are available, the solving routine still manages
to find a solution, though only up to a certain accuracy. However, giving
bounds to the search domain when less than 6 cables are taut is not simple.
The first reason is that since not all the platform degrees of freedom are
controlled, the end-effector cannot be easily bounded. The second reason is
that the cable tension needs a search domain too.
The third challenging aspect related to this algorithm is that its effec-
tiveness is highly dependent on both the choice of the reference points and
the problem scale. The first issue can be easily explained by observing for
example one equation from (2.27), i.e.

h
!2 h
!2 h
!2
X X X
βik xk − aix + βik yk − aiy + βik zk − aiz = ρ2i (4.2)
k=1 k=1 k=1

and the first element of the ith row of the Jacobian matrix of the system is
h
!
X
Ji1 = 2βi1 βik xk − aix (4.3)
k=1

Thus the first row of the Hessian matrix Hi of the ith equation will be, if
h = 4,
 2 
Hi1,1..12 = 2βi1 0 0 2βi1 βi2 0 0 2βi1 βi3 0 0 2βi1 βi4 0 0
(4.4)
The other rows of this hessian matrix contain the same elements but in
different positions. Thus the norm of Hi is given by
4
X
kHi k∞ = 2|βi1 | |βik | (4.5)
k=1
80 CHAPTER 4. EXPERIMENTS AND RESULTS

The maximum value of kHi k among all the equations in the system gives
the value of the constant p0 of the Kantorovitch theorem (see Sec. 3.2.1)
and, being considered the fourth condition required by the theorem, if p0
is small, the theorem has more chances to be successfully applied. With
the current parameterization, the lowest value for the constant p0 is 4, i.e.
the norm of the hessian matrices related to point to point distance relations
(2.28). Thus, considering also that coefficients β and α appear in the static
equations, it is important to choose among the different sets of reference
points the combination that gives the lowest coefficients β and α not only to
reduce overestimation in secondary points evaluation but also to minimize
p0 .
Here, instead, an example is proposed that shows how Kantorovitch suc-
cess is affected also by the problem scaling. Let a robots with two cables be
considered. By the notation presented in Sec. 2.3 the coordinates of its exit
points are a1x = 0, a1z = 0, a2x = 7 and a2z = 1.5 and platform geometry
is defined by d12 = 3, d1g = 2, d2g = 2. All the measurements are in meters.
For cable lengths ρ1 = 3 and ρ2 = 5.5 the system has a solution Y0 in
            
x1 1.1197 x2 1.6640 xg 0.0909 τ1 0.9795
= = = =
z1 −2.7832 z2 0.1679 zg −1.0681 τ2 0.3768
(4.6)
Suppose that new cable lengths are now given by motors, and they are
ρ1 = 3.003 and ρ2 = 5.503. So, the inverse of J(Y0 ) is computed and its norm
is kΓY0 k∞ = 3.5418 = s0 . Constant r0 is computed as kΓY0 f (Y0 )k∞ =
0.00704. Since constant p0 = 4 in this case and the dimension of the problem
is 8 the following inequality does not hold

16p0 s0 r0 = 1.5957  1 (4.7)

However, if the geometric data are all multiplied by a scale factor sf = 1/2,
an equivalent problem is obtained. But if all the computations are repeated,
since the following constants are now s0 = 4.3549, r0 = 0.00211 and p0 = 4,
the inequality in (4.7) is now satisfied

16p0 s0 r0 = 0.5886 ≤ 1 (4.8)

and the theorem now holds. However, how to determine the best scale factor
for a given direct geometrico-static problem is a subject that deserves more
investigations, and up to now we have dealt with it only by an empirical
approach.
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 81

On-line computing of Wrench Feasible Workspace boundaries

In the perspective of developing a user interface which includes several tools

for the analysis of the teaching prototype presented previously, a module for
the on-line analysis of its constant-orientation workspace was been imple-
mented. However, this is still at an early stage of development and the vast
literature devoted to this subject suggests a more in-depth investigation of
that topic [Pusey et al. 2004, Gouttefarde et al. 2006, Perreault et al. 2010,
Ruiz et al. 2015, Gouttefarde et al. 2011, Bruckmann et al. 2008c].
Let the platform pose be defined by the six parameters [xO′ , yO′ , zO′ ,
ϕx , ϕy , ϕz ]. Another real-time module that can be implemented in the con-
trol routine for the configuration with m = 6 cables in tension consists in a
procedure that computes, for a given set of pose parameters [zO′ , ϕx , ϕy , ϕz ],
the boundaries for xO′ and yO′ of the wrench feasible workspace (WFW).
Let us consider the 6 equilibrium equations in (2.3)
 
t1
   .. 

Le  .  = 0 ,

L1 . . . Lm (4.9)
| {z }  t6 
M
Q

where, since ρi > 0, ti = τi /ρi ≥ 0, i = 1, . . . , 6. Once [zO′ , ϕx , ϕy , ϕz ] are

assigned, the system has 6 equations in the 8 unknowns [xO′ , yO′ , t1 , . . . , t6 ].
By observing that, if a given pose has at least one cable tension equal to zero,
it is outside the WFW, when only a given unknown ti = 0, the considered
pose is lying on the frontier of the WFW.
By recalling that the equations (4.9) are linear in the unknowns [t1 , . . . , t6 ]
and by setting one unknown ti = 0, it is possible to choose 5 equations among
the 6 in (4.9), solve them for the 5 unknowns [t1 , . . . , ti−1 , ti+1 , . . . , t6 ] and
then substitute them back to the remaining equation.
This function γi will depend on the sole variables xO′ and yO′ and it
represents a part of the WFW boundary. By repeating this approach for ti ,
with i = 1, . . . , 6, six curves γi can be analytically computed and they are
all polynomials of the form

γi = ci0 + ci1 x + ci2 y + ci3 xy + ci4 xy 2 + ci5 x2 y + ci6 y 3 + ci7 x3 + ci8 x2 + ci9 y 2
(4.10)
where i = 1, . . . , 6 and coefficients cik contain the geometric data of the
robot and the assigned pose parameters. Each curve γi (as shown in Fig.
4.4) divides the plane xO′ yO′ in two regions Ψi− and Ψi+ where ti is negative
82 CHAPTER 4. EXPERIMENTS AND RESULTS

Figure 4.4 – Wrench feasible workspace determination for assigned orienta-

tion and zO′ of the moving platform.

or positive, respectively. Thus the WFW is given by the intersection of all

the Ψi regions, with i = 1, . . . , 6.
The first step of the adopted strategy consists in computing all the inter-
sections λij between each pair of curves γi and γj , with i, j = 1, . . . , 6, j > i.
This can be done by eliminating, through software of symbolic algebra, vari-
able yO′ (or xO′ ) from each pair of functions γi and γj , obtaining C26 = 15
univariate polynomials pij in the only variable x0′ (or yO′ ). These polyno-
mials are all of degree 10 and their coefficients depend on cik and cjk .
Once all the real intersections λ are computed, it is possible to check all
the parts of curves comprised between two intersections to determine whether
they are part of the WFW boundaries. Let an intersection λij = (xij , yij )
be considered (see Fig. 4.5) and rit and rjt be the lines tangent to γi and γj
in λij . The expression of rit is, for example:

γix (x − xij ) + γiy (y − yij ) = 0 (4.11)

where
∂γi
γix = (4.12)
∂x xij ,yij
∂γi
γiy = (4.13)
∂y xij ,yij
4.1. DESCRIPTION OF THE TEACHING PROTOTYPE 83

Figure 4.5 – Detail of the two lines bisecting the angles formed by the two
tangents of γi and γj in λij .

′ and r ′′ bisecting the angles formed by the inter-

Then the two lines rij ij
section of rit and rjt are derived and their expressions in parametric form
are, respectively
(
′ x = xij − b′ δ ′
rij : (4.14)
y = yij + a′ δ ′
(
′′ x = xij − b′′ δ ′′
rij : (4.15)
y = yij + a′′ δ ′′

where δ ′ and δ ′′ are real parameters and

γix γjx γiy γjy
a′ = q +q b′ = q +q (4.16)
2 + γ2
γix 2 + γ2
γjx 2 + γ2
γix 2 + γ2
γjx
iy jy iy jy
γ ix γ jx γ iy γ jy
a′′ = q −q b′′ = q −q (4.17)
2 2
γix + γiy 2 2
γjx + γjy 2 + γ2
γix 2 + γ2
γjx
iy jy

Then, variables [t1 , . . . , t6 ] are computed in the two pairs of points lying
on rij′ and r ′′ obtained by substituting δ ′ = ±ǫ and δ ′′ = ±ǫ in (4.15), where
ij
ǫ is an arbitrary small real number. If for one point, ti > 0, i = 1, . . . , 6,
it is possible to consequently determine the corresponding sections of curves
γi and γj constituting part of the WFW boundary. The algorithm proceeds
84 CHAPTER 4. EXPERIMENTS AND RESULTS

analyzing all the intersections λ and determines whether all the sections of
curves γ deemed part of the WFW boundary form closed loops.
This procedure runs in less than 1.5ms so it is manageable in real-time.
It is possible, thus, to assign platform orientation and zO′ and computes at
each cycle time the boundaries of the WFW of xO′ and yO′ . However, this
algorithm is at its preliminary stage of development and its implementation
will be further improved in the future. For example, since the expressions
of curves γ are known, the area of a section of the WFW determined by
following this approach can be computed fairly quickly. It would thus be
possible to assess whether the current trajectory of the robot or the external
wrenches applied on the moving platform are leading to a dangerous con-
traction of the WFW or if the current pose of the end-effector is getting near
to the WFW boundaries.

4.2 Test of the DGP solving procedure

The results of the tests made with the two versions of the code developed
for finding the complete solution of the DGP of cable robots are presented
here. These teste are representative of the mean performances of the code.
In all examples, lengths are expressed in meters and angles in radians. Each
solution reported in the tables at the end of chapter is given as

• a configuration number;

• the number m of taut cables;

• 3 parameters xO′ , yO′ and zO′ describing the position of the origin O′
of the mobile frame in Oxyz;

• 3 angles ϕx , ϕy and ϕz , representing the orientation of the platform

by three consecutive rotations along the x, y and z fixed axes;

• n values of cable tensions;

• a symbol that denotes whether the solution is stable (>) or not (<).

All tests were performed on a PC with an Intel R Xeon W3520 2.67 GHz
CPU with 4 cores.
4.2. TEST OF THE DGP SOLVING PROCEDURE 85

4.2.1 Test on a CDPR with 3 cables in tension

The first test (Table 4.3) is taken from Berti et al. [2013]. The aim is to
compare the performances of the current code with respect to the original
implementation made at the beginning of the Ph.D. program.
The algorithm presented in [Berti et al. 2013], with the same parame-
terization described in Section 2.3, finds all equilibrium configurations with
3 cables in tension in 202 seconds, but it does not verify whether solutions
with slack cables exist and it does not assess stability. The current version
of the code analyzes all possible equilibrium configurations with m cables in
tension, m ≤ 3, in roughly 25 seconds, including stability analysis. With the
beta version, computation time is 22 seconds. All computation times refer
to single-core executions. The same problem with the same parameteriza-
tion was solved with HOM4PS 2,0, a software package which implements the
polyhedral homotopy continuation method for solving polynomial systems of
equations, in 29 seconds [Lee et al. 2008]. It must be noted that the solu-
tions obtained by HOM4PS also include unfeasible and complex ones, and
sub-DGPs for CDPRs with less than 3 cables in tension are not considered.

4.2.2 Test on the Cogiro prototype geometry

For the second test (Table 4.4), the geometry of the giant cable suspended
prototype Cogiro developed by the LIRMM in Montpellier was used. This
robot has 8 cables in crane configuration and it belongs to the over con-
strained family. However, it is possible to show that, under the assumption
that cables are perfectly stiff, the maximum number of cables that can be
under tension at the same time is 6 [Merlet 2012b]. The code to solve the
cable robot DGP was developed on this assumption so it analyzes all the

6
X 8!
NDGP = = 246 (4.18)
k! (8 − k)!
k=1

sub-DGPs. Table 4.4 reports all solutions found by running the code for
all the 246 sets of taut cables and the number of solutions retrieved is 46,
i.e. 6 with 6 cables in tension, 16 with 5 cables in tension, 15 with 4 cables
in tension, 4 with 3 cables in tension and 5 with 2 cables in tension. The
computation time is 1343 seconds by a single-core execution, and 348 seconds
if the code is distributed over 4 CPUs. The beta version of the code shown
better performances: 1219 seconds for single core execution and 314 for the
parallel implementation.
86 CHAPTER 4. EXPERIMENTS AND RESULTS

4.2.3 Test on the Marionet-VR prototype

The third test (Table 4.5) refers to the large-scale robot MARIONET-VR
[Merlet 2010]. This robot uses linear actuators instead of winches to change
the cable lengths, and it is employed in a virtual-reality environment as a
motion provider and haptic device. The moving platform was designed for
demonstration purposes and its anchor points do not belong to the same
plane. Table 4.5 reports all solutions found by running the code for all
possible sets of taut cables, i.e. for 63 sub-problems. The number of solutions
retrieved is 17, i.e. 3 with 6 cables in tension, 1 with 5 cables in tension,
7 with 4 cables in tension, 4 with 3 cables in tension and 2 with 2 cables
in tension. The computation time is 702 seconds by a single-core execution,
and 178 seconds if the code is distributed over 4 CPUs. Performances of the
beta version of the code are practically the same. Feasibility was checked on
the prototype shown in Fig. 4.6.

(a) (b)

Figure 4.6 – Validation of Test 3 on the large scale robot MARIONET-VR:

(a) configuration No. 1 with 6 cables in tension; (b) configuration No. 3
with 6 cables in tension.

4.2.4 Test on the teaching prototype

This fourth test (Table 4.4) shows the case of a 6-6 CDPR having all cable
anchor points on the moving platform lying on the same plane. As shown
in Section 2.3, a specific parameterization can be used in this case. The
algorithm finds 2 solutions with 6 cables in tension, 11 solutions with 5 taut
cables, 12 with 4 and 9 with 3 in 187 seconds by a single-core execution, and
in 49 seconds if the code is distributed over 4 CPUs. The beta version of the
code finds all the solution in almost the same amount of time, 179 seconds.
4.2. TEST OF THE DGP SOLVING PROCEDURE 87

Five out of 34 solutions are stable and they are reported in Table 4.4, for the
sake of conciseness. Their feasibility was checked on the teaching prototype
(Fig. 4.7).

(a) (b)

(c) (d)

Figure 4.7 – From Table 4.6: (a) configuration No. 1, (b) Configuration No.
3, (c) configuration No. 4, (d) configuration No. 5.

4.2.5 Discussion of results

As previously stated, the tests reported above are representative of the av-
erage performance of the problem-solving code. The first test shows that
the code described in this thesis with the refinements introduced with re-
spect to the first version presented in [Berti et al. 2013] has 4 times lower
computation times. Moreover, it computes all the sub-DGPs and assesses
stability. The third and the fourth tests present the complete analysis of
the DGP of a 6-6 CDPR. For the latter test, thanks to planar disposition
of the anchor points on the platform, the code takes advantage of simpler
parameterizations for the sub-problems with 4, 5 and 6 cables, as explained
88 CHAPTER 4. EXPERIMENTS AND RESULTS

in Section 2.3, with a strong effect on computation times. The second and
the third tests analyzes the DGP of robots with spatial disposition of the
anchor points on the respective moving platforms. For this kind of geometry,
the code needs higher computation times. However, performances are still
particularly interesting. For the second example, without considering the 36
subproblems with 1 and 2 active cables, that are processed in less than 2
seconds, the code solves 56 sub-DGPs with 3 cables in tension, 70 with 4
cables in tension, 56 with 5 cables in tension and 28 with 6 cables in tension
in 1343 with the standard version of the code (1219 with the beta version),
with a mean computation time of nearly 7 (6) seconds for each problem.
For the third example, without considering the 21 subproblems with 1 and
2 active cables (they are generally processed in less than 1 second), the code
solves 20 sub-DGP with 3 cables in tension, 15 with 4 cables in tension, 6
with 5 cables in tension and 1 with 6 cables in tension in 702 seconds, with
a mean computation time of roughly 17 seconds for each problem.
It is worth mentioning that the obtained computation times are not nec-
essarily the lowest ones. It can happen that the procedure that selects the
best cable numbering for each DGP fails to find the best combination. In-
deed, the ranking criterion expressed in (3.16) is an empirical relation coming
from the data gathered during tests and on the results obtained in [Merlet
2004], which deserve further investigation.
In tests 2, 3 and 4, multiple feasible solutions are obtained for a given
set of cable lengths. In the third example, the robot (Fig. 4.6a) can easily
switch from configuration No. 1 to configuration No. 3 (Fig. 4.6b) (passing
through an unstable configuration with 4 taut cables) and vice-versa, while
configuration No. 2 cannot be directly reached without detaching cables.
Configuration No. 3 with 6 cables in tension is almost coincident with con-
figuration No. 4, with 5 taut cables, as can be noted by the small value
of tension in the fourth cable. Accordingly, in this part of the workspace,
the robot may easily act as an underconstrained one in an uncontrolled way.
In the fourth test, if the platform is in configuration No. 1 (Fig. 4.7a), a
potential external disturbance can cause the robot to fall in one of the 3
solutions with 5 cables in tension (Figs. 4.7b, 4.7c and 4.7d show configura-
tions No. 3, 4 and 5), and without additional external sensors it is difficult
to determine in which solution branch the robot is operating. Conversely,
configuration No. 2, with 6 cables in tension, is not directly achievable with
a finite movement of the platform without detaching and reattaching cables.
The current version of the code can determine the stability of equilibrium
only after the computation of all statically-admissible solutions. The results
obtained during the experimental campaign show that stable solutions are
4.2. TEST OF THE DGP SOLVING PROCEDURE 89

usually a small subset of statically-admissible poses. Therefore, the imple-

mentation of a stability filter during the computation would allow the domain
regions where unstable configurations may happen to be discarded, thus al-
lowing the direct computation of stable solutions. A few attempts have been
made in order to develop such a filter, but the overestimation introduced
by interval-analysis computation severely compromises its effectiveness. On
the other hand, the analysis of unstable configurations may be of practical
interest. Indeed, an unstable solution may provide a bridge towards stable
configurations. The possibility of configuration changes without crossing sin-
gularities is indeed specific to CDPRs. In this situation, the set of physical
constraints represented by the interference between the platform and taut
cables may be reduced, thus disclosing new paths for other operating modes.
 90
DATA
a1 = [0, 0, 0] a2 = [10, 0, 0] a3 = [0, 12, 0]
b′1 = [0.817, 0, 0] b′2 = [−0.408, 0.707, 0] b′3 = [−0.408, −0.707, 0]
g′ = [0, 0, −0.577] (ρ1 , ρ2 , ρ3 ) = [7.5, 10, 9.5] Q = 1
Conf. m xO ′ yO ′ zO ′ ϕx ϕy ϕz τ1 τ2 τ3 Stability
1 3 2.745 3.979 5.506 3.007 0.340 0.109 0.526 0.511 0.581 >
2 3 1.700 3.687 5.809 0.339 −1.036 −2.596 0.676 0.251 0.486 <

CHAPTER 4. EXPERIMENTS AND RESULTS

3 3 3.020 4.757 3.879 −0.038 0.027 0.776 0.590 0.783 0.956 <
4 3 1.846 4.074 5.322 2.146 −0.708 2.423 0.684 0.305 0.614 <
5 3 2.138 4.287 6.030 −0.482 −0.360 −2.211 0.546 0.325 0.550 <
6 3 3.499 5.369 4.709 −2.908 −0.174 −2.659 0.289 0.787 0.912 <
− 2 - - - - - - - - - -
− 1 - - - - - - - - - -

Table 4.3 – Solution set for a CDPR suspended by 3 cables.

 4.2. TEST OF THE DGP SOLVING PROCEDURE
DATA
a1 = [−7.175, −5.244, 5.462] a2 = [−7.316, −5.103, 5.472] a3 = [−7.303, 5.236, 5.476] a4 = [−7.161, 5.373, 5.485]
a5 = [7.182, 5.348, 5.488] a6 = [7.323, 5.206, 5.499] a7 = [7.302, −5.133, 5.489] a8 = [7.161, −5.270, 5.497]
b1 = [0.503, −0.493, 0.000] b′2 = [−0.510, 0.351, 0.998] b′3 = [−0.503, −0.270, 0.000] b′4 = [0.496, 0.356, 1.000]
′

b′5 = [−0.503, 0.493, 0.000] b′6 = [0.500, −0.340, 0.999] b′7 = [0.502, 0.275, 0.000] b′8 = [−0.505, −0.346, 0.998]
g′ = [0, 0, 0.5] (ρ1 , ρ2 , ρ3 , ρ4 , ρ5 , ρ6 , ρ7 , ρ8 ) = [10.482, 9.839, 10.160, 9.968, 10.310, 9.422, 9.663, 9.656] Q = 1
Conf. m xO ′ yO ′ zO ′ ϕx ϕy ϕz τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8 Stability
1 6 −0.465 −0.309 1.988 2.177 −0.505 1.499 0 0.655 0 0.680 0.309 0.311 0.632 0.136 >
2 6 −0.182 −0.481 1.631 −0.027 −0.718 −3.121 0 0.874 0.396 0.071 0.445 0.279 0.683 0 >
3 6 0.519 1.329 2.460 1.048 1.339 −0.765 0.454 0.279 0 0.761 0 0.879 0.340 0.305 >
4 6 0.635 0.381 1.697 0.396 0.646 −2.999 0.389 0.269 0.710 0 0 0.818 0.471 0.091 >
5 6 0.921 0.236 1.989 −2.193 0.243 1.421 0.358 0.219 0.617 0.154 0 0.628 0 0.724 >
6 6 0.416 0.564 1.627 2.761 −1.363 −1.210 0.670 0.085 0.027 0.589 0 0.741 0.697 0 >
7 5 0.245 −0.521 1.418 −2.147 1.260 −0.514 0 0 0.570 0.599 0 0.012 0.634 0.543 >
8 5 0.255 −0.039 0.828 0.033 −0.099 0.055 0 0 0.622 0.546 0 0.011 0.626 0.551 >
9 5 −0.836 −0.182 2.462 −1.369 −0.222 −1.386 0 0.236 0 1.459 0 0.147 0.591 0.895 <
10 5 0.339 1.246 2.387 0.688 1.067 −0.883 0 0.205 0 1.422 0 0.381 0.446 0.916 <
11 5 −0.405 −0.664 1.902 1.131 −0.659 2.580 0 1.101 0.328 0 0.728 0.227 0.543 0 <
12 5 0.250 −1.140 2.361 −1.079 −1.201 −0.403 0 0.198 0.607 0.770 0 0.101 0 1.434 <
13 5 −0.111 −0.304 1.747 0.124 −0.892 2.799 0 0.711 0.365 0.295 0.570 0 0.855 0 <
14 5 0.469 1.295 2.572 1.868 1.330 0.053 0.651 0 0 0.773 0 0.826 0.543 0.080 <
15 5 0.577 0.242 1.789 0.462 0.814 −3.137 0.399 0 0.963 0 0 0.585 0.448 0.384 <
16 5 0.588 0.084 1.779 −1.430 0.391 1.450 0.408 0 1.112 0 0 0.436 0.106 0.956 <
17 5 0.712 −0.492 1.990 −1.970 0.959 0.693 0.033 0 1.485 0.079 0 0.034 0 1.579 <
18 5 0.109 0.045 1.346 0.829 −0.476 0.568 0.682 0 0.274 0.306 0 0.660 0.678 0 <

91
19 5 1.198 0.503 2.442 1.086 0.461 −1.454 0.617 0.720 0 0.308 0 1.370 0 0.352 <
 92
20 5 0.574 1.289 2.541 1.799 1.509 −0.065 0.717 0.121 0 0.604 0 0.993 0.448 0 <
21 5 0.976 0.674 2.135 −1.322 0.548 2.532 0.798 0.233 0.497 0 0 1.176 0.385 0 <
22 5 0.378 −0.289 1.218 −0.981 0.610 0.341 0.435 0.696 0.082 0 1.138 0 0.110 0 <
23 4 0.188 1.161 2.654 1.703 1.021 0.109 0 0 0 1.554 0 0.177 0.705 0.790 <
24 4 −0.831 −0.141 2.560 −1.555 −0.276 −1.367 0 0.111 0 1.558 0 0 0.769 0.820 <
25 4 −0.269 −0.504 2.419 −0.622 −0.445 −1.262 0 0.476 0 1.597 0 0.490 0 1.611 <
26 4 0.219 0.092 2.744 −3.098 −0.006 −1.393 0 0.624 0 0.652 0 0.655 0 0.667 <
27 4 0.748 0.696 2.428 0.628 0.447 −1.260 0 0.438 0 1.601 0 0.549 0 1.582 <
28 4 0.230 0.094 2.685 3.137 −0.016 1.380 0 0.621 0 0.641 0 0.648 0 0.647 <
29 4 −0.163 −0.612 2.814 1.784 0.414 2.682 0 0.460 1.442 0 0 0.258 1.526 0 <

CHAPTER 4. EXPERIMENTS AND RESULTS

30 4 0.751 0.638 2.900 −1.810 −0.389 2.693 0 0.232 1.556 0 0 0.524 1.485 0 <
31 4 0.275 −0.003 1.747 0.069 0.006 −3.058 0 0.721 0.766 0 0 0.733 0.788 0 <
32 4 −0.256 −0.460 1.875 0.862 −0.728 2.460 0 1.030 0.361 0 0.891 0 0.612 0 <
33 4 0.318 −1.088 2.441 −1.510 −1.140 0.042 0 0.099 0.737 0.720 0 0 0 1.509 <
34 4 0.962 0.456 3.039 1.717 0.820 −0.948 1.540 0 0 0.025 0 1.549 0 0.065 <
35 4 0.509 1.290 2.573 1.965 1.400 0.129 0.718 0 0 0.702 0 0.891 0.541 0 <
36 4 0.795 0.425 2.084 −1.002 0.601 2.378 0.954 0 0.598 0 0 1.096 0.437 0 <
37 4 1.301 0.319 2.574 1.365 0.301 −1.528 0.868 0.686 0 0 0 1.557 0 0.092 <
38 3 −0.414 0.404 2.333 2.198 −0.937 0.499 0 0 0 1.717 0 0 1.519 0.198 <
39 3 0.644 −0.429 1.929 −1.728 0.999 0.879 0 0 1.631 0 0 0.015 0 1.637 <
40 3 0.072 0.031 1.854 0.882 −0.799 1.584 0 0 0.302 1.318 0 0 1.634 0 <
41 3 0.964 0.458 3.036 1.711 0.836 −0.959 1.561 0 0 0 0 1.571 0 0.037 <
42 2 0.123 −0.048 4.169 −3.134 −0.012 −0.027 0 0 0 2.110 0 0 0 2.115 <
43 2 −0.250 0.309 2.232 1.667 −1.008 0.934 0 0 0 1.822 0 0 1.826 0 <
44 2 0.215 −0.098 2.751 −0.007 0.005 2.028 0 0 1.808 0 0 0 1.817 0 <
45 2 0.176 0.171 3.362 0.006 0.002 −1.928 0 4.273 0 0 0 4.277 0 0 <
46 2 0.169 0.170 5.359 −3.139 0.004 −3.128 0 4.273 0 0 0 4.277 0 0 <

Table 4.4 – Complete solution set for the the 8 cable robot Cogiro. Platform anchor points are not coplanar.
 4.2. TEST OF THE DGP SOLVING PROCEDURE
DATA
a1 = [1.886, 0.558, 2.590] a2 = [1.921, −2.876, 2.750] a3 = [−1.498, 0.809, 2.745]
a4 = [1.567, 0.895, 2.600] a5 = [−1.581, −3.015, 2.740] a6 = [−1.692, 0.372, 2.580]
b′1 = [0.63, 0.59, 0.585] b′2 = [0.63, 0.585, 0.015] b′3 = [−0.03, 0.05, 0.585]
b′4 = [0.63, 0.05, 0.585] b′5 = [−0.03, 0.585, 0.015] b′6 = [−0.03, 0.59, 0.585]
g′ = [0.3, 0.3, 0.3] (ρ1 , ρ2 , ρ3 , ρ4 , ρ5 , ρ6 ) = [2.755, 3.519, 2.849, 2.837, 3.489, 2.609] Q = 1
Conf. m xO ′ yO ′ zO ′ ϕx ϕy ϕz τ1 τ2 τ3 τ4 τ5 τ6 Stability
1 6 −0.270 0.235 0.778 2.554 0.124 0.080 0.398 0.226 0.248 0.078 0.244 0.268 >
2 6 0.253 −0.520 0.338 0.960 −0.105 −3.077 0.262 0.291 0.293 0.278 0.314 0.283 >
3 6 −0.278 −1.470 0.549 −0.670 0.014 −0.043 0.374 0.271 0.156 0.004 0.376 0.220 >
4 5 −0.279 −1.470 0.549 −0.669 0.016 −0.046 0.381 0.267 0.161 0 0.380 0.213 >
5 4 −0.502 −0.452 0.351 0.817 −0.913 −2.873 0 0.300 0 0.492 0.243 0.679 <
6 4 −0.289 −0.592 0.553 0.940 0.069 0.008 0 0.395 0.496 0.464 0.432 0 <
7 4 0.530 −0.505 2.234 −1.986 −0.324 −3.011 1.403 0 0.353 0 1.332 0.134 <
8 4 −0.217 0.025 1.237 −2.420 0.068 0.058 0.545 0.082 0 0 0.177 0.506 <
9 4 0.334 −0.380 2.325 −1.695 0.067 3.134 0.464 1.200 0 0 0.031 1.352 <
10 4 0.151 −0.579 2.411 −2.009 0.333 3.075 0.054 1.337 0 0.395 0 1.409 <
11 4 0.879 −0.455 0.751 0.862 0.757 3.036 0.658 0.229 0.524 0 0.339 0 <
12 3 0.132 −0.597 2.425 −2.065 0.376 3.057 0 1.355 0 0.471 0 1.419 <
13 3 0.386 −0.364 2.294 −1.703 −0.087 −3.059 1.393 0 0 0 1.256 0.480 <
14 3 0.552 −0.525 2.211 −2.131 −0.431 −2.963 1.421 0 0.546 0 1.374 0 <
15 3 0.332 −0.383 2.328 −1.700 0.073 3.132 0.442 1.229 0 0 0 1.373 <
16 2 −0.218 −1.192 1.499 1.220 0.379 1.539 0 1.569 1.612 0 0 0 <
17 2 0.342 −0.676 1.244 1.194 −0.403 −1.458 0 0 0 1.691 1.681 0 <

Table 4.5 – Solution set for the CDPR suspended by 6 cables MARIONET-VR: cable anchor points on the platform

93
are not coplanar.
 94
DATA
a1 = [0.238, 0.063, 0.785] a2 = [0.736, 0.551, 0.785] a3 = [0.738, 0.551, 0.781]
a4 = [0.062, 0.729, 0.785] a5 = [0.064, 0.729, 0.789] a6 = [0.238, 0.062, 0.785]
b′1 = [0.369, −0.002, 0.025] b′2 = [0.369, 0.002, 0.025] b′3 = [−0.017, 0.033, 0.025]
b′4 = [−0.020, 0.031, 0.025] b′5 = [−0.020, −0.031, 0.025] b′6 = [−0.017, −0.033, 0.025]
g′ = [0, 0, 0.019] (ρ1 , ρ2 , ρ3 , ρ4 , ρ5 , ρ6 ) = [0.799, 0.790, 0.796, 0.787, 0.798, 0.791] Q = 1
Conf. m xO ′ yO ′ zO ′ ϕx ϕy ϕz τ1 τ2 τ3 τ4 τ5 τ6 Stability

CHAPTER 4. EXPERIMENTS AND RESULTS

1 6 0.346 0.453 0.066 −0.003 0.005 −0.518 0.153 0.228 0.156 0.225 0.163 0.219 >
2 6 0.348 0.453 0.088 −0.019 0.004 2.081 0.231 0.165 0.233 0.162 0.239 0.155 >
3 5 0.363 0.454 0.122 −2.420 0.406 −1.341 0.202 0.333 0.105 0.362 0 0.151 >
4 5 0.341 0.437 0.122 2.419 0.345 0.449 0.100 0.362 0 0.159 0.204 0.327 >
5 5 0.335 0.464 0.122 −3.124 −0.823 2.636 0 0.149 0.210 0.326 0.116 0.352 >
6 5 0.366 0.401 0.098 −0.384 0.623 1.265 0.445 0.026 0.382 0.005 0.334 0 <
7 5 0.341 0.539 0.118 −1.401 −0.157 3.104 0.066 0.237 0.180 0.483 0 0.206 <
8 5 0.356 0.436 0.130 3.119 0.448 −0.087 0.023 0.336 0.045 0.373 0 0.392 <
9 5 0.294 0.463 0.099 0.710 −0.028 1.369 0.384 0.002 0.333 0 0.459 0.013 <
10 5 0.424 0.416 0.117 1.019 1.163 −2.600 0.173 0.480 0 0.216 0.072 0.232 <
11 5 0.332 0.456 0.131 2.809 −0.256 2.131 0.022 0.381 0 0.409 0.019 0.338 <
12 5 0.356 0.470 0.131 −2.731 −0.202 −2.250 0 0.388 0.037 0.327 0.058 0.360 <
13 5 0.277 0.404 0.116 1.217 −0.877 2.383 0 0.223 0.067 0.252 0.165 0.469 <
14 4 0.349 0.467 0.117 −1.979 0.416 −1.395 0.331 0.232 0.196 0.387 0 0 <
15 4 0.347 0.463 0.137 −1.899 0.370 1.567 0.384 0.134 0.234 0.429 0 0 <
16 4 0.414 0.396 0.121 0.399 1.451 2.627 0.424 0.461 0 0.142 0.134 0 <
17 4 0.370 0.440 0.138 3.034 0.894 2.441 0.397 0.439 0 0.160 0.173 0 <
 4.2. TEST OF THE DGP SOLVING PROCEDURE
18 4 0.361 0.443 0.117 2.019 0.578 0.036 0.201 0.386 0 0 0.341 0.215 <
19 4 0.354 0.449 0.138 1.951 0.612 2.940 0.234 0.426 0 0 0.401 0.120 <
20 4 0.363 0.539 0.122 −1.466 −0.471 2.666 0.127 0 0.433 0.466 0 0.135 <
21 4 0.347 0.480 0.138 −2.262 −0.341 0.231 0.167 0 0.409 0.442 0 0.152 <
22 4 0.324 0.440 0.139 2.339 −0.467 −1.519 0 0.159 0.172 0 0.414 0.426 <
23 4 0.264 0.424 0.122 1.437 −0.567 1.751 0 0.142 0.133 0 0.437 0.448 <
24 4 0.339 0.448 0.137 2.837 −1.247 −0.602 0 0 0.396 0.141 0.228 0.417 <
25 4 0.330 0.445 0.117 2.891 −1.187 2.648 0 0 0.336 0.223 0.210 0.376 <
26 3 0.367 0.476 0.142 −1.991 0.768 2.158 0.364 0.413 0 0.406 0 0 <
27 3 0.349 0.440 0.139 2.006 0.809 2.651 0.376 0.417 0 0 0.387 0 <
28 3 0.359 0.462 0.138 −1.864 0.152 1.207 0.371 0 0.390 0.420 0 0 <
29 3 0.311 0.462 0.096 0.472 −0.036 1.353 0.392 0 0.358 0 0.445 0 <
30 3 0.378 0.429 0.106 0.355 0.320 −2.882 0 0.435 0 0.354 0 0.412 <
31 3 0.295 0.444 0.108 −0.051 −0.653 −2.915 0 0.335 0 0.442 0 0.421 <
32 3 0.355 0.425 0.142 1.877 0.215 −2.738 0 0.404 0 0 0.381 0.398 <
33 3 0.319 0.459 0.142 −2.344 −1.141 −1.026 0 0 0.378 0.412 0 0.394 <
34 3 0.334 0.459 0.139 2.297 −1.121 −0.474 0 0 0.381 0 0.395 0.406 <

Table 4.6 – Solution set for a CDPR suspended by 6 cables: all cable anchor points on the platform are coplanar.

95
Chapter 5

Conclusions

This thesis presents an approach to deal with displacement analysis prob-

lems of cable-driven parallel robots based on interval-analysis methods. In
particular, the developed methods aim to solve the direct geometrico-static
problem in a comprehensive and guaranteed way.
The first chapter gives an overview on the cable-driven parallel robots
state of the art and describes the main advantages offered by that this par-
ticular architecture of parallel robots. Indeed, they can provide very high
velocities to the end-effector and may have very large workspaces thanks to
the possibility of coiling cables on actuated winches. On the other hand,
cable flexibility also causes the main issues on their kinematic analysis. In
particular, determining the end-effector pose being given the cable lengths is
a problem that must be carefully dealt with. Since cables cannot withstand
compressive loads, the current robot platform pose may have only a subset
m over the n cables that are taut. Moreover, when the number of taut cables
m < 6, not all the platform degrees of freedom can be restrained and the
robot end-effector still preserves some freedoms. This implies that its pose is
ultimately determined by static equilibrium. Therefore the direct kinematic
problem of cable robots is strictly related to their static analysis and it is
called a direct geomtrico-static problem (DGP).
These aspects motivate the development of the codes and procedures
presented in this thesis. The main tool used for the implementation of the
described algorithm was interval analysis. Interval analysis is a numerical
technique that comprehends methods to deal with a large variety of problems
and to obtain results guaranteed against elimination and round-off errors.
But in the case of cable robot DGPs, interval analysis offers other important
advantages. For a better comprehension of the subsequent chapters, the

97
98 CHAPTER 5. CONCLUSIONS

second part of the first chapter describes some basics elements of this theory,
and lists some of its possible applications in robotics.
In the second chapter, the main definitions used through all the manuscript
are introduced and the assumptions made in order to model cable-robots and
develop the procedures are stated and discussed. Then the geometrico-static
model for CDPRs having m = 1, . . . , 6 cables in tension is presented along
with the different systems of equations used to solve the related DGPs. The
second chapter reports also a brief overview of the different parameteriza-
tions used during the development of the code but that turned out to be
ineffective.
In the third chapter, the code structure was described in detail together
with its main procedures. Then particular attention was devoted to some
refinements that are fundamental for the reliability of the code. The last part
of the third chapter describes a new version, currently still in a development
phase.
The code was tested on real prototypes, and during the Ph.D. program,
in order to simplify testing operations, a small teaching prototype was built.
Thanks to this prototype it was possible to develop a version of the DGP
solving routine that can be used in real-time together with an online al-
gorithm to compute the wrench feasible workspaces boundaries for a given
orientation and height of the moving platform. Then, in the second part of
the fourth chapter, numerical results of the most relevant tests conducted
on the offline algorithm for finding the complete set of DGP solutions were
presented and discussed.

5.1 Future perspectives

The first version of the code presented in the third chapter can be considered
both a good framework and a starting point for further developments. Its
performances are good and it was proven to be reliable, but it can be further
improved in several ways.
The first one consists in completing and refining the beta version. Indeed,
in order to measure its real effectiveness, the complete revision of all the
procedures implemented up to now has to continue in order to avoid all the
possible multiple computations of the same functions, code blocks or boxes.
The second family of possible improvements consists in adding further
modules to implement more realistic cable models. A first version of such a
code was presented in [Merlet & Alexandre-dit Sandretto 2015] that offers
several hints for further enhancements of the methods described in this thesis.
5.1. FUTURE PERSPECTIVES 99

Then, more attention will be devoted to the development of real-time pro-

cedures. First of all, an interval-based algorithm for the inverse geometrico-
static problem is currently under study, combined with a trajectory planning
module that could take into account paths with different numbers of taut ca-
bles. Second, cable interferences must also be considered [Blanchet & Merlet
2014]. Indeed, cable-cable, cable-platform and cable-environment collision
analysis must necessarily be integrated into the trajectory planning and it
can also significantly improve the DGP solving procedures. Third, up to
now a real-time compliant procedure that computes all the sub-DGPs is not
currently available but it remains a long-term objective. In order to achieve
that result, the following issues must be addressed first:

• determine the smallest search domain of the platform pose variables

after a single cycle-time, including the unknowns on cable tensions
when there are less than 6 taut cables;

• investigate if the Kantorovitch theorem and the inflation procedure

could be further specialized for cable robot kinematic problems.

Moreover, the study of how uncertainties on robot parameters affects

robot control will be dealt with. Finally, as explained at the end of the
fourth chapter, implementing the adopted stability check in filter procedures
is not manageable. However, especially for real-time control, this is a very
important issue so other strategies will be investigated and tested.
Bibliography

G. Abbasnejad & M. Carricato (2012). ‘Real solutions of the di-

rect geometrico-static problem of under-constrained cable-driven parallel
robots with 3 cables: a numerical investigation’. Meccanica 47(7):1761–
1773.

G. Abbasnejad & M. Carricato (2015). ‘Direct Geometrico-Static Problem

of Underconstrained Cable-Driven Parallel Robots with n Cables’. IEEE
Transactions on Robotics, DOI:10.1109/TRO.2015.2393173 31(2).

Airbus (2010). Simulation | Airbus, a leading aircraft manufacturer.

J. Albus, et al. (1993). ‘The Nist robocrane’. Journal of Robotic Systems

10(5):709–724.

J. Alexandre Dit Sandretto (2013). Certified calibration of parallel cable-

driven robots. Theses, Université Nice Sophia Antipolis.

J. Angeles (2002). Fundamentals of robotic mechanical systems, vol. 2.

Springer.

S. Behzadipour & A. Khajepour (2006). ‘Stiffness of cable-based parallel

manipulators with application to stability analysis’. Journal of mechanical
design 128(1):303–310.

A. Berti, et al. (2013). ‘Solving the Direct Geometrico-Static Problem of

3-3 Cable-Driven Parallel Robots by Interval Analysis: Preliminary Re-
sults’. In T. Bruckmann & A. Pott (eds.), Cable-Driven Parallel Robots,
vol. 12 of Mechanisms and Machine Science, pp. 251–268. Springer Berlin
Heidelberg.

L. Blanchet & J.-P. Merlet (2014). ‘Interference detection for cable-driven

parallel robots (CDPRs)’. In Advanced Intelligent Mechatronics (AIM),
2014 IEEE/ASME International Conference, pp. 1413–1418.

101
102 BIBLIOGRAPHY

P. Bosscher, et al. (2007). ‘Cable-suspended robotic contour crafting system’.

Automation in Construction 17(1):45–55.

R. Bostelman, et al. (1994). ‘Applications of the NIST RoboCrane’. In

Proceedings of the 5th International Symposium on Robotics and Manu-
facturing, pp. 14–18.

R. Bostelman, et al. (1999). ‘Delivery of an Advanced Double-Hull Ship

Welding System using RoboCrane.’. In IIA/SOCO. Citeseer.

R. Bostelman, et al. (2002). ‘The flying carpet: a tool to improve ship

repair efficiency’. In American Society of Naval Engineers Symposium,
Bremerton, WA, Sept, pp. 10–12.

S. Bouchard, et al. (2010). ‘On the ability of a cable-driven robot to gen-

erate a prescribed set of wrenches’. Journal of Mechanisms and Robotics
2(1):011010.

T. Bruckmann, et al. (2008a). ‘Wire Robots Part I: Kinematics, Analysis

& Design’. In J.-H. Ryu (ed.), Parallel Manipulators, New Developments.
I-Tech Education and Publishing.

T. Bruckmann, et al. (2008b). ‘Wire robots part II-dynamics, control &

application’. Parallel Manipulators-New Developments pp. 133–152.

T. Bruckmann, et al. (2008c). ‘Continuous workspace analysis, synthesis and

optimization of wire robots’. In ASME 2008 International Design Engi-
neering Technical Conferences and Computers and Information in Engi-
neering Conference, pp. 59–66. American Society of Mechanical Engineers.

T. Bruckmann, et al. (2006). ‘Calculating force distributions for redundantly

actuated tendon-based Stewart platforms’. In J. Lennarčič & B. Roth
(eds.), Advances in Robot Kinematics, pp. 403–412. Springer Netherlands.

T. Bruckmann, et al. (2013). ‘An energy-efficient wire-based storage and

retrieval system’. In Advanced Intelligent Mechatronics (AIM), 2013
IEEE/ASME International Conference on, pp. 631–636.

M. Carricato (2013). ‘Direct geometrico-static problem of underconstrained

cable-driven parallel robots with three cables’. Journal of Mechanisms and
Robotics 5(3):031008.

M. Carricato & J.-P. Merlet (2010). ‘Geometrico-Static Analysis of Under-

Constrained Cable-Driven Parallel Robots’. In J. Lenarčič & M. M.
BIBLIOGRAPHY 103

Stanisic (eds.), Advances in Robot Kinematics: Motion in Man and Ma-

chine, pp. 309–319. Springer Netherlands.
M. Carricato & J.-P. Merlet (2011). ‘Direct geometrico-static problem
of under-constrained cable-driven parallel robots with three cables’. In
IEEE International Conference on Robotics and Automation, ICRA 2011,
Shanghai, China, 9-13 May 2011, pp. 3011–3017.
M. Carricato & J.-P. Merlet (2013). ‘Stability Analysis of Underconstrained
Cable-Driven Parallel Robots’. IEEE Transactions on Robotics 29(1):288–
296.
R. Clavel (1988). ‘A fast robot with parallel geometry’. In Proc. Int. Sym-
posium on Industrial Robots, pp. 91–100.
J.-F. Collard & P. Cardou (2013). ‘Computing the lowest equilibrium pose of
a cable-suspended rigid body’. Optimization and Engineering 14(3):457–
476.
T. Dallej, et al. (2011). ‘Towards vision-based control of cable-driven par-
allel robots’. In Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ
International Conference on, pp. 2855–2860. IEEE.
D. Daney, et al. (2006). ‘Interval method for calibration of parallel robots:
Vision-based experiments’. Mechanism and Machine Theory 41(8):929–
944.
D. Daney, et al. (2004). ‘Interval methods for certification of the kinematic
calibration of parallel robots’. In Robotics and Automation, 2004. Pro-
ceedings. ICRA’04. 2004 IEEE International Conference on, vol. 2, pp.
1913–1918. IEEE.
R. Dekker, et al. (2006). ‘Design and Testing of an Ultra-high-speed Cable
Robot’. Int. J. Robot. Autom. 21(1):25–34.
N. Douglas (1998). Numerical computing disasters.
S. Ehardt (2015). Bowden cable.
A. Fattah & S. K. Agrawal (2005). ‘On the design of cable-suspended planar
parallel robots’. Journal of mechanical design 127(5):1021–1028.
J. Fink, et al. (2011). ‘Planning and control for cooperative manipulation and
transportation with aerial robots’. The International Journal of Robotics
Research 30(3):324–334.
104 BIBLIOGRAPHY

P. Gallina, et al. (2001). ‘3-dof wire driven planar haptic interface’. Journal
of Intelligent and Robotic Systems 32(1):23–36.

B. Gao, et al. (2012). ‘Combined inverse kinematic and static analysis and
optimal design of a cable-driven mechanism with a spring spine’. Advanced
Robotics 26(8-9):923–946.

V. Gough (1956). ‘Contribution to discussion of papers on research in auto-

mobile stability, control and tyre performance’. In Proc. Auto Div. Inst.
Mech. Eng, vol. 171, pp. 392–394.

M. Gouttefarde, et al. (2011). ‘Interval-analysis-based determination of the

wrench-feasible workspace of parallel cable-driven robots’. IEEE Transac-
tions on Robotics 27(1):1–13.

M. Gouttefarde, et al. (2006). ‘Determination of the wrench-closure

workspace of 6-DOF parallel cable-driven mechanisms’. In J. Lennarčič
& B. Roth (eds.), Advances in Robot Kinematics, pp. 315–322. Springer
Netherlands.

M. Gouttefarde, et al. (2014). ‘Kinetostatic Analysis of Cable-Driven Parallel

Robots with Consideration of Sagging and Pulleys’. In J. Lenarčič &
O. Khatib (eds.), Advances in Robot Kinematics, pp. 213–221. Springer
International Publishing.

M. Grebenstein, et al. (2012). ‘The hand of the DLR Hand Arm System:
Designed for interaction’. The International Journal of Robotics Research
31(13):1531–1555.

E. Hansen & G. W. Walster (2003). Global optimization using interval anal-

ysis: revised and expanded. CRC Press.

T. Heyden & C. Woernle (2006). ‘Dynamics and flatness-based control of

a kinematically undetermined cable suspension manipulator’. Multibody
System Dynamics 16(2):155–177.

M. Hiller, et al. (2005). ‘Design, analysis and realization of tendon-based

parallel manipulators’. Mechanism and Machine Theory 40(4):429–445.

L. Jaulin (2001). Applied interval analysis: with examples in parameter and

state estimation, robust control and robotics. Springer.

Q. Jiang & V. Kumar (2010a). ‘The direct kinematics of objects suspended

from cables’. In ASME 2010 International Design Engineering Technical
BIBLIOGRAPHY 105

Conferences and Computers and Information in Engineering Conference,

pp. 193–202. American Society of Mechanical Engineers.

Q. Jiang & V. Kumar (2010b). ‘The inverse kinematics of 3-D towing’. In

Advances in Robot Kinematics: Motion in Man and Machine, pp. 321–328.
Springer.

S. Kawamura, et al. (2000). ‘High-speed manipulation by using parallel

wire-driven robots’. Robotica 18(01):13–21.

C. Kossowski & L. Notash (2002). ‘CAT4 (Cable Actuated Truss–4 Degrees

of Freedom): A novel 4 DOF cable actuated parallel manipulator’. Journal
of Robotic Systems 19(12):605–615.

K. Kozak, et al. (2006). ‘Static analysis of cable-driven manipulators with

non-negligible cable mass’. IEEE Transactions on Robotics 22(3):425–433.

J. Lamaury & M. Gouttefarde (2013). ‘Control of a large redundantly actu-

ated cable-suspended parallel robot’. In Robotics and Automation (ICRA),
2013 IEEE International Conference on, pp. 4659–4664. IEEE.

J. Lamaury, et al. (2012). ‘Design and Control of a Redundant Suspended

Cable-Driven Parallel Robots’. In J. Lenarcic & M. Husty (eds.), Latest
Advances in Robot Kinematics, pp. 237–244. Springer Netherlands.

D. E. Lee (1992). ‘Automated all-weather cargo transfer system’. US Patent

5,154,561.

T. Lee, et al. (2008). ‘HOM4PS-2.0: a software package for solving polyno-

mial systems by the polyhedral homotopy continuation method’. Comput-
ing 83(2-3):109–133.

J.-P. Merlet (2004). ‘Solving the forward kinematics of a Gough-type parallel

manipulator with interval analysis’. The International Journal of robotics
research 23(3):221–235.

J.-P. Merlet (2006). Parallel robots, vol. 128. Springer.

J.-P. Merlet (2007). ‘ALIAS-C++’. Available on line at: http://www-

sop.inria.fr/coprin/logiciels/ALIAS/ALIAS-C++/ALIAS-C++.html.

J.-P. Merlet (2008). ‘Kinematics of the wire-driven parallel robot MARI-

ONET using linear actuators’. In Robotics and Automation, 2008. ICRA
2008. IEEE International Conference on, pp. 3857–3862.
106 BIBLIOGRAPHY

J.-P. Merlet (2009). ‘Interval analysis for certified numerical solution of

problems in robotics’. International Journal of Applied Mathematics and
Computer Science 19(3):399–412.

J.-P. Merlet (2010). ‘MARIONET, A Family of Modular Wire-Driven Par-

allel Robots’. In J. Lenarčič & M. M. Stanisic (eds.), Advances in Robot
Kinematics: Motion in Man and Machine, pp. 53–61. Springer Nether-
lands.

J.-P. Merlet (2012a). ‘The kinematics of the redundant N–1 wire driven par-
allel robot’. In (ICRA), 2012 IEEE International Conference on Robotics
and Automation, St. Paul, MN, USA, 14-18 May 2012, pp. 2313–2318.

J.-P. Merlet (2012b). ‘The kinematics of the redundant n-1 wire driven paral-
lel robot’. In IEEE International Conference on Robotics and Automation
(ICRA), 2012, pp. 2313–2318.

J.-P. Merlet & J. Alexandre-dit Sandretto (2015). ‘The Forward Kinematics

of Cable-Driven Parallel Robots with Sagging Cables’. In Cable-Driven
Parallel Robots, pp. 3–15. Springer.

J.-P. Merlet & D. Daney (2010). ‘A portable, modular parallel wire crane for
rescue operations’. In (ICRA), 2010 IEEE International Conference on
Robotics and Automation, Anchorage, Alaska, 3-8 May 2010., pp. 2834–
2839.

A. Ming & T. Higuchi (1994). ‘Study on multiple degree-of-freedom posi-

tioning mechanism using wires. II: Development of a planar completely
restrained positioning mechanism’. International Journal of the Japan
Society for Precision Engineering 28(3):235–242.

R. E. Moore (1966). Interval analysis. Prentice-Hall Englewood Cliffs.

R. E. Moore (1979). Methods and applications of interval analysis. SIAM.

J. Murayama, et al. (2004). ‘SPIDAR G&G: A two-handed haptic interface

for bimanual VR interaction’. In Proceedings of EuroHaptics, pp. 138–146.
Citeseer.

R. Nan (2006). ‘Five hundred meter aperture spherical radio telescope

(FAST)’. Science in China Series G 49(2):129–148.

A. Neumaier (1990). Interval methods for systems of equations, vol. 37.

Cambridge university press.
BIBLIOGRAPHY 107

A. Neumaier (2001). Introduction to numerical analysis. Cambridge Univer-

sity Press.

NIST (2006). RoboCrane for Aircraft Maintenance.

C. Paul, et al. (2006). ‘Design and control of tensegrity robots for locomo-
tion’. IEEE Transactions on Robotics 22(5):944–957.

R. Penrose (1955). ‘A generalized inverse for matrices’. Mathematical Pro-

ceedings of the Cambridge Philosophical Society 51(3):406–413.

S. Perreault, et al. (2010). ‘Geometric determination of the interference-

free constant-orientation workspace of parallel cable-driven mechanisms’.
Journal of Mechanisms and Robotics 2(3):031016.

A. Pott (2010). ‘An Algorithm for Real-Time Forward Kinematics of Cable-

Driven Parallel Robots’. In J. Lenarcic & M. M. Stanisic (eds.), Advances
in Robot Kinematics: Motion in Man and Machine, pp. 529–538. Springer
Netherlands.

A. Pott, et al. (2010). ‘Large-scale assembly of solar power plants with par-
allel cable robots’. In Robotics (ISR), 2010 41st International Symposium
on and 2010 6th German Conference on Robotics (ROBOTIK), pp. 1–6.

A. Pott, et al. (2013). ‘IPAnema: A family of Cable-Driven Parallel Robots

for Industrial Applications’. In T. Bruckmann & A. Pott (eds.), Cable-
Driven Parallel Robots, vol. 12 of Mechanisms and Machine Science, pp.
119–134. Springer Berlin Heidelberg.

J. Pusey, et al. (2004). ‘Design and workspace analysis of a 6–6 cable-

suspended parallel robot’. Mechanism and machine theory 39(7):761–778.

F. Quadvlieg, et al. (2011). ‘Using dynamic positioning for side-by-side cargo

transfer operations’. In 2nd International conference on ship manoeuvring
in shallow and confined water: ship to ship interaction, Trontheim.

H. Ratschek & J. Rokne (1995). ‘Interval methods’. In Handbook of global

optimization, pp. 751–828. Springer.

N. Riehl, et al. (2009). ‘Effects of non-negligible cable mass on the static

behavior of large workspace cable-driven parallel mechanisms’. In Robotics
and Automation, 2009. ICRA ’09. IEEE International Conference on, pp.
2193–2198.
108 BIBLIOGRAPHY

J. Rohn (1996). ‘Enclosing solutions of overdetermined systems of linear

interval equations’. Reliable Computing 2(2):167–171.

G. Rosati, et al. (2007). ‘Design, implementation and clinical tests of a

wire-based robot for neurorehabilitation’. IEEE Transactions on Neural
Systems and Rehabilitation Engineering 15(4):560–569.

A. Ruiz, et al. (2015). ‘ARACHNIS: Analysis of Robots Actuated by Cables

with Handy and Neat Interface Software’. In A. Pott & T. Bruckmann
(eds.), Cable-Driven Parallel Robots, vol. 32 of Mechanisms and Machine
Science, pp. 293–305. Springer International Publishing.

V. Schmidt, et al. (2015). ‘Investigating the Effect of Cable Force on Winch

Winding Accuracy for Cable-Driven Parallel Robots’. In P. Flores & F. Vi-
adero (eds.), New Trends in Mechanism and Machine Science, vol. 24 of
Mechanisms and Machine Science, pp. 315–323. Springer International
Publishing.

SkyCam (2015). SkyCam.

D. Surdilovic, et al. (2007). ‘STRING-MAN: Wire-robot technology for

safe, flexible and human-friendly gait rehabilitation’. In ICORR 2007.
IEEE 10th International Conference on Rehabilitation Robotics, Noord-
wijk, Netherlands, 13-15 June 2007, pp. 446–453.

S. Tadokoro, et al. (2002). ‘A motion base with 6-DOF by parallel cable drive
architecture’. IEEE/ASME Transactions on Mechatronics 7(2):115–123.

R. A. Tapia (1971). ‘The Kantorovich theorem for Newton’s method’. Amer-

ican Mathematical Monthly pp. 389–392.

B. Wilcox (2012). ‘ATHLETE: A limbed vehicle for solar system explo-

ration’. In Aerospace Conference, 2012 IEEE, pp. 1–9.

R. L. Williams II (1998). ‘Cable-suspended haptic interface’. International

Journal of Virtual Reality 3(3):13–21.

W. Wu & S. Rao (2007). ‘Uncertainty analysis and allocation of joint toler-

ances in robot manipulators based on interval analysis’. Reliability Engi-
neering & System Safety 92(1):54–64.

M. Yamamoto, et al. (2004). ‘Trajectory control of incompletely restrained

parallel-wire-suspended mechanism based on inverse dynamics’. IEEE
Transactions on Robotics 20(5):840–850.
BIBLIOGRAPHY 109

W. Yang Ho, et al. (2015). ‘Haptic Interaction with a Cable-Driven Parallel

Robot Using Admittance Control’. In A. Pott & T. Bruckmann (eds.),
Cable-Driven Parallel Robots, vol. 32 of Mechanisms and Machine Science,
pp. 201–212. Springer International Publishing.
Acknowledgments

First of all, I would like to thank my supervisors Marco Carricato and Jean-
Pierre Merlet, for their support and guidance throughout all these years.
Their advice on both research as well as on my career have been priceless.
I would also like to express my deepest gratitude to Tobias Bruckmann,
Andreas Pott and Stéphane Caro, who accepted to be the referees and re-
viewed my manuscript, giving me precious comments and suggestions.
I would also like to thank professor Vincenzo Parenti-Castelli that gave
me the pleasure to be president of the jury, and dr. David Daney for being
member of the examining committee.
A special thank to my friends and colleagues Claudio, Margherita, Fab-
rizio, Federica, Irene, Michele, Nicola, Antonio, Daniele, Glauco, Julien,
Laurent, Rémy, Karim, Ting.
Finally, words cannot express how grateful I am to my girlfriend Giulia,
to my family and to all my friends who always encouraged me and provided
unending support during these years.

111