Amehri Walid (1)

Workspace Estimation and Design Optimization of Soft
Robots
Walid Amehri
To cite this version:

Walid Amehri. Workspace Estimation and Design Optimization of Soft Robots. Computer science.
Centrale Lille Institut, 2022. English. �NNT : 2022CLIL0005�. �tel-04028186v2�
HAL Id: tel-04028186

https://theses.hal.science/tel-04028186v2
Submitted on 14 Mar 2023
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
CENTRALE LILLE
THESE
Présentée en vue
d’obtenir le grade de
DOCTEUR
En
Spécialité : Automatique, Génie Informatique,
Traitement du Signal et des Images
Par
Walid AMEHRI
DOCTORAT DELIVRE PAR CENTRALE LILLE
Titre de la thèse :
Estimation de l’Espace de Travail et l’Optimisation de la Conception

des Robots Déformables.
Workspace Estimation and Design Optimization of Soft Robots.
Soutenue le 28 Février 2022 devant le jury d’examen :
Président Philippe Fraisse Professeur, Université de Montpellier

Rapporteur Sébastien Briot Chargé de Recherche CNRS, HDR, Centrale Nantes
Rapporteur Guillaume Laurent Maître de Conférences, HDR, ENSMM
Examinateur Philippe Fraisse Professeur, Université de Montpellier
Examinateur Jessica Burgner-Kahrs Associate Professor, Université de Toronto
Examinateur Pierre Renaud Professeur, INSA Strasbourg
Directeur de thèse Gang Zheng Chargé de Recherche INRIA, HDR, INRIA Lille
Co-Directeur de thèse Alexandre Kruszewski Professeur, Centrale Lille
Thèse préparée dans le Centre de Recherche en Informatique, Signal et Automatique de Lille,

CRIStAL, CNRS UMR 9189
Ecole Doctorale MADIS-631
Membre du Groupe des Écoles Centrale

Centrale Lille
Cité Scientifique – CS20048 – 59651 Villeneuve d’Ascq Cedex – France
Tél. + 33 (0)3 20 33 53 53 – http://centralelille.fr
1
2
Contents
1 Introduction 11
1.1 General Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Contributions of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2 Modeling and Forward Methods 27

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Piece-wise Constant Strain (PCS) . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Finite Element Method (FEM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Workspace Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 Forward Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3 Optimization-Based Approach for Workspace Estimation 45

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Optimization Approach for the PCS case . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Optimization Approach for the FEM case . . . . . . . . . . . . . . . . . . . . . . 59
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4 Interval Analysis Approach for Workspace Estimation 67

4.1 Interval Analysis Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Forward-Backward Approach Introduction . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Forward-Backward Approach for the PCS case . . . . . . . . . . . . . . . . . . . 70
4.4 Forward-Backward Approach for the FEM case . . . . . . . . . . . . . . . . . . . 81
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5 Continuation Approach for Workspace Estimation: PCS case 89

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 Implementation of the Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3 Validation and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6 Design Optimization: PCS case 115

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2 Implementation of the Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.3 Validation and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
3
7 Conclusions and Perspectives 131
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4
Acknowledgments
First of all, I would like to express my thanks to my two supervisors, Gang Zheng and Alexandre
Kruszewski for presenting me with the opportunity to work on such an interesting subject, and
also for their guidance, help, and sharing their experience with me. I am grateful.
Thanks to the Region Hauts-de-France, the Project Inventor (I-SITE ULNE, le programme
d’Investissements d’Avenir, Métropole Européenne de Lille), and the Project ROBOCOP [ANR
PRCE 19 CE19] for the financial supporting of this thesis.
Many thanks to all members of DEFROST team, and especially Ke Wu for the help with
the platform. Thanks to Shijie Li for his help with the validation part.
Many thanks to the reviewers of my manuscript and jury members.
I would like also to express my gratitude to my family and friends for their care and support.
5
6
Nomenclature
Model
Piece-wise Constant Strain Finite Element Method
Symbol
DoFs Degrees of Freedom
z Exclusion
b Kronecker product
.
Derivative with respect to time
’ Derivative with respect to space —
p Converts R6 in sep3q —
r Converts R3 in sop3q —
n Total number of DoFs
N Total number of segments/bodies Dimension of the finite element mesh
nτ Actuation force vector (Input) dimension
nu “ 3 End-effector (Output) dimension
T T
“ T ‰ n
q “ q1 ¨ ¨ ¨ qN P R P R3n Displacement vector
Generalized strain vector
T
τ “ rτ1 ¨ ¨ ¨ τnτ s P Rnτ Soft robot actuators force vector.
nˆn
K “KPR Generalized stiffness matrix “ Kpqq P R3nˆ3n Generalized stiffness matrix
nˆnτ
Hpqq PR Generalized actuation matrix P R3nˆnτ Generalized actuation matrix
F pqq P Rn Generalized external force vector P R3n Generalized external force vector
δτ Actuation vector with respect to prior configuration.
Generalized strain vector with respect Displacement vector with respect
δq
to prior configuration to prior configuration
“ ‰T
τ “ τ 1 ¨ ¨ ¨ τ nτ Minimum actuators bound vector
T
τ “ rτ 1 ¨ ¨ ¨ τ nτ s Maximum
“ actuators
‰ bound vector
T “ rτ 1 , τ 1 s ˆ ¨ ¨ ¨ ˆ τ nτ , τ nτ Actuators bounds
u P Rnu End-effector position vector in the inertial frame.
WE P Rnu Workspace of a soft robot
nu ´1
BWE PR Soft robot workspace boundary
WEd Discretized space of the current configuration
WEs Feasible small displacement neighborhood space around the current configuration
B “ rB, Bs Lower B and upper B bounds of a domain B
R P SOp3q Orientation matrix —
X P r0, Ls Ă R Abscissa along the robot arm —
Continued on next page
7
Table 1 – continued from previous page
Model
Symbol ˆ ˙
R u
pXq “ P SEp3q
g 0T 1 —
: Configuration matrix
v, w pXq P R3 Linear and angular velocity —
p, k pXq P Rˆ3 Linear and
˙ angular strain —
w̃ v
pXq “ P sep3q
η̂ 0T 0 —
: Velocity twist matrix
` ˘T
η pXq “ ˆvT , wT ˙P R6 Velocity vector —
k̃ p
pXq “ P sep3q
ξˆ 0T 0 —
: Strain twist matrix
` ˘T
ξ pXq “ ˜kT , pT P R6¸Strain vector —
k, w
r r 03ˆ3
pXq “ P R6ˆ6 : Adjoint
adξ,η q, r
r v r k, w
r —
representation
˜ of the strain
¸
k, w
r q, r
v
P R6ˆ6 : coAdjoint
r r
pXq “
ad˚ξ,η 03ˆ3 r k, w
r —
representation
ˆ of the˙strain
R 03ˆ3
pXq “ P R6ˆ6 :
Adg uR
r R —
Adjointˆrepresentation ˙ of g
R uR
P R6ˆ6 :
r
pXq “
Ad˚g 03ˆ3 R —
coAdjoint representation of g
şX
pXq “ 0 Adg : Tangent operator
Tg —
of the exponential map
Exponential map of X ξp P SEp3q
eX ξ —
p
with θ2 “ kT k
T
“ rα1 ¨ ¨ ¨ αnτ s P Rnτ
α —
: Actuators length parameter
“ ‰T
α “ α1 ¨ ¨ ¨ αnτ —
: Minimum actuators length bounds
T
“ rα1 ¨ ¨ ¨ αnτ s
α —
: Maximum actuators “ length‰ bounds
“ rα1 , α1 s ˆ ¨ ¨ ¨ ˆ αnτ , αnτ
A —
: Actuators length bounds
T
“ rL1 ¨ ¨ ¨ LN s P RN
L —
: Segments length parameter
T
“ rL1 ¨ ¨ ¨ LN s
L —
: Minimum segments length bounds
Continued on next page
8
Table 1 – continued from previous page
Model
Symbol
“ ‰T
L “ L1 ¨ ¨ ¨ LN —
: Maximum
“ ‰ segments
“ length ‰ bounds
“ L1 , L1 ˆ ¨ ¨ ¨ ˆ LN , LN
L —
: Segments length bounds
Table 1: Acronyms and Glossary for the adopted mathematical modeling methods.
9
10
Chapter 1
Introduction
1.1 General Introduction

1.1.1 Robotics: Definitions and Classes
In an etymological sense, and according to [137], the word robotics originated from the word
robot, which is derived from the Czech noun robota meaning labor. Its original appearance
was in the RUR play of the first fictional automatons called robots published in 1920 [137].
Scientifically, robotics is a multidisciplinary field that aims at the design, modeling, control, and
applications of robots for the purpose of increasing productivity and assisting human beings [32].
In accordance with [44, 52, 121], robots can be classified based on their fabrication material
and their degrees of freedom. Consistently, there are two main classes of robots (as shown by
Fig. 1.1): rigid and hyper-redundant (it is also important to state that some rigid robots can
also be hyper-redundant, e.g., serpentine robots).
Figure 1.1: Classification of robots [44, 52, 121].
1.1.2 Rigid Robotics and the Rise of a New Robot Class

Rigid robots are a class of robots made of rigid materials with a Young’s modulus greater than
109 P a [72]
The first appearance of a rigid robot can be traced back to the third century B.C. and earlier
by the mechanical engineer Yan Shi that created the first human-shaped figure of his mechanical
11
handiwork [85]. The first mechanical humanoid robot Mechanical Knight was designed by
Leonardo da Vinci (see A in Fig. 1.2). Mechanical robots have then continued to emerge until
1738 where a mechanical duck called Digesting Duck that was able to eat and flap its wings was
created by the inventor and artist Jacques de Vaucanson (see B in Fig. 1.2). Following the birth
of electricity, robotics have evolved first via the Teleautomaton device that was created in 1898
by Nicola Tesla, which was a radio-controlled vessel (see C in Fig. 1.2), followed by the first
electronic humanoid robot Elektro in 1930 and created by the Westinghouse Electric Corporation
(see D in Fig. 1.2). The first appearance of rigid robots imitating biological behavior is traced
back to 1948 by William Grey Walter [112] (see E in Fig. 1.2). Next, the first commercial
robot Unimate was created in 1956 by the Unimation company [126], the same company that
introduced the first installed industrial robot in 1961 (see F in Fig. 1.2). After that, the first
full-scale humanoid robot WABOT-1 [67] was created by the Waseda University (see G in
Fig. 1.2). Then, the famous six electro-mechanically driven axis robot Famulus [115] (see H in
Fig. 1.2) was developed by the KUKA Robot Group in 1973. Next, in 1975, a programmable
universal manipulation arm PUMA was created by Victor Scheinman (see I in Fig. 1.2). In
1983, the creation of the first parallel programming language used for the robot’s control was
performed by Stevo Bozinovski and Mihail Sestakov.
Figure 1.2: Rigid Robots Evolution: A- Mechanical Knight. B- Digesting Duck. C- Teleautoma-
ton. D- Elektro. E- Elsie and Elmer [112]. F- Unimate [126]. G- WABOT-1 [67]. H- Famulus
[115]. I- PUMA. J- Delta and ABB Flex-Pitcher robots [41, 43, 51].
Since then, researchers made a huge progress on rigid robotics, and owing to their stiff
material and high-torque joints which make them exceptionally stable and accurate, rigid robots
have become widely and successfully deployed especially in the industrial environment [120] and
12
have been proved to increase productivity (e.g., Delta and ABB Flex-Pitcher robots in J of Fig.
1.2) [41, 43, 51].
However, as new and more complex applications arise, traditional rigid robots have shown
many drawbacks due to their high stiffness, lack of compliance in conventional actuation
mechanisms, and limited degrees of freedom [42]. Moreover, rigid robots have been considered
less practical when performing operations in dynamic environments [30] and not safe when
interacting in fragile environments and also with humans [32, 131].
In the last decades, as a consequence of these drawbacks, many researchers have been
motivated to seek a novel design of robots that can be flexible and safe in order to cope with
new applications, leading to the rise of soft robotics.
1.1.3 Soft Robotics
Hyper-redundant robots consist of a large number of DoFs [121]. Continuum robots are capable
of continuous deformations, but not all continuum robots are necessarily soft [121]. Soft robots
are a sub-class of continuum robots that are hyper-redundant robots [82].
In a general sense, soft robotics is a branch of robotics that consists of the design, modeling,
and control of robots that are fabricated from flexible and compliant materials equivalent to
those found in living organisms such as silicone rubber and other materials that can be readily
found in nature [65, 70, 111, 121].
The history of soft robotics is provided in detail in [59, 70, 88, 111, 125]. In the following, we
present only the main points of the history of soft robotics. The first appearance of soft robots
is traced back to the late 1970s where robot grippers based on granular materials were first
published in [28, 90, 113]. After, a continuously-deforming body using elastomers was controlled
by a pneumatic actuation in 1984 [128]. Next, robot grippers using electrorheological fluid were
proposed in 1989 [64]. In 1991, [118] developed a flexible micro-actuator and later extended its
applications to soft robotic grasping and also mobile mechanisms [117, 119] (see A in Fig. 1.3).
Ever since the field of soft robotics continued to make outstanding progress and has gained
impressive attention at the beginning of the 21st century. Consequently, a huge portion of new
soft material has been synthesized and made commercially accessible [13, 59, 65, 68, 70, 88, 111,
121], diverse fabrication techniques for soft robots have been developed and made available
[13, 59, 65, 68, 70, 88, 111, 121], a growing number of works showing the use of soft robotics has
been published in high-profile scientific journals [59, 65, 68, 70, 111, 121], and researchers generally
agree that soft robots should be used in robotic applications in the future as they are safer and
more adaptive in dynamic environments compared to the conventional rigid robots.
Being made of soft and deformable material, soft robots provide many benefits such as high
adaptability and dexterity, and safe collisions [35], which are very useful for various applications,
especially the manipulation of fragile objects [53, 59, 69, 76] (see B, C, and D in Fig. 1.3),
environment exploration [11, 37, 46, 80, 87, 136] (see E, F, G, and H in Fig. 1.3), and medical
operations [23, 25, 33, 34, 54, 74, 81, 114, 122] (see I, J, K in Fig. 1.3).
13
Figure 1.3: Soft Robots Applications: A- grasping and mobile mechanisms [117,119]. B- octArm
[76]. C- elephant’s trunk [53]. D- octopus arm [69]. E- terrestrial exploration [80]. F- tuna
robotics [136]. G- multi-modal underwater [46]. H- planetary explorations [87]. I- transoral
robotic surgery [74]. J- laryngeal surgery [114]. K- distal lung sampling [81].
1.1.4 Challenges and Motivations

Even though soft robots have made an interesting progress during the last few years, the actual
process of designing a soft robot is still mainly inspired by a range of biological systems [111]
including the structure and behavior of animal species such as the trunk of elephants [53],
and the arms of an octopus [69]. Such a design procedure can be useful for initial tests and
experimentation of the designed soft robot. However, when faced with specific performance
objectives such as path planning and control [35, 129], it is possible that such a soft robot’s
design may not be able to achieve its objectives, especially because its range of reachability (i.e.,
its workspace) may be restricted.
Therefore, it is of benefit to the soft robotics community to propose generic methodologies
in order to evaluate the workspace of soft robots, and also to assist and guide the systematic
design of soft robots for the purpose of optimizing specific performance objectives.
The workspace estimation in soft robotics is still an open subject, and its importance is due
to the numerous benefits that it can provide in order to solve different soft robotics’ scientific
challenges, mainly related to the robot’s mechanical design and controller synthesis. On the
one hand, the result of workspace evaluation can provide information about the accessibility of
the object to be controlled [35, 38, 129], i.e. by identifying if the object’s position belongs to
the soft robot’s workspace, sparing the controller’s design time that can be wasted in trying
14
to control an unreachable object outside a robot’s workspace. On the other hand, workspace
information is crucial as well when planning feasible trajectory to be tracked by soft robots
since the start and finish positions of a feasible trajectory should belong to a robot’s workspace
[129]. Besides, workspace analysis is also a necessary step to assist an effective, purpose-driven,
and optimal design of soft robots [58], improving their stability, and increasing the range of
their manipulability and reachability.
To achieve this aim, the actual thesis adopts two different mathematical models to describe
the deformation of soft robots both with a slender shape and general form. Based on those
adopted mathematical models, several approaches to estimate the workspace of soft robots are
proposed. Finally, an effective method to optimize the design of soft robots for the purpose of
achieving specific performance objectives is presented.
1.2 State of the Art
This work is concerned with the workspace evaluation and design optimization of soft robots
based on their adopted mathematical models. Accordingly, a concise state of the art targeting
the modeling, workspace evaluation, and design optimization of soft robots will be presented.
1.2.1 Mathematical Modeling of Soft Robots
Owing to their natural conformity and compliance, soft robots consists of a large number of
DoFs and are characterized by different mechanical laws [50] compared to that of rigid robots,
which makes their kinematic and dynamic modeling highly nonlinear and more complex.
In the literature, several modeling methods were proposed to describe the deformation of soft
robots, and the most important ones are the Piece-wise Constant Curvature (PCC) [127], the
Finite Element Method (FEM) [39], and the Piece-wise Constant Strain (PCS) based Discrete
Cosserat method [103].
In the following, we will briefly investigate each modeling method.
1.2.1.1 Piece-wise Constant Curvature
The PCC approach [127] was initially established for kinematics studies and later extended to
the dynamics of soft robots. It describes the investigated soft robot via a fixed number of arcs
characterized by the curvature’s radius, the arc’s angle, and the bending plan. The robot is
kinematically described through ns successive transformations Tij (as shown in Fig. 1.4), with
ns being the total number of segments that constitutes the investigated robot.
The PCC has been widely and successfully applied to several continuum robots controlled by
various methods of actuation [36, 63, 109, 127] . However, when the studied soft robot is subject
to external loads, the assumption of a constant curvature does not always hold.
15
Figure 1.4: PCC modeling approach [127].
1.2.1.2 Finite Element Method

In [39], a modeling approach based on FEM was developed to describe the deformation of soft
robots with general geometries (as shown by Fig. 1.5). In this method, the deformable domain
of the structure is discretized into smaller finite elements using a specific meshing geometry,
composed of a finite number of DoFs, which is then used to interpolate the behavior of the
deformable domain by measuring the variation of the associated nodal displacements.
This modeling technique was successfully applied to different geometries of soft robots
controlled by different means of actuation [15, 39, 47, 107]. However, to obtain a good modeling
precision, this approach requires a fine mesh when discretizing the robot’s spatial space, which
in turn increases the number of DoFs and results in high computation complexity.
To solve this problem, model-order reduction (MOR) techniques have been introduced into
the FEM approach [47] in order to reduce the number of DoFs and eventually decrease the
computation complexity, but this improvement slightly affects the modeling precision.
Figure 1.5: FEM modeling approach [39, 47].
1.2.1.3 Piece-wise Constant Strain

The Discrete Cosserat method [97, 98, 102–105] is a Piece-wise Constant Strain (PCS) method
that can be viewed as an extension of the PCC technique.
This approach (see Fig. 1.6) describes the designed soft robot with a finite set of stain
vectors (linear and angular) and provides the benefit of achieving a higher modeling precision
with a lower number of DoFs. Besides, this modeling technique considers the material properties
and geometric nonlinearity, thus taking into account large deformations. However, like the PCC,
this approach can only be applied to slender-shaped soft robots. Moreover, the PCS assumes
that all cross-sections (which are perpendicular to the centroidal line of slender soft robots) are
16
rigid and undeformable. Such an assumption is imposed for the sake of modeling simplicity, but
inevitably it may affect the modeling precision. To solve this problem, a generalization of the
PCS method was recently investigated in [20].
Figure 1.6: PCS modeling approach [103, 104].
1.2.1.4 Modeling Methods Comparison

The PCC modeling approach provides an effective and simple method for the modeling of soft
robots. This method was largely deployed in the community of soft robotics, however, the
adopted assumptions are not always valid which is very important for modeling the dynamics of
soft robots.
Alternatively, the PCS method takes into account both the material properties of the studied
robot and geometric nonlinearity, but with the limitation that it can only be applied to slender-
shaped soft robots. Accordingly, the PCS modeling method is adopted by the present thesis in
order to achieve the mathematical model of soft robots with slender forms.
On the other hand, and although it is computationally expensive, the FEM modeling method
is adopted by this thesis in order to provide the mathematical model of soft robots with general
form.
1.2.2 Workspace Evaluation

By definition, the workspace (also known as the accessible output set) of a robot is an area/volume
that contains the whole of its achievable equilibrium configurations.
1.2.2.1 Rigid Robotics Workspace Methodologies

The subject of workspace determination has been widely investigated by the community of rigid
robotics. In the literature, we can find many contributions to this topic using different methods
to represent the workspace of rigid serial/parallel manipulators.
As to analytical Jacobian-based methods, the method proposed by [2, 3] (see A in Fig. 1.7)
uses an explicit input-output function and consists of analytically determining singularities that
make the sub-Jacobian row-rank deficient in order to map the boundaries of rigid manipulators.
17
This method was indeed successful in determining the singularities for a wide variety of rigid
mechanisms [1]. However, this method has its limitations [2]: it requires that the output is an
explicit function of the input (i.e., x “ Ωpqq), thus making it impossible to be used for soft robots.
Moreover, a higher dimension of inputs increases the complexity of the analytical formulation
of the sub-Jacobian which as a result makes it near impossible to analytically determine all
singularities of such mechanisms.
In [55, 56] a Jacobian-based numerical method was presented for mapping the boundaries of
rigid manipulator’s workspaces using a continuation method based on the theory of bifurcation
(see B in Fig. 1.7). To demonstrate the broadness of the applicability of this method, it was
used to determine the workspace of planar and spatial Stewart platforms [73]. However, while
seeking an initial boundary point, starting from a feasible initial point inside the workspace,
this method may fail to identify enclosure curves (namely voids/holes) inside the workspace due
to the discretization [2]. Moreover, it may also not be possible to distinguish an internal and an
external boundary to the workspace when the ray emanating from an initial configuration passes
through a bifurcation point [2]. These problems were properly handled in [18] where the methods
presented in [55] were extended to treat more complex workspaces. Moreover, continuation
methods were also used in [17] for the computation of tensegrity mechanism workspace, and
a general high-order continuation method [57] with automatic differentiation was applied to a
planar RRR mechanism and a three-dimensional Orthoglide parallel mechanism.
Regarding numerical methods, an optimization approach to determine the boundaries of rigid

manipulator’s workspaces has been presented in [116] (see C in Fig. 1.7). This numerical method
consists of finding a suitable radiating point in the output coordinate space and then determining
the points of intersection of a representative pencil of rays, emanating from the radiating point,
with the boundary of the accessible set. This method is based on an optimization approach
in which the accessibility of the workspace boundary is formulated in terms of a constrained
optimization problem. However, the difficulties of this method lie in defining an appropriate
objective function, non-convexity issues, and the inability to identify interior knowledge of the
workspace.
Also, the wrench-feasible workspace of parallel rigid robots driven by cables has been analyzed
in [48] by applying an interval analysis bisection method. Such a method has been then extended
to analyze the workspace of cable-driven parallel mechanisms in [14, 79] (see D in Fig. 1.7).
The advantage of this approach lies in the fact that it explores all feasible configurations in the
workspace. However, this approach is in a way exhaustive since it consists of bisecting the whole
space in order to find the feasible configurations which represent the workspace.
Finally, a numerical probabilistic method, Monte Carlo, was presented in [4, 27] to find the
2D and 3D workspace of a 3-DOF serial manipulator (see E in Fig. 1.7). This numerical method
yields a robust result since it explores all actuators combinations, and has the advantage that it
does not rely on the inverse problem of the robot. However, this technique is computationally
expensive for high DoFs mechanisms.
18
Figure 1.7: Rigid workspace methods: A- Jacobian-based analytical methods [3]. B- Jacobian-
based numerical methods [55]. C- optimization methods [116]. D- interval analysis methods [48].
E- Monte-Carlo methods [27].
Although the above methods have been successfully implemented in determining the
workspace of several types of rigid robots. However, they cannot be directly extended to
analyze soft robots because of the modeling complexity for soft robots. In fact, the input-output
relation for open-loop rigid chains (also called serial-link rigid robots) can normally be explicitly
defined through its direct kinematic model (u “ f pqq). However, such a relation for soft robots
is generally defined in an implicit way (f pu, τ q “ 0) through a complex static/kinematic model
(which will be explained in Chapter 2) for the purpose of modeling soft material deformations.
On the other hand, closed-loop rigid chains (also called parallel rigid robots) necessitate the use
of numerical schemes in order to solve the direct kinematics of the robot [12, 21, 22, 78], which is
the same problem for soft robots. Hence, the workspace methodologies explored for the different
kinds of rigid mechanisms need to be adapted in order to treat the same problem for soft robots
since they consist of a large number of DoFs due to the compliance and the flexibility of the
used material, and also because the modeling of soft robots is established through different
mechanical laws, i.e., principles of continuum mechanics [50] (which will be explained in Section
2.5.2 for the PCS model, and Section 2.5.3 for the FEM model).
1.2.2.2 Continuum Robotics Workspace Approaches

On the other side, in the community of continuum robotics, the workspace subject has been
rarely investigated. To our best knowledge, three main contributions were provided and all of
them consist of discretizing the inputs (actuators) space in different fashions.
The paper of [24] (see A in Fig. 1.8) presents the workspace characterization for concentric
19
tube continuum robots by extending Monte-Carlo algorithms developed in [27]. Also, in [132]
(see B in Fig. 1.8), the workspace analysis of cable-driven continuum manipulators based on the
static model was investigated using an approach that consists of discretizing the input domain
and finding the set of all reachable equilibrium positions of the manipulator’s end-effector.
The limitations of the above-stated approaches are: firstly, they consist of discretizing the
inputs space, which varies depending on the configuration of the investigated continuum robot;
secondly, the computational complexity (which is equal to nnd τ ) of these approaches can explode
exponentially when increasing the dimension of the actuators nτ or when we want to get higher
precision, i.e., increasing nd ; thirdly, valuable information about the robot’s singularities and
interior/exterior boundaries cannot be identified using these classical approaches since they only
focus on solving the direct model of the investigated continuum robots in order to find the set
of poses that constitutes the workspace.
In [26] (see C in Fig. 1.8), an approximation method was used to determine the workspace
boundaries of continuum robots by sweeping the robot from the neutral configuration to the
fully curved configuration for the purpose of deducing the workspace surface from the robot’s tip
trajectory. The proposed approximation method consists of estimating the workspace boundary
by searching only the points distributed on the surface of the workspace. However, this method
still consists of exploring the inputs space in a different manner by varying one configuration
from its minimum to its maximum value while keeping the other configurations at a fixed
neutral value. This strategy can map the workspace boundaries of the investigated robot and at
the same time reduce the computational time of the pre-proposed scanning method from nnd τ
combinations to 2nτ combinations. Nevertheless, for robots controlled by a higher number of
actuators, the approximation method will still be computationally expensive. Another limitation
of this approach lies in failing to identify some workspace boundaries because singularities result
not only from actuators saturation, but also due to physical limitations [18, 55] such as the
mechanism’s length, self-collision, or actuation instabilities, because this approach doesn’t apply
the idea that the workspace interior/exterior boundaries of the investigated continuum robots
are extracted from the set of their output singularities ignoring the row-rank deficiency of the
Jacobian of the generalized mechanical constraints which doesn’t comply with what was studied
in [3, 18, 55].
1.2.2.3 Soft Robotics Workspace Methods
For a given soft robot’s configuration, till the starting date (1 October 2018) of the present
thesis, no work has yet been investigated about how to comprehensively evaluate its workspace.
1.2.2.4 Workspace Methods Summary
As stated above, different methodologies were investigated for the workspace evaluation of rigid
robots, few for continuum robots, and almost none for soft robots.
The methods investigated for continuum robots are not efficient since they are based on the
discretization of the inputs space, whose dimension depends on the studied configuration of the
robot, but such methods can still be used as a comparative reference since they exhaustively
explore all possible achievable configurations of the investigated robot’s workspace.
On the other hand, it is more appropriate to extend the applicability of the main workspace
approaches developed in the rigid robotics community in order to treat the soft robotics case.
However, this cannot be done in a direct way since soft robots are modeled in a more complex
manner, yielding an implicit input-output relation.
20
Figure 1.8: Continuum workspace methods: A- Monte-Carlo algorithms [24]. B- direct method
[132]. C- approximation method [26].
1.2.3 Design Optimization of Soft Robots
Several design optimization methods were proposed for specific soft robotics applications, which
can be classified into three main categories: heuristic methods, model-based, and hybrid.
1.2.3.1 Heuristic Methods
Genetic algorithms were explored in [31, 58] for the design optimization of soft robots (see A in
Fig. 1.9). These methods are particularly useful for treating highly nonlinear problems since
they don’t require the calculation of costly gradients. However, defining a proper fitness function
is difficult.
Evolutionary algorithms were also used in [16] (see B in Fig. 1.9) where it was proposed
to optimize a small number of size parameters of pneumatic-actuated inflatable manipulators,
and in [110] (see C in Fig. 1.9) for the design optimization of soft pneumatic actuators. In
[10] (see D in Fig. 1.9), the central concept was to develop an assistive manipulator that can
automate the bathing task for elderly citizens where the development of a novel algorithm based
on cooperative multi-agent reinforcement learning that simultaneously optimizes stiffness and
position was proposed.
21
Figure 1.9: Heuristic Methods: A- Automatic design of soft robots [58]. B- Multi-objective
design optimization of a soft pneumatic robot [16]. C- Design optimization of soft pneumatic
actuators [110]. D- Multi-objective optimization for stiffness and position control in a soft robot
arm [10].
1.2.3.2 Model-Based Approaches
As to model-based approaches, in [133] (see A in Fig. 1.10), a density-based topology optimization

method was employed for the design of the exterior shell of a soft pneumatic gripper with a
given shape and position in order to grasp a wide range of objects using the FEM model. Next,
improving the mechanical performance of soft pneumatic robots through parameterized layout
and shape optimization of air chambers was studied in [49] (see B in Fig. 1.10). In [84] (see
C in Fig. 1.10), a model-based optimization for the design of a rotational soft robotic system
driven by double cone dielectric elastomer actuators was investigated. The work of [29] (see D
in Fig. 1.10) studied the design optimization of a soft robotic device for neurorehabilitation by
computing a set of optimal parameters which, in turn, could be used to customize the device
according to the patient’s need. In [40] (see E in Fig. 1.10), the design of a pneumatically
actuated silicone module with three pneumatic chambers is considered and optimized with
the aim of using it in a soft robot arm for robotic surgery applications. In [45] (see F in Fig.
1.10), the design optimization of soft crawling robots (inchworm and quadrupedal) for maximum
bending motion with respect to input pressure was evaluated using FEM. The work of [75] (see
G in Fig. 1.10) developed a methodology for converting caterpillar-like soft-bodied robot from
a lumped-parameter form into a Computer-Aided Design (CAD) model that could be easily
manufactured using the results of coarse dynamic simulations. Also, the subject of optimal
actuator location was investigated in [83], where it is proposed to minimize a linear-quadratic
(LQ) function to find the optimal actuators location. This method was then extended in [130]
(see H in Fig. 1.10) to achieve the optimal actuator location for a medical endoscope controlled
by electro-active polymers.
22
Figure 1.10: Model-based approaches: A- topology optimized design of a multimaterial soft
gripper [133]. B- soft pneumatic robots chamber layout design optimization [49]. C- rotational
soft robotic system design [84]. D- soft robotic device design optimization [29]. E- design
optimization of a pneumatically actuating silicone module [40]. F- design optimization of a soft
crawling robot [45]. G- dynamics-based design of soft robots [75]. H- optimal actuator location
for electro-active polymer actuated endoscope [130].
1.2.3.3 Hybrid Approaches
In [94] (see Fig. 1.11), two frameworks, a model-based optimization, and deep reinforcement
learning were proposed for the design optimization of a multi-chamber pneumatic-driven soft
actuator in order to optimize its mechanical performance. The design goal was to achieve
maximal horizontal motion of the top surface of the actuator with a minimum effect on its
vertical motion.
Figure 1.11: Hybrid approaches [94].
23
1.2.3.4 Conclusion
As stated above, different strategies (heuristic, model-based, and hybrid) were fully investigated
for the design optimization of different soft robots. Those methods have helped achieve the
optimal design of the investigated soft robots which were successfully deployed in real-life
scenarios.
However, despite the significant progress in the conceptual design of soft robots, the mentioned
methods are in a sense specific to a particular type of robot with a predefined concept.
On the other hand, generic methods focusing on the systematic design optimization of soft
robots that are controlled via a generic number of actuators for the purpose of optimizing specific
performance objectives have rarely been addressed in the literature.
This is due to the difficulty and complexity that lies in establishing an appropriate cost function
combining multiple desired design objectives, particularly when these objectives are competing
with each other.
To our best knowledge, no former work has been investigated on the optimization of the
design of soft robots for the purpose of optimizing their workspace attainability and reachability
which is of high importance and benefit to the design optimization of soft robots.
1.3 Contributions of the Thesis
1.3.1 Contributions and Organization of the Manuscript
For soft robotic applications such as design optimization, path planning, and pick-place, it is
necessary and useful to evaluate the workspace of the studied soft robot. In this thesis, we
investigate such a problem from two views, a direct and an inverse one.
On the one hand, the direct problem of the workspace can be formulated by the following
question: given the designed configuration of the investigated soft robot, how should we
efficiently estimate its workspace? (which will be studied in Chapter 3, Chapter 4, and Chapter
5). Estimating the workspace of a given soft robot’s prototype is a crucial step for assessing
the reachability and accessibility of particular locations which relates to the controller synthesis
of the robot. Also, solving this problem helps in determining the robot’s singularities and
impediments of motion which relates to the design of the robot.
On the other hand, the following question formulates the inverse problem: how should we
optimize the design of a soft robot in order to achieve specific performance objectives? (which
will be investigated in Chapter 6). Given specific performance objectives (e.g., reaching specific
locations), it is valuable for many reasons (both economic and scientific) to determine the soft
robot’s optimal design such that its workspace contains those desired locations in a virtual
environment before proceeding to the physical design of the final prototype of the robot.
Accordingly, to answer these two questions, the manuscript was organized into 7 chapters as
follows (See Fig. 1.12):
24
Figure 1.12: Manuscript organization.
Chapter 2 introduces the dynamic model of the adopted PCS and FEM modeling methods.
Also, in this chapter, we establish the workspace definition of soft robots. Finally, we present
a classic approach (named the forward approach) that consists of discretizing the actuators
space in order to provide the workspace of soft robots. This approach was then validated via a
trunk-like soft robot for both the PCS and FEM cases.
In Chapter 3, we propose an optimization-based approach [9, 123] that consists of estimating
the workspace of soft robots by mapping their workspace’s exterior boundaries. This approach
was applied to both the PCS [123] and FEM [9] modeling methods.
Next, an interval analysis-based method [7] (named the forward-backward approach) is
outlined in Chapter 4 in order to achieve the workspace of soft robots by exploring all feasible
configurations. The proposed approach was applied to both the PCS and FEM modeling
methods.
Chapter 5 presents a continuation method [8] based on the theory of bifurcation in order to
map the interior and exterior boundaries of soft robots. This approach was only applied to the
PCS modeling method.
After, an optimization approach based on the adopted PCS model was established in Chapter
6 in order to optimize the design of soft robots for the purpose of achieving specific performance
objectives.
Finally, Chapter 7 provides the conclusions of the present thesis and presents the perspective
of future works, especially, the orientation aspect of the workspace, the generalization of the
continuation approach, and the possibility of its extension to treat the FEM case, and also the
design optimization of soft robots based on the FEM method.
1.3.2 Publications
The contributions of the present thesis are outlined in the following:
International Conference:
1. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. ”Fem based workspace estimation
for soft robots: A forward-backward interval analysis approach.” In 2020 3rd IEEE
International Conference on Soft Robotics (RoboSoft), pp. 170-175. IEEE, 2020.
Summary: In this paper, the preliminary result of an interval analysis-based approach
in order to estimate the workspace of soft robots using the FEM modeling method was
submitted.
25
Journal Articles:
1. Walid, Amehri, Gang Zheng, Alexandre Kruszewski, and Federico Renda. ”Discrete
Cosserat Method for Soft Manipulators Workspace Estimation: An Optimization-Based
Approach.” Journal of Mechanisms and Robotics 14, no. 1 (2021): 011012.
Summary: In this article, we presented an optimization-based approach in order to estimate
the exterior workspace boundary of slender-shaped soft robots using the PCS modeling
method.
2. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. ”Workspace Boundary Estimation
for Soft Manipulators Using a Continuation Approach.” IEEE Robotics and Automation
Letters 6, no. 4 (2021): 7169-7176.
Summary: In this letter, we provided a continuation approach based on the theory of
bifurcation in order to estimate the interior and exterior boundaries slender-shaped soft
robots using the PCS modeling method.
3. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. ”FEM-based Exterior Workspace
Boundary Estimation for Soft Robots via Optimization.” IEEE Robotics and Automation
Letters (2022).
Summary: In this paper, we extended the applicability of the optimization-based approach
by applying it to the FEM modeling method in order to estimate the exterior workspace
boundary of soft robots.
4. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. Soft Robotics SORO, FEM-based
Reachable Workspace Estimation of Soft Robots using an Interval Analysis approach.
Summary: In this journal, the full version of the proposed interval analysis approach was
presented which contains new results for FEM model with external forces, self-contained
explanation and results on interval analysis, novel uniform spatial grid discretization
strategy, a detailed algorithm with an improved stop condition, and comprehensive
configuration simulation scenarios to emphasize the effectiveness of the proposed interval
analysis based approach.
Submitted Articles:
1. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. ASME JMR, Position-Access
Workspace of Soft Manipulators using an Interval Analysis Method.
Summary: In this article, we extended the applicability of the interval analysis-based
approach by applying it to the PCS modeling method in order to estimate the workspace
of soft robots with slender shape.
2. Amehri, Walid, Gang Zheng, and Alexandre Kruszewski. IJRR, Cosserat-based Optimiza-
tion Design for Slender Soft Manipulators.
Summary: In this article, we proposed an optimization approach based on the adopted
PCS model in order to optimize the design of soft robots with slender shape for the purpose
of achieving specific performance objectives.
26
Chapter 2
Modeling and Forward Methods
2.1 Introduction
As stated in Chapter 1, the PCS method provides a good modeling precision with a lower number
of DoFs which can only be applied to slender-shaped soft robots. On the other hand, the FEM
method provides a precise model of soft robots with the general form, but it is computationally
expensive.
Accordingly, the first step of this chapter is to present the mathematical models of the PCS
and FEM methods for the purpose of modeling soft robots with a slender shape and general
form, respectively. Next, we establish the definition of a soft robot’s workspace. Finally, we
propose a forward approach to estimate the workspace of soft robots. This approach was then
validated using a trunk-like soft robot [129] for both the PCS and FEM cases.
In the following, we do not consider the material’s nonlinearity, instead, we consider that
the material is isotropic, and the deformation is entirely elastic.
2.2 Piece-wise Constant Strain (PCS)

The dynamic model based on the Piece-wise Constant Strain (PCS) approach was developed in
[101] for a soft-rigid multibody system. In the following, all the quantities are expressed in the
local (body) coordinate frame if not specified. The superscript 1 represents the partial derivative
with respect to the space variable.
In the following, we illustrate the mathematical model of the PCS model using the example of
a tendon-driven soft robot. The method could be as well applied to other manners of actuation
[102].
2.2.1 Kinematics
The configuration of a soft body i with respect to its predecessor in the chain (as illustrated by
Fig. 2.1) is defined as a curve gi p¨q : X P r0, Li s ÞÑ gi pXq P SEp3q with:
27
Figure 2.1: Schematic of the PCS kinematics.
ˆ ˙
Ri ui
gi pXq “ (2.1a)
0T 1
where Ri is the orientation matrix, ui is the end-effector position, Li is the length, ξî is the
strain matrix, of a body i, and X P r0, Ls is the abscissa along the robot arm.
The continuous models of the position, velocity and acceleration of a soft body can be derived
from the Cosserat rod theory, which gives [19]:
gi1 “ gi ξî (2.2a)

ηi1 “ ξ9i ´ adξi ηi (2.2b)
η9 i1 “ ξ:i ´ adξ9i ηi ´ adξi η9 i (2.2c)
where adp¨q P R6ˆ6 is the adjoint operator of the Lie algebra, and
ˆ ˙
k̃i pi ˘T
ξî pXq “
`
P sep3q ξi pXq “ kTi , pi T P R6 (2.3)
0T 0
defines the strain state with k̃i pXq P sop3q, ki pXq P R3 , and pi pXq P R3 respectively the angular
and linear strain (when compared to the reference values k˚i and p˚i );
ˆ ˙
w̃i vi ` ˘T
η̂i pXq “ P sep3q ηi pXq “ wiT , viT P R6 . (2.4)
0T 0
is the cross-section velocity twist with w̃i pXq P sop3q, wi pXq P R3 , and vi pXq P R3 respectively
the angular and linear velocity. To model constrained rod, such as the Kirchhoff-Love case with
angular strain only, the strain field is specified as:
ξi “ Bqi qi ` ξi˚ (2.5)
where Bqi P R6ˆni forms a basis for the allowed motion subspace, qi P Rni contains the values of
the allowed strains and, ξi˚ P R6 is the initial twist modeling the initial shape.
28
Assuming piece-wise constant strains [98], (2.2) can be analytically integrated using the
matrix exponential method, leading to:
gi pXq “eX ξ̂i (2.6a)

ηi pXq “Ad´1
gi ηh ` Ad´1
gi Tgi Bqi q9i (2.6b)
żX
η9 i pXq “Ad´1
gi η9 h ` Ad´1
gi Adgi psq adηi psq dsBqi q9i ` Ad´1
gi Tgi Bqi q
:i (2.6c)
0
where Adgi pXq P R6ˆ6 is the Adjoint operator of SEp3q, and Tgi pXq is the tangent operator of
the exponential map.
Successive applications of the kinematics (2.6) for all the bodies of the system yields the
definition of the geometric Jacobian Ji pq, Xq P R6ˆn and its derivative J9i pq, q,
9 Xq P R6ˆn (n being
T T
“ T T ‰
the total number of DOFs), which links the generalized strain vector q “ q1 q2 ¨ ¨ ¨ qN P Rn
(N being the total number of bodies) and the velocity twist ηi pXq, for each soft body i, as
shown below:
i
ÿ
ηi pXq “ Ad´1
gh ¨¨¨gi Tgh Bqh q9h
h“0
(2.7a)
i
ÿ
i
“ Sh q9h “ Ji pq, Xqq9
h“0
i
ÿ żX
i
η9 i pXq “ Sh q:h ` Ad´1
gh ¨¨¨gi Adgh psq adηh psq dsBqh q9h
h“0 0
(2.7b)
i
ÿ
i i
“ Sh q:h ` S9 h q9h “ Ji pq, Xq:
q ` J9i pq, q,
9 Xqq9
h“0
η̂i pXq “gi´1 g9 i (2.7c)

th i 6ˆnp¨q
where the block elements of the i Jacobian Sp¨q P R and its derivative S9 p¨q P R6ˆnp¨q
i
have been defined in [103].
2.2.2 Continuous Dynamics

Once a Jacobian is found, the generalized dynamics of the system can be obtained by projecting
the free dynamics of each soft body by virtue of the D’Alembert’s principle. The free dynamic
equation of a soft body, with its boundary conditions, is given by [97]:
` ˘1 ` ˘
Mi η9 i ` ad˚
ηi Mi ηi “ Fii ´ Fai ` ad˚
ξi Fii ´ Fai ` F̄ei
` ˘ ` ˘ (2.8)
Fii ´ Fai p0q “ ´FJi Fii ´ Fai pLi q “ Ád˚ gij FJj ;
where Mi pXq “ diagpJxi , Jyi , Jzi , Ai , Ai , Ai qρi P R6ˆ6 is the screw inertia matrix of the cross-
section (Jï pXq being the second moment of the area about the axis ¨ and Ai pXq being the
area of the cross-section); F̄ei pXq P R6 is the distributed external load; Fai pXq P R6 is the
internal wrench due to the distributed actuation [99]; Fii pXq P R6 is the internal wrench due to
the elasticity of the soft body; FJp¨q P R6 is the wrench transmitted across joint p¨q and ad˚p¨q
(respectively Ad˚p¨q ) P R6ˆ6 is the co-adjoint (respectively co-Adjoint) map of the Lie algebra
(respectively Lie group).
2.2.3 Elasticity and Actuation Forces

Regarding the internal elastic force, a linear viscoelastic constitutive model is usually chosen:
Fii pXq “ Σi pξi ´ ξ ˚ q ` Υi ξ9i “ Σi Bqi qi ` Υi Bqi q9i (2.9)
29
where
Σi pXq “ diagpGi Jxi , Ei Jyi , Ei Jzi , Ei Ai , Gi Ai , Gi Ai q
Υi pXq “ diagpJxi , 3Jyi , 3Jzi , 3Ai , Ai , Ai qνi P R6ˆ6
are the screw stiffness and viscosity matrix (Ei being the Young’s modulus, Gi the shear modulus
and νi the shear viscosity).
With regards to the actuation load (as depicted in Fig 2.2), by computing the force and
moment exerted by the internal tendon on the mid-line of the rod [99].
Figure 2.2: Schematic of the tendon-driven actuation of one body.
This is given by:

nτ „
d˜i pXqλi pXq
ÿ 
Fa pXq “ τi “ Hτ pq, Xqτ (2.10)
i“1
λi pXq
where di pXq P R3 represents the distance from the mid-line to the internal ith actuator, d˜i
denotes di in sop3q, λi pXq P R3 is the unit vector tangent to the actuator path, and τ P Rnτ
is the vector of magnitude of the actuators force given by the negative of the tendons tension
(nτ being the total number of actuators). While the distance di pXq is fixed by design, the
tangent vector λi pXq depends, in general, on the deformation ξi pXq of the soft robot body [108],
[100]. Going further into details, the unit tangent vector λi pXq can be obtained by spatial
differentiation of the position vector of the actuator, thus:
” ı
1‰ ξî di pXq ` d1i pXq
“ ´1
gi pgi di q 3 3
λi pXq “ 1 “ . (2.11)
kgi´1 pgi di q k kξî di pXq ` d1 pXqk
i
where di pXq is expressed in homogeneous coordinates, r¨s3 extracts the first three rows of a
homogeneous vector and k¨k takes the Euclidean norm.
Remark 1. Considering di as a small parameter, a useful approximation can be obtained, which

allows for a unit tangent vector independent of the deformation with a slight loss of accuracy
T
[97]. In this case, one can consider λi “ r1 0 0s . As a result, the generalized actuation matrix
Hpqq may be approximated by Hpqq « H.
2.2.4 Discrete Dynamics

By Jacobian projection of the free dynamics (2.8), we obtain the generalized dynamics in its
classical form:
M pqq q: ` pC pq, qq
9 ` Dq q9 ` Kq “ H pqq τ ` F pqq (2.12)
30
where M P Rnˆn is the generalized mass matrix, C P Rnˆn is the generalized Coriolis matrix,
D P Rnˆn is the block-diagonal generalized damping matrix, K P Rnˆn is the block-diagonal
generalized stiffness matrix, H P Rnˆnτ is the generalized actuation matrix, F P Rn is the vector
of generalized position-dependent external force and τ P Rnτ is the vector of applied actuators
force. Going further into details, the coefficient matrices take the form:
N ż Li
ÿ
M pqq “ JiT Mi Ji dX (2.13a)
i“1 0
N ż Li
ÿ ´ ¯
C pq, qq
9 “ JiT ad˚Ji q9 Mi Ji ` Mi J9i dX (2.13b)
i“1 0
˜ ż L1 ż LN ¸
D “ diag BqT1 Υ1 dXBq1 , ¨ ¨ ¨ , BqTN ΥN dXBqN (2.13c)
0 0
˜ ż L1 ż LN ¸
K “ diag BqT1 Σ1 dXBq1 , ¨ ¨ ¨ , BqTN ΣN dXBqN (2.13d)
0 0
»˜
ż L1 ¸T ˜ ż LN ¸T fiT
H pqq “ – BqT1 Hτ dX , ¨ ¨ ¨ , BqTN Hτ dX fl (2.13e)
0 0
N ż Li
ÿ
F pqq “ JiT F̄ei (2.13f)
i“1 0
2.3 Finite Element Method (FEM)

2.3.1 FEM Introduction
The dynamic model of soft robots can be obtained based on Galerkin method of FEM and the
Euler Lagrange equation [95].
FEM consists firstly of spatially discretizing the deformable domain of the structure into
smaller finite elements N through the creation of a corresponding mesh which contains a finite
number of DoFs n to interpolate the behavior of the deformable domain by measuring the
variation of the associated nodal displacements. Then, the Euler-Lagrange method is applied to
deduce the dynamics of soft robots.
In the literature, various geometries are proposed to discretize the spatial domain of a given
structure, such as prisms, beams, hexahedrons and tetrahedrons [96, p. 732]. For the purpose
of generality and convenience for element matrix integration, we select the linear tetrahedron
element (shown by Fig. 2.3) as the appropriate meshing structure in this thesis.
31
Figure 2.3: Linear tetrahedron element.
Take the trunk-like soft robot as an example, Fig. 2.4 shows the corresponding discretized
version in the domain.
Figure 2.4: Discretized mesh model for trunk-like soft robot using the linear tetrahedron element.
Based on the discretized mesh model, the deduction of its dynamics can be realized element
by element. In the following, we apply the modeling approach on a single linear tetrahedron
element to get its motion equation, and then we perform an assembly process of all elements to
obtain the motion equation for the whole structure.
2.3.2 Displacement Model

After having created the mesh, we have to define the primary variables of a single element,
which corresponds to the displacement of the nodes. According to [96, p. 93], the displacement
vector can be defined by the following formula:
δe px, y, z, tq “ Ne qe (2.14)
where the index e denotes the investigated element, Ne is the basis functions matrix, and qe
represents the nodal (1, 2, 3, 4) displacement vector for a single element, with:
» fi
N1 0 0 ¨ ¨ ¨ N4 0 0
Ne “ – 0 N1 0 ¨ ¨ ¨ 0 N4 0 fl (2.15)
0 0 N1 ¨ ¨ ¨ 0 0 N4 e
And
qe “ rq1x q1y q1z q2x q2y q2z q3x q3y q3z q4x q4y q4z sTe (2.16)
The basis functions Ne represents the mathematical model of the displacement over space,
which needs to satisfy some requirements such as the continuity over a specific order and the
Kronecker property where each basis function has a value of one at its own node and zero at the
other nodes in the corresponding element [95, p. 82].
In the literature, we can find different polynomial families to define the basis functions. In
the following, we present the Lagrange family [96] because it is accurate and straightforward. It
has been stated in [71, p. 252] that the order p of the polynomial is defined via the number of
nodes of the element ne by the following equation:
1
ne “ pp ` 1qpp ` 2qpp ` 3q “ 4 (2.17)
6
Using the linear tetrahedron element which contains four nodes, the polynomial order for
the basis functions should be equal to 1. According to this, the basis functions can be written
in the following polynomial form:
“ ‰
Ni px, y, zq “ αp1,iq ` αp2,iq x ` αp3,iq y ` αp4,iq z e (2.18)
32
where αi for 1 ď i ď 4 are the coefficients to be determined. For this, we define Vequ as the
volume of the element in an equilibrium state, then the basis functions of the linear tetrahedron
element can be formulated as:
„ 
1
Ni px, y, zq “ pai ` bi x ` ci y ` di zq (2.19)
6Vequ e
ai bi ci di
with αp1,iq “ , αp2,iq “ , αp3,iq “ , αp4,iq “ , and
6Vequ 6Vequ 6Vequ 6Vequ
1 x1 y1 z1
1 1 x2 y2 z2
Vequ “
6 1 x3 y3 z3
1 x4 y4 z4 e
where pxi , yi , zi q are the coordinates of the node i in the corresponding element. The variables
ai , bi , ci and di in (2.19) are constant coefficients which are computed when the element is in
equilibrium point, with the following form:
xj yj zj 1 yj zj
ai “ p´1qm`1 xk yk zk , bi “ p´1qm 1 yk zk
xl yl zl e 1 yl zl e
xj 1 zj xj yj 1
ci “ p´1qm xk 1 zk , di “ p´1qm xk yk 1
xl 1 zl e xl yl 1 e
where m “ j ` k ` l, and i, j, k, l P t1, 2, 3, 4u.
2.3.3 Strain-Displacement Model

The linear strain-displacement relation is defined by the following form [95]:
ε e “ Be q e (2.20)
where εe is the strain vector and Be is the strain matrix calculated by:
» fi
BN1 BN4
0 0 ¨¨¨ 0 0
— Bx Bx ffi
— ffi
— ffi
—
— 0 BN1 BN4 ffi
0 ¨¨¨ 0 0 ffi
—
— By By ffi
ffi
— ffi
— ffi
— BN1 BN4 ffi
— 0 0 ¨¨¨ 0 0 ffi
— Bz Bz ffi
Be “ — (2.21)
— ffi
ffi
— BN1 BN1 BN4 BN4 ffi
—
— By 0 ¨¨¨ 0 ffi
— Bx By Bx ffi
ffi
— ffi
— ffi
—
— 0 BN1 BN1 BN4 BN4 ffi
¨¨¨ 0 ffi
—
— Bz By Bz By ffi
ffi
— ffi
– fl
BN1 BN1 BN4 BN4
0 ¨¨¨ 0
Bz Bx Bz Bx e
33
By substituting (2.19) into the above equation, we get:
» fi
b1 0 0 ¨¨¨ b4 0 0
—0 c1 0 ¨¨¨ 0 c4 0 ffi
— ffi
1 — —0 0 d1 ¨¨¨ 0 0 d4 ffi
Be “ ffi (2.22)
— c1
6Vequ — b1 0 ¨¨¨ c4 b4 0 ffi
ffi
–0 d1 c1 ¨¨¨ 0 d4 c4 fl
d1 0 b1 ¨¨¨ d4 0 b4 e
2.3.4 Stress-Displacement Model

The relation between the stress and the nodal displacement vector is defined using the constitutive
relation by the following representation [95]:
σe “ CBe qe (2.23)
where σe represents the stress vector.

For an isotropic material with given Young’s Modulus E and Poisson’s ratio ν, the elasticity
matrix is defined by: » fi
c11 c12 c13 0 0 0
—c12 c22 c23 0 0 0 ffi
— ffi
—c13 c23 c33 0 0 0 ffi
C“— —0
ffi (2.24)
— 0 0 c44 0 0 ffi
ffi
–0 0 0 0 c55 0 fl
0 0 0 0 0 c66
where
E p1 ´ νq
c11 “ c22 “ c33 “
p1 ` νq p1 ´ 2νq
Eν
c12 “ c13 “ c23 “
p1 ` νq p1 ´ 2νq
E
c44 “ c55 “ c66 “
2 p1 ` νq
With the deduced displacement relation, strain-displacement relation and stress-displacement
relation, we are ready to apply the Euler-Lagrange method in order to derive the motion equation
of the corresponding element.
2.3.5 Euler-Lagrange Equation

The Euler-Lagrange method is based on energy formulation, and it allows us to find the motion
equation of the corresponding element of the soft robot’s mesh. In general, its form can be
written as follows [95]: „ 
d BL BED BL
` “ (2.25)
dt B q9e B q9e Bqe
where ED is the dissipation energy, and L is the Lagrangian defined by [95]
Lpqe , q9e , tq “ EK ´ EP (2.26)
with EK and EP representing the kinetic energy and the potential energy, respectively.
34
2.3.5.1 Potential Energy
The potential energy is defined by the difference between the strain energy and the potential
work [95]:
EP “ ES ´ WP (2.27)
2.3.5.2 Strain Energy

It is known that the strain energy is the energy stored in the body under loading, and it can be
modeled as follows [95]:
1
ES “ qeT Ke pqe qqe (2.28)
2
where Ke represents the stiffness matrix for the analyzed element.
The integral form of the strain energy has the following form [95]:
ż
1
ES “ εT σe dv (2.29)
2 Ve e
where ε and σ are the strain and the stress defined in (2.20) and (2.23), respectively.
By substituting (2.20) and (2.23) into (2.29), we get:
„ż 
1
ES “ qeT BeT CBe dv qe (2.30)
2 Ve
Then, from (2.28) and (2.30) we can directly get the stiffness matrix formula as follows:
ż
Ke pqe q “ BeT CBe dv (2.31)
Ve
2.3.5.3 Potential Work

The potential work [95] represents the sum of the work done by the actuator forces Feb and the
work done by the external (surface) forces Fes .
“ ‰
WP “ Feb ` Fes qe (2.32)
with:
ż
Feb “ NeT feb dv “ He pqe qτ (2.33a)
Ve
ż
Fes “ NeT fes ds “ Fe pqe q (2.33b)
Se
where feb and fes are the unit body and surface forces, He pqe q is the actuation matrix, τ is the
applied actuators force, and Fe pqe q is the external applied forces, for the corresponding element.
2.3.5.4 Kinetic Energy

The kinetic energy is the energy of the body resulting from its motion, it can be modeled as
follows [95]:
1
EK “ q9e T Me pqe qq9e (2.34)
2
35
where Me represents the mass matrix for the related element.
The integral form of this kinetic energy has the following form [95]:
ż
1 T
EK “ ρe δ9e δ9e dv (2.35)
2 Ve
where δe is the displacement defined in (2.14).

By substituting (2.14) in (2.35), we get:
„ż 
1
EK “ q9e T ρe NeT Ne dv q9e (2.36)
2 Ve
Then, from (2.34) and (2.36) we can directly get the mass matrix formula as follows:
ż
Me pqe q “ ρe NeT Ne dv (2.37)
Ve
where ρe is the volumetric mass density.
2.3.5.5 Dissipation Energy

The dissipation energy is the energy due to the phenomena of damping, it can be modeled as
follows [95]:
1
ED “ q9e T De pqqq9e (2.38)
2
where De represents the damping matrix for the analyzed element.
The integral form of this damping energy has the following form [95]:
ż
1 T
ED “ µe δ9e δ9e dv (2.39)
2 Ve
where µe represents the damping coefficient, and δe is the displacement defined in (2.14).
By substituting (2.14) in (2.39), we get:
„ż 
1
ED “ q9e T µe NeT Ne dv q9e (2.40)
2 Ve
Then, from (2.38) and (2.40) we can directly get the damping matrix formula as follows:
ż
De pqe q “ µe NeT Ne dv (2.41)
Ve
2.3.5.6 Motion Equation

To obtain the dynamics of a single element via FEM, we replace the equations of the potential
energy (2.27) and the kinetic energy (2.34) in the Lagrangian (2.26), and using (2.33), we obtain:
1 T 1
Lpqe , q9e , tq “ q9e Me pqe qq9e ´ qeT Ke pqe qqe ` rHe pqe qτ ` Fe pqe qs qe (2.42)
2 2
Since Ke is a symmetric matrix, thus the right term of (2.25) becomes:
BL
“ ´Ke pqe qqe ` He pqe qτ ` Fe pqe q (2.43)
Bqe
36
The terms on the left of (2.25) depends only on q,
9 since the mass matrix Me and the damping
matrix De are symmetric, therefore we obtain:
„ 
d BL
“ Me pqe qq:e
dt B q9e
(2.44)
BED
“ De pqe qq9e
B q9e
Finally, the motion equation of the analyzed element can be written as follows:
Me pqe qq:e ` De pqe qq9e ` Ke pqqq “ He pqe qτ ` Fe pqe q (2.45)
Similarly, we apply the same procedure for each element of the mesh model, then we can
assemble all elements using the connectivity matrix [96] to obtain the motion equation for the
whole structure:
q ` Dpqqq9 ` Kpqqq “ Hpqqτ ` F pqq
M pqq: (2.46)
where M pqq is the global mass matrix, Kpqq is the global stiffness matrix and Dpqq is the global
damping matrix, and formulated below:
N
ÿ
M pqq “ Me pqq (2.47a)
e“1
N
ÿ
D pqq “ De pqq (2.47b)
e“1
N
ÿ
K pqq “ Ke pqq (2.47c)
e“1
N
ÿ
H pqq “ He pqq (2.47d)
e“1
N
ÿ
F pqq “ Fe pqq (2.47e)
e“1
with Me pqq, De pqq, Ke pqq, He pqq, and Fe pqq are defined respectively in (2.37), (2.41), (2.31),
(2.41), (2.33a), and (2.33b).
2.4 Workspace Definition

The workspace of a soft robot contains all its achievable equilibrium configurations where the
end-effector can reach and stay there. In this thesis, we focus on the position aspect of the
workspace, and we discuss the orientation aspect of the workspace in Chapter 7, Section 7.2.1.
For the purpose of defining the workspace of soft robots, let us note u P Rnu as the position
of the end-effector in the inertial frame, and q P Rn (n being the total number of DoFs) has a
different meaning according to the appropriate modeling approach (it denotes the generalized
strain vector in PCS, and the displacement vector in FEM). Then, we can define its workspace
as follows:
37
Definition 1. Consider a soft robot dynamically modeled by (2.12) for the PCS case or (2.46)
for the FEM case, controlled by bounded actuators τ P T , the workspace WE of the end-effector
u, is a subspace of Rnu , defined below:
WE “ tu “ ϕpqq | Dpq P Rn , τ P T q s.t. Φpτ, qq “ 0u
where ϕ represents the geometric model of the studied soft robot, Φpτ, qq “ Kq ´ Hpqqτ ´ F pqq
for PCS, and Φpτ, qq “ Kpqqq ´ Hpqqτ ´ F pqq for FEM.
Remark 2. Using the above definition, It is important to state that the workspace contains all
equilibrium positions where the soft robot’s end-effector can reach and stay there. Moreover, the
stability of a calculated configuration in the workspace can be asserted by verifying the eigenvalues
of the stiffness matrix [134]. Finally, the elastic limit can be asserted by verifying that the
corresponding constraint Φpτ, qq “ 0 is valid for the corresponding τ P T .
In this thesis, we adopt two strategies to estimate the workspace of soft robots;
• Discretizing the inputs (actuators) space T [4, 24, 26, 27, 132].
• Discretizing the outputs (end-effector) space [3, 48, 55, 78, 116].
The first strategy is manifested through the so-called forward approach [4, 24, 27, 132]
(see Section 2.5), and the second strategy is manifested through three different approaches;
optimization [123] (see Chapter 3), interval analysis (see Chapter 4) [7], and continuation [8]
(see Chapter 5), which will be established in what follows.
2.5 Forward Approach

2.5.1 Forward Approach Introduction
The most basic approach to determine the workspace of soft robots is to iteratively approximate
the input-output map φ : T Ñ WE through the discretization of the actuators bound T with a
prescribed precision nd , and then determine for each actuator value its corresponding end-effector
position, as illustrated by Fig. 2.5.
Figure 2.5: Forward Approach
In the following, we will explain the details of the forward approach for both the PCS and
the FEM modeling methods.
38
2.5.2 Forward Approach for Workspace Estimation: PCS case
According to Definition 1, the workspace of a soft robot contains all equilibrium points. Applying
this property to equation (2.12), the workspace of the studied soft robot is determined by the
following equations:
K q “ Hpqqτ ` F pqq (2.48a)

upqq “ E1 gpqq ET2 (2.48b)
with
gpqq “ eL1 ξ̂1 ¨ ¨ ¨ eLN ξ̂N (2.49a)

ξî “ B
z q qi ; i P 1 ¨ ¨ ¨ N
i (2.49b)
The purpose of E1 and E2 in (2.48b) is to extract the end-effector position u from the configuration
matrix g, where E2 is a constant elementary vector defined as E2 “ r01ˆ3 1s, and E1 is a
constant elementary matrix defined by the following: when nu “ 2, it may take the form of
E1 “ rI2 , 02 s, or E1 “ rr1, 0, 0, 0sT , r0, 0, 1, 0sT sT (depending on the actuators location); when
nu “ 3, E1 “ rI3 03ˆ1 s (with I and 0 are respectively the identity and zero matrices).
Using the PCS framework, the problem of estimating WE is equivalent to, for all admissible
τ P T , measure all the possible configurations of its end-effector u defined in (2.48b). However,
this is not trivial, mainly due to the reason that the mapping φ : T Ñ WE is implicit and
nonlinear, thus no analytic φ can be found such that u “ φpτ q.
In practice, for the purpose of estimating WE from the bounded actuators τ P T , we might
use numerical methods to approximate the mapping φ. In the following, we may use the gradient
of the generalized strain vector q with respect to the actuators force vector τ , denoted as ∇τ pqq,
which will be analyzed hereafter.
By definition, the gradient ∇τ pqq is calculated by:
” Bq ıT
∇τ pqq “
Bτ
Thus, by calculating the partial derivative of (2.48a) with respect to the actuators vector τ , we
have:
Bq BrHpqqτ s BrF pqqs
K “ `
Bτ Bτ Bτ
Using the principle of variable separation, the above equation is equivalent to:
Bq BrHpqqs Bq BrF pqqs Bq

K “ pIn b τ q ` Hpqq `
Bτ Bq Bτ Bq Bτ
From this we get: ˜ ¸
BrHpqqs BrF pqqs Bq
K´ pIn b τ q ´ “ Bpqq
Bq Bq Bτ
where the symbol b is the Kronecker tensor product. Let us define the matrix B as follows for
the sake of simplicity:
Bpτ, qq “ K ´ Hq pqqpIn b τ q ´ Fq pqq (2.50)
where
BrHpqqs BrF pqqs
Hq pqq “ , Fq pqq “
Bq Bq
39
Obviously, the calculation of ∇τ pqq depends on the invertibility of the matrix B in (2.50),
which in turns is indirectly determined by the generalized actuation matrix H in (2.10), and the
vector of generalized position-dependent external force F in (2.13f).
Therefore, the following assumption has been imposed in this thesis.
Assumption 1. For the studied soft robot, it is supposed that the actuators are installed in
such a way ensuring that the matrix B defined in (2.50) is invertible for all q P WE .
Remark 3. Physically, the above assumption means that the actuators of the investigated soft
robots are arranged in a way that a small actuators variation for a feasible configuration q P WE
might locally and uniquely determine the relative variation of position. In other words, it implies
that the mounted actuators are installed such that they are not mounted on each other. If such
an assumption is violated, i.e., several actuators are mounted in a redundant way, then it is
necessary to keep only 1 actuator and remove other ones.
Hence, if Assumption 1 is satisfied, we can then obtain:
˜ ¸´1
Bq BrF pqqs
“ K ´ Hq pqqpIn b τ q ´ Hpqq
Bτ Bq
“ ‰T
which means ∇τ pqq “ H T pqq B´1 pτ, qq .
As stated earlier in the introduction of this chapter (Section 2.5.1), the straightforward
approach to estimate the workspace via (2.48) is to firstly discretize the actuators bounds T (as
indicated in Fig. 2.5), and then approximate the mapping φ in an iterative manner.
Precisely, after discretizing the actuators force bounds T , noted as Td , then for each τ P Td ,
with the knowledge of the end-effector position upq pj´1q q and the generalized strain vector q pj´1q
(for the case j “ 1, upq p0q q, q p0q represents the end-effector’s initial position and the initial
generalized strain vector of the studied soft robot respectively), the following proposition enables
us to calculate the next generalized strain vector q pjq .
Proposition 1. If Assumption 1 is satisfied, then for a given τ P Td the following dynamics:
q pjq “ q pj´1q ´ B´1 pτ, q pj´1q q Φpτ, q pj´1q q (2.51)
exponentially converges to the corresponding solution of q in (2.48a), with B defined in (2.50)

and
Φpτ, q pj´1q q “ Kq pj´1q ´ Hpq pj´1q qτ ´ F pq pj´1q q.
Remark 4. It is important to state that convergence towards the closest solution with reference
to the initial solution can only be guaranteed if the conditions resulting from Kantorovich’s
theorem [89] are respected.
Proof. Solving (2.48a) numerically can be achieved by minimizing the following quadratic cost
function Spτ, q ˚ q:
Spτ, q ˚ q “ ΦT pτ, q ˚ qΦpτ, q ˚ q, where

Φpτ, q ˚ q “ Kq ˚ ´ Hpq ˚ qτ ´ F pq ˚ q
First, we compute the time derivative of the cost function:

9 q ˚ q “ 2ΦT pτ, q ˚ qΦpτ,
Spτ, 9 q˚ q
40
Next, we apply the principle of variable separation, and we obtain:
˚
9 q ˚ q “ 2ΦT pτ, q ˚ q BrΦpτ, q qs q9˚
Spτ,
Bq ˚
T
“ 2Φ pτ, q qBpτ, q ˚ qq9˚
˚
with B defined in (2.50).

In order to minimize the cost function Spτ, q ˚ q, we need to define an adaptive law q9˚ such
9 q ˚ q is negative. For this, and in accordance with Assumption 1, if the following
that Spτ,
adaptive law is chosen:
q9˚ “ ´γB´1 pτ, q ˚ qΦpτ, q ˚ q (2.52)
where γ is a positive parameter, then we can obtain:
9 q ˚ q “ ´2γSpτ, q ˚ q
Spτ,
which implies that Spτ, q ˚ q exponentially converges to 0. Equivalently, it means that q ˚

exponentially converges to q with the proposed continuous-time adaptive law (2.52). Finally,
the discrete-time version of (2.52) can be written as:
q pjq “ q pj´1q ´ ∆tγB´1 pτ, q pj´1q q Φpτ, q pj´1q q

1
and we prove Proposition 1 by choosing γ “ .
∆t
After, with q pjq calculated via (2.51), we compute the strain twist ξî related to the strain
pjq
pjq
vector qi of each body i as follows:
ξî “ Bqi qi ; i P 1 ¨ ¨ ¨ N
pjq { pjq
with which the configuration matrix gpq pjq q associated to the generalized strain vector q pjq can
be calculated as: pjq pjq
gpq pjq q “ eL1 ξ̂1 ¨ ¨ ¨ eLN ξ̂N
Then, we can derive the corresponding end-effector position as follows:
upq pjq q “ E1 gpq pjq q ET2
Finally the estimation of WE can be successfully accomplished by iterating the above

procedure for all feasible inputs τ P Td from the initial state j “ 1 to the last iteration j “ nd .
2.5.3 Forward Approach for Workspace Estimation: FEM case

In accordance with the workspace definition (see Definition 1), the workspace of a soft robot
can finally be defined by the following set of equations:
Kpqq q “ Hpqqτ ` F pqq (2.53a)

qE “ Cq (2.53b)
u “ qE ` up0q (2.53c)
where qE is the displacement vector of the end-effector with respect to its known initial position
vector up0q , u represents the position vector of the end-effector, and C P R3ˆ3n is a selection
matrix associated with the end-effector node coordinates.
41
To achieve the straightforward approach, given the information of the end-effector’s position
vector upj´1q and the vector of nodal displacements q pj´1q (for the case j “ 1, up0q and q p0q are
respectively the initial position of the end-effector and the initial nodal displacement vectors
of the investigated soft robot), we apply Newton-Raphson method to obtain the next nodal
displacement q pjq corresponding to its actuators value τ P Td , with Td being the discretized set
of T , which can be realized as follows:
q pjq “ q pj´1q ´ Φq pu, q pj´1q q Φpτ, q pj´1q q, @τ P Td
where Φq pτ, qq is the gradient of Φpτ, qq with respect to q, and Φpτ, q pj´1q q “ Kpq pj´1q qq pj´1q ´
Hpq pj´1q qτ ´ F pq pj´1q q. Then, the corresponding end-effector position vector can be deduced
by:
upjq “ Cq pjq ` up0q
Finally, as depicted by Fig. 2.5, the estimation of the reachable workspace WE can be
achieved by iterating the above process for all τ P Td .
2.5.4 Experimental Validation

The goal of this section is to validate the workspace estimation obtained by applying the forward
approach on the PCS and FEM models with the real workspace obtained from the experimental
setup using a trunk-like soft robot [130] (as depicted by Fig. 2.6) for both the PCS and FEM
models.
The exact geometric parameters of the investigated trunk-like soft robot are illustrated in
Fig. 2.6. The studied soft robot is controlled by 4 independent symmetric tendons (2 upward
and 2 lateral) mounted through it from the base to the tip, and the tendons are actuated by 4
independent stepper motors. To obtain the value of the end-effector position, a sensor position
was positioned on the tip of the robot, and for this purpose, a long and uniform hole was made
along the whole length of the robot. Moreover, 3D-printed rigid rings (blue rings in Fig. 2.6)
are mounted along the soft robot to minimize friction between the tendons and the soft rubber.
Finally, the casting material of the investigated trunk-like robot is an isotropic silicone rubber
with its Young’s modulus E “ 6.5 ˆ 105 P a and a Poisson’s ration ν “ 0.45.
Figure 2.6: Trunk-like soft robot [129].
42
Using the forward approach, the real workspace can be achieved by reading the value of the
position of the end-effector for each value of τ . The actuators bounds were within a magnitude
T “ r0, 5s1 ˆ ¨ ¨ ¨ ˆ r0, 5s4 N , and the obtained experimentation result of the workspace is depicted
by gray-colored points in Fig. 2.7.
On the other hand, in a simulation environment, we apply both the PCS and FEM methods
to model the investigated trunk-like soft robot. The studied soft robot was modeled by the
PCS method using N “ 15 cylindrical bodies with equal lengths and decreasing radius from
the base to the tip of the robot. Using the FEM method with N “ 5024 tetrahedral elements
and n “ 1484 nodes, the investigated soft robot was modeled. The tendons in the simulation
environment were approximately positioned using their real coordinates. Based on the adopted
mathematical models, we apply the forward approach to get the corresponding workspace
estimation. The obtained simulation results of the workspace for both the PCS and FEM
cases are depicted by red-colored points in Figs. 2.7a and 2.7b, respectively. The result of the
workspace obtained from the real experimentation took approximately 360 seconds, while the
results of the workspace obtained via the simulations took approximately 50 seconds.
(a) WE : Real (gray) and PCS-simulation (red). (b) WE : Real (gray) and FEM-simulation (red).
Figure 2.7: Forward approach - WE obtained from the real experimentation (gray) and WE
achieved by the simulations of the adopted mathematical models (red).
Visually, we can observe from Fig .2.7, the workspace estimation results estimated by applying
the forward approach on the PCS and FEM case approximately matches the workspace result
obtained from the real experimentation. Numerically, we compare the volume of the set of
points obtained from the real experiment Vr with the set of points obtained via the PCS (noted
VsP CS ) and FEM (noted VsF EM ) simulation results. The volume calculated from the set of
points obtained by the real robot is Vr « 0.0021m3 , and the volumes obtained the PCS and
FEM simulation results are VsP CS « 0.0028m3 and VsF EM « 0.0025m3 , respectively. Then, by
comparing both results, we find |Vr ´ VsP CS | « 7.3 ˆ 10´4 m3 for the accuracy of the PCS result
and |Vr ´ VsF EM | « 4.8 ˆ 10´4 m3 for the accuracy of the FEM result.
Based on this, and in the following chapters, the workspace obtained via the forward approach
will be considered as a reference. Moreover, the workspace estimation achieved by the approaches
that will be presented in the next chapters (Chapter 3, Chapter 4, and Chapter 5) will be
superimposed and compared with the workspace estimation obtained from the forward approach
43
in order to highlight their validity and efficiency.
2.6 Conclusion
In this chapter, we have presented the mathematical models of two adopted modeling methods,
PCS and FEM. Next, we have presented the definition of the workspace of a soft robot. Finally,
we have proposed a forward approach to estimate the workspace of soft robots based on PCS and
FEM, and we have used a trunk-like soft robot to validate the workspace estimation obtained
via the forward approach for the PCS and FEM with the real workspace result.
However, the main disadvantage of the forward approach lies in its computation complexity
which can be very high.
Given nd subsets (nd P N), since τ P T Ă Rnτ (nτ P N), for each entry of τ we will obtain nnd τ
possible combinations of input for τ . Based on this perspective, the computation complexity of
the forward approach depends on two parameters, nd and nτ . To obtain a correct estimation of
WE , the discretization precision nd needs to be sufficiently big, which will result in a significant
increase in the computation complexity. Moreover, it will be exponentially exploded when the
number of inputs nτ becomes larger.
Note that the base of the computation complexity for the forward approach equals nτ (the
inputs’ dimension), which varies according to different configurations of soft robots. Whereas,
the dimension of the end-effector space is constant (nu ď 3; for the position case) because it is
independent of any configuration of the investigated soft robot. Based on this observation, to
reduce the computation complexity when estimating the workspace of soft robots, it is therefore,
reasonable to propose novel methods that are based on discretizing the output space, i.e. the
end-effector space u, which is invariant in dimension, instead of discretizing the input space, i.e.,
the actuators space T with variant dimension.
As stated in Section 2.5.4, the forward approach will be used as a reference to validate the
workspace methodologies presented in the following chapters.
44
Chapter 3
Optimization-Based Approach for

Workspace Estimation
3.1 Introduction
Due to the high computation complexity of the forward approach when estimating the workspace
of soft robots, it is logical to propose an approach that can discretize the end-effector space
since its dimension is smaller and constant compared to the actuators space for hyper-redundant
robots, such as soft robots. In addition, it is more efficient to only map the exterior boundary
of the workspace and avoid the heavy computation of its interior points.
Such a method is called the optimization-based approach [116] in this thesis and it consists
of discretizing the end-effector space in order to map the exterior boundary of WE , noted as
BWE , and deducing that the area/volume enclosed by BWE is the reachable workspace WE .
Generally speaking, this method consists of numerical algorithms used to map the surface
enveloping soft robots’ workspaces [9, 123], which will be realized by the following steps:
1. In the first step, we establish the definition of a soft robot’s workspace boundary, denoted
as BWE .
2. Next, consistent with the workspace boundary definition, we seek to find a boundary point
ub on BWE . The strategy proposed consists of selecting a radiating point v, then from
this point, a ray is emanated along a certain direction until BWE is met (as illustrated by
Fig 3.1).
Figure 3.1: For a radiating vector v determine its associated boundary point ub .
45
3. Finally, a map of BWE can be achieved by solving the procedure that consists of finding
ub for multiple successive rays with respective directions emanating at angular intervals.
Unlike the forward method (Section 2.5), the computation complexity of the proposed
optimization-based approach depends on two factors: the dimension of the space to be discretized,
and the discretization precision. By managing these two parameters, the optimization-based
approach can reduce the complexity of the workspace estimation.
1. It can reduce the dimension of the space to be discretized in the forward approach. In
the forward approach, we discretize the actuators space whose dimension nτ is varying
according to the soft robot structure, but the space to be discretized in the optimization-
based approach is always less than or equal to 3 for the position case;
2. When discretizing the space with a prescribed precision (with nd sub-spaces), the forward
method cannot reduce the value of nd , while the optimization-based approach, which
consists only in finding then mapping the boundary points ub , results in reducing the total
number of iterations.
Briefly, the proposed optimization-based approach can be summarized into the following
procedure:
[Step 1] Defining the workspace boundary BWE .
[Step 2] Finding a point on BWE .
[Step 3] Mapping the workspace boundary BWE .
In the following, we present the detailed procedure of the optimization-based approach for
both the PCS [123] and the FEM [9] modeling methods.
3.2 Optimization Approach for the PCS case

3.2.1 Workspace Boundary BWE Definition
Given a configuration of a soft robot, with its workspace WE defined in Definition 1, and
τ P T “ rτ , τ s, @u P WE , we can then denote the neighborhood of u with a specified positive
radius ε by:
Bε puq “ t uε P Rnu | ku ´ uε k ă ε u (3.1)
based on which the workspace boundary BWE according to [116] can be defined below.
Definition 2. The workspace boundary BWE is a subset of its workspace WE , and defined by
the following:
BWE “ tu P WE | @ε ą 0, Duε P Bε puq, s.t. uε R WE u
Remark 5. To address the mapping of BWE , the strategy proposed is based on solving a
constrained minimization problem defined in Section 3.2.2 of the optimization-based approach
procedure. For this, we first need to calculate the
“ gradient
‰ of the end-effector position upqq with
respect to the actuators vector τ , denoted as ∇τ upqq , and established in the following theorem.
46
Theorem 1. For a given controllable configuration of the soft robot, with ∇τ pqq defined in
Assumption 1, the geometric Jacobian Jpqq defined in (2.7a), and the orientation matrix Rpqq
computed via the definition of the configuration matrix gpqq through the following:
Rpqq “ E1 gpqq ET1 (3.2)
then the
“ gradient
‰ of the end-effector position upqq with respect to the actuators vector τ , denoted
as ∇τ upqq , is established by:
“ ‰
∇τ upqq “ ∇τ pqqJ T pqqr03 RpqqsT (3.3)
“ ‰T
with ∇τ pqq “ H T pqq B´1 pτ, qq being defined in Assumption 1, and Bpτ, q, L, αq “ KpL, αq ´
Hq pq, L, αqpIn b τ q ´ Fq pq, L, αq.
Proof.
“ Using
‰ the principle of variable separation, the gradient of the end-effector position
∇τ upqq can be written as follows:
” ı ” Bupqq ıT ” “ ‰T ıT
∇τ upqq “ “ uq pqq ∇τ pqq
Bτ
Based on the velocity twist vector η̂pqq defined in (2.7c), we can derive the linear velocity vpqq
via the following:
vpqq “ upqq
9 “ E1 η̂pqq ET2 , with η̂pqq “ g ´1 pqq gpqq
9
Next, we develop the term η̂pqqET2 :
Bgpqq
η̂pqq ET2 “ g ´1 pqq gpqq
9 ET2 “ g ´1 pqq 9 ET2
pI4 b qq
Bq
Using the definition of the configuration matrix gpqq, we develop the above equation, and we
obtain:
» fi
„ ´1 ´1
 BRpqq Bupqq „ 
R pqq R pqqupqq 0
η̂pqq ET2 “ – Bq Bq fl q9
0 1
0 0
” Bupqq ıT
“ R´1 pqq q9 0
Bq
Substituting the above equation of η̂pqqET2 in the linear velocity upqq,
9 we achieve the following
expression of upqq:
9
Bupqq
upqq
9 “ E1 η̂pqq ET2 “ R´1 pqq q9 (3.4)
Bq
Using the relation between the velocity vector ηpqq and the geometric Jacobian Jpqq defined by
(2.7a), we derive another form of the linear velocity upqq,
9 which can be formulated as follows:
“ ‰T
ηpqq “ wpqq upqq
9 “ Jpqq q9
which gives:
upqq
9 “ r03 I3 s Jpqq q9 (3.5)
Using the expressions (3.4) and (3.5), we deduce the following equation:
Bupqq
R´1 pqq q9 “ r03 I3 s Jpqq q9
Bq
47
Then, we get:
Bupqq
uq pqq “ “ Rpqq r03 I3 s Jpqq “ r03 Rpqqs Jpqq
Bq
“ ‰
Finally, by substituting uq pqq in ∇τ upqq , we deduce the following:
“ ‰ “ ‰T
∇τ upqq “ ∇τ pqq uq pqq “ ∇τ pqqJ T pqqr03 RpqqsT
3.2.2 Finding a point on BWE

In order to map the surface enveloping the workspace WE , we need first to figure out how to
determine a boundary point ub . For concrete applications, it is a nontrivial task to find a point
on the workspace boundary BWE (as shown in Fig. 3.2). To clearly visualize this problem, we
present the following simplified illustration (see Fig. 3.2).
Figure 3.2: For a radiating vector v, calculate its corresponding actuators τ b via (3.6), and
deduce its associated boundary point ub via (2.48).
Assume that a radiating point v is selected, and that v is exterior to the workspace WE . It
is now proposed that, consistent with Definition 2, a boundary point ub in a particular direction
emanating from the radiating point v (as depicted by Fig. 3.2), can be determined by solving
the following constrained optimization problem [116]:
τ b “ arg min kupqq ´ vk22 “ f pqq

τ
s.t. τ P T (3.6)
K q “ Hpqqτ ` F pqq
upqq “ E1 gpqq ET2
where τ b is the optimal and feasible actuators vector (since it satisfies the bounded constraint)
for the purpose of minimizing the nonlinear cost function f pqq “ kupqq ´ vk22 , which represents
the distance between v and upqq.
Based on the solution of τ b , and in accordance with Proposition 1, we calculate the cor-
responding generalized strain vector q b through (2.51). Then, we compute the strain twist ξîb
related to the strain vector qib of each segment i using (2.49b). Next, we derive the configuration
matrix gpq b q associated to the generalized strain vector q b from (2.49a). Finally, using (2.48b),
we deduce the boundary point position ub .
48
Regarding the implementation of the methodology, we `need ˘to compute the gradient of the
objective function f pqq with respect to τ , denoted by ∇τ f pqq , which can be formulated as
follows: ` ˘ “ ‰` ˘
∇τ f pqq “ 2∇τ upqq upqq ´ v (3.7)
“ ‰
with ∇τ upqq defined in (3.3) .
Another additional question linked to the implementation of the proposed methodology is
how a radiating point v may be generated. As depicted by Fig. 3.2, the radiating point v must
be exterior to the workspace. In practice, since the length L of a soft robot is bounded (L ď L),
~ ą L, where O “ r0, 0, 0sT is the origin of the inertial
we only need to choose v such that ||Ov||
frame.
3.2.3 Mapping the Workspace Boundary BWE

In this section, we present a numerical methodology used for mapping the workspace boundary
BWE . Solving this problem is equivalent to finding a boundary point ub for multiple successive
directions (as indicated in Fig. 3.3). In the following, we represent a simplified illustration of
mapping the workspace boundary (see Fig. 3.3).
Figure 3.3: Numerical map of BWE - for each radiating vector v pjq , compute its corresponding
actuators τ bj via (3.8), and deduce its associated boundary point ubj via (2.48).
The workspace boundary BWE may be numerically mapped by solving the optimization
problem (3.8) for nd successive rays, with respective directions v pjq , j “ 0, 1, 2, . . . , nd emanating
at angular intervals of the angles δz , δx and δy (with δz,x,y “ 2π{nd ) from a radiating point v0
[116]:
τ bj “ min }upqq ´ v pjq }22
τ
s.t. τ P T
(3.8)
K q “ Bpqqτ ` F pqq
upqq “ E1 gpqq ET2
where τ bj is the optimal and feasible actuators force vector in order to minimize the distance
between v pjq and upqq.
From the solution τ bj , and according to Proposition 1, we calculate the corresponding
b
generalized strain vector q bj (2.51). Then, we compute the strain twist ξî j related to the strain
b
vector qi j of each segment i using (2.49b), and based on which we derive the configuration
49
matrix gpq bj q associated to the generalized strain vector q bj from (2.49a). Finally, we deduce
the boundary point position ubj using (2.48b).
The direction vector v pjq is expressed by the following:
pjq
v pjq “ Eδ v0
where the initial radiating point v0 is exterior to the workspace and can be calculated using
pjq pjq
the same technique suggested in Section 3.2.2. The rotation matrix Eδ is defined as: Eδ “
Eδpjq Eδpjq Eδpjq , and Eδpjq represents the basic rotation matrices about the z, x and y axis
z x y z,x,y
respectively [52].
Up to this point, the procedure of the optimization-based approach has been completely
described in each step with its related specifics.
3.2.4 Convexity Issues during the Mapping of BWE

Going a step further, difficulties may now arise in relation to some details of the methodology
implementation, and it concerns the strategy to be adopted if non-convexity of the workspace
WE interferes with the mapping of the workspace boundary BWE as illustrated in Fig. 3.4,
where as a result of the non-convexity problem, the workspace may be over-estimated.
Figure 3.4: Non-convexity problem of BWE - sudden leap between the deduced boundary points
ubn1 and ubn2 .
Eventually, in this particular situation, mapping the boundary BWE may result in an over-
estimation of the workspace since all cyan colored boundary points in Fig. 3.4 between the
successfully determined boundary points ubn1 and ubn2 , may not be identified, due to fact that,
for all admissible direction vectors v pjq between v pn1 q and v pn2 q , the only possible solutions that
can be achieved by the minimization problem (3.8) will be either ubn1 or ubn2 , i.e.,
@v pjq P rv pn1 q v pn2 q s, (3.8) ùñ ubn1 _ ubn2
To solve this specific problem, we need first to understand the phenomenon leading to the
occurrence of this situation. Actually, this particular behavior is mainly due to a sudden leap
between two successive iterations, either gradual or declining, of at least one actuators force
b
τk j P τ bj of the optimization problem (3.8).
50
In this thesis, and consistent with the above observation, the strategy proposed to solve
this particular problem is firstly based on discretizing the space between the boundary points
ubn1 and ubn2 , based on which we obtain a list of radiating points vl . In the next step, for each
radiating point vl P rubn1 ubn2 s we proceed by solving the minimization problem (3.6), which
allows us to obtain the feasible actuators force vector τ b corresponding to each radiating point
vl . Next, by solving (2.48), we can identify the boundary points ub on the non-convex part
of the workspace (See the cyan colored boundary points in Fig. 3.4). Finally, by mapping all
boundary points, we obtain the correct estimation of the workspace (as shown in Fig. 3.5).
Briefly, the proposed strategy to solve the non-convexity problem can be summarized into
the following procedure:
[Step. A] Calculate vl by discretizing the space between the boundary points ubn1 and ubn2 .
[Step. B] For each discretized point vl , find its related feasible actuators vector by solving (3.6),
then deduce its corresponding boundary point from (2.48).
In the following illustration (see Fig. 3.5), we present a simplified figure describing two iterations
of the proposed strategy for solving the non-convexity problem.
Figure 3.5: Non-convexity solution - for each point vl , calculate its related actuators via (3.6),
and deduce its associated boundary point via (2.48).
In the following, a brief algorithm is presented to describe the main steps of the optimization-
based approach along with the adopted strategy to solve the non-convexity problem (See
Algorithm 1).
3.2.5 Validation and Analysis

In this section, we will illustrate the presented results by evaluating the workspace for different
configurations of tendon-actuated soft robots. We consider various scenarios, and we implement
both the forward and optimization-based approaches to get WE and BWE respectively.
51
Algorithm 1 Calculate BWE
Require: v0 , T , F̄ei , F p0q , E1 , E2 , K, Ei , Gi , Ai , Li , nd , ns , r
δ Ð 2π{nd ; BWE Ð ∅; Ź initialization
Y Ð rT , E1 , E2 , K, Ei , Gi , Ai , Li s; F Ð F p0q
for j Ð 1 to nd do
pjq
v pjq Ð Eδ v0 Ź direction vector
pjq
τb Ð solvep3.8q pv pjq , Y q Ź optimal solution
pjq pjq pjq
rub , qb s Ð solve(2.48) pτb , Y, F q Ź boundary point
pjq
F Ð solve(2.13f) pqb , F̄ei q
pjq
BWE Ð BWE ‘ ub
Ź append
pj´1q pjq
if }ub ´ ub }2 ą r then Ź non-convexity
pjq pj´1q
vl Ð dispub , ub , ns q Ź discretize
for s Ð 1 to ns do
psq psq
τb Ð solve(3.6) pvl , Y q Ź optimal solution
psq psq psq
rub , qb s Ð solve(2.48) pτb , Y, F q
psq
F Ð solve(2.13f) pqb , F̄ei q
psq
BWE Ð BWE ‘ ub Ź append
end for
end if
end for
In the following table (see Table 3.1), we give the numerical values of the Young’s modulus
Ei , shear modulus Gi , the Poisson’s ratio ν, the inertial length Li , the radius ri , the second
moments of area Jpx,y,zqi and the cross-section area Ai for each segment i P t1, 2, 3u.
Soft robot segments

Physical
First Second Third
parameters
Li (m) 0.3 0.2 0.1
ri (m) 0.06 0.04 0.02
Jxi (m4 ) 2.03 ˆ 10´5 4.02 ˆ 10´6 2.51 ˆ 10´7
Jpy,zqi (m4 ) 1.07 ˆ 10´5 2.01 ˆ 10´6 1.25 ˆ 10´7
Ai (m2 ) 0.0113 0.0050 0.0013
E (Pa) 18 ˆ 104
ν 0.45
G (Pa) 62.069 ˆ 103
Table 3.1: Numerical values of the soft robot physical parameters.
For the optimization-based approach we propose to discretize the angles with a discretization
step size of 0.05 Radian (unit).
In the following scenarios and also for the scenarios of the following workspace estimation
chapters (Chapter 4 and Chapter 5), the obtained results of the workspace estimation via the
proposed approaches are superimposed with the forward method results, and the 3D results are
presented via 2D-views (Oxz, Oyx, and Oyz). Also, since in this thesis we only focus on the
52
position aspect of the workspace (see Definition 1), the obtained simulation results represent only
the feasible positions of the workspace but not the orientation aspect. Besides, please note that
one operation is a complete iteration that contains all the necessary steps to achieve one single
feasible point in the workspace. Also, the presented simulation scenarios were implemented
using MATLAB (In the forward approach, we used the function ”fsolve” to solve the forward
statics, while the function ”fmincon” was used to solve the optimization problems formulated in
this thesis) on an Intel Xeon(R) with a 16-GB RAM and a 3.50 GHz processor.
Finally, for the forward approach, we propose to discretize the actuators force vector with
a discretization step size of 1 Newton (unit). In addition, since we know from simulation the
cost of a single operation of the forward approach in order to find a feasible configuration in the
workspace (0.00165 second for a single operation), then, we only gave an approximation to the
computational time for the scenarios where the investigated soft robot is actuated by more than
4 actuators since the workspace estimation of these scenarios via the forward approach requires
a long time.
3.2.5.1 Scenario 1: planar 2-segments 2-tendons soft robot
In the first scenario, we consider a soft robot composed of two segments and actuated by
two tendons (See Fig. 3.6a). The two tendons are installed on the Ozx-plan, and parallel to
each other. The first tendon is fixed at the position p0, 0, ´r1 {2q and extends along the first
segment length L1 . The second tendon is fixed at the position p0, 0, ´r2 {2q and extends to
pL1 ` L2 , 0, ´r2 {2q. Such a chosen structure allows us to obtain a 2D workspace. We choose
a tension magnitude within T1,2 “ r0, 200sN that will allow us to apply a couple magnitude
within CT “ r0, 6sN ¨ m and CT “ r0, 4sN ¨ m.
1 2
First, we use the forward approach to estimate the workspace WE of this soft robot. We
choose nd “ 200, and we obtain the following estimation of WE (See Fig. 3.6b). The computation
cost of the forward approach for the two segments - two tendons soft robot is equal to 2002
operations, and the workspace estimation took 66.35 seconds to obtain the full estimation of its
workspace WE .
Next, we apply the optimization-based approach with nd “ 120, and based on (3.8) we
calculate all boundary points on the surface enveloping the workspace (See Fig. 3.6b). Finally,
as illustrated by Fig. 3.6b, we achieve BWE of this particular soft robot configuration. The
computation cost of the optimization-based approach for the studied soft robot is equal to 168
operations, and it required nearly 1.82 seconds to fully estimate its BWE .
53
(a) Soft robot structure. (b) Forward and Optimization approach.
Figure 3.6: Scenario 1 - workspace (gray) WE , boundary points (green), and workspace boundary
(red) BWE .
In the second scenario, we consider the soft robot of the first scenario and we add two symmetric
tendons (See Fig. 3.7a). We choose a tension magnitude within T1,¨¨¨ ,4 “ r0, 100sN that will
allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m and CT “ r0, 2sN ¨ m.
1,3 2,4
cost of the forward approach for the two segments - four tendons soft robot is equal to 1004
operations, and the workspace estimation will take 1.65ˆ105 seconds to obtain the full estimation
of its workspace WE .
54
Figure 3.7: Scenario 2 - Workspace (gray) WE , boundary points (green), and workspace boundary
(red) BWE .
In the third scenario, we consider adding another segment and a tendon to the first scenario
(See Fig. 3.8a). The additional tendon is mounted in a parallel manner to the tendons in the
first scenario: it is fixed at the position p0, 0, ´r3 {2q and extends along the soft robot length
L1 ` L2 ` L3 . Also, this configuration enables us to achieve a larger 2D workspace estimation
compared to the first scenario. We choose a tension magnitude within T1,2,3 “ r0, 100sN that
will allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m, CT “ r0, 2sN ¨ m, and
1 2
CT “ r0, 1sN ¨ m.
1
cost of the forward approach for the three segments - three tendons soft robot is equal to 1003
operations, and the workspace estimation took 2962 seconds to obtain the full estimation of its
workspace WE .
55
(red) BWE .
In the fourth scenario, three additional tendons are routed in a fashion where they are symmetric
to the tendon configuration in the third scenario (See Fig. 3.9a). We choose a tension magnitude
within T1,¨¨¨ ,6 “ r0, 100sN that will allow us to apply a couple magnitude within CT “ r0, 3sN ¨m,
1,4
CT “ r0, 2sN ¨ m, and CT “ r0, 1sN ¨ m.
2,5 3,6
cost of the forward approach for the three segments - six tendons soft robot is equal to 1006
operations, and the workspace estimation will take 1.65ˆ109 seconds to obtain the full estimation
of its workspace WE .
Next, we apply the optimization-based approach with nd “ 120 , and based on (3.8) we
56
(red) BWE .
3.2.5.5 Scenario 5: spatial 3-segments 9-tendons soft robot
In the final scenario, three additional tendons are added to the configurations of the above
scenario in order to obtain a 3D workspace. The three additional tendons are mounted in
the following way; the seventh, eighth and ninth tendons are routed in a manner where the
seventh tendon is fixed at the position p0, ´r1 {2, 0q and extends along the first segment length
L1 , the eighth tendon is fixed at the position p0, ´r2 {2, 0q and extends to pL1 ` L2 , ´r2 {2, 0q,
and the ninth tendon is fixed at the position p0, ´r3 {2, 0q and extends along the soft robot
length L1 ` L2 ` L3 (See Fig. 3.10a). With the tendons being mounted in this fashion, the
actual configuration allows us to have a 3D workspace. We choose a tension magnitude within
T1,...,9 “ r0, 50sN that will allow us to apply a couple magnitude within CT “ r0, 1.5sN ¨ m,
1,4,7
CT “ r0, 1sN ¨ m, and CT “ r0, 0.5sN ¨ m.
2,5,8 3,6,9
choose nd “ 100, and we obtain the following estimation of WE (See Figs. 3.10b, 3.10c, 3.10d).
The computation cost of the forward approach for the three segments - nine tendons soft robot
is equal to 1009 operations, and the workspace estimation will take 1.65 ˆ 1012 seconds to obtain
the full estimation of its workspace WE .
Next, we apply the optimization-based approach with ndy “ 120 and ndz “ 72, and based
on (3.8) we calculate all boundary points on the surface enveloping the workspace (See Figs.
3.10b, 3.10c, 3.10d), and we finally achieve BWE of this particular soft robot configuration. The
operations, and it required nearly 510 seconds to fully estimate its BWE .
57
(a) Soft robot structure. (b) Oxz View.
(d) Oyz View.

(c) Oyx View.
Figure 3.10: Scenario 5 - Workspace (gray) WE , boundary points (green), and workspace
boundary (red) BWE .
3.2.5.6 Notes on scenarios results
Based on the achieved results, we can clearly observe that the workspace boundary estimation
result achieved via the optimization-based method encloses the estimated workspace obtained
via the forward approach.
In addition, the proposed optimization-based approach can yield equivalent workspace
estimation precision compared to the forward method, and simultaneously reduce both the
number of operations and the computational time required for the workspace estimation for
each scenario.
3.2.5.7 Notes on computational efficiency
For each particular configuration of a tendon-actuated soft robot, we showed the estimation of
its workspace using both the forward and optimization-based approach. From the simulation
results, we observe that the proposed optimization-based approach consisting of mapping the
58
exterior boundaries of soft robots’ workspaces has significantly reduced the computation cost of
the workspace estimation for each soft robot configuration.
In the following table (see Table. 3.2), we summarize the operations complexity and the
computational time for each scenario.
Table 3.2: Operations Complexity and computational time: forward vs optimization.
Operations Time (seconds)

Approach
Forward Optimization Forward Optimization
Inputs
2
nτ “ 2 200 168 66.35 1.82
nτ “ 3 1003 220 2962 3.52
nτ “ 4 1004 168 « 1.65 ˆ 105 3.92
nτ “ 6 1006 220 « 1.65 ˆ 109 13.58
nτ “ 9 1009 8640 « 1.65 ˆ 1012 510
Visually, the results of the above table are depicted in Figs. 3.11a and 3.11b, and we can
clearly observe that, when the size of the actuators, (i.e., the number of tendons) increases,
the computation complexity of the forward approach explodes exponentially, whereas the
computation complexity of the optimization-based approach remains almost linearly stable.
(a) Operations computation complexity. (b) Time computation complexity
Figure 3.11: Operations complexity and computational time of the forward and optimization
approach for the workspace estimation of soft robots using the PCS model.
3.3 Optimization Approach for the FEM case

3.3.1 Implementation of the Methodology
The implementation of the optimization-based approach on the FEM model [9] follows the same
steps described for the PCS model and detailed in Section 3.2.
In the first step, the definition of the workspace boundary for the FEM model is similar to
that of the PCS method (see Definition 2 in Section 3.2.1).
59
In the second step, the goal is the same, which is to find a point on BWE . However, the
optimal problem to find a boundary point is formulated under different constraints for the FEM
model, and established as follows:
rτ b , q b s “ arg min }upqq ´ v}22

τ,q
s.t. τ P T (3.9)
Kpqqq “ Hpqqτ ` F pqq
u “ Cq ` up0q
where the obtained actuators vector τ b with its associated displacement vector q b represents the
optimal and feasible solution of (3.9).
Finally, we deduce the related end-effector boundary position using the following:
ub “ Cq b ` up0q
Similar to the PCS case, the third step consists of mapping the workspace boundary BWE .
However, the optimal problem needs to be adapted for the FEM model, and formulated as
follows:
rτ bj , q bj s “ arg min }upqq ´ v pjq }22
τ,q
s.t. τ P T (3.10)
u “ Cq ` up0q
where the achieved actuators vector ubj with its associated displacement vector q bj is the optimal
and feasible solution to minimize the distance between v pjq and u .
Accordingly, the associated end-effector boundary position can be deduced as:
ubj “ Cq bj ` up0q
Finally, the solution to non-convexity problems that may arise using the FEM model can be
solved following the same procedure described in Section 3.2.4 for the PCS case. However, we
need to use the corresponding optimization equations (3.9) and the related FEM static model
established in (2.53) during Step. B of the proposed strategy.
The algorithm of the optimization-based approach for the FEM case is presented in the
following (See Algorithm 2):

In this section, we illustrate the presented results by evaluating the reachable workspace of a
trunk-like soft robot with a Young’s modulus E “ 1.8 ˆ 107 P a, and a Poisson’s ratio ν “ 0.45.
We consider three different scenarios where the trunk-like soft robot is actuated by different
cable routing configurations, then we implement both the forward and optimization-based
approach in order to estimate the reachable workspace WE and the workspace boundary BWE ,
respectively, for each scenario.
For the optimization-based approach we propose to discretize the angles with a discretization
step size of 0.08 Radian (unit).
60
Require: v0 , C, T , nd , ns , r, up0q
δ Ð 2π{nd ; BWE Ð ∅; Ź Initialization
for j Ð 1 to nd do
pjq
v pjq Ð Rδ v0 Ź direction vector
rτ bj , q bj s Ð solve(3.10) pv pjq , C, T , up0q q Ź optimal solution
ubj Ð Cq bj ` up0q Ź boundary point
BWE Ð BWE ‘ ubj Ź append
if }ubj´1 ´ ubj }2 ą r then Ź non-convexity
vl Ð dispubj´1 , ubj , ns q Ź discretize
for s Ð 1 to ns do
psq
rτ bs , q bs s Ð solve(3.9) pvl , C, T , up0q q
ubs Ð Cq bs ` up0q Ź boundary point
BWE Ð BWE ‘ ubs Ź append
end for
end if
end for
3.3.2.1 Scenario 1: trunk-like soft robot actuated by two tendons.
In the first scenario, we consider a cable routing configuration where the trunk-like soft robot is
actuated by two tendons (as shown by Fig. 3.12a). We apply a force whose magnitude is within
a specific range T p0q “ T “ r0 100s ˆ r0 100s.
Following the procedure presented in Section 2.5.3, we firstly apply the forward approach
with a discretization precision nd “ 100, and we obtain the corresponding estimation of
WE (illustrated by blue colored points in Figs. 3.12b, 3.12c and 3.12d) for this soft robot’s
configuration. In term of operations’ computation complexity, the forward approach requires
nnd τ “ 1002 operations, and in term of time’s computation, the workspace estimation took 1463
seconds to obtain the full estimation of this scenario’s workspace.
Next, we use (3.10) (See Fig. 3.3) of the optimization-based approach with ndz “ 72 and
ndx “ 12 to map the boundary points of this particular configuration.
For this scenario, the optimization-based approach required a total of 864 operations with a
time computation of 984 seconds. The obtained boundary points are depicted by green colored
points in Figs. 3.12b, 3.12c and 3.12d, which represent the workspace boundary BWE of this
soft robot’s configuration.
61
(a) Soft Robot Structure nτ “ 2. (b) Oxy plan view.
(c) Ozy plan view. (d) Ozx plan view.
Figure 3.12: Scenario 1 - WE estimation via forward approach (blue points) and BWE estimated
via optimization approach (green points).
3.3.2.2 Scenario 2: trunk-like soft robot actuated by three tendons.
In this scenario, we consider a cable routing configuration where the trunk-like soft robot is
actuated by three tendons (as shown by Fig. 3.13a). We apply a force whose magnitude is
within a specific range T p0q “ T “ r0 100s ˆ r0 100s ˆ r0 100s.
First, we apply the forward approach with a discretization number equal nd “ 100, and to
achieve the workspace WE estimation (illustrated by blue colored points in Figs. 3.13b, 3.13c
and 3.13d) of this particular configuration. In term of operations’ computation complexity,
the forward approach requires nnd τ “ 1003 operations, and in term of time’s computation, the
workspace estimation will take 1.4 ˆ 105 seconds to obtain the full estimation of this scenario’s
workspace.
Next, using (3.10) of the optimization-based approach with ndz “ 72 and ndx “ 12, we map
the boundary points of this particular configuration.
For this scenario, the proposed optimization-based approach required a total of 864 operations
with a time computation of 1034 seconds. The achieved boundary points are illustrated by green
colored points in Figs. 3.13b, 3.13c and 3.13d, which represent the workspace boundary BWE of
this particular configuration.
62
3.3.2.3 Scenario 3: trunk-like soft robot actuated by four tendons.
In the final scenario, we consider a cable routing configuration where the trunk-like soft robot
is actuated by four symmetric tendons (as shown by Fig. 3.14a). We apply a force whose
magnitude is within a specific range T p0q “ T “ r0 100s1 ˆ ¨ ¨ ¨ ˆ r0 100s4 .
First, we apply the forward approach with a discretization precision equal nd “ 100, and we
obtain the corresponding estimation of WE (illustrated by blue colored points in Figs. 3.14b, 3.14c
and 3.14d) for this soft robot’s configuration. In term of operations’ computation complexity,
the forward approach requires nnd τ “ 1004 operations, and in term of time’s computation, the
workspace estimation will take 1.4 ˆ 107 seconds to achieve the full estimation of this scenario’s
workspace.
Next, we use (3.10) of the optimization-based approach with ndz “ 72 and ndx “ 12 to map
the boundary points of this particular configuration.
For this scenario, the optimization-based approach required a total of 864 operations with a
time computation of 1039 seconds. The obtained boundary points are depicted by green colored
points in Figs. 3.14b, 3.14c and 3.14d, which represent the workspace boundary BWE of the
studied soft robot’s configuration.
63
In the following table (see Table 3.3), and Figures. 3.15a and 3.15b, we summarize the operations
complexity and time computation for each scenario.
Table 3.3: Operations and time computational complexity for the FEM case: forward vs
optimization-based.

Approach
Forward Optimization Forward Optimization
Inputs
nτ “ 2 1002 864 1463 984
nτ “ 3 1003 864 « 1.4 ˆ 105 1034
nτ “ 4 1004 864 « 1.4 ˆ 107 1039
64
(a) Operations computation complexity. (b) Time computation complexity
Figure 3.15: Operations complexity and computational time of the forward and optimization
approach for the workspace estimation of soft robots using the FEM model.
3.4 Conclusion
In this chapter, we have proposed an optimization-based approach that consists of mapping the
exterior boundaries of the workspace of soft robots.
The proposed approach was successfully implemented to both the PCS and the FEM models,
where we have shown its efficiency in reducing the computation complexity and computational
time necessary to estimate the workspace of soft robots in contrast to the forward approach
which explodes exponentially when increasing the dimension of the actuators.
On the other hand, since this approach consists only in mapping the exterior boundary of the
workspace by determining the boundary points lying on the surface enclosing the workspace, it
surely will not be able to identify inside information of the workspace such as interior boundaries.
65
66
Chapter 4
Interval Analysis Approach for

Workspace Estimation
Due to the limitations encountered by the optimization approach that cannot provide knowledge
on the interior configurations of the workspace, it is therefore necessary to provide a method
that can overcome this limitation, but is also based on the second strategy, i.e., of discretizing
the end-effector’s space.
Such a method is called the forward-backward approach [7] that is based on interval analysis
techniques [48, 61, 77–79], and it consists of discretizing the end-effector’s space by starting from
an initial reachable configuration then exploring the whole possible reachable space in order to
finally estimate the workspace of a soft robot.
In the following, we will first provide a brief introduction about interval analysis techniques
used by this approach, and then an introduction of the approach itself followed by its realization
for both the PCS and FEM models.
4.1 Interval Analysis Introduction

The proposed forward-backward approach uses interval analysis techniques in order to estimate
the workspace of soft robots. Thus, a brief introduction to some interval analysis techniques
used in this thesis is presented in the following. More details can be found in [61].
Consider two real intervals X1 “ rX 1 , X 1 s and X2 “ rX 2 , X 2 s, then according to [61], the
product of X1 and X2 can be calculated by the following:
“ ` ˘ ` ˘‰
X1 X2 “ min CpX1 , X2 q , max CpX1 , X2 q
with CpX1 , X2 q “ tX 1 X 2 , X 1 X 2 , X 1 X 2 , X 1 X 2 u.
The product of a real matrix A P R by a real interval X “ rX , X s with appropriate dimension
is calculated as follows: #“ ‰
AX , AX , A ą 0
AX “ “ ‰
AX , AX , A ă 0
However, due to the wrapping effect ( AX Ą tAx | x P X u ) [61], this technique yields an
over-estimation of the resulted space which can be explained by considering the following
example: ˆ ˙ ˆ ˙
1 1 r´1, 0s
A“ , X “
0 1 r1, 2s
67
then ˆ ˙
r0, 2s
AX “
r1, 2s
which implies that p0, 2qT belongs to AX , whereas it does not belong to the actual value set
B “ tAx | x P X u, as shown by Fig. 4.1.
Figure 4.1: Over-estimation introduced by the wrapping effect.
Therefore, it is necessary to propose a technique to eliminate the resulted over-estimation in

order to achieve the correct and accurate space, and this will be implemented in the proposed
forward-backward approach and explained in the following.
4.2 Forward-Backward Approach Introduction

In this section, we present a brief overview of how the proposed forward-backward approach
operates.
Firstly, we need to discretize the end-effector space. To realize this, we need to have an
idea about its bounds. In this approach, we approximate the relation between q and τ at the
pj ´ 1qth iteration as follows:
Kδq pjq “ Hpq pj´1q qδτ pj´1q ` F pq pj´1q q, PCS
and
Kpq pj´1q qδq pjq “ Hpq pj´1q qδτ pj´1q ` F pq pj´1q q, FEM
with δq pjq “ q ´ q pj´1q and δτ pj´1q “ τ ´ τ pj´1q , where δq pjq and δτ pj´1q represent respectively
the generalized strain and the actuators vectors with respect to the prior configuration (for
the initial state j “ 1, upq p0q q and q p0q are the end-effector’s initial position and the initial
generalized strain vector respectively, and τ p0q is the initial actuators force vector).
Based on the above result, a forward estimation of the end-effector workspace related to δq pjq
can be obtained (see Step 1 in Fig. 4.2), noted as WEpjq , by using interval analysis techniques
[61] (which will be detailed in Section 4.3.1 for the PCS case and Section 4.4 for the FEM
case). Once it is estimated, we can discretize WEpjq Ă R3 with a prescribed precision and obtain,
as shown by Step 2 in Figs. 4.2 and 4.3, the associated discretized space WEpjq (subscript “d”
d
stands for discretization). Next, based on the discretized space WEpjq , we calculate a feasible
d
68
small neighborhood, noted as WEpjq s
(subscript “s” stands for small neighborhood), around the
pj´1q
current end-effector position u (see Step 2 in Figs. 4.2 and 4.3). After that, we proceed
backwardly to the actuators domain T in order to compute the corresponding actuators τ pjq (see
Step 3 in Figs. 4.2 and 4.4), with which the investigated soft robot may reach WEpjq s
. This step,
i.e. Step 3, is necessary to obtain the accurate corresponding end-effector position upjq which
is approximately equal to the associated point in the set WEpjq s
. In addition, this step is also
pjq
necessary for the computation of the new actuators bounds T in Step 4. After, by utilizing the
actuators vector τ pjq inferred from Step 3, we calculate the new actuators bound T pjq (shown
as Step 4 in Figs. 4.2 and 4.5), allowing us to forwardly construct the new end-effector bound
WEpj`1q which will be required for the next iteration of the proposed approach.
However, since interval analysis techniques lead to an over-estimation of WEpjq (due to the
wrapping effect [61]), this thesis proposes a simple strategy to accurately reduce the dimension
of WEpjq . The strategy proposed to reduce the overestimation lies within two steps of the
forward-backward approach. In part, Step 2 indeed allows us to reduce the over-estimation
by discarding certain unfeasible small neighborhoods because the feasible small neighborhoods
WEpjq
s
must belong to the discretized space WEpjq and the possible small neighborhoods (called
d
the η-neighborhood in Step 2) around the current configuration upj´1q (which is illustrated
by Fig. 4.3, and will be explained in detail in Step 2, Section 4.3.2). On the other hand, the
over-estimation correction occurs mainly at the level of Step 5 of the approach where we dismiss
pre-explored feasible small neighborhoods, and therefore efficiently and precisely reduce the
over-estimated workspace (which is illustrated by Fig. 4.6, and will be explained in detail in
Step 5, Section 4.3.5).
Eventually, to achieve the whole estimation of WE using the forward-backward method, we
iterate the above procedure starting from the initial state j “ 1 until the stop condition (which
will be detailed in Step 5 and depicted by Fig. 4.7) is verified.
Figure 4.2: Forward-Backward approach.
Compared to the forward method, the forward-backward approach has the following two
advantages:
1. It reduces the space’s dimension that must be discretized (3 for the position case);
2. It eliminates wrong sub-spaces by the interval estimation technique, and discards pre-
explored configurations, yielding a significant decrease of the necessary iteration number
to explore the whole workspace, which results in reducing the required computational time
to achieve WE .
69
Compared to the optimization-based approach, the forward-backward approach is exhaustive
since it seeks to explore the whole workspace providing its interior and exterior points.
4.3 Forward-Backward Approach for the PCS case

In this section, we present the detailed realization of the forward-backward approach applied to
the PCS model following the portrayed scheme in Fig. 4.2.
4.3.1 Step 1: Forward estimation of WEpjq

In the first step, we use interval analysis techniques to forwardly estimate the workspace of the
end-effector.
Precisely, suppose that at the pj ´ 1qth iteration we have the information of the end-effector
position upj´1q , the generalized strain vector q pj´1q and the corresponding actuators vector
τ pj´1q P T pj´1q (for the case j “ 1, it corresponds to the initial configuration of the studied soft
robot). Then, it is now proposed that (2.48) can be approximated by the following:
Kδq pjq “ Hpq pj´1q qδτ pj´1q ` F pq pj´1q q (4.1a)
upjq
“ E1 gpδq pjq
q ET2 (4.1b)
with δq pjq “ q ´ q pj´1q and δτ pj´1q “ τ ´ τ pj´1q .

Let us note the matrix A as:
Apq pj´1q q “ K ´1 Hpq pj´1q q
Define now
A` “ maxt0, Au
where the maxtü operation is applied element-wise [135], and
A´ “ A ´ A`
then using the following theorem, we achieve an over-estimation of the end-effector workspace
WEpjq corresponding to the given actuators bound T pj´1q .
Theorem 2. For a given controllable configuration of the soft robot, with the actuators bound
T pj´1q defined as
pj´1q
T pj´1q “ rδτ pj´1q , δτ s
pj´1q
where δτ pj´1q and δτ represent the lower and upper bounds of the actuators vector with
respect to the prior configuration respectively.
An over-estimated end-effector workspace WEpjq is given by
! ” ı)
WEpjq “ upjq P Rnu |upjq P upjq , upjq (4.2)
with
pjq
upjq “ E1 gpδq pjq qET2 , and upjq “ E1 gpδq qET2
pjq
the configuration matrix bounds gpδq pjq q P rgpδq pjq q, gpδq qs , and the generalized strain vector
pjq pjq pjq
bounds δq P rδq , δq s are determined by:
pj´1q
δq pjq “ A` δτ pj´1q ` A´ δτ ` K ´1 F pq pj´1q q
70
and
pjq pj´1q
δq “ A` δτ ` A´ δτ pj´1q ` K ´1 F pq pj´1q q
Proof. From (4.1a) we can get the generalized strain vector as follows:
δq pjq “ K ´1 Hpq pj´1q qδτ pj´1q ` K ´1 F pq pj´1q q

“ Apq pj´1q qδτ pj´1q ` K ´1 F pq pj´1q q
Using Lemma 3 in [135], we get the following:

pj´1q pj´1q
A` δτ pj´1q ` A´ δτ ď Aδτ pj´1q ď A` δτ ` A´ δτ pj´1q
and deduce the generalized strain vector bounds as:

pjq
δq pjq ď δq pjq ď δq
with:
pj´1q
δq pjq “ A` δτ pj´1q ` A´ δτ ` K ´1 F pq pj´1q q
and
pjq pj´1q
δq “ A` δτ ` A´ δτ pj´1q ` K ´1 F pq pj´1q q
Next, applying the same lemma to (2.49b), and since Bqi ě 0, we obtain the strain twist bounds
as:
pjq pjq
ξˆ ď ξˆ ď ξˆ
pjq
i i i
After, by implementing the above result in (2.49a), and according to [61], we get the configuration
matrix bounds:
pjq
gpδq pjq q ď gpδq pjq q ď gpδq q
Finally, we deduce the bounds of the end-effector position:
upjq ď upjq ď upjq
with
pjq
upjq “ E1 gpδq pjq qET2 and upjq “ E1 gpδq qET2
which defines the over-estimated workspace of the end-effector WEpjq in (4.2).
4.3.2 Step 2: Computing the Feasible Small Neighborhood WEpjq

s
With WEpjq being estimated in Step 1, the second step consists of discretizing WEpjq with a
prescribed precision, and then selecting a feasible small neighborhood WEpjq
s
around the current
pj´1q
end-effector position u .
The proposed strategy is to discretize WEpjq using a globally uniform spatial grid with a
global small pre-defined precision sEd in each direction x, y and z, i.e.,
! )
WEpjq “ upjq P WEpjq |upjq “ upj´1q ` sEd Γ, for Γ P Z3 (4.3)
d
Next, an η-neighborhood of the configuration upj´1q can be defined as:

! )
Sη pupj´1q q “ ũ P R3 |}upjq ´ upj´1q }8 “ η
71
with η “ sEd . Consequently, with the discretized space WEpjq being established, we can then
d
define the feasible small neighborhood of the end-effector position upj´1q as follows:
WEpjq
s
“ WEpjq X Sη pupj´1q q (4.4)
d
The relationship between the estimated workspace WEpjq , its discretization space WEpjq and the
d
relative feasible small neighborhood space WEpjq
s
are illustrated by Fig. 4.3.
Figure 4.3: Feasible small neighborhood WEpjq

s
.
4.3.3 Step 3: Backward Estimation of the Corresponding Configura-

tion
The third step’s purpose is to choose the discretized points in the feasible small neighborhood
space WEpjq
s
, except the point upj´1q , and then backwardly calculate the feasible actuators vector
pjq
τ which can drive the soft robot’s end-effector to reach those points. The corresponding
scheme is depicted by Fig. 4.4.
Figure 4.4: Backward estimation of τ pjq .
Precisely, since any point ûpjq P WEpjq

s
zupj´1q is located inside the feasible small region of
pj´1q
u , then we can use this value to compute the associated configuration by solving the
72
following nonlinear constrained optimization problem:
τ pjq “ arg min kupqq ´ ûpjq k22

τ
s.t. τ P T (4.5)
upqq “ E1 gpδq pjq q ET2
where τ pjq is the optimal and feasible solution (since it satisfies the bounded and nonlinear
constraints) for the purpose of minimizing the nonlinear cost function kupqq ´ ûpjq k22 , which
represents the distance between ûpjq and upqq.
With the calculated τ pjq , and in accordance with Proposition 1, we can calculate the
corresponding generalized strain vector q pjq through (2.51). Then, we compute its related strain
twist ξî using (2.49b). Next, we derive the associated configuration matrix gpq pjq q via (2.49a).
pjq
Finally, using (2.48b), we may deduce the corresponding position upjq .

Regarding the
` implementation
˘ of the methodology, the gradient of f pqq with respect to τ ,
denoted by ∇τ f pqq can be analytically formulated as:
` ˘ “ ‰` ˘
∇τ f pqq “ 2∇τ upqq upqq ´ ûpjq
“ ‰
with ∇τ upqq “ ∇τ pqqJ T pqqr03 RpqqsT being already established in Theorem 1, and ∇τ pqq “
“ ‰T
H T pqq B´1 pτ, qq being defined in Assumption 1, with Bpτ, q, L, αq “ KpL, αq´Hq pq, L, αqpIn b
τ q ´ Fq pq, L, αq.
4.3.4 Step 4: Calculation of New Actuators Bound
In this step, we use the value of τ pjq to compute the associated new bound, noted as T pjq , from
T pj´1q .
pjq
For this, denote τk for 1 ď k ď nτ as the kth value of τ pjq , and diagpτ pjq q as the related
nτ ˆ nτ diagonal matrix whose entries are the nτ elements of the vector τ pjq , i.e.,
! )
pjq
diagpτ pjq q “ diag τ1 , ¨ ¨ ¨ , τnpjq
τ
Then, the new bound T pjq , illustrated by Fig. 4.5, can be determined using the value of τ pjq
from Step 3 and the value of T pj´1q as follows:
T pjq “ T pj´1q ´ diagpτ pjq q ˆ 1nτ ˆ2 (4.6)
where 1nτ ˆ2 represents the nτ ˆ 2 matrix with all entries equal to 1.
73
Figure 4.5: New actuators bound T pjq .
4.3.5 Step 5: Iteration and Stop Condition

For a particular configuration upj´1q with its corresponding over-estimated workspace WEpjq
calculated via Step 1, its related discretized space WEpjq computed via Step 2, and its associated
d
feasible small neighborhood WEpjq
s
deduced from (4.4). An important property of Step 2 allows
us to not take into account certain non-feasible neighborhoods around upj´1q , which results
in diminishing the number of feasible small neighborhoods WEpjq s
to be explored for the next
iterations of the forward-backward approach, as illustrated in Fig. 4.3, by verifying the condition
(4.4).
At the same time, given the information of WEpj´1q
s
, and the newly calculated feasible small
pjq
neighborhood WEs from (4.4), an additional characteristics allowing us to further shrink the
space to be explored is that some of the points belonging to WEpjq s
have already been explored
by the former iterations, noted as WEpj´1q
Y
, and defined by recurrence as follows:
WEpj´1q
Y
“ Yj´1 pkq
k“1 WEs
The above elimination procedure is executed in Step 5, as depicted by Fig. 4.6, by verifying the
following condition:
` ˘
WEpjq
s
Ð Sη pupj´1q qz WEpj´1q
Y
X Sη pupj´1q q (4.7)
Figure 4.6: Discard pre-explored feasible neighborhoods.
74
Combining these two conditions (Eq. (4.4) in Step 2 and Eq. (4.7) in Step 5) enables us to
shrink the space to be explored efficiently and accurately, resulting in a significant decrease of
the computation complexity of the proposed forward-backward approach (as it will be shown in
the simulation results in Section 4.3.6).
Consequently, the stop condition of the proposed forward-backward method can be formulated
as follows:
Dj P N, such that ũ P WEpj´1q

Y
, @ũ P WEpjq
s
or equivalently
Dj P N, such that WEpjq

Y
“ WEpj´1q
Y
implying that when backwardly estimating the related actuators vector τ pj´1q of each point
ũ P WEpjq
s
, and then, if the associated end-effector position upjq belongs to the union of all feasible
small neighborhood set of the last iterations WEpj´1q
Y
(as shown in Fig. 4.7), we can then conclude
that these configurations have already been explored.
Figure 4.7: Stop condition.
Finally, we deduce that all feasible small neighborhoods were explored and no more reduction
of the over-estimated workspace is possible. At this level, the estimated workspace is:
WE “ WEpj´1q
Y
.
In the following, a brief algorithm is presented in the table below to describe the main steps
of the forward-backward approach.
75
Algorithm 3 Calculate tu, q, τ, T , WEs , WEY upjq , WE
Require: Γ, sEd , tu, q, τ, T , WEs , WEY upj´1q
while WEpjq
s
R WEpj´1q
Y
do Ź Stop Condition.
WE Ð f1 ptu, q, T upj´1q q
pjq
Ź Step 1: Forward estimation of WEpjq , eq. (4.2)
WEpjq Ð f2 pupj´1q , Γ, sEd q Ź Step 2: Discretized space WEpjq , eq. (4.3)
d d
WEpjq
s
Ð WEpjq X Sη pupj´1q q Ź Step 2: Feasible neighborhood WEpjq
s
, eq. (4.4)
d
ûpjq Ð WEpjq s
zupj´1q Ź Exclusion.
pjq pjq pj´1q
tu, q, τ u Ð f3 pû , q q Ź Step 3: Backward estimation of upjq , eq. (4.5)
pjq pjq pj´1q
T Ð f4 pτ , T q Ź Step 4: New actuators bound T pjq , eq. (4.6)
pjq pj´1q
WEs Ð Sη pu qzpWEY X Sη pupj´1q qq Ź Step 5: Dismiss explored area, eq. (4.7)
pj´1q
WEY Ð WEY Y WEpjq

pjq pj´1q
s
Ź Append.
pj´1q
tu, q, τ, T , WEY u Ð tu, q, τ, T , WEY upjq Ź Update Values.
j Ðj`1 Ź Next iteration.
end while
WE Ð WEpj´1q Y
Ź The Estimated Reachable Workspace.

In this section, the same soft robots’ configurations presented in the scenarios of Section 3.2.5
in Chapter 3 are considered, where we implement both the forward and forward-backward
approaches to deduce the workspace estimation WE . Table 3.1 presents the numerical values of
the investigated soft robot.
In the following scenarios, the workspace obtained from the forward approach is depicted by
black-colored points. We suggest discretizing the end-effector space with a discretization step
size of sEd “ 0.02 Meter (unit).
In the first scenario, we consider a soft robot composed of two segments and actuated by two
tendons (See Fig. 4.8a). The two tendons are installed as the first scenario of Section 3.2.5.1 in
Chapter 3. We choose a tension magnitude within T1,2 “ r0, 200sN that will allow us to apply a
couple magnitude within CT “ r0, 6sN ¨ m and CT “ r0, 4sN ¨ m.
1 2
Next, we apply the proposed forward-backward interval analysis approach by following the
procedure presented in Section 4.3. Using T p0q we can forwardly estimate WEp1q as stated in Step.
1 of Fig. 4.2, based on which a feasible small neighborhood WEp1q s
is determined as mentioned
in Step. 2 of Fig. 4.2. Next, as indicated in Step. 3 of Fig. 4.2, we backwardly compute the
feasible actuators vector τ p1q of each configuration in the feasible small neighborhood WEp1q s
.
p1q
Finally, we find the new actuators bound T as demonstrated in Step. 4 of Fig. 4.2. The same
process is pursued in the next iterations until we meet the stop condition as stated in Step. 5
and shown by Fig. 4.7.
For this scenario, the forward-backward approach took 36 complete iterations and found a
total of 1820 feasible points in the workspace with a time computation of 17.5 seconds. Those
obtained feasible points are illustrated by red-colored points in Fig. 4.8b, and the union of those
feasible points represents the workspace WE estimation, which is depicted by the gray zone with
green contour. Moreover, for a complete 36 iterations it is expected to have WEd “ 636 « 1028
76
points to explore. However, what we actually explored was a total of WEd “ 6732 points, and a
final total of WEs “ 1820 feasible points.
(a) Soft robot structure. (b) Oxz plan view of WE .
Figure 4.8: Scenario 1 - WE estimation via forward approach (black points) and forward-backward
approach (gray area).
tendons (See Fig. 4.9a). We choose a tension magnitude within T1,2 “ r0, 150sN that will allow
us to apply a couple magnitude within CT “ r0, 4.5sN ¨ m and CT “ r0, 3sN ¨ m.
1,3 2,4
procedure presented in Section 4.3. The forward-backward approach in this scenario took 47
complete iterations and computed a total of 3640 feasible points in 187.63 seconds. Those
obtained feasible points are illustrated by red-colored points in Fig. 4.9b, and their union
represents the workspace WE estimation, which is illustrated by the gray zone with green
contour. In addition, we actually explored a total of WEd “ 13464 points, and a final total of
WEs “ 3640 feasible points, instead of an expected WEd “ 647 « 3 ˆ 1036 points to explore.
77
Figure 4.9: Scenario 2 - WE estimation via forward approach (black points) and forward-backward
approach (gray area).

In the third scenario (See Fig. 4.10a), we consider the same tendons routing of the third scenario
in Section 3.2.5.3 from Chapter 3. We choose a tension magnitude within T1,2,3 “ r0, 100sN
that will allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m, CT “ r0, 2sN ¨ m, and
1 2
CT “ r0, 1sN ¨ m.
1
Next, following the procedure presented in Section 4.3, we apply the proposed interval
analysis approach. For this scenario, the forward-backward approach took 32 complete iterations
and computed a total of 2275 feasible points in a duration of 72.84 seconds. Those obtained
feasible points are depicted by red-colored points in Fig. 4.10b, and their union is illustrated
by the gray zone with green contour, which gives the workspace WE estimation. Besides, we
actually explored a total of WEd “ 8415 points, and a final total of WEs “ 2275 feasible points,
instead of an expected WEd “ 632 « 7 ˆ 1024 points to explore.
Figure 4.10: Scenario 3 - WE estimation via forward approach (black points) and forward-
backward approach (gray area).
78
In this scenario, three additional tendons are routed in a fashion where they are symmetric to the
tendons configuration in the third scenario (See Fig. 3.9a). We choose a tension magnitude within
T1,¨¨¨ ,6 “ r0, 100sN that will allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m,
1,4
2,5 3,6
Next, the proposed interval analysis approach is applied by following the procedure presented
in Section 4.3. The forward-backward approach of this scenario took 55 complete iterations and
computed a total of 4550 feasible points in the workspace within 2436 seconds. The red-colored
points in Fig. 4.11b depict those obtained feasible points, where their union represents the
workspace WE estimation, which is illustrated by the gray zone with green contour. Furthermore,
for a complete 55 iterations it is expected to have WEd “ 655 « 6 ˆ 1042 points to explore.
However, we actually explored a total of WEd “ 16830 points, and a final total of WEs “ 4550
feasible points.
In the final scenario (See Fig. 4.12a), we consider the same tendons routing of the last scenario
in Section 3.2.5.5 from Chapter 3, but in this scenario, the three additional tendons are mounted
upward (i.e., on the positive level of the Oy axis). We choose a tension magnitude within
T1,...,9 “ r0, 50sN that will allow us to apply a couple magnitude within CT “ r0, 1.5sN ¨ m,
1,4,7
CT “ r0, 1sN ¨ m, and CT “ r0, 0.5sN ¨ m.
2,5,8 3,6,9
procedure presented in Section 4.3. This approach took 35 complete iterations and computed
a total of 13350 feasible points with a computational time of 4679 seconds. Those obtained
feasible points are depicted by red-colored points in Figs.4.12b, 4.12c and 4.12d, and their union
of represents the workspace WE estimation, which is depicted by the gray area with green
contour. Furthermore, we actually explored is a total of WEd “ 50490 points, and a final total
of WEs “ 13350 feasible points, instead of the expected WEd “ 635 « 1027 points to explore.
79
(c) Oyx plan view of WE . (d) Oyz plan view of WE .
4.3.6.6 Notes on scenarios results
The proposed forward-backward approach yields comparable workspace estimation precision

compared to the forward method, and at the same time reduces the complexity required to
estimate the workspace of soft robots.
In fact, this significant reduction introduced by the proposed forward-backward approach
was partially due to the elimination process of inappropriate configurations from Step. 2 (See
Fig. 4.3), but mainly due to the dismissal of all feasible configurations that have already been
explored during Step. 5 (See Fig. 4.6).
In Table 4.1, and Figures. 4.13a and 4.13b, we summarize the operations’ complexity and
computational time of each scenario.
80
Table 4.1: Operations Complexity and computational time: forward vs forward-backward.

Approach Forward- Forward-
Forward Forward
Inputs Backward Backward
nτ “ 2 2002 1820 66.35 17
nτ “ 3 1003 2275 2962 72
nτ “ 4 1004 3640 « 1.65 ˆ 105 187
6
nτ “ 6 100 4550 « 1.65 ˆ 109 2436
nτ “ 9 1009 13350 « 1.65 ˆ 1012 4679
(a) Operations. (b) Time
Figure 4.13: Operations complexity and computational time of the forward and forward-backward
approaches for the workspace estimation of the investigated scenarios using the PCS model.
4.4 Forward-Backward Approach for the FEM case

4.4.1 Implementation of the Methodology
The implementation of the forward-backward approach for the FEM model follows the same
steps described for the PCS model (which was detailed in Section 4.3, and illustrated by Fig.
4.2).
However, Step 1 and Step 3 of the forward-backward needs to be adapted for the FEM case.
For the first step of the this approach, we firstly need to linearly approximate the workspace
defined by (2.53) with the knowledge of the position vector of the end-effector upj´1q , the
displacement vector q pj´1q , and the corresponding actuators vector τ pj´1q P T pj´1q , as follows:
Kpq pj´1q qδq pjq “ Hpq pj´1q qδτ pj´1q ` F pq pj´1q q
(4.8)
upjq “ Cδq pjq ` upj´1q
with δq pjq “ q ´ q pj´1q and δupj´1q “ u ´ upj´1q .
Next, suppose that the actuators are well installed in a manner that the investigated soft robot
81
is always controllable [134], then for any given value of q pj´1q the matrix Kpq pj´1q q is always
invertible. Let us introduce the matrix A as:
Apq pj´1q q “ CK ´1 pq pj´1q qHpq pj´1q q
and define the matrix A` as before by
A` “ maxt0, Au
and A´ “ A ´ A`
With the above necessary elements, for the related displacement δq pjq , the over-estimated
workspace WEpjq corresponding to the actuators bound T pj´1q can be calculated via the following
theorem:
Theorem 3. Given a soft robot’s controllable configuration, with the actuators bound T pj´1q
defined as
pj´1q
T pj´1q “ rδτ pj´1q , δτ s
pj´1q
with δτ pj´1q and δτ represents the lower and upper bounds of the actuators vector with
respect to the prior configuration respectively. Then the over-estimated end-effector workspace
WEpjq is defined as ! ” ı)
WEpjq “ upjq P Rnu |upjq P upjq , upjq (4.9)
with
pj´1q
upjq “ A` δτ pj´1q ` A´ δτ ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q
and
pj´1q
upjq “ A` δτ ` A´ δτ pj´1q ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q .
Proof. Using the linear approximation defined in (4.8), the nodal displacement vector can be
directly deduced as:
“ ‰
δq pjq “ K ´1 pq pj´1q q Hpq pj´1q qδτ pj´1q ` F pq pj´1q q
With the defined matrix A, we can then use the above equation to write the end-effector
displacement in function of the matrix A, which yields:
upjq “ Cδq pjq ` upj´1q“ ‰
“ CK ´1 pq pj´1q q Hpq pj´1q qδτ pj´1q ` F pq pj´1q q ` upj´1q
“ CK ´1 pq pj´1q qHpq pj´1q qδτ pj´1q ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q
“ Apq pj´1q qδτ pj´1q ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q
Next, according to Lemma 3 in [135], the following over-estimation is obtained:
pj´1q pj´1q
A` δτ pj´1q ` A´ δτ ď Aδτ ď A` δτ ` A´ δτ pj´1q
By simply defining
pj´1q
upjq “ A` δτ pj´1q ` A´ δτ ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q
and
pj´1q
upjq “ A` δτ ` A´ δτ pj´1q ` CK ´1 pq pj´1q qF pq pj´1q q ` upj´1q
that leads to:
upjq ď upjq ď upjq
Finally, we obtain the over-estimated workspace WEpjq defined by (4.9).
82
The third step of the forward-backward approach needs also to be adapted in order to treat
the FEM case. Therefore, as explained in Step 3 of Section 4.3.3 for the PCS model, we follow
the same procedure for the FEM model, and we determine the associated displacement vector
of the end-effector as:
q̂Epjq “ ûpjq ´ upj´1q
Consequently, we can calculate the equivalent actuators and displacement vector by solving the
optimization problem formulated below:
rτ pjq , q pjq s “ arg min }qE ´ q̂Epjq }22

τ,q
s.t. τ P T (4.10)
qE “ Cq
Solving the backward mechanism (4.10) yields the feasible actuators vector τ pjq with its related
feasible displacement vector q pjq , which will eventually drive the soft robot’s end-effector to
reach ûpjq .
Finally, we deduce the end-effector position associated with the optimal displacement vector
q pjq as
upjq “ Cq pjq ` upj´1q .
The algorithm of the forward-backward approach for the FEM case consists of the same
main steps that was described for the PCS model in Table 3. However, the functions of Steps 1
and 3 need to be adapted for the FEM case as explained above.

This section presents the comparative results between the forward and forward-backward
approaches to the same soft robots’ configurations presented in the scenarios of Section 3.3.2 in
Chapter 3.
In the following scenarios, the workspace obtained from the forward approach is depicted by
blue-colored points. We propose to discretize the end-effector space with a discretization step
size of sEd “ 0.005 Meter (unit).
4.4.2.1 Scenario 1: trunk-like soft robot actuated by two cables.

In the first scenario (as shown by Fig. 4.14a), we consider the same configuration of the
first scenario in Section 3.3.2.1 from Chapter 3. The applied force is of the following range:
T p0q “ T “ r0 100s ˆ r0 100s.
For this scenario, the forward-backward approach took 37 complete iterations and found a
total of 635 feasible points in the workspace with a time computation complexity of 757 seconds.
Those obtained feasible points are depicted by red colored points in Figs. 4.14b, 4.14c and
4.14d), and the union of all small neighborhood of those feasible points gives the estimation of
workspace WE , which is represented by the gray zone with green contour. It is clear to see from
the three-view drawing that the workspace estimation result obtained via forward-backward
method coincides with that obtained via forward approach.
Moreover, for a complete 37 iterations it is expected to have WEd “ 637 “ 6.18 ˆ 1028 points
to explore. However, what we actually explored is a total of WEd “ 1680 points, and a final total
of WEs “ 635 feasible points. This significant reduction was partially due to the elimination
83
process of inappropriate configurations from Step. 2 (See Fig. 4.3), but mainly due to the
elimination of all feasible configurations that were already explored during Step. 5 (See Fig.
4.6).
Figure 4.14: Scenario 1 - WE estimation via forward approach (blue points) and forward-
backward (gray area) approach.
4.4.2.2 Scenario 2: trunk-like soft robot actuated by three cables.

In this scenario (as shown by Fig. 4.15a), we consider the same configuration of the second
scenario in Section 3.3.2.2 from Chapter 3. The applied force is of the following range: T p0q “
T “ r0 100s ˆ r0 100s ˆ r0 150s.
Next, we apply the forward-backward interval analysis approach using the equations presented
in Section 4.4. For this scenario, the forward-backward approach took 44 complete iterations and
calculated a total of 1941 feasible points in the workspace with a time computation complexity
of 2063 seconds. Those achieved feasible points are illustrated with red colored points in Figs.
4.15b, 4.15c and 4.15d), and the union of all small neighborhood of those feasible points yield
the estimated workspace WE , which is represented by the gray zone within the green contour.
Using the three-view illustration, we can clearly observe that the estimated workspace calculated
via the proposed forward-backward approach conforms with that obtained through the forward
approach.
Furthermore, for a complete 44 iterations it is expected to get WEd “ 644 “ 1.17 ˆ 1034
points to investigate. Yet, the total of points that we actually explored is WEd “ 5466, with a
final total of WEs “ 1941 feasible points. This substantial reduction was slightly due to Step. 2
(refer to Fig. 4.3) elimination process of irrelevant configurations, but considerably due to the
exclusion of already explored feasible configurations throughout Step. 5 (See Fig. 4.6).
84
4.4.2.3 Scenario 3: trunk-like soft robot actuated by four cables.
In this scenario (as shown by Fig. 4.16a), we consider the same configuration of the last scenario
in Section 3.3.2.3 from Chapter 3. The applied force is of the following range: T p0q “ T “
r0 100s ˆ r0 100s ˆ r0 150s ˆ r0 150s.
procedure presented in Section 4.4. For this scenario, the forward-backward approach took 42
complete iterations and computed a total 3667 feasible points in the workspace with a time
computation complexity of 3733 seconds. Those obtained feasible points are depicted by red
colored points in Figs. 4.16b, 4.16c and 4.16d, and the union of all small neighborhood of those
feasible points gives the estimation of the workspace WE , which is represented by the gray zone
with green contour. It is clear to see from the three-view drawing that the workspace estimation
result obtained via forward-backward method coincides with that obtained via forward approach.
In addition, for a complete 42 iterations it is expected to have WEd “ 642 “ 4.81 ˆ 1032
points to explore. However, what we actually explored is a total of WEd “ 10626 points, and a
final total of WEs “ 3667 feasible points. This significant reduction was partially due to the
elimination process of inappropriate configurations from Step. 2 (See Fig. 4.3), but mainly due
to the elimination of all feasible configurations that were already explored during Step. 5 (See
Fig. 4.6).
85
In the following table (see Table 4.2) and illustrations (see Fig. 4.17a and Fig. 4.17b), we
summarize the operations and time computation complexity of each scenario.
Table 4.2: Operations and time computational complexity: forward vs forward-backward.

Approach Forward - Forward -
Forward Forward
Inputs Backward Backward
nτ “ 2 1002 635 1463 757
nτ “ 3 1003 1941 « 1.4 ˆ 105 2063
nτ “ 4 1004 3667 « 1.4 ˆ 107 3733
86
Figure 4.17: Operations complexity and computational time of the forward and forward-backward
approaches for the workspace estimation of the investigated scenarios using the FEM model.
4.5 Conclusion
In this chapter, we have proposed an interval analysis approach, named the forward-backward
approach after its procedure, which consists of exploring the whole feasible configurations of the
workspace by discretizing the end-effector’s domain.
The proposed approach was implemented for both the PCS and the FEM models, where we
have shown its effectiveness in reducing the computation complexity and computational time
necessary to estimate the workspace of soft robots in contrast to the forward approach which
explodes exponentially when increasing the dimension of the actuators.
On the other hand, since this approach consists of determining all the possible configurations
that are feasible, it is useful in identifying interior and exterior information about the workspace.
However, this approach is exhaustive in the sense that it explores all feasible configurations of
the workspace instead of mapping only its interior and exterior boundaries.
87
88
Chapter 5
Continuation Approach for

Workspace Estimation: PCS case
5.1 Introduction
Due to the limitations encountered by the optimization approach in Chapter 3 that cannot
provide knowledge of the interior configurations and boundaries of the workspace of soft robots,
and also due to the limitations of the interval analysis based forward-backward approach in
Chapter 4 is exhaustive by exploring the whole workspace, it is logical to provide an alternative
methodology allowing us to determine only but all boundaries of soft robots workspaces, both
interior and exterior.
Such a method is called the continuation approach [8] that is based on the theory of bifurcation.
This method is founded on the mathematical theory of bifurcation that was originally introduced
by the French mathematician Henri Poincaré [91], and it consists of studying the solutions of
nonlinear differential equations, in addition to the study of possible variations in the topological
structure of a vector field family. Subsequently, Henri Poincaré classified various types of
bifurcation points [92]. Accordingly, this approach takes into account possible bifurcation
behavior in continuation calculation while mapping the interior and exterior boundaries of soft
robots workspaces.
Figure 5.1: Overall view of the continuation method.
The continuation method [55] applies the idea that the workspace boundary (noted as BWE )
of a soft robot can be extracted from the set of its output singularities. Starting from an initial
89
point on the boundary (i.e., point A as shown in Fig. 5.1), this method firstly classifies the
status of the boundary point, based on which it calculates the associated tangent vectors, and
then numerically predicts its evolution tendency which gives point B. Note that this predicted
point B might not be still on BWE , thus a correction procedure is applied to seek a point of
BWE by minimizing the distance to B which yields point C P BWE . Such a procedure is for the
purpose of mapping the point B to one-dimensional solution curves.
Briefly speaking, the proposed continuation approach can be summarized in the following
procedure:
1. Formulating the analytical criteria of the workspace boundary.
2. Finding an initial point on the workspace boundary.
3. Calculating tangent vectors to continuation curves at boundary points.
4. Mapping one-dimensional solution curves.
In the following, we will outline the realization of the continuation approach for the PCS
model and later discuss the possibility of its extension to the FEM model.
5.2 Implementation of the Methodology

In this section, we present the specific details of the continuation approach applied to the PCS
model.
5.2.1 Formulating the Analytical Criteria of the Workspace Boundary

Intuitively, a workspace boundary might be caused by the physical constraint (such as bounded
length) of the soft robots, or by the boundedness of actuators, i.e., τk P rτ k , τ k s where τ k and
τ k represent respectively the lower and upper bound of the kth actuator. In order to take into
account actuator boundedness, we introduce two intermediate variables s2k´1 P R and s2k P R
1 1
satisfying τ k ´ τk ` s22k´1 “ 0 (equivalent to τk ě τ k ), and ´τ k ` τk ` s22k “ 0 (equivalent to
2 2
τk ď τ k ). The purpose of the introduction of the variable s P R2nτ is to treat the boundedness of
the actuators in the same way as the mathematical model (e.g., (2.48)) via algebraic equations.
Following the above thoughts, in order to formulate the analytical criteria of the workspace
boundary by taking into account the saturation of actuators, let us introduce the following
generalized coordinates vector:
“ ‰T “ ‰T
Q “ uT , τ T , sT ” uT , Z T (5.1)
nτ
where the actuators
“ Tvector
‰ τ P Rn and the intermediate variable vector s P Rns are combined
T T nQ
in the vector Z “ τ , s P R Z, Q P R with nQ “ nu ` nZ , nZ “ nτ ` ns , and ns “ 2nτ .
With the introduced intermediate variable vector s, the bounded inequality constraints of
each actuator can be then transformed into equality constraints, and this enables us to achieve
a generalized formulation of the studied soft robot with the following mechanical constraints
90
ΨpQq: ` ˘
u ´ E1 g qpτ q ET2
» fi
1
— τ 1 ´ τ1 ` s21 ffi
— ffi
— 2 ffi
— ´τ 1 ` τ1 ` 1 s22 ffi
— ffi
— 2 ffi
—
— .. ffi
ffi
— . ffi
— 1 2 ffi
— τ k ´ τk ` s2k´1 ffi
— ffi
ΨpQq “ Ψpu, Zq “ — 2 ffi “ 0 (5.2)
1
— ´τ k ` τk ` s22k ffi
— ffi
— 2 ffi
— .. ffi
.
— ffi
— ffi
—τ ´ τ ` 1 s2
— ffi
ffi
— nτ nτ ns ´1 ffi
— 2 ffi
— ffi
– 1 2 fl
´τ nτ ` τnτ ` sns
2
where Ψpu, Zq P Rm is the generalized mechanical constraints, with m “ nu ` ns , and the
generalized strain vector q corresponding to the actuators vector τ is calculated via (2.51).
Remark 6. Note that for a given actuator configuration τ “ rτ1 , ¨ ¨ ¨ , τnτ sT with τk “ rτ k , τ k s
for 1 ď k ď nτ , its associated intermediate variables s2k´1 and s2k can be uniquely determined
by: a a
s2k´1 “ 2|τk ´ τ k |, s2k “ 2|τ k ´ τk | (5.3)
“ T T ‰T
i.e., Z “ τ , s can be uniquely determined by τ . Furthermore, by using numerical approach
to solve Ψpu, Zq “ 0, like Newton method, we can then obtain the associated value of u.
Considering the above formulation of the generalized mechanical constraints (5.2), the
generalized form of a soft robot WE can therefore be explicitly reformulated as follows:
WE “ t u P Rnu | Ψpu, Zq “ 0, for some Z u (5.4)
Now, consider a soft robot configuration in which its end-effector position lies on the
workspace boundary of (5.4). Physically, this situation is equivalent to a singular behavior
in which the soft robot end-effector cannot move in certain directions. Mathematically, we
can deduce that the soft robot’s workspace boundary is a subset of its workspace at which
the sub-Jacobian matrix of the generalized mechanical constraints Ψpu, Zq with respect to the
actuators vector τ and the intermediate variable s, denoted as ΨZ pu, Zq P RmˆnZ , is row-rank
deficient [56], and it can be defined by the following:
Definition 3. The workspace boundary BWE of a soft robot is a subset of its workspace WE , at
which the sub-Jacobian ΨZ pu, Zq P RmˆnZ is row-rank deficient, i.e.,
BWE “ t u P WE | rank ΨZ pu, Zq ă m, for Z P RnZ with Ψpu, Zq “ 0 u
Equivalently, the row-rank deficient of ΨZ pu, Zq implies that there exists an orthonormal
vector γ P Rm forming the nullspace of ΨTZ , i.e., ΨTZ pu, Zqγ “ 0. Thus, an analytical condition
[62] for a soft robot’s workspace boundary can be reformulated as:
BWE “ t u P WE | ΨTZ pu, Zqγ “ 0, γ T γ “ 1, Ψpu, Zq “ 0u (5.5)
91
In accordance with (5.5), a composite vector is introduced:
“ ‰T “ ‰T
Y “ uT , Z T , γ T ” QT , γ T (5.6)
where Y P RnY , nY “ nu ` nZ ` m “ nQ ` m.
Since boundary points must be characterized by the generalized vector Q, that satisfy (5.5),
then, stated in terms of the composite vector Y , points on BWE can be characterized by the
following composite set of equations:
» fi
Ψpu, Zq
— T
GpY q “ – ΨZ pu, Zqγ fl “ 0 (5.7)
ffi
1 T
pγ γ ´ 1q
2
Clearly, the above equation is a system of m ` nZ ` 1 equations with nu ` nZ ` m variables.
Thus the dimension of its solutions, equivalently the dimension of BWE , is equal to nu ´ 1.
Accordingly, when the end-effector position is two-dimensional u P R2 , then BWE is a
one-dimensional curve. On the other hand, in the case where soft robots are used in spatial ap-
plications, the output vector is three-dimensional u P R3 , implying that BWE is two-dimensional.
Remark 7. In spatial applications, this thesis proposes an approach to calculating sets that are
of dimension two by mapping the achievable one-dimensional BWE curves on the Oxy, Oyz,
and Oxz plans of the workspace boundary solution set (5.7).
Moreover, a practical approach [62] to map two-dimensional workspace boundary sets has
been proposed, where the purpose is to define a linear relation within the output coordinates by
introducing a plan intersecting with the workspace boundary in a one-dimensional curve.
5.2.2 Finding an Initial Point on the Workspace Boundary

In order to map the workspace boundary, we need first to figure out how to determine an initial
boundary point ub0 . For concrete applications, it is a nontrivial task to find a point on the
workspace boundary BWE (as shown in Fig. 5.2).
Figure 5.2: A boundary point ub0 .
Assume that a radiating point u0 is selected, and that u0 is exterior to WE . It is now

proposed that, consistent with Definition 3, a boundary point ub0 in a particular direction
emanating from the radiating point u0 (as depicted by Fig. 5.2), can be determined by solving
92
the following constrained optimization problem:
τ b0 “ min kupqq ´ u0 k22

τ
s.t. τ P T
(5.8)
` ˘
upqq “ E1 g qpτ q ET2
where τ b0 is the optimal and feasible actuators vector (since it satisfies the bounded constraint)
for the purpose of minimizing the nonlinear cost function f pqq “ kupqq ´ u0 k22 , which represents
the distance between u0 and upqq.
From the solution τ b0 , and according to Proposition 1, we calculate the corresponding
generalized strain vector q b0 (2.51). Then, we compute the strain twist ξîb0 related to the strain
b
vector qi j of each body i using (2.49b), and based on which we derive the configuration matrix
gpq q associated to the generalized strain vector q b0 from (2.49a). Finally, we deduce the
b0
boundary point position ub0 using (2.48b).

Using the values of ub0 , τ b0 , and sb0 , we establish the corresponding generalized vector
b0
Q using (5.2). Since the obtained solution verifies the workspace boundary definition (see
Definition 3), what remains now is to find a vector γ b0 that satisfies (5.5), which can be done by
calculating the nullspace of ΨTZ pQb0 q.
Finally, the obtained vector γ b0 with the generalized vector Qb0 yields the desired initial
composite vector Y b0 (5.6) from which the algorithm that consists of mapping the workspace
boundary solution curves (5.7) will start.
Regarding the implementation of the methodology, we need to calculate the sub-Jacobian
ΨZ pu, Zq, which can be formulated as follows:
„ “ ‰ 
´∇τ upqq 0nu ˆns
ΨZ pu, Zq “ (5.9)
∆˘1 ∆s
“ ‰
with ∇τ upqq “ ∇τ pqqJ T pqqr03 RpqqsT being already established in Theorem 1, and ∇τ pqq “
“ ‰T
H T pqq B´1 pτ, qq being defined in Assumption 1, with Bpτ, q, L, αq “ KpL, αq´Hq pq, L, αqpIn b
τ q ´ Fq pq, L, αq.
The matrices ∆˘1 and ∆s are equal to the following:
» fi
´1 ¨¨¨ 0 » fi
—`1
— ¨¨¨ 0 ffi
ffi s1
∆˘1 “ — ... .. ..
ffi , ∆s “ – .. (5.10)
— ffi — ffi
— . . ffi . fl
–0 ¨¨¨ ´1fl sns
0 ¨¨¨ `1
The initial radiating point u0 in 5.2, may be generated using the same approach explained
in Section 3.2.2 of Chapter 3.
Alternatively, we could also directly select a limit point, i.e., a point for which the related
actuators vector is equal to a combination set of upper and lower values of each actuator, as an
initial boundary point. Such a point is always located on BWE , according to Definition 3, since
its associated sub-Jacobian is row-rank deficient due to the actuators saturation (because one or
multiple variables s will be zero when some actuators reach their upper or lower bounds, and
thus the matrix ∆s in (5.10) will cause the sub-Jacobian ΨZ to be row-rank deficient).
93
5.2.3 Calculating Tangent Vectors to Continuation Curves at Bound-
ary Points
As pointed out by Fig. 5.1, mapping BWE consists of finding tangent vectors hpY q to solution
curves at each boundary point met during the continuation process. But before tangent
calculation, we need to identify the status of each boundary point met along the mapping of the
workspace boundaries.
Actually, there exist two different types of boundary points [55], a regular boundary point,
and a bifurcation point. A boundary point can be classified as regular [55] if the Jacobian matrix
of the composite set that characterizes candidate points on the workspace boundary GpY q
defined in (5.7), denoted as GY pY q P RnY ´1ˆnY , has full row-rank, and ΨZ pu, Zq is row-rank
deficient 1.
On the other hand, a bifurcation point is where at least two solution curves can cross, and it
can be classified into two different categories, simple and multiple [55]. A bifurcation point may
be classified as simple [124] if both GY pY q and ΨZ pu, Zq are row-rank deficient 1. Finally, a
multiple bifurcation point [93] can be identified by two conditions, the first is that GY pY q is
row-rank deficient of degree 1, and the second is that ΨZ pu, Zq is row-rank deficient of degree 2.
These possible cases are summarized in Table 5.1.
Row-Rank
ΨZ pu, Zq GY pY q
Boundary point
Regular m´1 nY ´ 1
Simple m´1 nY ´ 2
Bifurcation
Multiple m´2 nY ´ 2
Table 5.1: Classification of the possible categories of boundary points.
In the following, we will explain in detail how, in each case, tangent vectors to continuation
curves may be calculated.
5.2.3.1 Regular boundary points

Within the solution curve (5.7), a regular point ubr characterized by the composite vector Y br
(as shown by Fig. 5.3), where at most a single solution curve is passing through, may be met.
At such a point, the row-rank deficiency of ΨZ is 1, and GY pY br q has full row-rank.
Figure 5.3: Regular boundary point - tangent vector.
At such a point, a tangent vector hpY br q (as illustrated by Fig. 5.3) to the set (5.7) may be
calculated based on the following theorem [6, 106]:
Theorem 4. Let G: S Ă RnY ÞÑ RnY ´1 , be of class C r , r ě 1, on an open set S Ă RnY . Thus,

at a point Y br , if GY has full row-rank, then a unit tangent vector hpY br q P RnY is uniquely
94
defined by:
GY pY br qhpY br q “ 0
hT pY br qhpY br q “ 1
(5.11)
GY pY br q
ą0
hT pY br q
The tangent vector hpY br q determined by (5.11) can be calculated at each point along the
workspace boundary solution set (5.7) as long as GY pY br q has full row-rank.
However, difficulties may arise when multiple solution curves cross at bifurcation points
of (5.7). According to Table 5.1, GY is row-rank deficient at such configurations, and (5.11)
is unable to determine a unique tangent vector. Therefore, new techniques must be adopted
[93, 124] in order to compute tangent vectors hpY q to continuation curves at both simple and
multiple bifurcation points.
5.2.3.2 Bifurcation points - Simple bifurcation point

While proceeding along the workspace boundary solution set defined by (5.7), a point ubs
characterized by its composite vector Y bs (as depicted by Fig. 5.4) may be found at which the
Jacobian GY pY bs q becomes row-rank deficient of degree 1, and the rank deficiency of ΨZ is 1.
Such a point is classified as a simple bifurcation point [124], where a pair of solution curves (See
Fig. 5.4) characterized by a pair of tangent vectors hpY bs q satisfying (5.11) are traversing it.
For the purpose of identifying the nature of a singular behavior at such a point Y bs , we
perform an elementary row decomposition, e.g. singular value decomposition (SVD) or QR
decomposition, to simplify GY pY bs q into the following form:
„ 
F pY bs q
EGY pY bs q “ Y (5.12)
0
where E is an elementary non-singular matrix, and FY pY bs q is an pnY ´ 2q ˆ nY matrix with

full row-rank. Applying elementary row operation to the workspace boundary solution set (5.7)
using E, we get the following equivalent equation:
„ bs

q bs q ” FpY b q “ EGpY bs q “ 0
GpY (5.13)
fpY s q
Using (5.12) and (5.13), we observe that FY pY bs q is the Jacobian matrix of FpY bs q, and the
Jacobian of the function fpY bs q is fY pY bs q “ 0.
Since FY pY bs q is full row-rank, i.e., rank FY pY bs q “ nY ´ 2, then the nullspace of FY pY bs q
(equivalently, the nullspace of GY pY bs q) is of dimension 2, which generates two different non-zero
solutions of the following equation:
FY pY bs qζ i “ 0, i “ 1, 2 (5.14)
that can be orthonormalized; ζ i ζ j “ δij , i, j “ 1, 2, where δ is the Dirac delta function.

The purpose now is to determine the pair ř of tangent vectors hpY bs q to continuation curves
at Y by writing it in the form of hpY q “ j αj ζ j . With ζ being calculated via (5.14), the
bs bs
remaining question now is about how to compute the vector α.

Using Taylor’s expansion, the tangent vectors of the solution curve can be calculated. Taylor’s
expansion for functions of nY variables is stated in [5] as follows:
95
Theorem 5. According to [5], let G: S Ă RnY ÞÑ RnY ´1 , nY ě 2, be of class C r , r ě 1, on
the open set S Ă RnY . Then, the quadratic Taylor approximation is defined by:
FY pY bs qαi ζ i
« ff
bs bs
GpY ` ∆Y q « GpY q ` 1
q q (5.15)
pαj ζ j qfY Y pY bs qpαk ζ l q
2
where the Hessian matrix of fpY q, evaluated at Y bs , is the matrix:
« ff
BfpY q
fY Y pY bs q ” (5.16)
BYi Yj b Y “Y s
Since Y bs ` ∆Y and Y bs must satisfy the boundary workspace solution set (5.7), then we
have:
q bs ` ∆Y q “ GpY
GpY q bs q “ 0
Taking that into consideration, (5.15) becomes:
FY pY bs qαi ζ i
« ff
1 “0 (5.17)
pαj ζ j qfY Y pY bs qpαk ζ k q
2
Noting that, for all αi , the upper sub-vector term FY pY bs qαi ζ i “ 0, since ζ i represents the
nullspace of FY pY bs q as stated in (5.14). Finally, (5.17) is reduced to the following:
“ ‰
αT ζ i fY Y pY bs qζ j α “ 0 (5.18)
where α “ rα1 , α2 sT .
Using elementary row decomposition, e.g. eigenvalue decomposition, the above quadratic
equation may be solved. Another way is to transform (5.18) to a quadratic equation in terms of
α1 and α2 and solving it using the quadratic formula. No matter which way is applied, solving
(5.18) yields two normalized solution vectors α1 and α2 . Finally, with ζ 1 and ζ 2 being calculated
via (5.14), we obtain a pair of tangent vectors hpY bs q “ rh1 pY bs qT h2 pY bs qT sT depicted by Fig.
5.4, and formulated by the following:
h1 pY bs q “ α11 ζ 1 ` α21 ζ 2
(5.19)
h2 pY bs q “ α12 ζ 1 ` α22 ζ 2
Figure 5.4: Simple bifurcation point - tangent vector.
96
5.2.3.3 Bifurcation points - Multiple bifurcation point
During the continuation calculation, a point ubm characterized by the composite vector Y bm (as
depicted by Fig. 5.5) may be encountered at which the Jacobian ΨZ becomes row-rank deficient
of degree 2, and the rank deficiency of GY pY bm q is 1. Such a point is classified as a multiple
bifurcation point [93], where numerous branches may occur (See Fig. 5.5) depending on the
design of the soft robot.
This thesis uses a numerical method [93] that consists of finding, for all branches intersecting
at Qbm , all vectors γ such that GY pQbm , γq for each branch is row-rank deficient.
The adopted strategy can be separated into two major steps:
[Step 1] Calculating all solutions for γ at a multiple bifurcation point.
[Step 2] Calculating all tangent vectors hpQbm , γq to continuation curves.
Figure 5.5: Multiple bifurcation point - tangent vector.
Step 1: Calculating all solutions for γ at a multiple bifurcation point

In the first step, vector γ for which GY pQbm , γq is row-rank deficient will be determined.
Since the row-rank deficiency of ΨZ is 2, thus, we have 2 orthonormal solutions γ1 and γ2 such
that:
ΨTZ γi “ 0
(5.20)
γiT γj “ δij , i, j “ 1, 2
where γ can be represented as a linear combination of γ1 and γ2 as:
γ ” γpβq “ tanpβqγ1 ` γ2 (5.21)
with β P r0 πs. Hence, the purpose of this step is to solve the following problem:
Problem 1. Calculate all γ for which each corresponding GY pQbm , γq is row-rank deficient,
where: » fi
Ψu ” Ψτ ı ” Ψs ı 0
GY pQbm , γq “ – 0
— ΨTZ γ ΨTZ γ ΨTZ ffi
fl (5.22)
τ s
T
0 0 0 γ
97
Since the number of γ is finite, it can be assumed that neither γ1 nor γ2 are solutions.
Therefore, the solution of (5.21) must verify the condition sinpβq cospβq ‰ 0.
Lemma 1. GY pQbm , γq is row-rank deficient if and only if there exists a non-zero vector
c “ rcT1 , cT2 , c3 sT , where c1 P Rm , c2 P RnZ , and c1 P R:
cT GY pQbm , γq “ 0 (5.23)
By developing (5.23) through (5.22), we obtain the following set of equations:

” ı
cT1 Ψτ ` cT2 ΨTZ γ “ 0 (5.24a)
” ıτ
T T T
c1 Ψs ` c2 ΨZ γ “ 0 (5.24b)
s
cT2 ΨTZ ` c3 γ T “ 0 (5.24c)
Introduce now µ as: µ ” µpβq “ tanpβq, then, the problem may be equivalently reformulated as
follows:
Problem 2. Assume that neither GY pQbm , γ1 q nor GY pQbm , γ2 q is row-rank deficient, the goal
now would be to find all µpβq such that GY pQbm , γpβqq is row-rank deficient.
A necessary condition for GY pQbm , γpβqq has already been outlined by the set of equations
(5.24a), (5.24b), and (5.24c). Multiplying (5.24c) by γ, we obtain:
cT2 ΨTZ γ ` c3 γ T γ “ 0
Using (5.20), we deduce:

ΨTZ γ “ 0 ùñ c3 “ 0
Taking that into consideration, (5.24c) finally becomes:
ΨZ c2 “ 0
And since the row-rank deficiency of ΨZ : RnZ ÞÑ Rm is 2, then there exists an orthonormal
basis σ of the nullspace of ΨTZ such that:
c2 “ d1 σ1 ` d2 σ2 ` ¨ ¨ ¨ ` dnσ σnσ ” σd (5.25)
where σ “ tσ1 , σ2 , ¨ ¨ ¨ , σnσ u, d “ td1 , d2 , ¨ ¨ ¨ , dnσ u, and nσ “ nZ ´ m ` 2.

From (5.24a), and (5.24b), we obtain the following:
« ff
” ı ” ı ” ı
T T T T
c1 Ψτ , Ψs “ ć2 ΨZ γ , ΨZ γ
τ s
“ ‰
For the purpose of formulating c1 in function of c2 , and since Ψτ , Ψs is right invertible, we
can write the above equation in the following form:
cT1 “ cT2 Mpγq (5.26)
with: « ff
” ı ” ı ” ı`
T T
Mpγq “ ´ ΨZ γ , ΨZ γ Ψτ , Ψs (5.27)
τ s
98
After, by substituting cT1 from (5.26) into (5.24b), we get:
˜ ¸
” ı
T T
c2 MpγqΨs ` ΨZ γ “0
s
Next, we replace (5.25) into the above equation, we obtain:
dT Hpγpβqq “ 0
with ˜ ¸
” ı
Hpγpβqq ” σ T MpγqΨs ` ΨTZ γ (5.28)
s
As stated in Lemma 2.3 of [93], GY pQbm , γq is row-rank deficient if and only if Hpγpβqq is
row-rank deficient. By substituting (5.21) into (5.28), we achieve:
Hpγpβqq “ tanpβqA ` B (5.29)
where:
˜ ¸
” ı
A “ σT Mpγ1 qΨs ` ΨTZ γ1 (5.30a)
s
˜ ¸
” ı
T
B“σ Mpγ2 qΨs ` ΨTZ γ2 (5.30b)
s
Based on Lemma 2.4 in [93], if both GY pQbm , γ1 q and GY pQbm , γ2 q have full row-rank, then
A is right invertible. Equivalently, according to Theorem 2.5 in [93], GY pQbm , γq is row-rank
deficient if only and if µ is a real eigenvalue of ´BA` .
Step 2: Calculating all tangent vectors hpQbm , γq to continuation curves.
In this step, we will describe how to calculate the tangent vectors hpQbm , γq to all solution
curves passing through a multiple bifurcation point.
Consistent with Lemma 2.4 in [93], denote χ as the eigenvector corresponding to the eigenvalue
µ of the matrix ´BA` . By definition, we obtain:
´BA` χ “ µχ (5.31)
Let us define ρ2 P Rns as the following [93]:
ρ2 “ A` χ (5.32)
From the above equation, we can easily deduce that χ “ Aρ2 . Using this, and substituting
(5.32) in (5.31), we obtain:
´Bρ2 “ µAρ2
which can be reformulated as:
pµA ` Bqρ2 “ 0
Since µ “ tanpβq, we get:
ptanpβqA ` Bqρ2 “ 0
Using (5.29), then (5.28), the above equation can finally be rewritten as follows:
σT S “ 0 (5.33)
99
˜ ¸
” ı
with S “ MpγqΨs ` ΨTZ γ ρ2 .
s
Since σ is an orthonormal basis of the nullspace of ΨTZ , and from (5.33) S is orthogonal to σ,
then S represents the column space of ΨTZ [86]. Therefore, there exists a unique vector ρ3 P Rm
in the column space of ΨTZ , for which the following is deduced:
S “ ´ΨTZ ρ3 (5.34)
Given that ρ3 and γ are respectively the column space and the nullspace of ΨTZ . Therefore,
using the first formula of S from (5.33), its second formula from (5.34), we obtain the following
set of equations:
˜ ¸
” ı
T
MpγqΨs ` ΨZ γ ρ2 “ ´ΨTZ ρ3 (5.35a)
s
0 “ γ T ρ3 (5.35b)
which finally allows us to obtain the final form of ρ3 P Rm as follows:

` « ” ı ff
MpγqΨs ` ΨTZ γ
„
´ΨTZ
ρ3 “ ρ2 (5.36)
γT 0
s
Next, we define ρ1 P Rnu `nτ by the following:

” ı`
ρ1 “ ´ Ψu , Ψτ Ψs ρ2 (5.37)
Substituting both Mpγq and ρ1 respectively from (5.26) and (5.37) into (5.35a), and using
(5.35b), we can deduce the following system of equations:
” ı
Ψu , Ψτ ρ1 ` Ψs ρ2 “ 0 (5.38a)
« ff
” ı ” ı
0, ΨTZ γ ρ1 ` ΨTZ γ ρ2 ` ΨTZ ρ3 “ 0 (5.38b)
τ s
γ T ρ3 “ 0 (5.38c)
Eventually, the above set can be reformulated into the following form:
» ” ı fi
Ψu , Ψτ Ψs 0 » fi
—« ff ffi ρ1
— ” ı ” ı ffi
0“— T T T ffi –ρ2 fl
— 0, ΨZ γ ΨZ γ ΨZ ffi
(5.39)
– τ s fl ρ3
0 0 γT
0 “ GY pQbm , γq hpQbm , γq
Finally, it is clear that:

hpQbm , γq “ rρT1 , ρT2 , ρT3 sT (5.40)
is the tangent vector of the continuation curve corresponding to its associated eigenvalue µ and
eigenvector χ.
100
5.2.4 Mapping One-Dimensional Solution Curves
With tangent vectors being calculated for each possible boundary point situation (as shown
in Fig. 5.3, Fig. 5.4, and Fig. 5.5), the final step of the continuation method is to map the
solution curves of (5.7).
Given a boundary point ub0 characterized by its composite vector Y b0 , where Y b0 is a solution
of the workspace boundary set (5.7), i.e. GpY b0 q “ 0, and its tangent vector hpY b0 q being
calculated depending on its category (regular boundary point, simple or multiple bifurcation
point), a ray is emanated from Y b0 , with an infinitesimally small discretization precision ε, on
p
which a prediction Y (as depicted by Fig. 5.6) of the appropriate solution is calculated as
follows:
p
Y “ Y b0 ˘ εhpY b0 q (5.41)
p
Next, based on the predicted state Y , its associated solution Y b (as shown in Fig. 5.6) must
verify the nonlinear constraint condition (5.7) of the workspace boundary solution set.
Figure 5.6: Mapping one-dimensional solution curves.
In this thesis, we propose to calculate Y b (as depicted in Fig. 5.6) using a constrained
optimization problem, subject to nonlinear constraints, and formulated by the following:
p
Y b “ min kY ´ Y k
Y (5.42)
s.t. GpY q “ 0
where GpY q defined in (5.7) to characterize the workspace boundary set.

With Y b being computed, we proceed by repeating the same process for each boundary point
met during the continuation calculation until all solution curves of the workspace boundary are
mapped.
An equivalent technique to (5.42) presented in [6] has also been proven to be an efficient
approach to find the next possible solution on the continuation curve.
The stop condition can be evaluated by verifying whether all limit points have been explored.
When it is the case, then the stop condition is true, and the continuation process is terminated
by yielding the inner and outer boundaries of the soft robot. If not, there might exist some
isolated boundaries which cannot be reached from the chosen initial point. In this situation, we
can choose another combination set of upper and lower values of each actuator, which has not
yet been explored by the former process, as a new initial point and restart the same process.
In following, we present a brief algorithm (see Algorithm 4) explaining the major steps of
the proposed continuation approach in order to map the interior and exterior boundaries of soft
robots.
101
Require: u0 , T , ε
BWE Ð ∅, stop Ð 0 Ź Initialization.
rτ b0 , ub0 s Ð solvep5.8q pu0 , T q Ź Initial boundary point.
Qb0 Ð solve(5.1) pu0 , T q Ź Initial boundary configuration.
γ b0 Ð nullpΨTZ pQb0 qq
T T
Y b0 Ð rQb0 , γ b0 sT Ź Initial composite vector.
while !stop do
f lag Ð T able5.1 pY b0 q Ź Boundary point class.
switch f lag do Ź Tangent vector.
case (Regular boundary point)
hpY b0 q Ð solve(5.11) pQb0 , Y b0 q
case (Simple bifurcation point)
case (Multiple bifurcation point)
Y p Ð Y b0 ˘ εhpY b0 q Ź Predict.
Y b Ð solve(5.42) pY b0 , Y p q Ź Correct.
BWE Ð BWE ‘ Y b Ź Append.
Y b0 Ð Y b Ź Update initial composite vector.
stop Ð updatepBWE , Y b0 q Ź Update stop condition.
end while
5.3 Validation and Analysis

We implement both the forward and the continuation approaches on planar and spatial soft
robots to deduce the workspace estimation WE , and the workspace boundary BWE . Table 3.1
presents the numerical values of the investigated soft robot.
In the following scenarios, the workspace estimation obtained from the forward approach is
depicted by gray-colored points. For the continuation approach, we suggest to proceed with a
step ε “ 0.01.
In the first scenario, we consider the same configuration of the first scenario in Section 3.2.5.1 of
Chapter 3. We choose a tension magnitude within T1,2 “ r0, 200sN that will allow us to apply a
couple magnitude within CT “ r0, 6sN ¨ m and CT “ r0, 4sN ¨ m.
1 2
Starting from an initial boundary configuration calculated via (5.8), which is a regular
boundary point in this case (blue point on Fig. 5.7b), we use (5.11) to calculate its corresponding
tangent vectors, which yields one tangent vector. Based on this, we calculate its associated
prediction via (5.41), then deduce its corresponding solution using (5.42). Proceeding along the
solution curve (5.7), we meet four bifurcation points (green points on Fig. 5.7b), between which
regular boundary points (red points on Fig. 5.7b) are met. Finally, workspace boundary curves
are mapped as shown in Fig. 5.7b.
For this scenario, the continuation approach computed a total of 87 boundary points and
took only 9.54 seconds to map the workspace boundaries for this scenario.
102
(a) Soft robot structure. (b) Oxz plan view of WE and BWE .
Figure 5.7: Scenario 1 - WE estimation via forward approach (gray points) and workspace
boundaries BWE computed via the continuation approach (red curves and green bifurcation
points).
tendons (See Fig. 5.8a). We choose a tension magnitude within T1,2 “ r0, 150sN that will allow
us to apply a couple magnitude within CT “ r0, 4.5sN ¨ m and CT “ r0, 3sN ¨ m.
1,3 2,4
The numerical continuation algorithm starts from the blue point on Fig. 5.8b, which is a
limit configuration in this case. Next, using (5.40) we obtain 4 distinct tangent vectors, each
one corresponding to its associated solution curve since we have 4 that crosses through this
point. After, we calculate the prediction corresponding to each tangent vector (5.41), and we
obtain for each curve its associated solution (5.42). Regular boundary points (red points in Fig.
5.8b) are computed between each of the sixteen total bifurcation points, including the initial
point (green points in Fig. 5.8b) found for this particular soft robot structure. Finally, interior
and exterior boundaries to the planar two segments - four tendons soft robot’s workspace are
mapped and illustrated by Fig. 5.8b.
103
points).
In the third scenario (See Fig. 4.10a), we consider the same tendons routing of the third scenario
in Section 3.2.5.3 from Chapter 3. We choose a tension magnitude within T1,2,3 “ r0, 100sN
that will allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m, CT “ r0, 2sN ¨ m, and
1 2
CT “ r0, 1sN ¨ m.
1
Interior and exterior workspace boundaries of this particular soft robot structure are mapped
(as shown by Fig. 5.9b), using (5.8) we find an initial boundary configuration (blue point on Fig.
5.9b) that is a regular boundary point, based on which, we calculate its unique corresponding
tangent vector through (5.11). Next, through (5.41) we predict the possible solution, and we
determine its associated solution (5.42), proceeding all the way along solution curves we identify
8 bifurcation points (green points in Fig. 5.9b), with regular boundary points (magenta points
in Fig. 5.9b) lying between them.
104
Figure 5.9: Scenario 3: WE estimation via forward approach (gray points) and workspace
points).
In this scenario, three additional tendons are routed in a fashion where they are symmetric to the
tendons configuration in the third scenario (See Fig. 3.9a). We choose a tension magnitude within
T1,¨¨¨ ,6 “ r0, 100sN that will allow us to apply a couple magnitude within CT “ r0, 3sN ¨ m,
1,4
2,5 3,6
Solution curves of the workspace boundary set (5.7) are mapped beginning from an initial
boundary point corresponding to a limit point calculated via a random combination set of lower
and upper values of each actuators vector bounds and is located in the interior boundary. From
this initial point, relevant tangent vectors are calculated (5.40), yielding 6 possible directions to
explore. Next, we predict the solution corresponding to each boundary curve from (5.41), and
deduce their appropriate boundary configuration (5.42) (see Fig. 5.10b). Finally, solution sets
are calculated and exterior and interior workspace boundaries are mapped (as shown by Fig.
5.10b).
took only 1096 seconds to map the workspace boundaries for this scenario.
105
points).
In this scenario, we consider a soft robot composed of two segments and actuated by four
tendons (See Fig. 5.11a). The two lateral tendons are installed as in the first scenario. The
two upward tendons are installed on the Oxy plan, and parallel to each other. The first tendon
is fixed at the position p0, r1 {2, 0q and extends along the first segment length L1 . The second
tendon is fixed at the position p0, r2 {2, 0q and extends to the end-effector position. We choose
the bounded tension magnitude T1,2,3,4 “ r0, 150sN enabling us to apply a couple magnitude
CT “ r0, 4.5s, and CT “ r0, 3sN ¨ m.
1,3 2,4
Interior and exterior workspace boundaries of this particular structure are mapped starting
from an initial limit point. Next, using (5.40) we compute all possible tangent vectors for each
plan Oxy, Oyz and Oxz. Based on this, we calculate a prediction corresponding to each tangent
vector and we obtain each associated solution (5.42). The obtained result is depicted by red
curves in a three-view drawing (Oxz, Oyx, and Oyz plans, in Figs. 5.11b, 5.11c, and 5.11d
respectively).
took only 107 seconds to map the workspace boundaries for this scenario.
106
(c) Oyx plan view of WE and BWE . (d) Oyz plan view of WE and BWE .
points).
In this scenario, we consider a soft robot composed of two segments and actuated by six tendons
(See Fig. 5.12a). There are four lateral symmetric tendons (as in Fig. 5.8a), and two upward
tendons as in the above scenario. We choose the bounded tension magnitude T1,¨¨¨ ,6 “ r0, 100sN
enabling us to apply a couple magnitude CT “ r0, 3s, and CT “ r0, 2sN ¨ m.
1,2,3 4,5,6
The obtained result is depicted by red curves in a three-view drawing (Oxz, Oyx, and Oyz
plans, in Figs. 5.12b, 5.12c, and 5.12d respectively). For this scenario, the continuation approach
computed a total of 4107 boundary points and took only 980 seconds to map the workspace
boundaries for this scenario.
107
points).
In this scenario, we consider a soft robot composed of three segments and actuated by four
tendons (See Fig. 5.13a). The three lateral tendons are installed as in the third scenario (see Fig.
5.9a). The three upward tendons are installed on the Oxy plan, and are parallel to each other.
The first tendon is fixed at the position p0, r1 {2, 0q and extends along the first segment length
L1 . The second tendon is fixed at the position p0, r2 {2, 0q and extends to pL1 ` L2 , r2 {2, 0q, and
the third tendon is fixed at the position p0, r3 {2, 0q and extends all along to the end-effector
position. We choose a tension magnitude within T1,¨¨¨ ,6 “ r0, 100sN that will allow us to apply a
couple magnitude within CT “ r0, 3sN ¨ m, CT “ r0, 2sN ¨ m, and CT “ r0, 1sN ¨ m.
1,4 2,5 3,6
108
points).
In the final scenario (See Fig. 4.12a), we consider the same tendons routing of the last scenario
in Section 4.3.6.5 from Chapter 4. We choose a tension magnitude within T1,...,9 “ r0, 50sN that
will allow us to apply a couple magnitude within CT “ r0, 1.5sN ¨ m, CT “ r0, 1sN ¨ m,
1,4,7 2,5,8
and CT “ r0, 0.5sN ¨ m.
3,6,9
109
points).
In Table 5.2, and Figures 5.15a and 5.15b, we summarize the operations’ complexity and
computational time of each scenario.
Table 5.2: Operations Complexity and computational time: forward vs continuation.

Approach
Forward Continuation Forward Continuation
Inputs
nτ “ 2 (Sc1) 2002 87 66.35 9.54
nτ “ 3 (Sc3) 1003 327 2962.92 40.62
nτ “ 4 (Sc2, Sc5) 1004 p687, 688q « 1.65 ˆ 105 p96, 107q
nτ “ 6 (Sc4, Sc6, Sc7) 1006 p4126, 4107, 4122q « 1.65 ˆ 109 p1096, 980, 1392q
nτ “ 9 (Sc8) 1009 7621 « 1.65 ˆ 1012 2674
110
Figure 5.15: Operations complexity and computational time of the forward and continuation
approaches for the workspace and workspace boundaries estimation of the investigated scenarios
using the PCS model.
5.3.1 Discussion on Multi-Component Workspaces, Voids/Holes, and

Boundaries Characterization:
In this thesis, starting from one boundary point, the proposed continuation approach enables us
to map the exterior and interior workspace boundaries whether such boundaries result from
actuators saturation, or physical limitations such as the mechanism’s length, self-collision, or
elastic instabilities. However, as stated at the end of Section 5.2.4, isolated boundaries (see Fig.
5.16a) and voids (see Fig. 5.16b) might arise in complex soft robot configurations. In either
case, both BWEp1q and BWEp2q verify the workspace boundary Definition 3. Suppose that in the
first instance, the initial boundary point from which the continuation algorithm starts is on
BWEp1q as shown in Fig. 5.16. In this case, using this point as a starting configuration of the
continuation algorithm, we can surely map BWEp1q in the first process. Then, to map BWEp2q , we
need to verify if all limit points have been explored by the first process, and since we have a
second workspace boundary, it is evident that there exists at least one limit configuration that
was not explored by the former process. Then, using this configuration as a new initial point,
we can run a second continuation algorithm process allowing us to map BWEp2q .
111
(a) (b)
Figure 5.16: Multi-component workspaces (a); Holes/Voids (b).
Another important issue concerns the characterization of different types of boundaries:

exterior boundaries, interior boundaries, and voids. Firstly, we would like to emphasize that the
proposed continuation approach contains necessary information to calculate the normal vector of
each point on the obtained boundaries. In fact, for any boundary point Y b “ rQb , γ b s P BWE , its
normal vector can be calculated as [55]: η b ” ηpY b q “ ΨTu pQb qγ b . To explain the main idea, we
take Fig. 5.16 as an example. Consider, as depicted by Fig. 5.16b, a boundary point Y bE lying
on the exterior boundary BWEp1q , with its normal vector η bE , therefore in the normal direction
we can find a point (noted as upE ) which will not belong to the workspace. In summary, this
judgment can be formulated as the following condition:
DupE “ ubE ` εη bE , s.t. upE R WE
where ε P R is a pre-defined small value. Conversely, for a boundary point Y bI lying on the
interior curve, with its normal vector η bI as shown by Fig. 5.16b, an interior boundary may be
sorted by verifying:
@upI “ ubI ` εη bI , s.t. upI P WE
Finally, for Y bV lying on BWEp2q , with its normal vector η bV as illustrated by Fig. 5.16b, a void
may be distinguished by verifying the following condition:
DupV “ ubV ` εη bV , s.t. upV R WE
Note that the above condition is identical to that of the exterior boundary. However, the fact
that BWEp2q is enveloped by BWEp1q enables us to assert that BWEp2q is the void and BWEp1q is the
exterior boundary.
5.4 Conclusion
The continuation approach allows the mapping of interior and exterior boundaries of soft robots
with slender shapes using the PCS model.
However, it is important to state that this approach cannot provide information on all points
in the workspace, but only on boundary points. Moreover, this approach was only applied to
slender-shaped soft robots via the PCS model since this modeling method provides the analytical
form of all necessary parameters.
112
Thus, it is of interest to extend the applicability of this approach in order to also treat soft
robots with general shape which can be modeled by FEM. However, since the FEM model is
established differently from the PCS approach, some steps of the proposed continuation approach
will need to be adapted in order to consider the FEM case (which will be outlined in Chapter
7).
113
114
Chapter 6
Design Optimization: PCS case
6.1 Introduction
In Chapter 3, Chapter 4, and Chapter 5, we have proposed different approaches in order to

answer the direct problem of soft robots, i.e., to estimate the workspace of a given designed
configuration of a soft robot. On the other hand, this chapter treats the inverse problem, i.e.,
to optimize the design of a soft robot in order to guarantee that certain points belong to its
workspace. Solving such a problem is useful for robotic applications where we don’t have an
idea about the exact geometric parameters of the soft robot, but we do know what task is the
soft robot trying to achieve, such as Pick-and-Place, and trajectory planning.
The actual process of a soft robot’s design still follows intuition and the procedure of
trial-and-error. Considering specific performance objectives, the classical design approach of
soft robots is coupled with the uncertainty of achieving such objectives and also with the
substantial economical expenditures necessary for the trial-and-error endeavors. For the purpose
of accomplishing the desired objectives, it is therefore logical, for both economic and scientific
reasons, to optimize the design of soft robots in a virtual environment before proceeding to its
final physical conception.
This thesis follows the above thought and seeks to answer the following question: Given a
soft robot (illustrated by Fig. 6.1a), composed of a finite number of segments with bounded
length L P L, and driven by mounted actuators (i.e., tendons) with bounded actuators length
(i.e., tendons’ length) α P A, and bounded actuators magnitude (i.e., tendons’ tension) τ P T .
How should we optimize the design of a soft robot in order to reach the desired position (e.g.,
blue-colored points ud in Fig. 6.1a).
115
(a) Initial design. (b) Optimal Design.
Figure 6.1: Soft robot’s design.
In other words, we seek to determine the optimal length parameter L of each segment, and
the optimal actuators length parameter α such that the objective points (e.g., blue-colored
points ud in Figs. 6.1a) must be located inside the workspace of the optimally designed soft
robot, as depicted by Fig. 6.1b.
In the present work, we assume that the segments radius r and the distance of the actuators
to the mid-line d are both pre-determined and fixed parameters. However, the proposed method
can be generalized to treat the case where r and d are varying (which will be explained in
Chapter 7).
The reachable workspace of a soft robot has been defined in Definition 1 for the case that L
and α are constant parameters. However, since we now seek to achieve the optimal values of
L and α, the definition of the reachable workspace of a soft robot needs to be adapted, which
should be a function of L and α.
Definition 4. The workspace WE of the end-effector u of a soft robot modeled by Ψpτ, q, L, αq “ 0

(with q P Rn being the generalized strain vector), composed of bounded length of segments L P L,
and driven via bounded actuators τ P T , with bounded actuators length α P A, is a subspace of
R3 , defined by the following:
WE pL, αq “ tu “ ψpq, L, αq | Dpq P Rn , τ P T q, s.t. Ψpτ, q, L, αq “ 0u
where ψ represents the geometric model of the soft robot which depends not only on q, but
also on the parameters to be optimized L and α (more details will be discussed in 6.2.1), and
Ψpτ, q, L, αq “ KpL, αq q ´ Hpq, L, αqτ ´ F pq, L, αq.
To accomplish the vision depicted by Fig 6.1b, we propose an optimization approach, which
will be implemented on the PCS model and explained in the following.
6.2 Implementation of the Methodology

6.2.1 Influence of the Parameters to be Optimized on the PCS Static
Model
Before we step into the details of the proposed optimization approach, we need to clarify the
influence of the parameters to be optimized on the adopted mathematical model of soft robots.
In this work, we seek to optimize two parameters; the length Li of each segment i P 1 ¨ ¨ ¨ N , and
the actuators location, particularly the length parameter αk of each actuator k P 1 ¨ ¨ ¨ nτ .
116
To be consistent with Definition 4, the workspace of the studied soft robot is now governed
by:
KpL, αq q “ Hpq, L, αqτ ` F pq, L, αq (6.1a)

T
upq, L, αq “ E1 gpq, L, αq E2 (6.1b)
with
gpq, L, αq “ eα1 L1 ξ1 ep1´α1 qL1 ξ1 ¨ ¨ ¨ eαnτ LN ξN ep1´αnτ qLN ξN (6.2a)

p p p p
ξpi “ B
{ qi qi ; i P 1 ¨ ¨ ¨ N (6.2b)
The block-diagonal generalized stiffness matrix K, the generalized actuation matrix H, and the
vector of generalized position-dependent external forces F are now established as:
˜ «ż ż ff «ż ż ff ¸
KpL, αq “ diag BqT1 Σ1 dX, Σ1 dX Bq1 , ¨ ¨ ¨ , BqTN ΣN dX, ΣN dX BqN ,
X11 X12 1
XN 2
XN
(6.3a)
»˜ «ż ż ff¸T ˜ «ż ż ff¸T fiT
H pq, L, αq “ – BqT1 Hτ dX, Hτ dX , ¨ ¨ ¨ , BqTN Hτ dX, Hτ dX fl (6.3b)
X11 X12 1
XN 2
XN
«ż ff
N
ÿ ż
F pq, L, αq “ JiT F̄ei dX, JiT F̄ei dX (6.3c)
i“1 Xi1 Xi2
where Xi1 “ r0, αk Li s and Xi2 “ rαk Li , p1 ´ αk qLi s, for i P 1 ¨ ¨ ¨ N , and k P 1 ¨ ¨ ¨ nτ .

The configuration matrix gi of a body i and the strain twist ξpi of a body i can now be
defined as follows:
#
gis1 “ eαk Li ξi , X P Xi1
p
gi pX, L, αq “ (6.4a)
gis2 “ ep1´αk qLi ξi , X P Xi2
p
ξpi pX, L, αq “ gi´1 gi1 (6.4b)
The geometric Jacobian is now formulated as follows:

i
ÿ
Ji pq, L, α, Xq “ Ad´1 T B
g plq ¨¨¨gi g plq q plq
(6.5)
l“0
Remark 8. It is important to mention that, in addition to optimizing the segment length

parameter L and the actuators length parameter α, it is possible to optimize the parameter
representing the distance of the actuator to the mid-line d and also the segment radius parameter
r by following the same procedure that will be explained for the parameters L and α.
6.2.2 Optimization-based Method

In this section, we present an optimization approach for the purpose of optimizing the design of
the investigated slender soft robot in order to reach a single specific point and multiple target
points, as depicted in Figs. 6.2a and 6.2b.
To achieve one single point (as shown in Fig. 6.2a), the procedure consists of finding the
optimal parameters L and α in order that the soft robot design is optimized enough to the level
that its workspace will contain this point.
117
However, when trying to reach multiple target points (as shown in Figs. 6.2b), the parameters
must be optimized in order that the workspace of the designed soft robot contains all the desired
points, which is clearly more difficult since the cost function will be more complex than the first
case, because these objective points may be competing with each other.
(a) Optimal Design for a single point. (b) Optimal Design for multiple points.
Figure 6.2: Soft robot’s design to achieve a single point (a), and multiple points (b).
6.2.3 Optimal Design to Achieve One Single Point

The proposed method consists of determining the optimal values of both the length parameter
L of each segment denoted as L˚ , and the actuators length parameter α designated by α˚ , in
order that, given these optimal values, the end-effector’s workspace of the optimally designed
soft robot will contain the position to be accessed (as illustrated in Fig. 6.2a).
Denote u and ud as the end-effector’s position and the desired position to be reached,
according to the workspace Definition 4, we can then define the distance between u and ud as
f pq, L, αq “ 12 kupq, L, αq ´ ud k22
Hence, an optimal design of a soft robot in order to reach ud can be achieved through solving
the following nonlinear constrained optimization problem:
rL˚ , α˚ , τ ˚ s “ arg min f pq, L, αq

tL,α,τ u
s.t. τ P T , L P L, α P A (6.6)
KpL, αqq “ Hpq, L, αqτ ` F pq, L, αq
upq, L, αq “ E1 gpq, L, αqET2
where τ ˚ is the optimal actuators vector to reach the desired position ud . The vectors L˚ and
α˚ represent the optimal length of segments and the optimal actuators placement, respectively
(since they satisfy the bounded constraints and the nonlinear constraints) to minimize the
nonlinear cost function f pq, L, αq.
To solve the above optimization problem, gradient-based method [60] can be used which
depends on the knowledge of the gradient of f pq, L, αq with respect to the vectors L, α, and τ ,
noted as ∇rL,α,τ s f .
118
The gradient method to solve (6.6) can be established by the following [60]:
Ys`1 “ rL, α, τ ss ´ λ∇rL,α,τ s f prL, α, τ ss q
s.t. rL, α, τ ss`1 “ arg minkYs`1 ´ rL, α, τ sk
rL,α,τ s
τ P T , L P L, α P A (6.7)
KpL, αqq “ Hpq, L, αqτ ` F pq, L, αq
upq, L, αq “ E1 gpq, L, αqET2
with: rL, α, τ s0 is known, and λ is an infinitesimal small step size.

Or in an alternative form; The gradient method to solve (6.6) can be established by the
following [60]:
Ys`1 “ rL, α, τ ss ´ λ∇rL,α,τ s f prL, α, τ ss q
(6.8)
rL, α, τ ss`1 “ P rpYs`1 q, Project Ys`1 onto the feasible set
where P r is the projection operator.
Within the PCS framework, the following subsections are devoted to establishing the
analytical form: ” ı
∇rL,α,τ s f “ ∇L f, ∇α f, ∇τ f (6.9)
6.2.3.1 Calculation of ∇τ f
First, the analytical form of the gradient of the cost function with respect to the actuator vector
∇τ f has already been demonstrated and established in (3.7) and can be written in the following
form: ” Bf ıT
∇τ f “ “ ∇τ u rupq, L, αq ´ ud s (6.10)
Bτ
with ∇τ u represents the gradient of the end-effector with respect to τ , and is defined according
to Theorem 1 by:
” ıT
∇τ u “ ∇τ q ∇u q “ ∇τ q r03 Rpq, L, αqsJpq, L, αq (6.11)
where Jpq, L, αq is the geometric Jacobian defined in (6.5), the orientation matrix Rpq, L, αq
is computed via the definition of the configuration matrix gpq, L, αq through: Rpq, L, αq “
E1 gpq, L, αq ET1 , and ∇τ q is the gradient of the generalized strain vector with respect to the
actuators force vector and formulated according to Assumption 1 as:
“ ‰T
∇τ q “ H T pq, L, αq B´1 pτ, q, L, αq (6.12)
with
Bpτ, q, L, αq “ KpL, αq ´ Hq pq, L, αqpIn b τ q ´ Fq pq, L, αq
6.2.3.2 Calculation of ∇L f
The analytical form of the gradient of the cost function with respect to the segments length
parameter ∇L f can be calculated using the exponential map properties, but first let us write
the formula of ∇L f using the principle of variable separation as:
” Bf ıT ” Bf Bu ıT
∇L f “ “ (6.13)
BL Bu BL
119
By definition, the partial derivative of the cost function with respect to the end-effector position
vector is established by:
Bf T
“ rupq, L, αq ´ ud s (6.14)
Bu
Next, the partial derivative of the end-effector position vector with respect to the segment
length parameter can be decomposed using (6.1b) as follows:
Bu Bg Bu Bq
“ E1 E `
BL BL N Bq BL
with EN “ IN b ET2 , which can be written as:
Bu Bg Bq
“ E1 EN ` r∇u qsT (6.15)
BL BL BL
with
„ Bu T „ Bg T « ffT
∇u q “ “ E1 En “ r03 Rpq, L, αqsJpq, L, αq
Bq Bq
which is deduced from (6.11).
Bu Bg
Therefore, to formulate more explicitly, we need to find the analytical form of and
BL BL
Bq Bq
. For this, let us first calculate . Applying the partial derivative with respect to L to
BL BL
(6.1a), we obtain:
BK Bq BH BF
pIN b qq ` K “ τ`
BL BL BL BL
with K being always invertible, the above gives:
ˆ ˙
Bq BH BF BK
“ K ´1 τ` ´ pIN b qq (6.16)
BL BL BL BL
where the value of q corresponding to the value of τ can be calculated by solving (6.1a) using
Proposition 1.
Bg
Next, we calculate using the exponential map properties. First, we use (6.4a) in order to
BL
reformulate the exponential map gpq, L, αq from (6.2a) in the following form:
gpq, L, αq “ pg1s1 g1s2 q ¨ ¨ ¨ pgN

s1 s2
gN q (6.17)
Bg
with gis1 and gis2 being defined in (6.4a). Using the above exponential map formula (6.17),
BL
can be established as follows [66]:
Bg Bg s1 Bg s2
“ 1 rIN b pg1s2 ¨ ¨ ¨ gN s1 s2
gN qs ` g1s1 1 rIN b pg2s1 ¨ ¨ ¨ gN
s1 s2
gN qs ` ¨ ¨ ¨
BL BL BL
s1 s2 (6.18)
Bg s1 BgN
` g1s1 g1s2 ¨ ¨ ¨ N rIN b gN
s2
s ` g1s1 g1s2 ¨ ¨ ¨ gN
BL BL
where
Bgis1 Bg s1
„  ” ı
“ 04 , ¨ ¨ ¨ , i , ¨ ¨ ¨ , 04 “ 04 , ¨ ¨ ¨ , αk gi ξpi , ¨ ¨ ¨ , 04
BL BLi
and
Bgis2 Bg s2
„  ” ı
“ 04 , ¨ ¨ ¨ , i , ¨ ¨ ¨ , 04 “ 04 , ¨ ¨ ¨ , p1 ´ αk qgi ξpi , ¨ ¨ ¨ , 04
BL BLi
120
Bg Bq
With being calculated using (6.18), and from (6.16), we implement them into (6.15)
BL BL
Bu Bf
to obtain the analytical form of , and using from (6.14), we finally deduce ∇L f in (6.13).
BL Bu
6.2.3.3 Calculation of ∇α f
Last, with ∇τ f from (6.10), and ∇L f being calculated from (6.13), the remaining gradient to
be computed in order to complete (6.9) is ∇α f , which will be treated hereafter.
The analytical form of the gradient of the cost function with respect to the actuators length
parameter ∇α f can be calculated by following the same steps as those for ∇L f :
” Bf ıT ” Bf Bu ıT
∇α f “ “ (6.19)
Bα Bu Bα
Bf Bu
with being calculated from (6.14), thus, we only need to calculate the remaining term
Bu Bα
which can be decomposed using (6.1b) as follows:
Bu Bg Bq
“ E1 Enτ ` r∇u qsT (6.20)
Bα Bα Bα
with Enτ “ Inτ b ET2 . Then by applying the partial derivative with respect to α to (6.1a), we
obtain:
ˆ ˙
Bq ´1 BH BF BK
“K τ` ´ pInτ b qq (6.21)
Bα Bα Bα Bα
Bg
In the above equation, to calculate the partial derivative , we can use the properties of the
Bα
exponential map [66] from (6.17), which yields:
Bg Bg s1 Bg s2
“ 1 rInτ b pg1s2 ¨ ¨ ¨ gN s1 s2
gN qs ` g1s1 1 rInτ b pg2s1 ¨ ¨ ¨ gN
s1 s2
gN qs ` ¨ ¨ ¨
Bα Bα Bα
s1 s2 (6.22)
Bg s1 BgN
` g1s1 g1s2 ¨ ¨ ¨ N rInτ b gN
s2
s ` g1s1 g1s2 ¨ ¨ ¨ gN
Bα Bα
where
Bgis1 Bg s1
„  ” ı
“ 04 ¨ ¨ ¨ i ¨ ¨ ¨ 04 “ 04 ¨ ¨ ¨ Li gi ξpi ¨ ¨ ¨ 04
Bα Bαk
and
Bgis2 Bgis2
„  ” ı
“ 04 ¨ ¨ ¨ ¨ ¨ ¨ 04 “ 04 ¨ ¨ ¨ ´ Li gi ξpi ¨ ¨ ¨ 04
Bα Bαk
Bg Bq
With being calculated using (6.22), and from (6.21), we implement them into (6.20)
Bα Bα
Bu Bf
to obtain the analytical form of , and using from (6.14), we deduce ∇α f in (6.19).
Bα Bu
Finally, with ∇τ f being calculated from (6.10), ∇L f in (6.13), and ∇α f using (6.19), we
deduce the value of ∇rL,α,τ s f in (6.9), which is necessary for solving (6.6) via (6.7).
121
6.2.4 Optimal Design to Achieve Multiple Points
In a general sense, it is important to optimize the segments length parameter L and the actuators
length parameter α so that the end-effector of the studied soft robot can reach multiple target
points, which can be useful for various applications such as pick-and-place and trajectory
planning. Equivalently, doing this will result in optimizing the end-effector’s workspace in order
to contain the target points (as depicted by Fig. 6.2b).
p1q pn q
Given a set of target points, noted as Sd “ tud , ¨ ¨ ¨ , ud S u P R3ˆnS , for the obtained
optimized soft robot with optimized structure, it will exist j different configurations q pjq with the
pjq
corresponding end-effector position upjq for j “ 1 ¨ ¨ ¨ nS , such that kupjq ´ ud k22 “ 0. Therefore,
let us define the total distance of the end-effector’s position with respect to all target points in
Sd as:
nS
ÿ pjq
hpq, L, αq “ 12 kupjq pq pjq , L, αq ´ ud k22
j“1
then the optimal design of the investigated soft robot for the purpose of reaching all target points
in Sd can be obtained by solving the following nonlinear constrained optimization problem:
rL˚ ,α˚ , T ˚ s “ arg min hpq, L, αq

tL,α,T u
s.t. τ pjq P T , L P L, α P A
´ ¯ (6.23)
KpL, αq q pjq “ Hpq pjq , L, αqτ pjq ` F q pjq , L, α
upjq “ E1 g pjq pq pjq , L, αq ET2
” ıT
pjq pjq
with q pjq “ q1 ¨ ¨ ¨ qN P RN , T o “ rτ p1q ¨ ¨ ¨ τ pnS q so is the optimal actuators vector collection
” ıT
pjq pjq
corresponding to each target point ud P Sd , and τ pjq “ τ1 ¨ ¨ ¨ τnτ P Rnτ , for j P 1, ¨ ¨ ¨ , nS ,
is the actuators vector associated with the desired position udj . The vectors L˚ and α˚ represent
the optimal length of each segment of the robot and the optimal actuators length, respectively.
Similar to the case of a single point, the gradient-based method [60] can be used to solve the
above optimization problem which needs the calculation of hpq, L, αq with respect to the vectors
L, α, and τ , and formulated as:
” ı
∇rL,α,T s h “ ∇L h, ∇α h, ∇T h (6.24)
which will be detailed hereafter.
6.2.4.1 Calculation of ∇T h
According to the principle of variable separation, we have:
” Bh ıT ” Bh Bu ıT
∇T h “ “
BT Bu BT
” Bh Bup1q (6.25)
Bh BupnS q ıT
“ , ¨ ¨ ¨ ,
Bup1q Bτ p1q BupnS q Bτ pnS q
Then by definition, the partial derivative of h with respect to upjq can be written as:
Bh ” ıT
pjq pjq
“ u pq , L, αq ´ u dj (6.26)
Bupjq
122
Moreover, the partial derivative of the end-effector position upjq with respect to its associated
actuators vector τ pjq can be deduced from (6.11) and formulated by:
Bupjq Bupjq Bq pjq

“
Bτ pjq Bq pjq Bτ pjq
” ı Bq pjq (6.27)
“ r03 Rpq pjq , L, αqsJpq pjq , L, αq
Bτ pjq
Bq pjq
with deduced from (2.50) as:
Bτ pjq
Bq pjq
“ rBpjq pτ pjq , q pjq , L, αqs´1 H pjq pq pjq , L, αq (6.28)
Bτ pjq
where
Bpjq pτ pjq , q pjq , L, αq “ KpL, αq ´ Hqpjq pq pjq , L, αqpIn b τ pjq q ´ Fqpjq pq pjq q
Bq pjq
By substituting into (6.27), then (6.26) and (6.27) into (6.25), we finally obtain the
Bτ pjq
analytical form of ∇T h.
6.2.4.2 Calculation of ∇T L
Similarly, using the principle of variable separation gives us the following formula:
j“n
ÿS „ T
” Bh ıT ” Bh Bu ıT Bh Bupjq
∇L h “ “ “ (6.29)
BL Bu BL j“1
Bupjq BL
where the partial derivative of h with respect to upjq has already been established in (6.26).
Bupjq
Therefore, we need only to find the analytical form of . For this, using (6.15), we can write
BL
the following:
Bupjq Bg pjq Bq pjq
“ E1 EN ` r∇u q pjq sT (6.30)
BL BL BL
” ıT
with ∇u q pjq “ r03 Rpq pjq , L, αqsJpq pjq , L, αq being deduced from (6.27). Moreover, using
(6.16), we can get the following:
ˆ ˙
Bq pjq ´1 BH pjq pjq BF pjq BK pjq
“K τ ` ´ pIN b q q (6.31)
BL BL BL BL
where the value of q pjq corresponding to the value of τ pjq can be calculated by solving (6.1a)
using Proposition 1.
Bg
Next, we need to find the analytical form of using the exponential map properties. For
BL
pjq pjq
this, we reformulate the exponential map g pq , L, αq from (6.2a) as follows:
´ ¯ ´ ¯
pjq pjq pjq pjq
g pjq pq pjq , L, αq “ g1,s1 g1,s2 ¨ ¨ ¨ gN,s1 gN,s2 (6.32)
pjq ppjq pjq ppjq

with gi,s1 “ eαk Li ξi , and gi,s2 “ ep1´αk qLi ξi , for i “ 1 ¨ ¨ ¨ N , k “ 1 ¨ ¨ ¨ nτ , and j “ 1 ¨ ¨ ¨ nS .
123
Bg pjq
Using the above exponential map formula (6.32), can be calculated as follows [66]:
BL
pjq pjq ”
Bg pjq Bg1,s1 ” ´
pjq pjq pjq
¯ı
pjq Bg1,s2
´
pjq pjq pjq
¯ı
“ IN b g1,s2 ¨ ¨ ¨ gN,s1 gN,s2 ` g1,s1 IN b g2,s1 ¨ ¨ ¨ gN,s1 gN,s2 ` ¨ ¨ ¨
BL BL BL
pjq pjq
(6.33)
pjq pjq BgN,s1 ”
pjq
ı
pjq pjq pjq BgN,s2
` g1,s1 g1,s2 ¨ ¨ ¨ IN b gN,s2 ` g1,s1 g1,s2 ¨ ¨ ¨ gN,s1
BL BL
where
pjq
Bgi,s1 ”
pjq pjq
ı
“ 04 ¨ ¨ ¨ αk gi ξpi ¨ ¨ ¨ 04
BL
and
pjq
Bgi,s2 ”
pjq pjq
ı
“ 04 ¨ ¨ ¨ p1 ´ αk qgi ξpi ¨ ¨ ¨ 04
BL
Bg pjq Bq pjq
With being calculated using (6.33), and from (6.31), we implement both of them
BL BL
Bupjq
into (6.30) to obtain , then, using (6.26), we finally deduce ∇L h in (6.29).
BL
6.2.4.3 Calculation of ∇T α
As before, ∇α h can be formulated as:
j“n
ÿS „ T
” Bh ıT ” Bh Bu ıT Bh Bupjq
∇α h “ “ “ (6.34)
Bα Bu Bα j“1
Bupjq Bα
Bh Bu
with being calculated from (6.26), thus, we only need to find the analytical form of . For
Bu Bα
this, using (6.20), we can write the following:
Bupjq Bg pjq Bq pjq

“ E1 Enτ ` r∇u q pjq sT (6.35)
Bα Bα Bα
Moreover, using (6.21), we can then have:
ˆ ˙
Bq pjq ´1 BH pjq pjq BF pjq BK pjq
“K τ ` ´ pInτ b q q (6.36)
Bα Bα Bα Bα
Bg
Next, we have to find the analytical form of using the properties of the exponential map.
Bα
Bg pjq
For this, using (6.32), can be written as follows [66]:
Bα
pjq pjq ”
Bg pjq Bg1,s1 ” ´
pjq pjq pjq
¯ı
pjq Bg1,s2
´
pjq pjq pjq
¯ı
“ Inτ b g1,s2 ¨ ¨ ¨ gN,s1 gN,s2 ` g1,s1 Inτ b g2,s1 ¨ ¨ ¨ gN,s1 gN,s2 ` ¨ ¨ ¨
Bα Bα Bα
pjq pjq
(6.37)
pjq pjq BgN,s1
”
pjq
ı
pjq pjq pjq BgN,s2
` g1,s1 g1,s2 ¨ ¨ ¨ Inτ b gN,s2 ` g1,s1 g1,s2 ¨ ¨ ¨ gN,s1
Bα Bα
where
pjq
Bgi,s1 ”
pjq pjq
ı
“ 04 ¨ ¨ ¨ Li gi ξpi ¨ ¨ ¨ 04
Bα
124
and
pjq
Bgi,s2 ”
pjq pjq
ı
“ 04 ¨ ¨ ¨ ´ Li gi ξpi ¨ ¨ ¨ 04
Bα
Bg pjq Bq pjq
With being calculated using (6.37), and from (6.36), we implement both of them
Bα Bα
pjq
Bu
into (6.35) to obtain , then, using (6.26), we finally deduce ∇α h in (6.34).
Bα
Finally, with ∇T h being calculated from (6.25), ∇L h in (6.29), and ∇α h using (6.34), we
deduce the value of ∇rL,α,T s h in (6.24), which is necessary for solving (6.23).
6.3 Validation and Analysis
In this section, we illustrate the presented results by implementing the proposed optimization
approach on tendon-driven soft robots (as depicted by Figs. 6.1a). Table 3.1 presents the
numerical values of the investigated soft robot.
In the examined scenarios, WE is estimated using the forward approach (See Section 2.5.2) and
illustrated by the gray area. The workspace boundary BWE is estimated using the continuation
approach (See Section 5.2) and depicted by red curves and green bifurcation points. The
mounted tendons are depicted by black lines, and the desired points to reach are blue-colored.
Also, the presented simulation scenarios were implemented using MATLAB (the function
”fmincon” was used to solve the optimization problems formulated in this chapter) on an Intel
Xeon(R) with a 16-GB RAM and a 3.50 GHz processor.
6.3.0.1 Scenario 1: 2-tendons driven planar soft robot
In the first scenario, we investigate a soft robot consisting of two segments and driven by two
tendons installed on the Ozx plan (as shown in Fig. 6.3a).
In the initial state, the segments initial length vector is given L0 “ r0.2, 0.1sT m, and the
initial actuators length vector is α0 “ r0.15, 0.15sT . We select the bounded tension magnitude
T1,2 “ r0, 100sN allowing us to apply a couple magnitude within CT “ r0, 2.14sN ¨ m, and
1
CT “ r0, 1.42sN ¨ m. The segments length bounds are L “ r0, 0.3s ˆ r0, 0.2sm, and the
2
actuators length bounds are A “ r0, 1s2 .
The obtained results of WE and BWE for the initial state of the designed soft robot with the
given initial values of the parameters L0 and α0 are depicted by Fig. 6.3b where we can clearly
observe that the workspace does not contain the desired points.
125
(a) Initial design. (b) WE (gray) and BWE (red).
Figure 6.3: Soft robot’s initial design.
Firstly, we want to show that there are some cases where the optimal solution can only reach
a desired point due to the length constraint. For this, we impose the segments length bounds in
L “ r0, 0.24s ˆ r0, 0.16sm.
Using the proposed optimization approach, we apply (6.23) on the designed soft robot in
order to achieve the optimal values of L˚ and α˚ in order to reach the desired points. By doing
this, we find the optimal segments length parameter L˚ “ r0.225, 0.15sT , and the optimal
actuators length parameter α˚ “ r0.89, 0.41sT , which gives the optimal design shown by Fig.
6.4a. As illustrated by Fig. 6.4b, we can observe that the optimal solution allows only the
reachability of one single point due to the restrictions imposed on the segments length parameter
L.
(a) Optimal design. (b) WE (gray) and BWE (red).
Figure 6.4: Soft robot’s optimal design - achieving only one point.
To achieve the accessibility to multiple points, we can select a larger possibility for the length
bounds of each segment as L “ r0, 0.3s ˆ r0, 0.2sm.
Using the proposed optimization approach, we apply (6.23) on the designed soft robot. By
doing this, we find the optimal segments length parameter L˚ “ r0.28, 0.19sT , and the optimal
126
actuators length parameter α˚ “ r0.82, 0.80sT , which gives the optimal design shown by Fig.
6.5a. Accordingly, the workspace of the optimally designed soft robot now encloses the desired
points as illustrated by Fig. 6.5b, allowing the end-effector to reach these target locations.
Figure 6.5: Soft robot’s optimal design - achieving multiple points.
The computational time necessary to obtain the optimal results took approximately 0.445
seconds for this scenario.
6.3.0.2 Scenario 2: 3-tendons driven planar soft robot
In the second scenario, we investigate a soft robot comprised of three segments and controlled
by three tendons installed on the Ozx plan (as shown in Fig. 6.6a).
In the initial state, the segments initial length vector is given L0 “ r0.15, 0.1, 0.05sT m,
and the initial actuators length vector is α0 “ r0.15, 0.15, 0.15sT . We choose the bounded
tension magnitude T1,2,3 “ r0, 50sN allowing us to apply a couple magnitude within CT “
1
r0, 1.07sN ¨ m, and CT “ r0, 0.71sN ¨ m, and CT “ r0, 0.35s. The segments length bounds are
2 3
L “ r0, 0.3s ˆ r0, 0.2s ˆ r0, 0.1sm, and the actuators length bounds are A “ r0, 1s3 .
The obtained results of WE and BWE for the initial state of the designed soft robot with the
given initial values of the parameters L0 and α0 are depicted by Fig. 6.6b where we can clearly
observe that the workspace does not contain the desired points.
127
(a) Initial design. (b) WE (gray) and BWE (red).
Figure 6.6: Soft robot’s initial design.
Firstly, we want to show that there are some cases where the optimal solution can only reach
a desired point due to the length constraint. For this, we impose the segments length bounds in
L “ r0, 0.24s ˆ r0, 0.16s ˆ r0, 0.08sm.
Using the proposed optimization approach, we find the optimal segments length parameter
L˚ “ r0.225, 0.15, 0.075sT , and the optimal actuators length parameter α˚ “ r0.89, 0.76, 0.10sT ,
which gives the optimal design shown by Fig. 6.7a. As illustrated by Fig. 6.7b, we can observe
that the optimal solution allows only the reachability of one single point due to the restrictions
imposed on the segments length parameter L.
Figure 6.7: Soft robot’s optimal design - achieving only one point.
To achieve the accessibility to multiple points, we can select a larger possibility for the length
bounds of each segment as L “ r0, 0.3s ˆ r0, 0.2sm.
Using the proposed optimization approach, we find the optimal segments length parameter
L˚ “ r0.29, 0.19, 0.09sT , and the optimal actuators length parameter α˚ “ r0.88, 0.71, 0.17sT ,
which gives the optimal design as shown by Fig. 6.8a. Accordingly, the workspace of the
128
optimally designed soft robot now encompass the desired points as illustrated by Fig. 6.8b,
allowing the end-effector to reach these target locations.
Figure 6.8: Soft robot’s optimal design - achieving multiple points.
The computational time necessary to obtain the optimal results took approximately 0.612
seconds for this scenario.
6.4 Conclusion
In this chapter, we have proposed an optimization approach based on the adopted PCS model
for the purpose of optimizing the design of a soft robot with a slender shape. The approach was
then validated on different tendon-driven soft robot configurations.
However, it is important to mention that we can extend the proposed approach to also
optimize the radius of the segments and the distance of the actuator to the mid-line (which
will be detailed in Chapter 7). Moreover, the optimization approach can be extended to treat
different cases of actuation since the PCS method was already applied to different methods of
actuation, e.g., fluidic actuation [102].
On the other hand, to extend the applicability of the optimization approach in order to
treat the FEM case, we need to consider what parameters should be optimized in addition to
adapting the optimal problem for the FEM case (which will be explained in Chapter 7).
129
130
Chapter 7
Conclusions and Perspectives
7.1 Conclusions
In this thesis, we have investigated the workspace estimation and design optimization of soft
robots. On the one hand, the workspace estimation of soft robots provides crucial information
about the space where they can operate, which is useful for control applications (e.g., Pick-and-
Place, trajectory planning, etc). Moreover, evaluating the workspace also helps in determining the
singularities and impediments of motion to the robot which can be used as valuable information
for the design optimization of soft robots. On the other hand, the design optimization of soft
robots is a useful strategy to avoid the procedure of trial-and-error when designing soft robots.
Given specific performance objectives, it is more logical and interesting for both economic and
scientific reasons to determine the optimal design of a soft robot in order to achieve those
objectives in a virtual environment before designing the final prototype of the robot.
In Chapter 2, we have adopted two mathematical methods, the first is PCS which focuses on
slender-shaped soft robots, and the second is FEM which treats soft robots with general form.
To answer the direct problem of this thesis, we have firstly presented a classic approach
(forward approach), which consists of discretizing the actuators space, in order to estimate the
workspace of soft robots. However, due to its high complexity, alternative approaches were
proposed to efficiently estimate the workspace of soft robots.
Accordingly, we have proposed a first approach in Chapter 3 (optimization-based) that
consists of estimating the workspace of soft robots by mapping the exterior boundaries of
the workspace. This method successfully reduced the complexity of the workspace estimation
compared to the forward approach.
To address the challenge of estimating the interior/exterior points in the workspace by
discretizing the end-effector space, we have proposed a new approach based on interval analysis
(forward-backward) in Chapter 4 that consists of exploring all the feasible configurations in the
workspace of soft robots. Compared to the forward approach, this method successfully decreased
the complexity of the workspace estimation.
In a final approach to the workspace estimation problem, we have proposed in Chapter 5
a continuation approach, based on the PCS model, that consists of mapping the interior and
exterior boundaries of the workspace of soft robots. Notably, it provides information about the
singularities of soft robots.
To treat the inverse problem of this thesis, we have proposed an optimization method based
on the PCS model in Chapter 6 that consists of optimizing the design of soft robots in order to
achieve specific desired points.
131
In the following, we present a brief comparison of the proposed workspace estimation
approaches for both the PCS and FEM models.
7.1.1 PCS-based Workspace Estimation Approaches Comparison

Based on the PCS modeling method, three different approaches have been proposed to evaluate
the workspace of soft robots and were validated through the forward approach. Each approach
has its advantages and disadvantages, and the following summarizes their characteristics.
For the sake of clarity and simplicity, we consider one specific scenario (e.g., the planar
soft robot composed of three segments and actuated by three tendons). The structure of the
robot and its corresponding results are given in Fig. 7.1. Then, we summarize the advantages,
disadvantages, operations complexity, and computational time of each investigated workspace
approach in Table 7.1.
(a) Soft robot structure. (b) Optimization-based approach.
(c) Interval Analysis approach. (d) Continuation Approach.
Figure 7.1: Different workspace methodologies applied to the investigated soft robot modeled
through the PCS model.
132
Table 7.1: Advantages and limitations of the investigated workspace estimation approaches
based on the PCS model.
Approach
Optimization Interval Analysis Continuation
Qualities
Maps outer Explores
workspace boundaries. the whole Maps inner and outer
Avoids heavy workspace. workspace boundaries.
Advantages
calculation of Provide interior Ability to classify
interior points. as well as configurations.
Least expensive. exterior knowledge.
Inability to provide Inability to
Exhaustive.
interior knowledge. provide information
Disadvantages Inability to classify
Inability to classify on every point
configurations.
configurations. in the workspace.
Operations
220 2275 327
Complexity
Computational
3 72 40
Time (seconds)
In this specific scenario, the optimization-based approach is clearly the most efficient way to
estimate the workspace of soft robots since it consists only in mapping the exterior boundary
and avoiding the heavy computation of interior points, which also comes with the drawback
that it cannot provide interior information to the workspace.
On the other hand, the interval analysis approach consists of exploring the whole workspace
which allows us to have interior and exterior knowledge about the workspace, however, as a
result, this approach is exhaustive.
Finally, the continuation approach allows the mapping of both interior and exterior boundaries
to the workspace, this method allows us to identify the singularities of the robot which is useful
for the design. However, this approach does not seek to find all achievable configurations of the
workspace, but only the singularities of soft robots.
7.1.2 FEM-based Workspace Estimation Approaches Comparison
Based on the FEM modeling method, two different approaches have been proposed to evaluate
the workspace of soft robots and were validated through the forward approach. Each approach
has its advantages and disadvantages, and the following summarizes their characteristics.
For the sake of clarity and simplicity, we consider one specific scenario (e.g., the trunk-like
soft robot actuated by four symmetric tendons). The structure of the robot and its corresponding
results are given in Fig. 7.2. Then, we summarize the advantages, disadvantages, operations
complexity, and computational time of each investigated workspace approach in Table 7.2.
133
(a) Soft robot structure. (b) Optimization-based approach.
(c) Interval Analysis approach.
Figure 7.2: Different workspace methodologies applied to the investigated soft robot modeled
through the FEM model.
Table 7.2: Advantages and limitations of the investigated workspace estimation approaches
based on the FEM model.
Approach
Optimization Interval Analysis
Qualities
Maps outer
Explores the whole
workspace boundaries.
workspace .
Avoids heavy
Advantages Provide interior
calculation of
as well as
interior points.
exterior knowledge.
Least expensive.
Inability to provide Inability to provide
interior knowledge. interior knowledge.
Disadvantages
Inability to classify Inability to classify
configurations. configurations.
Operations
864 3667
Complexity
Computational
1039 3733
Time (seconds)
134
The interval analysis approach is computationally expensive compared to the optimization-
based approach. However, the interval analysis approach provides valuable interior and exterior
information of the workspace.
On the other hand, the optimization-based approach is more efficient. However, this approach
only maps the exterior boundary of the workspace ignoring its interior configurations.
In the following, several possible extensions of the obtained results will be discussed.
7.2 Perspectives
7.2.1 From Position-Access to Orientation-Access Workspace
This thesis has proposed different methodologies to estimate the position-access workspace of soft
robots. However, the proposed approaches can also be extended to treat the orientation aspect
of the workspace. Such an extension will come with necessary adaptations of the formulation of
the workspace definition and the workspace set.
For the PCS model, we can treat the orientation aspect of the workspace by taking into
consideration the orientation matrix R defined in (3.2) and then adapting the algorithms
established in Chapters 3, 4, and 5 in a proper way for the purpose of studying the orientation
aspect.
In the FEM model, we can treat the orientation aspect of the workspace by modifying
the selection matrix C in (2.53). By considering 3 independent nodal positions around the
end-effector qE1 , qE2 and qE3 , which can be used to define a unique plan of the end-effector, we
can always properly choose the matrix C P R6ˆ3np such that qE now is a function of those 3
points and contains both position and orientation information of the defined plan.
7.2.2 From Tendons-driven Soft Robots to other Methods of Actua-

tion
In this thesis, we have focused on tendon-driven soft robots in both the direct problem (i.e.,
workspace estimation) and inverse problem (i.e., design optimization).
However, since the PCS method was applied to different methods of actuation (e.g., fluidic
[102]), and that the FEM approach was also applied to different methods of actuation (e.g.,
pneumatic and hydraulic in [15, 107]), then, we can also use different methods of actuation by
adapting the actuation matrix H in both modeling methods to the adopted actuation manner.
7.2.3 Continuation Approach: Toward a Generalization of Tangent

Vectors Calculations
The continuation approach was explained in detail in Section 5.2 of Chapter 5, where tangent
vectors to boundary points met along the solution curve (5.7) are calculated in Section 5.2.3
based on a classification of those boundary points into three categories (see Table 5.1): regular
boundary points (see Fig. 5.3), simple bifurcation points (see Fig. 5.4), and multiple bifurcation
points (see Fig. 5.5).
The tangent vectors to each boundary point class were calculated using a different approach.
On the one hand, the tangent vector of regular boundary points were calculated using Theorem
4 in Section 5.2.3.1 which consists of calculating the nullspace of the workspace boundary set
Jacobian GY . On the other hand, tangent vectors to simple bifurcation points was calculated
135
using 2nd order Taylor expansion in Theorem 5 in Section 5.2.3.2 using the workspace boundary
set Hessian GY Y . Finally, tangent vectors to multiple bifurcation points were calculated by
formulating and solving the eigenvalue problem defined in Section 5.2.3.3 through the workspace
boundary set Jacobian GY . For the purpose of generality, it is of interest to provide a single
general-purpose methodology for calculating tangent vectors to all boundary point cases
met along the solution curve.
From the first point of view, Taylor expansions can indeed be used to treat all cases. In fact,
the 1st order Taylor expansion has already been used to calculate tangent vectors to regular
boundary points (see Theorem 4), and the 2nd order Taylor expansion is also used in Theorem 5
to calculate tangent vectors to simple bifurcation points. However, to treat multiple bifurcation
points, this methodology requires the calculation of partial derivatives of G up to the nb th-order,
with nb being the number of branches passing through the multiple bifurcation point must be
previously known. Thus, the main difficulty of this technique lies first in identifying nb , and
secondly in calculating partial derivatives of G up to the nb th-order.
In an alternative way, the eigenvalue problem to calculate tangent vectors of multiple
bifurcation points (see Section 5.2.3.3) can be extended to treat the simple bifurcation as well
as regular boundary point cases. However, the extension of this approach cannot be directly
implemented since it requires the existence and the knowledge of two orthonormal solutions γ1
and γ2 of ΨZ . However, this is difficult to obtain because for both simple bifurcation points
and regular boundary points becaause the row-rank deficiency of ΨZ is 1, i.e., we can only have
one orthonormal solution γ1 that verifies the property ΨZ γ1 “ 0. Therefore, the main challenge
using this proposed technique would be to determine the second solution γ2 such that ΨZ γ2 “ 0
for multiple bifurcation points as well as regular boundary points.
7.2.4 Continuation Approach: an Extension to treat FEM
As shown in Section 5.2, the continuation approach requires the analytical representation of the
parameters describing the static/kinematic model of soft robots. This is mainly due to the fact
that tangent vector calculation in Section (5.2.3) requires the computation of the Jacobians ΨZ
and GY , and also the Hessian matrix GY Y . With the PCS model, this is possible because we
have the analytical form of these mathematical parameters.
However, the FEM static model uses numerical approximations to establish its parameters,
therefore, in the FEM framework we don’t have the analytical form of those parameters. Hence,
extending the applicability of the continuation approach to treat the FEM model will not be
easy, and we think that the only possible way to perform this would be to calculate the Jacobians
ΨZ and GY , and also the Hessian matrix GY Y , via numerical techniques.
7.2.5 Design Optimization: Including more Parameters in PCS
In Chapter 6, Section 6.2, given specific performance objectives (e.g., achieving specific desired
points), we have explained how to optimize the segments length parameter L and the actuators
length parameter α, in order to achieve those objectives. What has been established for those
parameters can also be extended in order to optimize the segments radius parameter r of the
robot and the actuators distance to the mid-line d, as illustrated by Fig. 7.3.
136
Figure 7.3: Design Optimization - Including more Parameters in PCS.
For this, the PCS static/kinematic model must consider the variations of the radius parameter
r and the distance to the mid-line parameter d as what has been done for L and α in Section
6.2.1. With this, the next step is to take into account r and d in the optimal problem (6.23) by
reformulating it into the following:
nS
ÿ pjq
hpq, Xq “ 1
2 kupjq pq pjq , Xq ´ ud k22
j“1
and
rX˚ , T ˚ s “ arg min hpq, Xq
tX,τ u
s.t. τ pjq P T , L P L, α P A, r P R, d P D
´ ¯ (7.1)
KpXq q pjq “ Hpq pjq , Xqτ pjq ` F q pjq , X
upjq “ E1 g pjq pq pjq , Xq ET2
with X “ rL, α, r, ds.
One additional step that will be required for the implementation of the methodology is to
calculate the gradient of the objective function with respect to r and d, i.e., ∇r h and ∇d h.
Finally, solving (7.1) will yield the optimal segments length parameter L˚ , the optimal
actuators length parameter α˚ , the optimal segments radius parameter r˚ , and the optimal
actuators distance to the mid-line d˚ , allowing us to achieve a full design optimization of the
investigated soft robot based on the PCS model.
7.2.6 Design Optimization: From PCS to FEM

The optimization approach proposed to achieve the optimal design of soft robots modeled by
the PCS (see Chapter 6, section 6.2) can be extended to treat the FEM case as well.
For this, the same procedure followed for the PCS model might be applied to the FEM
framework with necessary adaptation. The first step is to establish the static/kinematic model
137
in function of the parameters to be optimized, e.g., the robots’ length L, the actuators length α,
the radius of the robot r, and the distance of the actuators to the mid-line d, as follows:
Kpq, Xq q “ Hpq, Xqτ ` F pq, Xq (7.2a)

qE “ Cq (7.2b)
u “ qE ` up0q (7.2c)
with X “ rL, α, r, ds.

The second step is to formulate the optimal problem in order to optimize the objective
parameters X:
nS
ÿ pjq
f pq, Xq “ 12 kupjq pq pjq , Xq ´ ud k22
j“1
and
rX˚ , T ˚ s “ arg min hpq, Xq
tX,τ u
s.t. τ pjq P T , L P L, α P A, r P R, d P D
´ ¯ (7.3)
Kpq pjq , Xq q pjq “ Hpq pjq , Xqτ pjq ` F q pjq , X
upjq “ Cq pjq ` up0q
Finally, solving (7.3) yields the optimal parameters to achieve the optimal design of the
investigated soft robot based on the FEM model.
138
Bibliography
[1] K Abdel-Malek and Nada Khairallah. Serial 5dof manipulators: workspace, void, and
volume determination. In Proc. of the 25th Design Automation in Mechanisms and
Robotics, Las Vegas. Citeseer, 1999.
[2] Karim Abdel-Malek, Frederick Adkins, Harn-Jou Yeh, and Edward Haug. On the deter-
mination of boundaries to manipulator workspaces. Robotics and Computer-Integrated
Manufacturing, 13(1):63–72, 1997.
[3] Karim Abdel-Malek and Harn-Jou Yeh. Analytical boundary of the workspace for general3-
dof mechanisms. The International Journal of Robotics Research, 16(2):198–213, 1997.
[4] D Alciatore and C Ng. Determining manipulator workspace boundaries using the monte
carlo method and least squares segmentation. ASME Robotics: Kinematics, Dynamics
and Controls, 72:141–146, 1994.
[5] Carl Barnett Allendoerfer. Calculus of several variables and differentiable manifolds.
Macmillan, 1974.
[6] Eugene L Allgower and Kurt Georg. Introduction to numerical continuation methods.
SIAM, 2003.
[7] Walid Amehri, Gang Zheng, and Alexandre Kruszewski. Fem based workspace estimation
for soft robots: a forward-backward interval analysis approach. In 2020 3rd IEEE
International Conference on Soft Robotics (RoboSoft), pages 170–175. IEEE, 2020.
[8] Walid Amehri, Gang Zheng, and Alexandre Kruszewski. Workspace boundary estimation
for soft manipulators using a continuation approach. IEEE Robotics and Automation
Letters, 6(4):7169–7176, 2021.
[9] Walid Amehri, Gang Zheng, and Alexandre Kruszewski. Fem-based exterior workspace
boundary estimation for soft robots via optimization. IEEE Robotics and Automation
Letters, 2022.
[10] Yasmin Ansari, Mariangela Manti, Egidio Falotico, Matteo Cianchetti, and Cecilia Laschi.
Multiobjective optimization for stiffness and position control in a soft robot arm module.
IEEE Robotics and Automation Letters, 3(1):108–115, 2017.
[11] Simona Aracri, Francesco Giorgio-Serchi, Giuseppe Suaria, Mohammed E Sayed, Markus P
Nemitz, Stephen Mahon, and Adam A Stokes. Soft robots for ocean exploration and
offshore operations: A perspective. Soft Robotics, 2021.
139
[12] Vigen Arakelian, Sébastien Briot, and Victor Glazunov. Increase of singularity-free zones in
the workspace of parallel manipulators using mechanisms of variable structure. Mechanism
and Machine Theory, 43(9):1129–1140, 2008.
[13] Siegfried Bauer, Simona Bauer-Gogonea, Ingrid Graz, Martin Kaltenbrunner, Christoph
Keplinger, and Reinhard Schwödiauer. 25th anniversary article: a soft future: from robots
and sensor skin to energy harvesters. Advanced Materials, 26(1):149–162, 2014.
[14] Alessandro Berti, Jean-Pierre Merlet, and Marco Carricato. Workspace analysis of
redundant cable-suspended parallel robots. In Cable-Driven Parallel Robots, pages 41–53.
Springer, 2015.
[15] Thor Morales Bieze, Frederick Largilliere, Alexandre Kruszewski, Zhongkai Zhang, Rochdi
Merzouki, and Christian Duriez. Finite element method-based kinematics and closed-loop
control of soft, continuum manipulators. Soft robotics, 5(3):348–364, 2018.
[16] Daniel M Bodily, Thomas F Allen, and Marc D Killpack. Multi-objective design optimiza-
tion of a soft, pneumatic robot. In 2017 IEEE International Conference on Robotics and
Automation (ICRA), pages 1864–1871. IEEE, 2017.
[17] Quentin Boehler, Isabelle Charpentier, Marc S Vedrines, and Pierre Renaud. Definition
and computation of tensegrity mechanism workspace. Journal of Mechanisms and Robotics,
7(4):044502, 2015.
[18] Oriol Bohigas, Montserrat Manubens, and Lluı́s Ros. A complete method for workspace
boundary determination on general structure manipulators. IEEE Transactions on Robotics,
28(5):993–1006, 2012.
[19] F. Boyer and F. Renda. Poincare’s equations for cosserat media: Application to shells.
Journal of Nonlinear Science, 2016.
[20] Frédéric Boyer, Vincent Lebastard, Fabien Candelier, and Federico Renda. Dynamics of
continuum and soft robots: A strain parameterization based approach. IEEE Transactions
on Robotics, 37(3):847–863, 2020.
[21] Sébastien Briot and Ilian A Bonev. Are parallel robots more accurate than serial robots?
Transactions of the Canadian Society for Mechanical Engineering, 31(4):445–455, 2007.
[22] Sébastien Briot, Wisama Khalil, et al. Dynamics of parallel robots. From rigid bodies to
flexible elements. Springer, 2015.
[23] Jessica Burgner, D Caleb Rucker, Hunter B Gilbert, Philip J Swaney, Paul T Russell,
Kyle D Weaver, and Robert J Webster. A telerobotic system for transnasal surgery.
IEEE/ASME Transactions on Mechatronics, 19(3):996–1006, 2013.
[24] Jessica Burgner-Kahrs, Hunter B Gilbert, Josephine Granna, Philip J Swaney, and
Robert J Webster. Workspace characterization for concentric tube continuum robots.
In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages
1269–1275. IEEE, 2014.
[25] Jessica Burgner-Kahrs, D Caleb Rucker, and Howie Choset. Continuum robots for medical
applications: A survey. IEEE Transactions on Robotics, 31(6):1261–1280, 2015.
140
[26] Kun Cao, Rongjie Kang, David T Branson III, Shineng Geng, Zhibin Song, and Jian S Dai.
Workspace analysis of tendon-driven continuum robots based on mechanical interference
identification. Journal of Mechanical Design, 139(6):062303, 2017.
[27] Yi Cao, Ke Lu, Xiujuan Li, and Yi Zang. Accurate numerical methods for computing
2d and 3d robot workspace. International Journal of Advanced Robotic Systems, 8(6):76,
2011.
[28] U Cardaun. Stand der greiferentwicklung, fordern und heben. Fuchleil Monlage Handhah-
ngstechnik, 28:40, 1978.
[29] TOMMASO Carella. Design optimization of a passive soft robotic device for neurorehabil-
itation. 2019.
[30] Sabri Cetinkunt and Wayne J Book. Performance limitations of joint variable-feedback
controllers due to manipulator structural flexibility. IEEE Transactions on Robotics and
Automation, 6(2):219–231, 1990.
[31] Frédéric Chapelle and Philippe Bidaud. Evaluation functions synthesis for optimal design
of hyper-redundant robotic systems. Mechanism and machine theory, 41(10):1196–1212,
2006.
[32] Andrea Cherubini, Robin Passama, André Crosnier, Antoine Lasnier, and Philippe
Fraisse. Collaborative manufacturing with physical human–robot interaction. Robotics
and Computer-Integrated Manufacturing, 40:1–13, 2016.
[33] Matteo Cianchetti, Cecilia Laschi, Arianna Menciassi, and Paolo Dario. Biomedical
applications of soft robotics. Nature Reviews Materials, 3(6):143–153, 2018.
[34] Matteo Cianchetti, Tommaso Ranzani, Giada Gerboni, Thrishantha Nanayakkara, Kaspar
Althoefer, Prokar Dasgupta, and Arianna Menciassi. Soft robotics technologies to address
shortcomings in today’s minimally invasive surgery: the stiff-flop approach. Soft robotics,
1(2):122–131, 2014.
[35] Anne Delettre, Guillaume J Laurent, Yassine Haddab, and Nadine Le Fort-Piat. Robust
control of a planar manipulator for flexible and contactless handling. Mechatronics,
22(6):852–861, 2012.
[36] Cosimo Della Santina, Antonio Bicchi, and Daniela Rus. On an improved state parametriza-
tion for soft robots with piecewise constant curvature and its use in model based control.
IEEE Robotics and Automation Letters, 5(2):1001–1008, 2020.
[37] Cosimo Della Santina, Manuel G Catalano, and Antonio Bicchi. Soft robots. Encyclopedia
of Robotics, 2020.
[38] Cosimo Della Santina, Robert K Katzschmann, Antonio Biechi, and Daniela Rus. Dynamic
control of soft robots interacting with the environment. In 2018 IEEE International
Conference on Soft Robotics (RoboSoft), pages 46–53. IEEE, 2018.
[39] Christian Duriez. Control of elastic soft robots based on real-time finite element method.
In 2013 IEEE International Conference on Robotics and Automation, pages 3982–3987.
IEEE, 2013.
141
[40] Yahya Elsayed, Augusto Vincensi, Constantina Lekakou, Tao Geng, CM Saaj, Tommaso
Ranzani, Matteo Cianchetti, and Arianna Menciassi. Finite element analysis and design
optimization of a pneumatically actuating silicone module for robotic surgery applications.
Soft Robotics, 1(4):255–262, 2014.
[41] Douglas R Ewing, Alessio Pigazzi, Yulun Wang, and Garth H Ballantyne. Robots in the
operating room—the history. In Seminars in laparoscopic Surgery, volume 11, pages 63–71.
Sage Publications Sage CA: Thousand Oaks, CA, 2004.
[42] Nima Fazeli, Samuel Zapolsky, Evan Drumwright, and Alberto Rodriguez. Fundamental
limitations in performance and interpretability of common planar rigid-body contact
models. In Robotics Research, pages 555–571. Springer, 2020.
[43] A Gasparetto and L Scalera. A brief history of industrial robotics in the 20th century.
Advances in Historical Studies, 8(1):24–35, 2019.
[44] Thomas George Thuruthel, Yasmin Ansari, Egidio Falotico, and Cecilia Laschi. Control
strategies for soft robotic manipulators: A survey. Soft robotics, 5(2):149–163, 2018.
[45] Mariam Md Ghazaly, Siti Norazlin Mohd Basar, and Muhammad Shadiq Lagani. Design
optimization & analysis of a soft crawling robot. International Journal of Integrated
Engineering, 13(6):285–298, 2021.
[46] Francesco Giorgio-Serchi, Andrea Arienti, Francesco Corucci, Michele Giorelli, and Ce-
cilia Laschi. Hybrid parameter identification of a multi-modal underwater soft robot.
Bioinspiration & biomimetics, 12(2):025007, 2017.
[47] Olivier Goury and Christian Duriez. Fast, generic, and reliable control and simulation of
soft robots using model order reduction. IEEE Transactions on Robotics, 34(6):1565–1576,
2018.
[48] Marc Gouttefarde, David Daney, and Jean-Pierre Merlet. Interval-analysis-based determi-
nation of the wrench-feasible workspace of parallel cable-driven robots. IEEE Transactions
on Robotics, 27(1):1–13, 2010.
[49] Di Guo and Zhan Kang. Chamber layout design optimization of soft pneumatic robots.
Smart Materials and Structures, 29(2):025017, 2020.
[50] Morton E Gurtin. An introduction to continuum mechanics. Academic press, 1982.
[51] Martin Hägele, Klas Nilsson, J Norberto Pires, and Rainer Bischoff. Industrial robotics.
In Springer handbook of robotics, pages 1385–1422. Springer, 2016.
[52] Brian C Hall et al. Lie groups, Lie algebras, and representations: an elementary introduc-
tion, volume 10. Springer, 2003.
[53] Michael W Hannan and Ian D Walker. Kinematics and the implementation of an elephant’s
trunk manipulator and other continuum style robots. Journal of robotic systems, 20(2):45–
63, 2003.
[54] Wissem Haouas, Redwan Dahmouche, Nadine Le Fort-Piat, and Guillaume J Laurent.
4-dof spherical parallel wrist with embedded grasping capability for minimally invasive
surgery. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), pages 2363–2368. IEEE, 2016.
142
[55] Edward J Haug, Chi-Mei Luh, Frederick A Adkins, and Jia-Yi Wang. Numerical algorithms
for mapping boundaries of manipulator workspaces. Journal of Mechanical Design,
118(2):228–234, 1996.
[56] EJ Haug, JY Wang, and JK Wu. Dextrous workspaces of manipulators. i. analytical

criteria. Journal of Structural Mechanics, 20(3):321–361, 1992.
[57] Gauthier Hentz, Isabelle Charpentier, and Pierre Renaud. Higher-order continuation for the
determination of robot workspace boundaries. Comptes Rendus Mécanique, 344(2):95–101,
2016.
[58] Jonathan Hiller and Hod Lipson. Automatic design and manufacture of soft robots. IEEE
Transactions on Robotics, 28(2):457–466, 2011.
[59] Josie Hughes, Utku Culha, Fabio Giardina, Fabian Guenther, Andre Rosendo, and Fumiya
Iida. Soft manipulators and grippers: a review. Frontiers in Robotics and AI, 3:69, 2016.
[60] Petros A Ioannou and Jing Sun. Robust adaptive control. Courier Corporation, 2012.
[61] Luc Jaulin, Michel Kieffer, Olivier Didrit, and Eric Walter. Interval analysis. In Applied
Interval Analysis, pages 11–43. Springer, 2001.
[62] Doo-Yearn Jo and EJ Haug. Workspace analysis of multibody mechanical systems using
continuation methods. 1989.
[63] Robert K Katzschmann, Cosimo Della Santina, Yasunori Toshimitsu, Antonio Bicchi,
and Daniela Rus. Dynamic motion control of multi-segment soft robots using piecewise
constant curvature matched with an augmented rigid body model. In 2019 2nd IEEE
[64] Gary L Kenaley and Mark R Cutkosky. Electrorheological fluid-based robotic fingers with
tactile sensing. In ICRA, pages 132–136, 1989.
[65] Sangbae Kim, Cecilia Laschi, and Barry Trimmer. Soft robotics: a bioinspired evolution
in robotics. Trends in biotechnology, 31(5):287–294, 2013.
[66] Pawel Kowal. A note on matrix differentiation. 2006.
[67] Chin-Hsing Kuo, Pei-Chun Lin, Terence Essomba, and Guan-Chen Chen. Robotics and
Mechatronics: Proceedings of the 6th IFToMM International Symposium on Robotics and
Mechatronics (ISRM 2019), volume 78. Springer Nature, 2019.
[68] Cecilia Laschi and Matteo Cianchetti. Soft robotics: new perspectives for robot bodyware
and control. Frontiers in bioengineering and biotechnology, 2:3, 2014.
[69] Cecilia Laschi, Matteo Cianchetti, Barbara Mazzolai, Laura Margheri, Maurizio Follador,
and Paolo Dario. Soft robot arm inspired by the octopus. Advanced Robotics, 26(7):709–727,
2012.
[70] Cecilia Laschi, Barbara Mazzolai, and Matteo Cianchetti. Soft robotics: Technologies and
systems pushing the boundaries of robot abilities. Science robotics, 1(1):eaah3690, 2016.
[71] Gui-Rong Liu and Siu S Quek. The finite element method: a practical course. Butterworth-
Heinemann, 2013.
143
[72] Carmel Majidi. Soft robotics: a perspective—current trends and prospects for the future.
Soft robotics, 1(1):5–11, 2014.
[73] Oren Masory and Jian Wang. Workspace evaluation of stewart platforms. Advanced
robotics, 9(4):443–461, 1994.
[74] Stefan Mattheis, Pia Hasskamp, Laura Holtmann, Christina Schäfer, Urban Geisthoff,
Nina Dominas, and Stephan Lang. Flex robotic system in transoral robotic surgery: the
first 40 patients. Head & neck, 39(3):471–475, 2017.
[75] Steven E McHugh, Frank J Saunders, and Jason H Rife. Dynamics-based design of a soft
robot. In ASME International Mechanical Engineering Congress and Exposition, volume
43833, pages 93–102, 2009.
[76] William McMahan, V Chitrakaran, M Csencsits, D Dawson, Ian D Walker, Bryan A Jones,
M Pritts, D Dienno, M Grissom, and Christopher D Rahn. Field trials and testing of the
octarm continuum manipulator. In Proceedings 2006 IEEE International Conference on
Robotics and Automation, 2006. ICRA 2006., pages 2336–2341. IEEE, 2006.
[77] J-P Merlet. Determination of 6d workspaces of gough-type parallel manipulator and
comparison between different geometries. The International Journal of Robotics Research,
18(9):902–916, 1999.
[78] Jean-Pierre Merlet. Parallel robots, volume 128. Springer Science & Business Media, 2005.
[79] Jean-Pierre Merlet. On the workspace of suspended cable-driven parallel robots. In
2016 IEEE International Conference on Robotics and Automation (ICRA), pages 841–846.
IEEE, 2016.
[80] Stefano Mintchev, Davide Zappetti, Jerome Willemin, and Dario Floreano. A soft robot
for random exploration of terrestrial environments. In 2018 IEEE International Conference
on Robotics and Automation (ICRA), pages 7492–7497. IEEE, 2018.
[81] Zisos Mitros, Balint Thamo, Christos Bergeles, Lyndon Da Cruz, Kevin Dhaliwal, and
Mohsen Khadem. Design and modelling of a continuum robot for distal lung sampling in
mechanically ventilated patients in critical care. Frontiers in Robotics and AI, 8, 2021.
[82] Thor Morales Bieze. Contribution to the kinematic modeling and control of soft manipula-
tors using computational mechanics. PhD thesis, Lille 1, 2017.
[83] Kirsten Morris. Linear-quadratic optimal actuator location. IEEE Transactions on
Automatic Control, 56(1):113–124, 2010.
[84] Sophie Nalbach, Rukmini Manoz Banda, Sipontina Croce, Gianluca Rizzello, David Naso,
and Stefan Seelecke. Modeling and design optimization of a rotational soft robotic system
driven by double cone dielectric elastomer actuators. Frontiers in Robotics and AI, 6:150,
2020.
[85] J Needham. Science and civilisation in china: Volume 2, history of scientific thought [en
lı́nea]. cambridge, uk: Cambridge university press. isbn 0-521-05800-7, 2005.
[86] Evar D Nering. Linear algebra and matrix theory. Technical report, 1970.
[87] Chelsea Shan Xian Ng and Guo Zhan Lum. Untethered soft robots for future planetary
explorations? Advanced Intelligent Systems, page 2100106, 2021.
144
[88] Surya G Nurzaman, Utku Culha, Luzius Brodbeck, Liyu Wang, and Fumiya Iida. Active
sensing system with in situ adjustable sensor morphology. PLoS One, 8(12):e84090, 2013.
[89] James M Ortega. The newton-kantorovich theorem. The American Mathematical Monthly,
75(6):658–660, 1968.
[90] AP Perovskii. Universal grippers for industrial robots. Russ Eng J, 60(8):3–4, 1980.
[91] Henri Poincaré. Sur l’équilibre d’une masse fluide animée d’un mouvement de rotation.
Acta mathematica, 7(1):259–380, 1885.
[92] Henri Poincaré. New methods of celestial mechanics, volume 13. Springer Science &
Business Media, 1992.
[93] Chaoxin Charles Qiu, Chi-Mei Luh, and Edward J Haug. Dextrous workspaces of
manipulators, part iii: Calculation of continuation curves at bifurcation points. Journal
of Structural Mechanics, 23(1):115–130, 1995.
[94] Mahsa Raeisinezhad, Nicholas Grant Pagliocca, Behrad Koohbor, and Mitja Trkov. Design
optimization of a pneumatic soft robotic actuator using model-based optimization and
deep reinforcement learning. Frontiers in Robotics and AI, 8:107, 2021.
[95] Singiresu S Rao. The finite element method in engineering. Butterworth-heinemann, 2017.
[96] JN Reddy. An introduction to the finite element method, volume 1221. McGraw-Hill New
York, USA, 2013.
[97] F. Renda, F. Boyer, J. Dias, and L. Seneviratne. Discrete cosserat approach for multisection
soft manipulator dynamics. IEEE Transactions on Robotics, 34(6):1518–1533, Dec 2018.
[98] F. Renda, V. Cacucciolo, J. Dias, and L. Seneviratne. Discrete cosserat approach for soft
robot dynamics: A new piece-wise constant strain model with torsion and shears. In 2016
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages
5495–5502, Oct 2016.
[99] F. Renda, M. Cianchetti, H. Abidi, J. Dias, and L. Seneviratne. Screw-based modeling of

soft manipulators with tendon and fluidic actuation. Journal of Mechanism and Robotics,
2017. doi:10.1115/1.4036579.
[100] F. Renda, M. Giorelli, M. Calisti, M. Cianchetti, and C. Laschi. Dynamic model of a

multibending soft robot arm driven by cables. Robotics, IEEE Transactions on, 30(5):1109–
1122, Oct 2014.
[101] F. Renda and L. Seneviratne. A geometric and unified approach for modeling soft-rigid
multi-body systems with lumped and distributed degrees of freedom. In 2018 IEEE
International Conference on Robotics and Automation (ICRA), pages 1567–1574, May
2018.
[102] Federico Renda, Costanza Armanini, Vincent Lebastard, Fabien Candelier, and Frederic
Boyer. A geometric variable-strain approach for static modeling of soft manipulators with
tendon and fluidic actuation. IEEE Robotics and Automation Letters, 5(3):4006–4013,
2020.
145
[103] Federico Renda, Frédéric Boyer, Jorge Dias, and Lakmal Seneviratne. Discrete cosserat
approach for multisection soft manipulator dynamics. IEEE Transactions on Robotics,
34(6):1518–1533, 2018.
[104] Federico Renda, Vito Cacucciolo, Jorge Dias, and Lakmal Seneviratne. Discrete cosserat
approach for soft robot dynamics: A new piece-wise constant strain model with torsion and
shears. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), pages 5495–5502. IEEE, 2016.
[105] Federico Renda, Michele Giorelli, Marcello Calisti, Matteo Cianchetti, and Cecilia Laschi.
Dynamic model of a multibending soft robot arm driven by cables. IEEE Transactions on
Robotics, 30(5):1109–1122, 2014.
[106] Werner C Rheinboldt. Numerical analysis of parametrized nonlinear equations. Wiley-

Interscience, 1986.
[107] Alejandro Rodrı́guez, Eulalie Coevoet, and Christian Duriez. Real-time simulation of
hydraulic components for interactive control of soft robots. In 2017 IEEE International
Conference on Robotics and Automation (ICRA), pages 4953–4958. IEEE, 2017.
[108] D.C. Rucker and R.J. Webster. Statics and dynamics of continuum robots with general
tendon routing and external loading. Robotics, IEEE Transactions on, 27(6):1033–1044,
Dec 2011.
[109] G Runge, M Wiese, L Günther, and A Raatz. A framework for the kinematic modeling of
soft material robots combining finite element analysis and piecewise constant curvature
kinematics. In 2017 3rd International Conference on Control, Automation and Robotics
(ICCAR), pages 7–14. IEEE, 2017.
[110] Gundula Runge, Jan Peters, and Annika Raatz. Design optimization of soft pneumatic
actuators using genetic algorithms. In 2017 IEEE International Conference on Robotics
and Biomimetics (ROBIO), pages 393–400. IEEE, 2017.
[111] Daniela Rus and Michael T Tolley. Design, fabrication and control of soft robots. Nature,
521(7553):467–475, 2015.
[112] Renato ME Sabbatini. Imitation of life: A history of the first robots. Brain and Mind.
Electronic Magazine on Neuroscience, (9), 1999.
[113] I Schmidt. Flexible moulding jaws for grippers. Industrial Robot: An International Journal,
1978.
[114] Nabil Simaan, Russell Taylor, and Paul Flint. A dexterous system for laryngeal surgery. In
IEEE International Conference on Robotics and Automation, 2004. Proceedings. ICRA’04.
2004, volume 1, pages 351–357. IEEE, 2004.
[115] Balkeshwar Singh, N Sellappan, and P Kumaradhas. Evolution of industrial robots and
their applications. International Journal of emerging technology and advanced engineering,
3(5):763–768, 2013.
[116] JA Snyman, LJ Du Plessis, and Joseph Duffy. An optimization approach to the deter-
mination of the boundaries of manipulator workspaces. Journal of Mechanical Design,
122(4):447–456, 2000.
146
[117] Koichi Suzumori. Elastic materials producing compliant robots. Robotics and Autonomous
systems, 18(1-2):135–140, 1996.
[118] Koichi Suzumori, Shoichi Iikura, and Hirohisa Tanaka. Development of flexible microactu-
ator and its applications to robotic mechanisms. In Proceedings. 1991 IEEE International
Conference on Robotics and Automation, pages 1622–1623. IEEE Computer Society, 1991.
[119] Koichi Suzumori, Shoichi Iikura, and Hiroshisa Tanaka. Applying a flexible microactuator
to robotic mechanisms. IEEE Control systems magazine, 12(1):21–27, 1992.
[120] Nahema Sylla, Vincent Bonnet, Frédéric Colledani, and Philippe Fraisse. Ergonomic
contribution of able exoskeleton in automotive industry. International Journal of Industrial
Ergonomics, 44(4):475–481, 2014.
[121] Deepak Trivedi, Christopher D Rahn, William M Kier, and Ian D Walker. Soft robotics: Bi-
ological inspiration, state of the art, and future research. Applied bionics and biomechanics,
5(3):99–117, 2008.
[122] Zion Tsz Ho Tse, Yue Chen, Sierra Hovet, Hongliang Ren, Kevin Cleary, Sheng Xu,
Bradford Wood, and Reza Monfaredi. Soft robotics in medical applications. Journal of
Medical Robotics Research, 3(03n04):1841006, 2018.
[123] Amehri Walid, Gang Zheng, Alexandre Kruszewski, and Federico Renda. Discrete cosserat
method for soft manipulators workspace estimation: An optimization-based approach.
Journal of Mechanisms and Robotics, 14(1):011012, 2021.
[124] JY Wang and JK Wu. Dextrous workspaces of manipulators, part 2: Computational

methods. Journal of Structural Mechanics, 21(4):471–506, 1993.
[125] Liyu Wang, Surya G Nurzaman, Fumiya Iida, et al. Soft-material robotics. Foundations
and Trends R in Robotics, 5(3):191–259, 2017.
[126] Patrick Waurzyniak. Masters of manufacturing: Richard morley. Manufacturing Engi-

neering, 131(1):35–35, 2003.
[127] Robert J Webster III and Bryan A Jones. Design and kinematic modeling of constant
curvature continuum robots: A review. The International Journal of Robotics Research,
29(13):1661–1683, 2010.
[128] James F Wilson. Robotic mechanics and animal morphology. In Robotics and Artificial
Intelligence, pages 419–443. Springer, 1984.
[129] Ke Wu and Gang Zheng. Fem-based gain-scheduling control of a soft trunk robot. IEEE
Robotics and Automation Letters, 6(2):3081–3088, 2021.
[130] Yongxin Wu and Yann Le Gorrec. Optimal actuator location for electro-active polymer
actuated endoscope. IFAC-PapersOnLine, 51(3):199–204, 2018.
[131] Holly A Yanco and Jill L Drury. A taxonomy for human-robot interaction. In Proceedings
of the AAAI Fall Symposium on Human-Robot Interaction, pages 111–119. sn, 2002.
[132] Han Yuan and Zheng Li. Workspace analysis of cable-driven continuum manipulators
based on static model. Robotics and Computer-Integrated Manufacturing, 49:240–252,
2018.
147
[133] Hongying Zhang, A Senthil Kumar, Jerry YH Fuh, and Michael Yu Wang. Topology
optimized design, fabrication and evaluation of a multimaterial soft gripper. In 2018 IEEE
[134] G Zheng, O Goury, M Thieffry, A Kruszewski, and C Duriez. Controllability pre-verification
of silicone soft robots based on finite-element method. In 2019 International Conference
on Robotics and Automation (ICRA), pages 7395–7400. IEEE, 2019.
[135] Gang Zheng, Denis Efimov, and Wilfrid Perruquetti. Design of interval observer for a
class of uncertain unobservable nonlinear systems. Automatica, 63:167–174, 2016.
[136] Joseph Zhu, Carl White, Dylan K Wainwright, Valentina Di Santo, George V Lauder, and
Hilary Bart-Smith. Tuna robotics: A high-frequency experimental platform exploring the
performance space of swimming fishes. Science Robotics, 4(34), 2019.
[137] Dominik Zunt. Who did actually invent the word “robot” and what does it mean?”. The
Karel Čapek website, 2002.
148
Résumé Substantiel
Introduction
Les robots rigides présentent de nombreux inconvénients lorsqu’ils fonctionnent dans des envi-
ronnements dynamiques et fragiles et par conséquent, les robots déformables ont été un engin
émergent qui a été graduellement étudié par les chercheurs afin de surmonter ces limitations et
de faire face à de nouvelles applications robotiques. Les robots déformables sont fabriqués à
partir de matériaux souples et flexibles, ce qui leur permet d’avoir plusieurs caractéristiques
telles qu’une grande dextérité, des collisions prudentes et sans danger ainsi qu’une flexibilité im-
portante. Toutes ces fonctionnalités offrent de nombreux avantages pour différentes applications,
notamment l’exploration de l’environnement [11, 37, 46, 80, 87, 136] et les opérations médicales
[23, 25, 33, 34, 54, 74, 81, 114, 122].
Cependant, en raison de leur conformité naturelle, la modélisation des robots souples est
plus complexe que celle des robots rigides, car les robots déformables comportent un nombre
élevé de degrés de liberté, leur déformation est non linéaire et ils sont caractérisés par des lois
mécaniques différentes de celles des robots rigides. Par conséquent, des problèmes scientifiques
tels que la détermination de l’espace de travail et l’optimisation de la conception des robots
déformables émergent et avec eux les possibilités de nouvelles contributions dans le domaine de
la robotique souple.
Même si les robots mous ont fait des progrès intéressants au cours des dernières années, le
processus de conception d’un robot mou s’inspire toujours principalement d’une série de systèmes
biologiques [111], notamment la structure et le comportement d’espèces animales telles que la
trompe des éléphants [53], et les bras d’une pieuvre [69]. Une telle procédure de conception peut
être utile pour les tests initiaux et l’expérimentation du robot souple conçu. Cependant, lorsqu’il
est confronté à des objectifs de performance spécifiques tels que la planification et le contrôle
de la trajectoire [35, 129], il est possible que la conception d’un tel robot souple ne soit pas en
mesure d’atteindre ses objectifs, notamment parce que sa plage d’accessibilité (c’est-à-dire son
espace de travail) peut être restreinte.
Par conséquent, il est utile pour la communauté de la robotique douce de proposer des
méthodologies génériques afin d’évaluer l’espace de travail des robots souples, et également
d’aider et de guider la conception systématique de robots souples dans le but d’optimiser des
objectifs de performance spécifiques.
L’estimation de l’espace de travail en robotique douce reste un sujet ouvert, et son importance
est due aux nombreux avantages qu’elle peut apporter pour résoudre différents défis scientifiques
de la robotique douce, principalement liés à la conception mécanique du robot et à la synthèse
du contrôleur. D’une part, le résultat de l’évaluation de l’espace de travail peut fournir des
informations sur l’accessibilité de l’objet à contrôler [35, 38, 129], c’est-à-dire en identifiant si la
position de l’objet appartient à l’espace de travail du robot souple, épargnant ainsi au contrôleur
149
le temps de conception qui peut être perdu à essayer de contrôler un objet inaccessible en
dehors de l’espace de travail du robot. D’autre part, les informations sur l’espace de travail sont
également cruciales lors de la planification d’une trajectoire réalisable à suivre par des robots
mous, car les positions de départ et d’arrivée d’une trajectoire réalisable doivent appartenir à
l’espace de travail d’un robot [129]. En outre, l’analyse de l’espace de travail est également une
étape nécessaire pour contribuer à une conception efficace, ciblée et optimale des robots mous
[58], en améliorant leur stabilité et en augmentant la portée de leur manipulabilité et de leur
accessibilité.
Pour atteindre cet objectif, la thèse actuelle adopte deux modèles mathématiques différents,
la méthode de déformation constante par morceaux (PCS) et la méthode des éléments finis
(FEM) pour décrire la déformation des robots mous de forme élancée et de forme générale,
respectivement. Sur la base de ces modèles mathématiques adoptés, plusieurs approches sont
proposées pour estimer l’espace de travail des robots mous. Enfin, une méthode efficace
d’optimisation de la conception des robots mous en vue d’atteindre des objectifs de performance
spécifiques est présentée.
Modélisation
Dans ce chapitre, nous avons présenté les modèles mathématiques des méthodes PCS et FEM
dans le but de modéliser les robots mous ayant une forme élancée et une forme générale,
respectivement.
Ensuite, nous avons établi la définition de l’espace de travail d’un robot souple.
Enfin, nous avons proposé une approche directe pour estimer l’espace de travail des robots
mous. Cette approche a ensuite été validée en utilisant un robot souple de type tronc [129] pour
les cas PCS et FEM.
Approche d’optimisation pour l’estimation de l’espace de

travail
En raison de la complexité de calcul élevée de l’approche directe lors de l’estimation de l’espace
de travail des robots mous, il est logique de proposer une approche qui peut discrétiser l’espace
de l’effecteur final puisque sa dimension est plus petite et constante par rapport à l’espace des
actionneurs pour les robots hyper-redondants, tels que les robots mous. De plus, il est plus
efficace de cartographier uniquement la limite extérieure de l’espace de travail et d’éviter le
calcul lourd de ses points intérieurs.
Une telle méthode [9,123] consiste à discrétiser l’espace de l’effecteur final afin de cartographier
la limite extérieure de l’espace de travail, et à déduire que la zone/volume délimitée par la limite
de l’espace de travail est l’espace de travail accessible.
L’approche proposée a été appliquée avec succès aux modèles PCS et FEM, où nous avons
montré son efficacité à réduire la complexité et le temps de calcul nécessaires pour estimer l’espace
de travail des robots mous, contrairement à l’approche directe qui explose exponentiellement
lorsque la dimension des actionneurs augmente.
D’autre part, comme cette approche ne consiste qu’à cartographier la limite extérieure de
l’espace de travail en déterminant les points limites situés sur la surface entourant l’espace de
travail, elle ne sera certainement pas en mesure d’identifier les informations internes de l’espace
de travail telles que les limites intérieures.
150
Approche d’analyse intervalle pour l’estimation de l’espace
de travail
En raison des limitations rencontrées par l’approche d’optimisation qui ne peut pas fournir de
connaissances sur les configurations intérieures de l’espace de travail, il est donc nécessaire de
fournir une méthode qui peut surmonter cette limitation, mais qui est également basée sur la
deuxième stratégie, c’est-à-dire la discrétisation de l’espace de l’effecteur.
Une telle méthode [7] est basée sur des techniques d’analyse d’intervalles [48, 61, 77–79], et
consiste à discrétiser l’espace de l’effecteur en partant d’une configuration initiale atteignable
puis en explorant l’ensemble de l’espace atteignable possible pour finalement estimer l’espace de
travail d’un robot souple.
L’approche proposée a été appliquée avec succès aux modèles PCS et FEM, où nous avons
montré son efficacité à réduire la complexité et le temps de calcul nécessaires pour estimer l’espace
de travail des robots mous, contrairement à l’approche directe qui explose exponentiellement
lorsque la dimension des actionneurs augmente.
D’autre part, comme cette approche consiste à déterminer toutes les configurations possibles
qui sont réalisables, elle est utile pour identifier les informations intérieures et extérieures de
l’espace de travail. Cependant, cette approche est exhaustive dans le sens où elle explore toutes
les configurations réalisables de l’espace de travail au lieu de cartographier uniquement ses
limites intérieures et extérieures.
Approche de continuité pour l’estimation de l’espace de

travail
En raison des limites rencontrées par l’approche d’optimisation qui ne peut pas fournir la
connaissance des configurations intérieures et des limites de l’espace de travail des robots mous,
et également en raison des limites de l’approche d’analyse d’intervalle qui est exhaustive en
explorant l’espace de travail entier, il est logique de fournir une méthodologie alternative nous
permettant de déterminer seulement mais toutes les limites des espaces de travail des robots
mous, à la fois intérieurs et extérieurs.
Une telle méthode est appelée l’approche de continuation [8] qui est basée sur la théorie de
la bifurcation. Cette méthode est fondée sur la théorie mathématique de la bifurcation qui a
été introduite à l’origine par le mathématicien français Henri Poincaré [91], et elle consiste à
étudier les solutions d’équations différentielles non linéaires, en plus de l’étude des variations
possibles de la structure topologique d’une famille de champs de vecteurs. Par la suite, Henri
Poincaré a classé différents types de points de bifurcation [92]. En conséquence, cette approche
prend en compte le comportement possible de bifurcation dans le calcul de continuation tout en
cartographiant les limites intérieures et extérieures des espaces de travail des robots logiciels.
L’approche de continuation permet de cartographier les frontières intérieures et extérieures
des robots mous aux formes élancées en utilisant le modèle PCS. Cependant, il est important de
préciser que cette approche ne peut pas fournir d’informations sur tous les points de l’espace de
travail, mais seulement sur les points limites.
Otpimisation de la conception des robots mous

Le processus actuel de conception d’un robot souple fait toujours appel à l’intuition et à la
procédure d’essai-erreur. Si l’on considère des objectifs de performance spécifiques, l’approche
151
classique de la conception des robots mous est associée à l’incertitude d’atteindre de tels objectifs
et aux dépenses économiques substantielles nécessaires pour les essais et les erreurs. Dans le
but d’atteindre les objectifs souhaités, il est donc logique, pour des raisons économiques et
scientifiques, d’optimiser la conception des robots mous dans un environnement virtuel avant de
procéder à sa conception physique finale.
Étant donné un robot souple composé d’un nombre fini de segments de longueur limitée, et
entraı̂né par des actionneurs montés de longueur limitée, et de magnitude limitée, ce chapitre
présente une méthode d’optimisation permettant d’optimiser la conception d’un robot souple
afin d’optimiser l’accessibilité de leur espace de travail.
L’approche proposée a ensuite été validée sur différentes configurations de robots souples
entraı̂nés par des tendons.
152
Estimation de l’Espace de Travail et l’Optimisation de la Conception des Robots
Déformables
Résumé
Les robots rigides présentent de nombreux inconvénients lorsqu'ils fonctionnent dans des
environnements dynamiques et fragiles et par conséquent, les robots déformables ont été un engin
émergent qui a été graduellement étudié par les chercheurs afin de surmonter ces limitations et de
faire face à de nouvelles applications robotiques.
Les robots déformables sont fabriqués à partir de matériaux souples et flexibles, ce qui leur permet
d'avoir plusieurs caractéristiques telles qu'une grande dextérité, des collisions prudentes et sans
danger ainsi qu'une flexibilité importante. Toutes ces fonctionnalités offrent de nombreux
avantages pour différentes applications, notamment l'exploration de l'environnement et les
opérations médicales. Cependant, en raison de leur conformité naturelle, la modélisation des
robots souples est plus complexe que celle des robots rigides, car les robots déformables
comportent un nombre élevé de degrés de liberté, leur déformation est non linéaire et ils sont
caractérisés par des lois mécaniques différentes de celles des robots rigides.
Par conséquent, des problèmes scientifiques tels que la détermination de l'espace de travail et
l'optimisation de la conception des robots déformables émergent et avec eux les possibilités de
nouvelles contributions dans le domaine de la robotique souple. L'évaluation de l'espace de travail
offre de nombreux avantages pour différentes applications de la robotique souple principalement
liées à leur conception et à leur contrôle. En conséquence, cette thèse étudie l'estimation de
l'espace de travail et l'optimisation de la conception des robots souples.
Afin d'accomplir cette tâche, on propose deux méthodes différentes pour la modélisation des
robots déformables, la première est la méthode de déformation constante par morceaux (PCS),
qui est utilisée pour la modélisation des robots déformables avec une géométrie continue et la
deuxième est la méthode des éléments finis (FEM), qui est utilisé pour la modélisation des robots
déformables avec une géométrie générale.
Ensuite, basée sur ces modèles mathématiques, cette thèse propose différentes méthodologies
pour estimer l'espace de travail des robots souples. Deux stratégies ont été proposées pour
l'estimation de l'espace de travail, la première consiste à discrétiser l'espace des entrées
(actionneurs), et la deuxième consiste à discrétiser l'espace des sorties (l'effecteur du robot).
Cependant, la première stratégie est inefficace, car elle dépend de la dimension des actionneurs
et celle-ci varie en fonction de la configuration du robot déformable étudiée.
En revanche, la seconde stratégie présente une méthodologie stable et efficace pour l'estimation
de l'espace de travail puisque l'espace de l'effecteur est toujours constant (inférieur ou égal à 3, si
nous nous concentrons sur l'aspect position de l'espace de travail), quelles que soient la
configuration du robot souple étudié et la dimension des actionneurs.
Les approches proposées pour l'estimation de l'espace de travail ont ensuite été appliquées aux
deux modèles mathématiques adoptés et validées à l'aide de différentes configurations de robots
déformables.
Enfin, cette thèse propose une approche d'optimisation basée sur les modèles mathématiques
adoptés pour optimiser la conception des robots souples afin d'atteindre certains objectifs
spécifiques.
Mots-clefs : Robot Mou, Modèle Cinématique/Dynamique, Espace de Travail, Optimisation.

Workspace Estimation and Design Optimization of Soft Robots
Abstract
Soft robots are an emergent instrument that has gradually been investigated by researchers in the
recent years to overcome limitations of traditional rigid robots as well as to propose novel robotic
applications. Rigid robots are challenged when operating in restricted and dynamic environments.
Being made from a soft and flexible material, soft robots provide many benefits such as high
dexterity, safe interactions, and increased adaptability, which are very useful for various
applications, especially the manipulation of fragile objects, environment exploration, and medical
operations.
However, owing to their natural conformity and compliance, soft robots consist of a large number
of degrees of freedom, their deformation is highly nonlinear, and they are characterized by
different mechanical laws compared to that of rigid robots. All these aspects make their modeling
more complex.
Consequently, scientific challenges such as workspace evaluation and design optimization of soft
robots arise and with them the opportunities of new contributions in the field of soft robotics.
The workspace evaluation provides many benefits for different soft robotic applications mainly
related to their design and control.
Accordingly, the present thesis investigates the workspace evaluation and design optimization of
soft robots.
For this, two different methods are adopted for the mathematical modeling of soft robots. The
first is the Piece-wise Constant Strain (PCS), which focuses on slender-shaped soft robots, and
the second is the Finite Element Method (FEM), which concerns soft robots with a general shape.
Based on the adopted mathematical models, this thesis proposes different methodologies to
estimate the workspace of soft robots.
In fact, two strategies can be followed for the workspace estimation of soft robots: the first
consists of discretizing the inputs (actuators) space, and the second consists of discretizing the
outputs (end-effector) space.
However, the first strategy is not efficient as it depends on the dimension of the actuators which
are used to control the investigated soft robot, and varies corresponding to the particular studied
configuration.
Conversely, the second strategy presents a stable and efficient way for the workspace estimation
since the end-effector's space is always constant (smaller or equal to 3, if we focus on the position
access) regardless of the studied soft robot's configuration and the dimension of actuators.
Each proposed workspace estimation approach was then applied to both the PCS and the FEM
models and validated via a variety of soft robots' configurations.
Finally, this thesis proposes a model-based optimization approach in order to optimize the design
of soft robots for the purpose of achieving specific performance objectives.
Keywords: Soft Robots, Kinematic/Dynamic Modeling, Workspace Estimation, Optimization.

Amehri Walid (1)

Uploaded by

Copyright:

Available Formats

Amehri Walid (1)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Amehri Walid (1)

Uploaded by

Copyright:

Available Formats

Workspace Estimation and Design Optimization of Soft

To cite this version:

HAL Id: tel-04028186

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

Traitement du Signal et des Images

Estimation de l’Espace de Travail et l’Optimisation de la Conception

Workspace Estimation and Design Optimization of Soft Robots.

Soutenue le 28 Février 2022 devant le jury d’examen :

Président Philippe Fraisse Professeur, Université de Montpellier

Thèse préparée dans le Centre de Recherche en Informatique, Signal et Automatique de Lille,

Ecole Doctorale MADIS-631

Membre du Groupe des Écoles Centrale

2 Modeling and Forward Methods 27

3 Optimization-Based Approach for Workspace Estimation 45

4 Interval Analysis Approach for Workspace Estimation 67

5 Continuation Approach for Workspace Estimation: PCS case 89

6 Design Optimization: PCS case 115

Many thanks to the reviewers of my manuscript and jury members.

1.1 General Introduction

Figure 1.1: Classification of robots [44, 52, 121].

1.1.2 Rigid Robotics and the Rise of a New Robot Class

1.1.3 Soft Robotics

1.1.4 Challenges and Motivations

1.2 State of the Art

1.2.1 Mathematical Modeling of Soft Robots

1.2.1.1 Piece-wise Constant Curvature

1.2.1.2 Finite Element Method

Figure 1.5: FEM modeling approach [39, 47].

1.2.1.3 Piece-wise Constant Strain

Figure 1.6: PCS modeling approach [103, 104].

1.2.1.4 Modeling Methods Comparison

1.2.2 Workspace Evaluation

1.2.2.1 Rigid Robotics Workspace Methodologies

Regarding numerical methods, an optimization approach to determine the boundaries of rigid

1.2.2.2 Continuum Robotics Workspace Approaches

1.2.2.3 Soft Robotics Workspace Methods

1.2.2.4 Workspace Methods Summary

1.2.3 Design Optimization of Soft Robots

1.2.3.1 Heuristic Methods

1.2.3.2 Model-Based Approaches

As to model-based approaches, in [133] (see A in Fig. 1.10), a density-based topology optimization

1.2.3.3 Hybrid Approaches

Figure 1.11: Hybrid approaches [94].

1.3 Contributions of the Thesis

1.3.1 Contributions and Organization of the Manuscript

Modeling and Forward Methods

2.2 Piece-wise Constant Strain (PCS)

gi1 “ gi ξˆi (2.2a)

ξi “ Bqi qi ` ξi˚ (2.5)

gi pXq “eX ξ̂i (2.6a)

η̂i pXq “gi´1 g9 i (2.7c)

have been defined in [103].

2.2.2 Continuous Dynamics

2.2.3 Elasticity and Actuation Forces

Figure 2.2: Schematic of the tendon-driven actuation of one body.

This is given by:

Remark 1. Considering di as a small parameter, a useful approximation can be obtained, which

2.2.4 Discrete Dynamics

2.3 Finite Element Method (FEM)

2.3.2 Displacement Model