1

Soft Robots Modeling: a Literature Unwinding

Costanza Armanini1 , Conor Messer2 , Anup Teejo Mathew1 , Frédéric Boyer3 , Christian Duriez4 and
Federico Renda∗1,5 ,

Abstract—The robotics community has seen an exponential An efficient theoretical framework is crucial for the pre-
growth in the level of complexity of the theoretical tools presented diction of the robot’s behaviour, the optimization of the
for the modeling of soft robotics devices. Different solutions have design and finally for control. While the kinematic modeling
been presented to overcome the difficulties related to the modeling
of soft robots, often leveraging on other scientific disciplines, of traditional rigid-link robots is fully defined by the link
dimensions and the joint coordinates, the virtually infinite
arXiv:2112.03645v1 [cs.RO] 7 Dec 2021

such as continuum mechanics and computer graphics. These

theoretical foundations are often taken for granted and this lead number of degrees of freedom of continuum robots drastically
to an intricate literature that, consequently, has never been the increases the complexity of their modeling. Analytic closed
subject of a complete review. Withing this scenario, the objective form solutions are available only for few cases characterized
of the presented paper is twofold. The common theoretical roots
that relate the different families of modeling techniques are by basic loading condition and mostly one dimensional cases.
highlighted, employing a unifying language that ease the analysis For this reason, one of the biggest effort in the modeling of soft
of their main connections and differences. Thus, the listing of the robots is the investigation of simplified assumptions and model
approaches naturally follows and a complete, untangled, review reductions that are accurate enough to predict the robot’s
of the main works on the field is finally provided. behaviour and, at the same time, that are computationally effi-
cient. The inevitable trade-offs between the model complexity,
I. I NTRODUCTION its computational cost and accuracy are the primary principles
The term soft robot appeared for the first time in a scientific that should be considered when modeling continuum robots.
paper in 2000, describing a McKibben prenumatic artificial Within this scenario, the contribution of this paper is
muscle [1], a family of braided pneumatic actuators developed twofold. On one side, we present a complete literature review
in the 50s to assist polio patients. Even though they were of the modeling techniques that have been presented so far,
not explicitly called soft robots, McKibben actuators probably spamming from the more general approaches such as Finite
represent the first example of a robotic device exploiting its Element Method to those that were specifically conceived for
compliance to achieve improved performances with respect to soft robotics applications.
their exclusively rigid counterpart. Since then, soft robotics On the other side, we present the general background
has been one of the fastest growing research community in and organize the defining characteristics of these modeling
the last decades [2], [3]. techniques presented in the literature. In this regard we go
Many different soft robotics devices and actuators have beyond a simple identification of already published work in
been presented, ranging in almost every possible techno- the field and attempt to untangle such an intricate research
logical field, from biomedical engineering to aerospace and topic. This is to provide the reader with the basic knowledge
underwater robotics. The increasing interest in soft robotics on the theoretical foundations of the published works.
is demonstrated by the vast number of review papers that The remaining parts of the paper are organized as follows.
have been published to summarize the employed techniques, In Section II some general concepts are drawn and the
the achievements and the future prospects of this promising classification employed in the paper to discuss all the mod-
research field [4], [5], [6], [2], [7], [8], [9], [10]. Among eling approaches is presented. In Section III we present the
the available reviews published with regards to the theoretical approaches that are based on continuum mechanics theories,
modeling components, some have been published for specific that are also briefly presented to the reader. In Section IV, we
groups of approaches, [11], [12], [13], [14], [15], some for present the so-called geometrical approaches, which rely on
specific application fields [16] and other for specific families of the hypothesis that the deformed shape of the robot follows
robot [17]. However, a comprehensive review on the modeling a specific geometrical curve. Proceeding with the level of re-
tools that have been presented to efficiently describe the static duction, in Section V the approaches that arise from a specific
and dynamics of soft robots is still lacking. discretization of the soft body are presented, while in Section
VI, surrogate (also called data driven) models are discussed.
* Corresponding author [email protected] Finally, Section VII presents the softwares implementation the
1 Department of Mechanical Engineering, Khalifa University of Science
proposed approaches, while in Section VIII the conclusion and
and Technology, Abu Dhabi, UAE.
2 Fulbright Student Researcher Program, Khalifa University of Science and the future perspectives are draw.
Technology, Abu Dhabi, UAE
3 LS2N Laboratory, Institut Mines Telecom Atlantique, Nantes, France
4 Inria, CNRS, Centrale Lille, Team DEFROST, University Lille, Lille, II. C HARACTERIZATION
France
5 KUCARS, Khalifa University of Science and Technology, Abu Dhabi, The modeling of a mechanical system is the mathematical
UAE. representation of the considered physical problem. In general,
 2

the model of a mechanical system is defined by the following which all rely on the assumption that the deformed shape
features [18], [19]: of the body resembles a theoretical space curve repre-
• Configuration: the set of position of the system’s particles
sented by a specific mathematical function. In this case,
compatible with the internal kinematic constraints (e.g. the generalized coordinates are usually absolute in nature.
joints) and external ones (geometric boundary conditions). The second group includes the widely known piecewise-
At any given time, the configuration of a mechanical system constant-curvature (PCC) models, which are based on the
defines a subset of the Euclidean space. discretization of the continuous soft body in a finite number
• Generalized Coordinates: a set of continuous or dis-
of sections having circular arc shape, with intrinsically
crete parameters describing any configuration. They are relative coordinates. This is by far the most numerous group
the coordinates of a unique point in an abstract space or of theoretical model proposed in the last decades, even
"configuration space". We call them absolute coordinates though they proved to be reliable only for specific loading
when they refer to a fixed inertial frame, relative coordinates conditions and geometries.
when they refer to a moving frame covaring together with 3) Discrete material models. As the name itself suggests,
the system. these models are based on a discretization of the continuous
• Kinematic map: it takes as input the generalized coordi-
body in a finite number of discrete material components.
nates and it returns the configuration of the system. This is As such, they are characterized by an a priori finite-
also called forward kinematics, while the inverse kinematics dimensional configuration space of absolute and/or relative
represents exactly the opposite process, i.e. the calculation coordinates, the relative ones being usually preferred in
of the coordinates (typically, the actuation ones) required to practice. Once the discretization is defined, different paths
obtain a specific configuration of the system. can be followed to obtain the kinematics of the system.
• Kinetic and Potential Energy: the functions that determine
Some models are based on traditional approaches employed
the fundamental dynamics of the system. In addition, ap- for the modeling of rigid link robots. Other models, on the
plied non-conservative forces and dissipation internal forces other hand, are based on the discretization of the soft body
may act on the system through their virtual work. In the in a finite number of discrete elastic rods.
context of soft robotics, the potential energy is of particular 4) Surrogate models. In this family of approaches, sometimes
importance as it can strongly characterize the approach. also referred to as ”data-driven”, the configuration of the
system is obtained employing sets of data and a learning
It should be noted that these features are not always clearly process. The great majority of the approaches falling in this
expressed in the developing of a theoretical model, as some- group use neural networks models and machine learning
times they might be implicitly posed by the assumptions of algorithms, or deep learning.
the model itself. One of the aims of this work is to shed light
on the underline structure of the models proposed in literature
using a unified language.
Based on these features, in the present paper, the theoretical
tools employed for the modeling of soft robots are classified
as follows:
1) Continuum Mechanics models. They are all characterized
by a continuous (infinite-dimensional) configuration space
of the system, and on physical considerations about the soft
bodies deformations. As such, the models in this category
benefit from a physically rigorous definition of the kinetic
and potential energy of the system. This family of models
can employ both absolute and relative coordinates and Fig. 1: A schematic of three representations provided by the different
families of approaches that were distinguished.
they are derived from continuum mechanics theories. When
no specific assumptions are made, they are based on the This classification has been defined to ease the explanation of
classical three-dimensional continuum mechanics theory. the models, in an effort to group them based on the different
On the other hand, other (reduced) approaches have been path to discretization that they employ to tackle the presence
presented for surface structures (shells, membranes) or slen- of the non-linearities intrinsically entailed in the modeling of
der structures (beam, rods). In particular, the latter include soft bodies. This could happen from the very first steps of
Cosserat, Kirchhoff and non-linear Euler Bernoulli beam the definition of the kinematics of the body, such as for the
theories, which are frequently employed in soft robotics. Discrete models, or when the involved equations are solved,
2) Geometrical models. These models are all based on some which is the case of continuum mechanics models. On the
sort of geometrical assumptions on the deformed shape other end, surrogate models often work around the problem of
undertaken by the soft body when specific loads are applied. the definition of a mechanical model through the employment
For this group, the central role is taken by the generalized of machine-learning algorithms. Clearly, this is only one of
coordinates, on which the system’s kinetic and potential the classifications that can be conceived and, in some cases,
energy are defined. In particular, two main groups emerged. it is possible to note some overlapping between the models,
The first one includes the so-called ”functional” models, that we will try to underline.
 3

III. C ONTINUUM M ECHANICS M ODELS move in Section III-B to the so-called ”director approaches”,
which replace the continuum medium to a rod as an oriented
Soft robotics is an extremely interdisciplinary field and
body [20]. This is characterized by a flexible curve and a
continuum mechanics represents the most influential com-
continuous set of rigid cross sections of infinitesimal thickness,
munity for the theoretical modeling of soft robots. Classical
represented by a triad of orthonormal basis vectors (directors).
elasticity theories have been used for centuries to precisely
These approaches are all based on the Cosserat rod theory
model the mechanics of continuum bodies and thus they offer
which leads, which encompasses the Kirchhoff and large
an established and general framework that is already available
deflections (non-linear) Euler Bernoulli theories. The main ad-
to the soft robotics community. Thus, one of the main goal
vantage of these approaches for the modeling of soft robots is
when employing these approaches in the robotics community
that, in principle, they do not require any hypotheses about the
consists in making them computationally efficient for the
through-the-thickness distributions of displacement and stress
considered purpose, avoiding any unnecessary redundancy.
fields [21]. For the same reason, however, the formulation of
Before going into the details of these theories, let us draw
general constitutive laws appears as more complicated with
some general considerations on the main steps that are required
respect to traditional three dimensional theories.
to obtain a model for a continuum body. The first important
remark that should be pointed out is that, while traditional rigid
robots can be fully represented by some finite discrete set of A. Classical 3D Models
frames, in soft robotics the medium is a continuum of particles. Defining a (soft) body as a set of material particles Ω
Even though this consideration is mostly taken for granted, labelled by their (material, in general curvilinear) coordinates
it represents the starting point of all the following theories. X, the aim of any three dimensional theory is to predict the
In contrast to the Eulerian view point of fluid mechanics, time-evolution of its configuration r defined by:
in the Lagrangian description of solid continuum mechanics,
the body’s configuration is parameterized through positional r(·) : X ∈ Ω 7→ r(X) ∈ R3 (1)
fields, which depend both on the material coordinates and on
In order to proceed, the first step consists in obtaining (and
time. In particular, the material coordinates X represent the
solving) the balance equations of the system. Shall we consider
coordinates of the particles of the body when it is in an ar-
a material subpart of the body B ⊂ Ω, having frontier ∂B and
bitrary configuration, which is called reference configuration.
outward unit normal n ∈ R3 . In general, volume and surface
The motion of a continuum body is defined by a continuous
forces are applied upon B, and modelled by some volume
sequence of configurations along time. Differentiating once the
and surface vector densities, b ∈ R3 and t ∈ R3 respectively.
positional field with respect to time provides the instantaneous
While the first represents the external forces acting on the
flow velocity of the material particle, while the acceleration
is obtained differentiating twice. A change in the configura-
tion of a continuum body results in a displacement, which
usually has two components: a rigid-body displacement and a
deformation. To describe the internal deformation state of the
body, it is necessary to define its strain (time) rate, which is a
combination of the gradient of the velocity fields. Definition
of strains needs to be objective, i.e., observer independent.
Finally, beyond kinematics, the closed formulation describing
the time evolution of a continuum medium is given by: Fig. 2: Scheme of the considered body
• a principle of the dynamics (Newton’s Law, d’Alembert’s volume, the latter represents the forces that B exchanges with
principle, Hamilton’s principle), providing the PDEs relating its surroundings (the other parts of Ω) across its boundary
the stress with the acceleration of the particles and the ∂B, which are also known as internal forces. In particular, t
external forces applied inside and across the boundaries of is called stress vector and we admit the Cauchy theorem, i.e.
the medium; t is a linear function of n:
• a set of geometric boundary conditions (BCs);
• a constitutive law that relate the time evolution of the stress t = σn (2)
to that of the strains.
where σ ∈ R3 × R3 is the Cauchy’s stress tensor. It is then
In general, such closed formulations are extremely nonlinear possible to express the balance equations for any part B of Ω,
and one of the biggest challenges is the characterization of the thus the force resultant is null:
simplifications strategies to solve them. In the following, we Z Z
will start considering in Section III-A the models that are ob- (b − ρv̇) dV + tdS = 0 (3)
tained from the three-dimensional elasticity theory, treating the B ∂B

soft body as a continuum medium. This represents the more where ρ is the mass density, while v is the body velocity.
general procedure, which however leads to high dimensional, Using Cauchy (2), and divergence theorems yields:
non-linear equations. In this context, Finite Element Method Z !
represents by far the most popular approach to solve these ∂σji
bi + − ρv̇i dV = 0 (4)
equations in a computationally effective way. We will then B ∂xj
 4

This relation should held for any volume B, and finally the in the de Saint Venant’s (1855) principle, which however are
Cauchy equilibrium equations are obtained: not usually employed within the soft robotics community. The
most popular numerical approach to solve three dimensional
∇ · σ + b = ρv̇ (5)
continuum problems is the Finite Element Method.
In order to obtain the configuration of the body from 1) Finite Element Method: The Finite Element Method
the solution of (5), it is necessary to define a constitutive (FEM) is a well-known and widely-spread numerical technique
relation characterizing the response of the body’s material for finding approximate solutions to Partial Differential Equa-
under external forces (or, more generally, any external stimuli) tions (PDEs). It was established by a set of scientific papers
and providing the relationship between the stress tensor σ and in the 40s and it soon became one of the most commonly
the strains. There are different ways to define a constitutive employed technique for the modeling of a wide range of
relation and the first and most famous example is the Hooke’s engineering problems, from structural mechanics to fluid flow
law for linear elastic materials, which in the 1D case reads and heat transfer. The main property of this technique is the
σ = E. One other more general constitutive relation has sub-division of the problem’s domain into a set of smaller
been proposed by Green (1839, 1842), which employs the parts, called finite elements. These are obtained by the con-
concept of strain energy function. More precisely, a Green struction of a mesh which represents the numerical domain for
elastic material (also known as hyperelastic material) is a type the solutions, and defines the discrete variables of the system
of constitutive model for which the stress–strain relationship as the value of the position field (and its derivatives) at the
can be obtained from a strain energy density function (or edges of the elements, called nodes. Once this discretization
stored-energy function). This function depends symmetrically is performed, the continuous field r is approximated on each
on the principal stretches of the deformation λ1 , λ2 and λ3 , element by a polynomial interpolation of its values at the
which, for incompressible materials satisfy the constrain I3 = nodes according to a Ritz reduction approach. Introducing
λ1 λ2 λ3 = 1, where I3 represents the third strain invariant. this discretization in the above continuous formulation and
It is worth to remind that the stretch ratios λi are defined as projecting it on the same trial (polynomials) basis, changes
the ratio between the stretched length Li and the undeformed the PDEs (5) into a set of algebraic discrete equations for
one L0,1 . Many soft robotics devices and components are steady state problems, or ODEs for transient problems. In the
realized with rubber-like materials, which, in the case of former case, the static equilibrium is obtained as a system of
static deformations, are often treated as hyperelastic (while, algebraic equations in the form:
in general, they can exhibit other nonlinear behaviors such as
hysteresis, viscoelasticity and stress softening). In particular, Qint (q) = Qext (q) (8)
the Ogden [22] material model represents one of the more where Qint (q) and Qext (q) represent the internal and external
general framework for the modeling of hyperelastic materials generalized forces, respectively, while q are the generalized
and it can reduce to the Neo-Hookean and Mooney–Rivlin coordinates, which in the FEM formulation are the nodes
solids. In the Ogden material the strain energy density is displacements. On the other hand, the dynamic equilibrium
expressed in terms of the principal stretches as: is given by a system of ODEs:
n
X µj αj α α
(
W (λ1 , λ2 , λ3 ) = (λ1 + λ2 j + λ3 j − 3) (6) q̇ = v
α (9)
j=1 j M (q)v̇ + Qint (q, v) = Qext (q, v)
where n, µj and αj are material constants. In particular, for
n = 1 and α = 2, the Neo-Hookean solid is obtained, while where M (q) is the generalized mass matrix. Note that the
the Mooney-Rivlin material for an incompressible material is terms appearing in (8) and (9) are computed integrating
obtained for n = 2, α1 = 2, α2 = −2. Once the strain energy the distributed variables element-by-element according to the
function is defined, the principal Cauchy stresses σ1 , σ2 and selected shape functions, which play the role of the kinematic
σ3 are related to the stretches through the equations: map, while the position of each element in the mesh is ensured
by the assembly process.
∂W Since nodal coordinates are of absolute nature, the non-
σi = λi − p, (7)
∂λi linearities appears in Fint (q, v). They are named geometric
where p represents the Lagrangian multiplier associated with or material, depending whether they come from the large dis-
the incompressibility constraint. It is worth to remind that placements (and namely the rotations) of the elements, or from
the Cauchy principal stresses are related to the corresponding the constitutive laws. To overcome the geometric non-linearity,
nominal (or engineering) stresses σe by the relationship σe,i = corotational approaches are used in three-dimensional FE
σi λ−1
i . The great majority of the materials that are employed softwares. One other solution to represent the non-linearities is
within the soft robotics community can be considered as hy- the employment of an absolute nodal coordinate formulation
perelastic. However, most soft robotics modeling approaches [23] .
still rely on a linear-elastic material assumption, as they are The FEM represent a standard for many popular simulation
mostly focused on the description of the large deformation of software, such as Abaqus, ANSYS, Comsol. The main ad-
the body, rather than the large strains. Many classical theories vantage of these software is that they provide a powerful and
of rods have been derived from (5) by employing simplifying general tool that can be applied to a wide spectrum of physical
assumptions on their average solution over the cross section, as problems ranging from structural dynamics, fluid-structure
 5

interactions, contact and thermodynamics. They provide a method, leading to the so-called directors approaches. Indeed
ready made and user-friendly framework that can be easily the prototype of a slender body is a cylinder whose diameter
used for specific studies on multiphysics systems. They have is much smaller when compared to its length. Thus, a solution
been widely used to model soft robots [24], [25] (Abaqus), to the problem can take the form of a limit of the three dimen-
[26] ANSYS. At the same time, their generality intrinsically sional solution when the diameter of the cylinder approaches
entail an increased computational cost and the simulations, zero, while the resultants of loads and the global stiffness of
especially in dynamics, can take a lot of time to converge, the body tend to finite limits [20]. In order to properly model
sometimes even days. bending and twisting, the line is endowed by directions that
There are also some popular examples of ad-hoc Finite can rotate independently of the deformation of the line itself.
Element approaches (and softwares) that have been specifically The directions and the associated axis are called directors
proposed for (soft) robotics applications. In 2007 [27], a group and a material curve combined with a collection of directors
of researchers from different institutes released Sofa (Sim- associated to each particle is called a Cosserat rod (Cosserat
ulation Open Framework Architecture) an open-source C++ and Cosserat, 1907 [35]). In other words, the Cosserat model
library which was originally presented as a computational en- treats the slender body as a deformable curve in which every
vironment for medical simulations [28], [29]. In the following particle is rigidly connected to a triad of orthonormal vectors
years, Sofa became a comprehensive high-perfomance library (directors) to characterize its orientation.
that have been widely implemented for different application The balance equations can be derived considering an arbi-
fields, and a SoftRobots plugin was also created for the design trary subinterval [a, b] ⊂ [0, L] of the rod, Figure 3. Across any
[30], modeling and control [31], [32], of soft robots, including
self-collision scenarios [33]. Moreover, a model combining the
FEM and the discrete Cosserat approach described in Section
III-B is presented in [34].
A discussion on the architecture of the software is addressed
in Section VII, while here we mostly focus on the modeling
part for the soft bodies. In Sofa, a deformable continuum is
modeled using a dynamic or quasi-static system of simulation
nodes. The node coordinates are the independent DOFs of
the object, and they are typically governed by equations of
the type (8) and (9). Some approaches have also introduced Fig. 3: Cosserat rod representation
reduced coordinates and these are presented in Section VI-B.
With regards to the actuation, a constraint-based approach section splitting the rod in two parts, the action of the material
is employed. Imposing a linearization of the internal forces on one side upon the other side is equivalent to the resultant
Qint (qi ) ≈ Qint (qi−1 ) + K(qi−1 )dq and considering the stress vector n and the resultant couple m in the global frame.
contribution of the actuation constraints, the static equilibrium Moreover, we suppose that ρ is the mass density of the rod, I
at each i-th step (8) becomes: is the inertia tensor, while v and ω are the linear and angular
velocity in the global frame. The rod is also subject to the
Qint (qi−1 ) + K(qi−1 )dq = Qext (qi−1 ) + B(qi−1 )T λ (10) external forces f¯ and moments l̄ distribution per unit of X
in the global frame. It is then possible to write the linear and
Above, the term B T λ models the contributions of the Lan- angular balance equation at the equilibrium [20]:
grange multipliers (i.e. the contribution of the actuators),  
 nb + b f¯ − ρAv̇ dS = 0 ,
R

where B represents the Jacobian of the constraint equations a a
imposed by the actuators. Three steps are then performed. In (11)
 mb + (r × n)b + R b l̄ − ∂ (ρI · ω)
dS = 0 ,
the first step, a free configuration q f ree is found for λ = 0 a a a t

and, for each constraint, the violation δ f ree is estimated. Considering also the boundary equations, we obtain the dy-
In the second step, the solver computes the value of λ namic equilibrium:
through a projection of the mechanics into the constraint 
 n + f¯ = ρAv̇,
 0
space, obtaining the smallest possible projection space δ =
BK −1 B T λ + δ f ree . In the last and third step, the final m0 + r 0 × f¯ + l̄ = ∂t (ρI · ω)
 n(0) = −f , n(L) = f m(0) = −l , m(L) = l ,

configuration is corrected using the value of the constraint 0 L 0 L
response q = q f ree + K −1 B T λ. (12)
Above, l0,L f0,L are the moments and forces applied in X =
0 and X = L, respectively. Finally, in order to obtain the
B. Directors Approaches solution to equations (12) and to obtain the configuration of
As discussed in Section III-A, in general, in order to the body, it is necessary to define a constitutive model which
determine the stresses in a generic body under the action of relates the internal forces and moments with the strains. A
given external loads, the solution of the equations of three system of six scalar equations is obtained, thus the involved
dimensional elasticity (5) appears as the canonical procedure. kinematical variables should be six or less.
Sometimes, however, the slenderness of the body (such in the Cosserat rod theory was introduced in bio-robotics [36]
cases often considered in soft robotics) suggests an alternative to study locomotion dynamics of hyper-redundant swimmers,
 6

in medical robotics for modeling continuum robots, and in respectively, the dynamic balance equations in the local frames
soft robotics for modeling soft arms inspired from octopus. are: (
The configuration of a Cosserat rod is defined by a vector M η̇ − adη T M η = Λ0 − adTξ Λ + F̄ ,
r(X, t) ∈ R3 representing the position of the center-line and (19)
Λ(0) = −F0 , Λ(1) = F1
a relative reference frame having orientation represented by
a rotation matrix R(X, t) ∈ SO(3). Both the mid-line and where ad is the adjoint representation of the Lie algebra and
the orientation matrix are defined as function of a curvilinear M is the cross sectional inertia tensor. Above, Λ models
abscissa X ∈ [0, L], where L is the length of the robot. Thus, the stress field along the beam (it is the dual counterpart
the configuration of a soft body is defined as a curve: of the strain field). This model is directly related to (12),
by using the relations between  vectors of components in the
g(·, t) : X ∈ [0, L] 7→ g(X, t) ∈ SE(3) (13)
T
T T
inertial and local frames: Λ = RT m , RT n and
which can be expressed in terms of r(X, t) and R(X, t) by 
T
T  T
the homogeneous matrix: F̄ = RT l̄ , RT f¯ . Note that equation (19) is a
! general framework that can support any non-linear constitutive
R(X, t) r(X, t) law beyond the linear Hooke model, although the process of
g(X, t) = (14)
0T 1 obtaining that constitutive relation is more complicated than
in the classical continuum setting.
Moving to the strain definition, the space variations of the field One of the most common procedure to reduce the number
g(X, t) can be entirely described by the vector field ξ(·, t) : of variables consists in employing Kirchhoff’s hypotheses
X ∈ [0, L] 7→ ξ(X, t) ∈ se(3) ∼ = R6 which is given by the (1859). The theory is based on the following four constitutive
boundary value problem: assumptions [20]: (1) the axis of the rod is inextensible; (2)
g 0 = g ξb , g(0) = g0 (15) the stress couple varies linearly with the curvature and the
twist; (3) there are no shear stresses between the cross-section
In particular, it is possible to define the strain field (ξ − ξ ∗ ), and the axis; (4) the cross-section is undeformable in its own
where ξ ∗ (X) is given by (15) in the reference configuration plane. In particular, one of the analytical consequences of these
g ∗ (X), while ξ can be expressed as: hypotheses is that the constitutive relations between the couple
T
! resultant m and the deformation of the rod are not simply
K̃ Γ

T T 6 linear, but they also have the special form [20]:
ξ = K ,Γ ∈R , ξb = ∈ se(3) (16)
0 0
m(X, t) = R(X)diag(A, B, C)K(X, t) (20)
Above, (Γ − Γ∗ ), (K − K ∗ ) ∈ R3 represent the linear
˜ represents the
and angular strains, respectively, while (·) where A(X), B(X) and C(X) depend on the material and
isomorphism between R3 and the algebra of skew-symmetric on the geometry of the cross-section. In 1944, the theory was
matrices so(3). further enhanced by Love [38] to also account for small axial
On the other end, the derivative of the configuration (14) tension and in 1972 [39] and 1981 [40] Reissner completed
with respect to time yields: the theory adding two deformation measures representing the
shear of the beam.
ġ = g ηb , (17) A special case of Kirchhoff’s theory is represented by
providing the velocity twist: non-linear Euler Bernoulli (EB) rod theory, whose linearized
T   version is still the most used method for the analysis of
w̃ v

η = wT , vT ∈ R6 , ηb = ∈ se(3) (18) bending elements in structural mechanics. Let us now consider
0 0 a rod which is straight in its undeformed configuration and
whose bending plane is the one for which the flexural rigidity
where w(X, t), v(X, t) ∈ R3 represent the angular and linear
coincides with the scalar B(X) in equations (20). As a
velocity in the local frame, respectively, which are related to
consequence of this assumption, κ1 and κ3 vanish, yielding to
their global counter-part through the rotation matrix R, i.e.
the so-called elastica:
w = RT ω, v = RT v.
Depending on the constraints considered by the definition B(X)θ0 (X) = Mext (X) (21)
(16), different non-linear rod theories can be obtained. In
the most general case, all six components of the vectors Γ where B(X) is the bending stiffness of the rod , θ0 = κ2 and
and K can be defined by an arbitrary time evolution which Mext (X) is the integral of the external bending moment about
corresponds to the Simo-Reissner beam’s model. On the other X acting on the remaining of the beam.
end, the Kirchhoff rod theory can be obtained by fixing the first In the following, we discuss the most important examples
component Γ1 of Γ to one, and the two others to zero, which of approaches employing the Cosserat rod theory for the mod-
prevents the beam from stretching and shearing respectively. eling of soft robots. In particular, the different approaches are
In [37], equations (12) are expressed in terms of wrenches grouped based on the numerical techniques that are employed
(i.e. 6×1 vectors). In particular, for a Cosserat beam subject to to solve (12). It should be noted that, while the great majority
a density of state-dependent external wrench F̄ on ]0, 1[ and of these approaches describe the soft robot as a rod, there are
two tip external wrenches F0 and F1 at X = 0 and X = 1 also some examples of modeling techniques which have been
 7

presented for 2D shells [41], [42] and combination of Cosserat (15) as soon as the rod’s configuration is parameterized by a set
rods, shells and rigid bodies [43]. of vector fields (e.g. the positional field along the backbone
1) GE-FEM: One of the most famous and powerfull r and the vector field of its 3 cross-sections Euler’s angles
method to capture in an exact manner the geometric nonlinear- θ). Applying the separation of variables, the components of
ities of soft robots is the geometrically-exact FEM (GE-FEM) these fields can be approximated on a truncated basis of Ritz
introduced by J.C. Simo [44] [45] [46]. In this context, the functions compatible with the boundary conditions:
model of rotations is introduced at the same level as that of P
r(X, t) ≈ i Φi (X)qi (t)
positions, by replacing the material points of classical medium P (22)
θ(X, t) ≈ i Ψi (X)qi (t)
by the rigid micro-structures of a Cosserat medium. The GE
character, then imposes to apply the FEM without resorting to where Φ(X) and Ψ(X) are the spatial shape functions, while
any simplifications on finite rotations, except the unavoidable q(t) are the time-dependent generalized coordinates.
space and time discretizations required by the numerical In this Section we gather the soft robotics models that
resolution. The GE-FEM was particularly developed for shells rely on this approximation, which, in some cases, is directly
[41], [42] [43] and rods [47], [48], [49], [50], where micro- applied to the configuration variables as in (22), while in
structures are rigid transverse fibers, and rigid cross-sections others, it is applied to the strain fields of the rod.
respectively. The method’s tour de force was to generalize In [54], the Ritz and Ritz-Galerkin methods are used to
all the key operations of the FEM (interpolation, space and derive a weak-form series solution for a Cosserat rod statics
time-integration. . . ), from the linear vector space of positional and dynamics. In particular, the kinematics of the rod is ob-
fields to the curved manifold SO(3). This approach, which is tained expressing equation (15) in a inertial reference system,
probably the most advanced for modeling mechanisms of rods placed at the base of the manipulator. Then, the backbone
subject to rigid motions and finite deformations, is today used kinematics can be approximated as in (22), while the rotation
as a reference for others, and commercialized in the SAMCEF field is parameterized with a unique further field of torsion
software through its plugin MECANO. In this software, the angle thanks to Kirchhoff assumption and Bishop frames.
model of rods is the full-Cosserat model of Reissner. However, Applying (22) to this field of angle and introducing all these
its dynamic resolution is ill-conditioned when the aspect ratio approximations into the balance equations projected onto the
of the rod increases, i.e. for very slender rods. To overcome same truncated basis, provides a set of algebraic equations (8)
this limitation, several FEM based on the non-linear Euler- similar to those obtained in the FEM approach which from
Bernoulli or Kirchhoff rod models, were proposed, such as this view point is no more than a piece-wise Ritz method.
the Kirchhoff GE-FEM of [49] and [51]. In the static case, a numerical solver is employed to find the
The GE-FEM has been rarely applied to soft robots mod- coefficients qi that minimize the approximation error, while
elled by Cosserat rods. This is probably due to the fact in the dynamic case the coefficients qi (t) are found from the
that it is computationally heavy and difficult of access for time numerical integration of the equations of motion, now in
researchers not experienced in non-linear structural dynamics. an ODE form (9). Based on an experimental observation, the
Nevertheless, in [52], a GE-FEM approach for the modeling of system states qi are assumed to be the coefficients of a series
inflatable robots is presented. In this case, the rotational dof of of Lagrange polynomial shape functions. This polynomial,
a rod are represented by quaternions without unity constraints given in the form (22), is finally fitted to three control points
q(X) ∈ H. It is then possible to define an additional strain (at the base, in the middle and at the tip of the manipulator).
variables Ξ(X) = Υ−1 Υ0 (X) where Υ(X) = ||q(X)||2 , Considering that at the base r(X, t) = [0, 0, 0], the final
which model the inflation of the rod and its space-rate, configuration of the continuum robot is reduced to the position
respectively. These can then be related to the other strain of only two physical points.
variables to obtain the kinematics, in the quaternion algebra, Alternatively to these reduction methods based on absolute
incorporating bending, twisting, extension, shear and inflation. coordinates, the approximated solution in the form (22) can
In this way, the configuration is isophormic to a subspace of also be obtained using the collocation method. The main idea
7D functions. Then, the weak form of the static equilibrium is is to replace the unknown strain field with a polynomial and
provided and solved employing a off-the-shelf FEM software setting the vanishing of the residual between the two in a finite
(COMSOL). set of points on the domain, also called collocation points.
Nevertheless, a geometrically-exact finite element approach Considering m collocation points, a total of m + 1 equations
for the modeling of soft robots is presented in [53]. Taking the is obtained, providing the parameters that are required to define
cue from [50], a helical shape function is employed in terms an m-th order polynomial. In [55], the ODEs describing the
of an exponential map to express the shape of the flexible kinematics of a Cosserat rod are directly solved in terms of
manipulator, which is given in terms of the initial position transformations on SE(3), using a combination of orthogonal
and orientation at the base (X = 0) and the strain ξ(X). collocation and forward integration through Magnus Expan-
The resulting dynamic equations take the general form (9) sion. In particular the unknown ξ(X) in (15) is expressed as
and is solved employing a geometric time integration, which a set of three Chebyshev polynomials (which would be six if
has the main benefit of preserving the Lie group structure and the shear strains would have been considered). To ensure that
guaranteeing the geometrical exactness of the approach. the interpolating polynomial satisfies the boundary conditions,
2) Ritz-Galerking Methods: In numerical analysis, the Ritz the minimization of the residual error is imposed, providing
method is a direct method to solve a BVP, such as (19) and a set of non-linear algebraic equations in the form (8). The
 8

computing of these equations require the integration of (15) Finally, the PVS approach has been implemented in a
to find the rotation matrix R(X) at the collocation points and Matlab Toolbox, SoRoSim, for the simulation of soft, rigid and
this is finally achieved using a Magnus Expansion. hybrid robots [57]. A special case of the PVS is the Piecewise
The Ritz approach has been recently applied to the strain Constant, Approach (PCS) [59], [60], where the strains in the
fields of Cosserat rods in the context of the Piece-wise Variable sections are assumed to be constant, which was also extended
Strain (PVS) approach, [56], [37], [57]. In this case, the Ritz to closed-chain geometries in [61].
coefficients of the strains, stand for the generalized coordinates 3) Shooting Methods: Depending on the considered exter-
of a set of homogeneous transformations along the soft robot, nal forces and initial conditions, the equations describing a
similar to the joint transformations for rigid ones [19]. This Cosserat rod (19) and its kinematics (15) represent a boundary
(relative) strain parametrization, produces a highly reduced set value problem (BVP). Here, we gather all the approaches that
of ODEs in the classical form (9), where q is then comparable employ standard numerical techniques, such as the shooting
to the vector of joint angles of a rigid robot. In [37], the PVS method, to solve this problem. In this way, the rod is treated as
was successfully validated against the GE-FEM with standard a continuous body and the numerical approximation is directly
benches in statics and dynamics. The results showed that the applied to solve the involved equations. The shooting method
approach can provide very good results in terms of accuracy represents one of the most popular numerical approach to solve
with a very few number of generalized coordinates. Moreover, a BVP, transforming it into an equivalent initial value problem.
in contrast to absolute Ritz methods, the strain functions do The initial value problem is iteratively solved for different
not need to fulfill any boundary conditions. However, the cost values of initial conditions, until a solution that satisfies the
to pay for this, is that it handles double space-integrals that are desired boundary condition is found. In other words, the
not easy to compute. In [37], a Newton-Euler inverse dynamic solver shoot the solution from the starting boundary condition
algorithm is used to achieve these computations. Recently, towards the final point and the iterations are continued until
other analytical methods were developed. They exploit the the error with respect to the targeted boundary condition is
fact that equation (15) is a system of homogeneous first order lower than the required precision.
ODEs that can be expressed in the matrix form Y 0 = Y A(X). Trivedi [62] presents the geometrically exact Cosserat model
Then, the exponential representation of the solution can be for an internally pressurized soft robotic manipulator. The
obtained through the Magnus Expansions, yielding, for a soft kinematics of the robot is obtained expressing the directors
manipulator, to the kinematic map in the form: as functions of quaternions and these are replaced in the
  definition of the space derivatives of both r and the direc-
g(X) = exp Ω(q, b X) (23) tors. In particular, in the planar case, the directors can be
expressed as functions of the cross section angle θ and an Euler
where Ω(q,
b X) is the Magnus expansion of ξ(X)b and exp Bernoulli model enhanced with shear is obtained. In the static
represents the exponential map in SE(3). equilibrium (12), the internal pressurization is represented as
With regards to the velocity twist (18), the equality of the a distributed force acting in the axial direction and subtracted
mixed partial derivative of g yields [42]: to the internal force n. The other considered external forces
are the weight of the robot and the concentrated forces and
η 0 (X) = ξ̇ − adξ η ,
(24) moments applied by the actuator at the end of each section.
η̇ 0 (X) = ξ̈ − adξ̇ η − adξ η̇ , Finally, the stress and the moment resultants n and m in
In [57], the analytical integration of equations (24) has been (12) are decomposed with respect to the directors to obtain a
provided. The strain discretization in the form (22) can then be system of six first order differential equations governing the
inserted in the differential kinematics for the system, yielding relationship between the curvatures, the torsion and the stress.
to the geometric Jacobian of the kinematic map and its time Even though this is not stated in the paper, we believe that
derivative η̇ = J (q)q̈ + J˙ (q)q̇ in se(3). The generalized the shooting method or a finite difference method is finally
dynamic equation of the system are finally obtained via applied to the obtained BVP.
D’Alembert’s principle. The continuous models is projected Rucker et al. [63] employ the continuum Cosserat theory
onto the space of generalized coordinates through the Jaco- for the static and the dynamic modeling of continuum robots
bians and integrating over the length of the soft links yields to with tendon actuators. In order to couple the rod and the
the equation of motion in the classical form (9). In particular, tendon actuation models, the external forces and moments
the actuation forces are now included in the internal forces, and distributions in (12) are represented as the sum of the action
not in the external ones as in other Cosserat rod approaches: of the actual external loads and the action of the tendons. The
" # tendons are modeled as ideal strings, which can take only a
na
X d˜i (X)tci (X) tension ui that is assumed to be constant along its length (i.e.
Λ = Fela (X) + ui (25) there is no friction between the tendon and the soft body). The
tci (X)
i=1
internal forces acting on the tendon can then be expressed as:
where na is the number of actuators, di is the distance from the
ni (X) = ui tci (X) , (26)
mid-line to the actuator, while tci (X) represents the tangent
unit vector to the actuator path. In [58], the PVS approach where tci is the unit vector tangent to the cable path. The dis-
is further extended to the modeling of concentric tube robot tributed force along the tendon can then be obtained imposing
systems, including the modeling of the tube’s insertion motion. the linear equilibrium (12)1 . Finally, the total distributed force
 9

given by the tendon is equal and opposite to the sum of the the displacement field of the rod in terms of other elliptic
individual force distributions on the tendons fi = −n0i (X). functions.
Similarly, the distributed moment at the backbone is the sum
Focusing on the application of EB rod theory for soft
of the cross products of each moment arm with each force.
robotics applications, in [69] it is employed to obtain the static
Expressing each tendon’s parametric space curve as function
and the dynamic behaviour of a soft arm with a continuously
of the strain variables and through some manipulations, the
rotating clamp and a dead load applied at the tip. In particular,
explicit model equations are obtained as a BVP and the
the equations have been used to analyse the a rise of a snap-
shooting method is applied to solve them numerically. In [64]
back phenomenon and the post buckling behaviour of the
the numerical solution of the dynamics of Kirchhoff rods is
system.
presented. In particular, a time discretization strategy is em-
ployed to transform the standard PDEs of a Kirchhof rod into In [70], multiple-backbone continuum robots are considered
BVPs in the spatial dimension. This is obtained replacing the and the kinematic and static modeling is obtained employing
time derivative terms with an implicit differentiation formula. EB analytical solutions to express the deflected shapes of
The same approach has been employed and extended in [65] the backbones within a sub-segment of the continuum robot.
for the dynamic modeling of tendon-driven robots. The same research group employed the EB rod theory for
In [66] and [67], the continuum Cosserat rod theory is the modeling of tendon driven continuum robots (TDCR)
applied to the modeling of concentric tubes robots. In both [71]. In particular, they can be composed of a series of
cases, the tubes are treated as Kirchhoff rods. Then, some segments that are connected by spacer disks where the cables
considerations are drawn based on the assumption that the can slide inside. Considering a planar bending problem, the
tubes are, in fact, concentric, which means that they all share static equilibrium of the end disk and the spacer disks are
the same center-line r(X) and each tube’s Ri (X) differs formulated. Finally, the geometric compatibility constraints
from the others by a rotation about the local tangent axis. relate the distances between the cables, the backbone and
In [67] the static equations (12) are modified to take into the disks, which are used in the resolution of the actuation
account the additional concentric tube constraint and then redundancy problem.
solved employing a shooting method. On the other end, in
In [72], EB rod theory is applied to the modeling of a fiber-
[66], instead of employing directly equations (12), the Euler-
reinforced bending actuator. The soft actuator is modeled as
Poincarè equations are applied on SE(3)×S(1)×... to provide
a Neo-Hookean material, where the strain energy functional
the constrained governing equations and, finally, the shooting
defined in (6) reduces to W = µ2 (I1 − 3), where µ is the
method is employed to obtain the solutions.
shear modulus and I1 = λ21 + λ22 + λ23 is the first invariant
In [68], a soft actuator is composed of a pliant hollow body of the principal stretch ratios. The principal nominal stresses
that is internally pneumatically actuated and whose swelling is can be obtained from the principal Cauchy stresses given by
constrained with rigid hoops placed along its length. On one (7), setting σe,i = σi λ−1 i . The bending model is obtained
side, the curvature of the actuator is described using the linear considering the fiber reinforcement on the circumferential
EB rod theory, while a visual tracking system is employed to direction (λ2 = 1) and the incompressibility of the material
obtain the actuator’s deformed shape at varying the pressure (λ1 λ2 λ3 = 1), yielding to the definition of the three nominal
load. stress and the Lagrangian multiplier p in terms of µ and
4) Analytical and Euler Bernoulli: The analytical solution λ1 . The bending moment can then be calculated from the
to equation (21), can be obtained only for specific loading integration of the effect of the stresses on the top and bottom
conditions (for example, only when the rod is subject to layers and, for equilibrium, this is finally set equal to the
concentrated external loads, and not distributed ones) and it is bending of the the internal air pressure against the distal
given in terms of elliptic functions. For example, considering cap of the actuator. Finally, a relationship between the input
the case of a simply supported beam subjected to a concen- air pressure and the bending angle in the free space, is
trated compressive force F applied at the tip, the analytical obtained. The actuator force can be obtained imposing the
solution of the Euler Bernoulli equation yields: torque equilibrium, under the hypothesis that the actuator is
constrained in a flat configuration and that no internal moments
! 2
 
were generated under pressurization. This approach is further

EI  θ(0)
F = 2 4 K sin  , (27) extended in [73], which also takes into account the pressure
L 2
on the lateral surface of the inner chamber of the actuator.
where an homoegenous cross section along the length of the In [74] a soft arm made of longitudinal pneumatic actuators
beam has been considered (i.e. B(X) = EI, where E is is described in terms of its curvature, given by (21), a stretch
the material’s Young’s Modulus and I is the area moment defining the extension or compression of the base curve, and
of inertia) and K is the complete elliptic integral of the first the bend plane defined by the bending angle at the base. The
kind. It should be noted that equation (27) provides an implicit equilibrium between the internal and the external loads is then
expression of the base angle θ(0) as a function of the applied formulated, together with the equilibrium on the cross section.
load F . Once the angle θ(0) is obtained for a given load The model is finally solved numerically, discretizing the arm
F , the rotational field θ(X) is obtained through integration into a finite number of sections and driving the residual of the
of (21). Similarly, the solution can be employed to obtain equilibrium equations to zero.
 10

IV. G EOMETRICAL M ODELS characteristic is that they all employ a chosen mathematical
In the previous Section we discussed the models which are function to describe the desired space curves representing
based on the application of continuum mechanics theories to the geometry of the robot. Thus, for these approaches, the
soft robotics. The presented models are characterized by the coordinates q in (28) are within a space of functions.
assumption of a continuous media and they are all, sometimes One of the first example falling in this category is the
implicitly, based on a strain energy assumption. In this Section, model developed in [75], which employs a serpinoid curve
we discuss a different family of approaches, which are not to describe the kinematics of snake-like robots. Then, the
based on specific materials considerations. More precisely, the kinematics of the robot is obtained by simple forces and
main characteristic of these approaches is that they all rely torques equilibrium, considering the action of the ground
on the assumption that the deformed shape of the soft body on the robot and the internal force components [76]. This
resemble a specific geometrical shape. As a consequence of approach has been employed to model a great variety of snake-
this, the constitutive laws modeling the elastic behaviour are like robots, but considering only planar cases, while the 3-
expressed as functions of the employed parametrization. All dimensional kinematics has not been addressed.
these approaches are based on a representation of the soft One other important example is the so-called modal ap-
body which falls in the definition of a Cosserat rod, thus proach, that have been originally presented for the modeling
the continuous body is represented by a backbone curve and of hyper-redundant robots both in statics [77] and in dynamics
the directors defining the orientation of the cross section. The [78], while a survey on this technique is presented in [12].
main difference from the approaches described in Section III-B In particular, the backbone curve r is parametrized using the
is that, when moving from the kinematics to the static and following integral representation:
dynamic equations, these approaches do not rely on the PDEs Z X
(12), but their are build starting from the generalized coordi- r(X, t) = λ(s, t)t(s, t)ds , (30)
0
nates specifically employed to represent the body geometry.
We can then conclude that, while the configuration of these where λ(X, t) represents a length scaling factor, while t(X, t)
models can vary, for instance being discrete or continuous, is the unit tangent vector, which can be obtained using any
all these approaches have their roots in some set of shape spherical kinematics representations (such as Euler angles or
coordinates mapped to the configuration by: quaternions). In other words, the back-bone curve is repre-
sented as a function growing from the base of the robot
q ∈ Rn 7→ g(q) (28) along its tangent, growing with a magnitude rate λ(X, t) and
The equilibrium is then given by the Euler-Lagrange equations with direction t(X, t), implicitly complying with the Kirchhoff
(or other equivalent principles of dynamics): assumptions. In order to obtain the complete description of
  the robot geometry, the backbone curve is finally associated
d ∂L(q, q̇) ∂L(q, q̇) to a set of orthonormal frames and a roll distribution that
− = Q(q, q̇) (29)
dt ∂ q̇ ∂q describes the twist of the robot (or, in other words, the twist
where L(q, q̇) = K(q, q̇) − U(q) is the Lagrangian of the of the set of reference frames). In conclusion, the geometry
system, K and U are, respectively, the kinetic and potential of the robot is defined by a small set of shape functions,
(including elastic) energy of the system and Q are the gener- describing the backbone curve itself, the chosen reference
alized external forces. Note that, depending on the definition of system and finally the twist. The number of independent shape
the Lagrangian, the resulting models can be equivalent to the functions that are required to uniquely define the backbone
ones derived in the previous section, with particular reference curve depends on the degrees of freedom of the considered
to the Ritz-based approach of section III-B2. problem. Moreover, the choice of the shape functions is not
The most popular approach falling in this category is the unique and it depends on the considered task or on the purpose
so called Piece-wise Constant Curvature (PCC) model, which of the modeling (for example, computational efficiency rather
discretizes a soft body in a finite number of sections having than model precision).
a circular arc shape. While some works present the Piecewise In [77], a set of four independent shape functions Si (X, t) =
Constant Curvature (PCC) and the Constant Curvature (CC) {λ θ(X) φ(X) ψ(X, t)} is employed. In particular, θ(X, t)
as two separate approaches, from the author’s point of view and φ(X, t) are the two angles defining t(X, t) and ψ(X, t) is
CC models can be seen as a specific case of PCC models, the roll distribution function. The inverse kinematic problem
when a single constant curvature section is employed to then consists in finding the set of shape functions satisfying the
represent the entire soft body. One other family of approaches task constraints, which, in most cases, are represented by the
based on geometrical assumption on the deformed shape of end-effector positioning. For this purpose, a modal approach
the soft body are the functional models, which describe the is employed and each Si (X, t) takes the form:
deformed configuration of the soft body through a theoretical NSi
X
geometrical curve. Si (X, t) = Φij (X)qij (t) (31)
j=1
A. Functional approaches where NSi is the number of modes distributed on the i-th
The functional approaches represent probably one of the shape function, qij are the modal participation factors and
first tentative to model soft robotics devices. Their main Φij (X) are the mode functions, which are chosen by the
 11

robot programmer. For example, an extensible planar backbone procedure is carried out to fit the cspline with the given
curve can be fully represented by the two shape functions backbone. To complete the description of the configuration
θ(X, t) and λ(X, t), which are restricted to the modal form of the robot, the orientation along the curve is finally obtained
(31). Thus, the backbone curve shape is fully described by assuming the minimizing of the torsion of the actuator’s
the modal factors qij , which are obtained to satisfy the task backbone.
constraints and which represent the generalized coordinates of In [84], elliptic Fourier descriptors are employed to describe
the system. soft deformable morphologies. With respect to the other meth-
In [11] it has been proved that, when the curvature of the ods described in this Section, this approach does not model
body is assumed to be constant, this approach provides a the soft robot through its backbone, but it employs a closed
transformation which is equivalent to (34), presented in the curve to represent the boundary of the 2-dimensional shape
following Section. In one other work by the same authors [12], (region) that it occupies. This is obtained employing elliptic
variational approaches are presented in order to compute the Fourier descriptor, through a procedure that fits a closed curve
optimal curve shapes that comply with joint constraints, while, to a set of two-dimensional points with arbitrary precision.
at the same time, satisfying the task constraints. The same In particular, the image of the region occupied by the robot
kinematical model was also employed in [79] and extended to is extracted from experimental recordings and its contour is
the dynamical modeling of coiling continuum robots. identified through a discrete representation. Finally, the four
One other example of a functional model is the one coefficients of the Fourier series are computed, providing the
presented in [80], where a polynomial curvature approach description of the shape.
is specifically conceived for control purposes, rather than
simulation. The approximated curvature k(X, t) is defined as
an finite expansion of monomials:
m
X B. Constant Curvature Models
k(X, t) = X i qi (t) , (32)
i=0
Constant curvature is often viewed as a desirable char-
where qi are the modal components of order i. When m = 0 acteristic in continuum robots, due to the simplifications it
the model coincides with the constant curvature one, which enables in kinematic modeling as well as in real-time control
is discussed in the following Section IV-B, and an analytical and other useful computations. This is motivated by the fact
solution can be obtained. that actuators with a path parallel to the mid-line produce
In [81] the polynomial curvature approach is further ex- a constant curvature shape. For these reasons, the constant-
tended to control a soft inverted pendulum, employing an curvature assumption has been successfully applied in a great
affine curvature function k(X, t) = q0 (t) + q1 (t)X. Thus, the number of continuum robots modeling approaches. In these
constant and linear terms, q0 and q1 respectively, represents the models, a soft body is represented as a finite number of
configuration variables. Both for the affine and the polynomial circular arcs, each having a curvature that is constant in
case, the local orientation of the robot is obtained by direct space (but not in time), Figure 1(b). For these approaches, the
integration, providing also the Cartesian space. coordinates q in (28) are specifically obtained to describe the
One other possible function that can be employed to constant curvature assumption and, in literature, it is possible
represent the robot’s geometry consists in the Pythagorean to find both continuum and discrete configuration spaces
hodograph (PH) curves [82]. The backbone is represented employed to describe soft bodies under the CC assumption.
using quadratic polynomials, yielding to We can distinguish two main families of approaches: some
5   models are developed from a kinematical relation between the
X 5
r(X) = qk (1 − X)5−k X k (33) actuator parameters and the arcs parameters (Kinematics-based
k
k=0 models), while other approaches are based on the mechanical
where qk are the chosen control points, which thus represent description of the problem (Mechanics-based models). Finally,
the generalized coordinates of the system. In order to obtain a survey on some of the main PCC approaches can be also
the optimal quadratic polynomials to compute the control found in [11] and [13].
points the minimizing of the potential energy is applied. 1) Kinematics-based models: Once the continuous body is
Finally, a neural network model is employed to predict the represented as a finite set of constant-curvature segments, each
effects of loads on the position of the robot, including the of these can be represented by a finite set of arc parameters
case of variable loads. and it is possible to obtain a map from them to the task
In [83] the kinematics of a soft pneumatic actuator is space of the robot. Different parameters can be employed to
modelled representing its backbone as a cubic Hermite spline describe a CC segment, yielding to different kinematics maps.
(also called cspline). In general, splines are special functions One of the most popular sets of arc parameters that have
which are piece-wise defined by polynomials and a cspline been proposed consists of triplets of curvature κ, the angle
is a cubic spline function whose polynomials are specified of the plane containing the arc Φ, and arc length L. Different
in Hermite form. In this way, the backbone is defined by approaches have been proposed to obtain the kinematic map
two control points and two control orientations (vectors). The from these arc parameters. In [11] it has been proved that they
control points are obtained experimentally and an optimization all provide an identical transformation from the arc base to any
 12

point s ∈ [0; L] of the robot arm: to describe the geometry of the soft body. One alternative is
T (κ, Φ, s) = provided by the exponential map of [102] [103] and [104],

1
 described above. In particular, the transformation (34) can
cos φ cos κs − sin φ cos φ sin κs k cos φ(1 − cos κs) be obtained employing the exponential coordinates from the
1
 sin φ cos κs cos φ sin φ sin κs sin φ(1 − cos κs) 
 
k
1
Lie Group theory, which was introduced in Section III-B.
 − sin κs 0 cos κs κ sin κs
 
 Considering that the transformation for a circular arc can be
0 0 0 1 decomposed in a rotation ζrot and in-plane ζinp transforma-
(34) tion, applying the product of exponential formula yields:
One way to obtain (34) is based on the employment of
Denavit-Hartenberg (D-H) parameters, [85], [86], [87]. In [85], T (κ, Φ, L) = exp(ζ̂rot Φ)exp(ζ̂inp (κ) L) (36)
the constant curvature assumptions and the Frenet-Serret for-
mulas are employed to obtain the complete evolution in space In [103], this approach is applied to the modeling of concen-
of the curve. The continuous backbone of the continuum robot tric tubes robot. It is assumed that the tubes directly apply
is then fitted with a virtual conventional rigid-link manipulator moments to one another which, given the arc assumption,
and modified D-H parameters are obtained to consider the can be considered as constant. The in-plane bending model is
coupling posed by the curvature in a continuum system. In this obtained employing Euler Bernoulli linear equation and the re-
way, the D-H table provides the standard homogenous transfor- sultant curvature of two overlapping tubes is obtained through
mation matrix (34) from the arc parameters to the task position a force balance in analogy with linear springs connected in
and orientation. In a following work by the same research team parallel. The axial tube rotation is obtained considering that
[86], a similar, improved, approach is presented. One prismatic two tubes have the same stiffness and assuming torsional
and four revolute joints (described by D-H parameters) are rigidity and, finally, the flexural and the torsional elastic
employed to provide the parameterization of a single constant energies are obtained. The mapping from the actuator (joint)
curvature section, and geometrical considerations are finally space to the arc parameters is obtained defining the length of
employed to provide the transformation from three parallel the different links composing the entire tube. The definition of
actuating tendons to the arc parameters: the lengths of each link is simply obtained from the position
l1 + l2 + l3 of the actuators on the tubes and by the lengths of the curved
L(q) = , portions. Finally, the active cannula joint space is parametrized
3 √
by the axial translations and rotations of each tube and the
!
3 l 2 + l3 − 2l 1
Φ(q) = tan−1 , (35) mapping from the joint space to the arc paramaters is obtained
3 l2 − l3
p with (36).
l12 + l22 + l32 − l1 l2 − l1 l3 − l2 l3 Similarly, in [105], the arc is expressed with an axis-rotation
κ(q) = d2
l1 + l2 + l3 parametrization. The origin of the axis ω is positioned at point
where li are the actuator’s lengths, while d is the distance from ρ on the horizontal plane and it is perpendicular to it. The
the center of the section to the actuator. This approach has magnitude of ρ is the radius of curvature of the arc, while the
been applied for the modeling of a great number of continuum magnitude of ω is equal to the angle between the proximal and
robots, such as multi-section bionic manipulators [88], [89], distal positions when ρ is used as the vertex of the angle. The
[90], [91], [92] and [93], [94]. Finally, an identical map (34) constant curvature arc is then represented by its length and
is also obtained in [95] and [96]. the two components of ω. The forward kinematics can thus
One of the main restriction of some PCC models is that the be obtained through the exponential maps. Straightforward ge-
employed parametrization and kinematic maps can implicitly ometry finally provides the mapping between the configuration
provide a numerical singularity which is encountered when parameters and the length of the actuating tendons (35). In
the curvature tends to vanish k → 0, resulting in an infinite essence, this is a screw theory method, since ω is exactly the
or undefined radius of curvature. In order to overcome this (constant) axis of the screw motion required to cover the arc
numerical limitation, different solutions have been introduced. bend, as shown in [106].
In [97] and [98], the rotational and position components One other example of a parametrization that does not entail
of the homogeneous transformation (34) are represented in the singularity in the proximity of a null curvature configura-
modal form, following a similar idea to the one behind the tion is provided in [107]. In particular, for each CC section
models presented in Section IV-A. In particular, the entries in composing the robot, the configuration of a segment is defined
(34) are numerically approximated using multivariate Taylor as linear combinations of four arcs included in the section’s
series expansions for the joint variables at 0. In this way, volume. In particular, considering the CC assumption, it is
the kinematics is computed directly in the joint space without possible to define the difference in length between opposite
requiring intermediate transformations. Based on this formu- arcs, ∆x,i and ∆y,i . Employing geometrical considerations,
lation, a Lagrangian approach has been developed in [99] and it is possible to obtain a transition map from the ”conven-
[100] for the spatial dynamics of a single section continuum tional” arc parameters (k, θ, δL) and the new parametrization
arm and this was further extended to multi-section arms in (∆x , ∆y , δL), where δL represents the change in the length
[101]. of the section with respect to the at rest position. Algebraic
Other approaches have been mostly focused on the defi- steps are finally used to obtain the explicit expression of the
nition of a different parametrization and transformation map analogous of (34), in terms of the new parameters.
 13

2) Mechanics-based models: In [108] both the forward by rigid disks and three tendon cables. The disks are treated as
and the inverse kinematics for a tendon driven manipulator rigid bodies, while the backbone and the cables sub-segment
are obtained using the PCC method. More precisely, the are treated as circular arcs. The position and orientation of a
manipulator configuration is explicitly given in terms of the rigid disk with respect to the adjacent one is given by the two
measured displacement of the actuation tendons. The soft orthogonal curvatures and the twist angle of the sub-segment.
manipulator is represented as infinite concentric arcs, which With regards to the actuation forces and moments, these can
can be described in terms of the usual arc parameters, (κ, Φ, s). be calculated taking as an input the tension on the cables,
Considering a single concentric tendon that can experience once the position of the holes in the disks is defined. For the
only a constant tension along its length, force and moment friction at each disk, assuming a continuous viscous model,
balances are employed to define the internal reaction forces an iterative approach is employed, where the contact friction
acting on a section of the beam. This model is then extended is updated at varying the cable tensions at the entrance and at
to consider the presence of a finite number of tendons (i.e. a the exit of the hole.
reduntant tendon model), employing an analogy with a system The modeling of concentric tubes robots using PCC ap-
of linear springs acting in parallel. Finally, the generalized proach was developed in [114], which extends the work
coordinates describing the configuration of the beam are the presented in [104] to account for the torsional twisting of the
curvature κ and the axial strain field. The static equilibrium tubes. The same results were also obtained in [115], where the
leads to the description of the mechanical response of the calculus of variations was employed. In particular, in [114],
manipulator in the form the relative twist angle between two tubes is defined as a
function of the arc length and then it is employed to write
Kq = B T u (37) the moment equilibrium and to impose the compatibility equa-
where B T represents the tendon moment arms and the axially tions, enforcing the coincidence of the tube centerlines. The
directed tangent. Linking the tendon tension to the tendon dis- torsional curvature is then obtained considering the Cosserat
placements is then possible to obtain the mechanical response Rod equilibrium equations (12). In particular, the analytical
of the tendons in terms of their displacements y: closed-form solutions are provided for the case of two con-
centric tubes in terms of Elliptic functions (as seen in Section
y = Cm u (38) III-B4), and the presence of multiple solutions, corresponding
to instability phenomena, is discussed. The stability analysis
where Cm is the compliace matrix, which is function of is further analyzed in [116] and [117].
B T and of the manipulator and tendon stiffness matrices and
lengths. Finally, combining (37) and (38) yields to
V. D ISCRETE M ODELS
y = A† q (39) In this Section we discuss the approaches where the configu-
ration of the system is discrete from the very beginning (i.e. the
where A = K −1 B T Cm −1 is the forward kinematics trans- system is not discretized at the equation level or through some
formation matrix. Equation (39) provides a mechanics-based assumptions on its geometry). In particular, we can distinguish
relationship between the beam and the tendon configuration, three main groups of models: the discrete rods, the pseudo-
analogous to transformations (35) that were obtained from rigid and the lumped-mass models.
geometric considerations. In a following work by the same
authors [109], the 3D static model is also developed with an
experimental validation. A. Discrete rods
In [110] and [111] the dynamics of a braided continuum The computer graphics research community represents an
manipulator is addressed. The model employs the usual arc important source of inspiration for soft robotics modeling.
parameters (κ, Φ, l) and the motion equation are obtained This is the case of the approaches that are described in this
using the principle of virtual powers. A Neo-Hookean model Section, which are all based, with different extent, on the
and the Cauchy-Green stretch tensor are employed to define discrete elastic rod (DER) approach, originally introduced in
a unit deformation energy, but the Hooke and the Euler- [118], [119] and [120] and sometimes also referred to as
Bernoulli equation are also presented as a simplification. To discrete differential geometry DDG [121]. In this formulation,
model the braided thread, a constrained volume is employed, the material curve of the rod is approximated by a discrete set
imposing that the volume of the solid is fully constrained by of lines connected at vertices. Thus, the configuration space
the thread. As a consequence of this, the axial and the radial takes the form:
deformation of the chamber are constrained on an helix, while
Q ⊂ SE(3) × SE(3) × SE(3) · · · × SE(3) (40)
its tangent deformation of free. In this way, the model defines
the deformation of the cross section caused by the constant In particular, in the DER formulation, the inextensible
curvature bending of one manipulator module under general Kirchhoff rod assumptions (20) are considered and, similarly
gravitational loads. to the Cosserat rod model, the configuration g of an elastic
In [112] and [113], the method of virtual power, equivalent rod is given by a curve r(X), representing the centerline, and
to the Euler-Lagrange equations (29), is employed to obtain the a material frame, where the first axis is always tangent to the
dynamics of TDCRs robots. The considered geometry for the curve. The Bishop (natural) frame {t, v, w} is then introduced,
robot is the same as in [71], i.e. an elastic backbone connected providing, for a given centerline, the most relaxed frame, i.e
 14

with zero twist. Finally, the parallel transport is employed to to tree-like architecture, which are frequently encountered in
define the evolution of the Bishop frame along the mid-line. the soft robotics community.
In [125], the DER-based formulation is further enhanced
to include frictional contact, inelastic collisions and inertial
effects. In particular, considering a planar case, a Rayleigh
damping matrix is defined to formulate the internal damping
forces vector which is added to the external forces. In order
to model the contact and the friction between the robot and
an unstructured ground, whose normal direction can vary with
the horizontal axis, a Coulomb law is employed. In particular,
Fig. 4: Representation of a rod in the DER approach all the possible contact constraints are gathered in a set which
is then introduced in the system dynamics using a modified
In this way, the model provides a simple parametrization mass-based method.
of the material frame, which is given by the curve r(X) and
an angle θ(X) between the Bishop and the material frame. B. Pseudo-rigid Models
This parametrization represents the basis of the discrete theory
employed for the kinematical description of the continuum When dealing with soft robotics modeling, the chance of
body. In particular, the rod is discretized into a series of finite exploiting well known and established rigid robotics theories is
nodes (or vertices) connected by straight segments (or edges). clearly intriguing. This is the idea that motivated the family of
Each node is characterized by a position vector ri , while each models that are described in this Section. Here, the soft bodies
segment is associated to the edge vector ei = ri+1 − ri and are represented as series of rigid links which are connected by
its tangent unit vector ti , Figure 4. In this way, the discrete revolute, universal or spherical joints, as seen in Figure 1(c).
curvature associate with the ith vertex takes the form: Thus, in general, the configuration space of the soft body is
given in the form:
2 sin φi φi
κi = = 2 tan , (41)
1 + cos φi 2 Q ⊂ SE(3) × SE(3) × SE(3) · · · × SE(3) (42)
where φi represents the turning angle between two consecutive While these approaches can provide satisfactory results for
edges. Similarly, it is possible to define the discrete curvature the modeling of hyper-redundant or snake-like robots, when
bi-normal. Each i-th vertex is also characterized by a total dealing with continuous elastic structures they provide a low
mass Mi , which is the average mass of the edges meeting order of spatial approximation accuracy.
in the vertex, while the mass moment of inertia for each edge In [126], a dexterous catheter manipulator is represented
can be obtained through volume integration. Finally, a discrete as a series of pin joints which are connected by a series
bending energy and a discrete twist energy can be obtained in or rigid links. Proceeding with approaches for biomedical
terms of the generalized coordinates (ri , θi , φi ). applications, in [127], a pseudo-rigid 3D approach is applied
Following the DER formulation, in [122] the three dimen- for the modeling of a steerable ablation catheter. The catheter
sional space of a soft filament is represented by a set of vertices is treated as a cantilever beam, which is represented as
ri (t) and a set of material frames Ri (t) (see (14)). Each four rigid links connected by three revolute joints and three
vertex is characterized by a linear velocity, a concentrated torsional springs. [128] presents a 6 dof pseudo-rigid model
mass and a set of concentrated external forces. Extension, for continuum manipulators subject to multiple external loads.
shear and axial deformations are considered. Finally, through In this case, a flexible manipulator is represented as four
the spatial integration, the discrete governing equations are rigid links of given length which are connected by three three
obtained and they are solved using a symplectic, second- joints having 2dof each. In [129], the dynamical modeling
order scheme. This approach can also consider the presence of compliant mechanism is addressed and a flexible beam is
of interaction forces, such as those related to solid and liquid represented as a mass-less rigid body with a torsion spring
friction or contact. Similarly to other approaches, the external attached at one end.
physical interactions are included in the external forces and In [130], the modeling of hyper-flexible manipulators is
moments vectors, while the internal physical effects (such as addressed using a serial rigid chain, where the number of
those modeling a muscular activity) are added to the internal kinematic dof goes to infinity. The backbone curve is first
forces and moments resultants. In a recent work [123], this described as a continuous curve with extended Frenet frames
model has been implemented in an open-source simulation attached to each point along its length. Then, for the numerical
environment, Elastica, for the modeling and simulation of the simulations, the backbone curve is approximated by a serial
dynamics of slender rods. chain of rigid bodies. In particular, the backbone is described
One other example of a soft robotics modeling technique in terms of a variable vector that corresponds to the joint angle
following the DER approach is presented in [124]. In particu- vector of a serial-chain manipulators.
lar, the planar case of the discrete eleastic rod formulation is In [131], a steering catheter is represented as a combination
considered, where the rod is free to move on a plane, while the of three sections: the virtual base of the distal shaft, which
torsional deformations are neglected. Moreover, the notion of is represented as a prismatic joint; the bending section of the
folding for a straight rod is introduced to apply the approach distal shaft is represented as two revolute joints, one prismatic
 15

joint and two revolute joints; the distal end of the catheter, the system and they are connected through linear spring and
which is treated as a rigid body. D-H parameters are then dampers. In particular, the presence of springs in the diagonal
employed to obtain the forward kinematics of the model. directions around each node, allows the modeling of the shear
In [132] a pseudo-rigid approach is employed for the model- effects. Assuming a uniformly stiff surface, the elastic constant
ing of a 8 arms swimmer inspired by the octopus. In particular, of the springs and the damping coefficients are assumed to be
each arm is modeled as a kinematic chain of cylindrical rigid constant. Considering an isotropic linear elastic material and
segments which are connected by planar rotatory joints. The assuming a straight undeformed configuration, the equation
first joints in the chain, connecting the arms of the octopus to of motion for each mass is obtained by combining direct
the main body, are modeled as actuated rotatory joints, while application of Newton’s second law and Hooke’s law.
the other joints are modeled as (unactuated) rotatory linear In [135], a discrete lumped-mass approach is presented for
spring and damper elements. the modeling of tendon driven medical robotic catheters. The
axial and the bending stiffness of the model are adjusted
C. Lumped-mass Models through the selection of the spring stiffness and their radial
location, while the smooth motion of the catheter is modeled
One of the most straightforward approach to model a con- through the dampers. Torsion is not considered, even though it
tinuum body is to represent it as an array of discrete masses, could be implemented adding torsional springs to the model.
dampers and springs, Figure 5. In this way, the governing Moreover, a modified Dahl friction formulation, is incorpo-
rated at the intersection of the control tendons and the mass
elements.
Because of their simplicity, lumped mass models are par-
ticularly suited for large robotics simulations library, such as
the one presented in [136]. Titan is a GPU-accelerated C++
software library for the modeling of soft bodies and multi-
agent robots at massive scales. The soft robotic body structure
is represented through spring-mass lattices and an interactive
approach is employed. In particular, the first iteration is applied
to all the springs to compute, with Hooke’s Law, the forces
acting on each corresponding mass and then a second iteration
is performed over all masses to update their accelerations and
positions. The main benefit of this approach is the possibility
to perform parallel computing, resulting in great computational
performances.
Fig. 5: The lumped mass representation of a continuous rod
VI. S URROGATE M ODELS
equations can be obtained by energy methods or applying
classical Newton equations to describe the propagation of the As extensively discussed in all the previous Sections, soft
forces from one particle to the other. Thus, the configurations robotics present challenging problems when dealing with
space is given in the form: modeling and they in general require great efforts and some
expertise in the continuum mechanics field. In order to over-
Q ⊂ R3 × R3 × R3 · · · × R 3 (43) come some of these difficulties, some authors presented a
complete different solution to tackle the problem, where the
The main advantage of lumped-parameter mechanics models
modeling of the system is obtained from large sets of data
is that they are characterized by a simple structure, which
(inputs and outputs) that are derived from various forms of
is easy to adapt to include complex phenomena such as
external sources. These approaches are often referred to as
nonlinear friction, material hysteresis, and inertial dynamics.
surrogate or data-driven models and a survey on some of
On the other end, to reach the same fidelity of the continuum
these approaches is presented in [14] and [15]. While one of
mechanics models or the FEM approach, they required a data-
the main benefit of these solutions is that they do not require
expensive system identification procedure.
a physical model, on the other end they rely on large amounts
In [133] and [134], mass-spring-damper arrays are used to of representative data, that are sometimes difficult to collect.
model the large deformation of a continuum surface (LDCS) Most of data-driven approaches presented for soft robotics
actuated by continuum arms. modeling are based on machine learning algorithms, mostly
While the vast majority of the soft robotics devices is made involving neural networks models.
out of parts that can be treated as rods, LDCS consists in
a flexible surface which is actuated into multiple curvature
profiles by integrated continuum arms. In particular, the above A. Neural Networks
mentioned works consider the 3D problem of a system com- Neural networks (NN) have been proved to be an effective
posed of two interconnected layers of LDCSs, where bending tool to solve many kinds of non linear problems in different
and shear deformations are allowed. The surface is modeled application fields, including robotics. In general, the goal is to
as a lattice where the masses are defined as the nodes of find some sort of underlying pattern in a set of given data. As
 16

the name suggests, they are inspired from the biological neural neurons propagate the input variables to the following layers,
networks that operates animal brains: the artificial neurons, while each neuron in the hidden layer is associated with a RBF
which represent the elementary units of NNs, can transmit kernel (usually Gaussian). The data are obtained experimen-
a signals to the other neurons. The signal is usually a real tally, measuring the arm’s tip position at varying the actuation
number and the output of each neuron is computed as a non- pressure, using a trilateration algorithm. In one other paper by
linear combination of all the inputs. Neurons are connected by the same authors [140], the same handling assistant trunk is
edges and both neurons and edges are characterized by weights modeled, both with a data-driven and a pseudo-rigid modeling
that are adjusted during the learning process. Usually, neurons approach. With regards to the latter, the trunk is modelled as a
are gathered in layers: the first layer is also called input series of rigid vertabrae connected by a total of four prismatic
layer, the last one is called output layer and the intermediate joints. For the data-driven approach, a modified Elman neural
layers are called hidden layers, Figure 6. NN results in a high network is employed. This network is also trained using the
experimental results based on a trilateration system.
FNNs are also employed in [141] and [142], which deal with
the inverse kinetics of a cable driven soft manipulator. While
the kinematical model describes the configuration of the robot
without considering the involved loads, the kinetics model
relates the motion of the robot with its actuation forces. In the
above mentioned papers, the direct kinetics model is obtained
using a geometrically-exact model that follows the approach
described in [143], which uses the continuous Cosserat rod
model (12). Thus, a FFN, taking the tip position as an input
and giving the cable tension as an output, is employed. After
fabricating the considered 3-cable driven manipulator, a data
Fig. 6: Scheme of a neural network
collection phase is carried out, using a set of cable tensions
dimensional set of nested functions: and obtaining the tip position with an infrared vision system.
These data were used for the optimization and training of the
y = fM (AM , . . . f2 (A2 , f1 (A1 , x)) . . . ) (44)
FNN. After the training phase, the performance of the FNN
where x are the values of the input nodes, Ai are the edges are measured on test sets, using the output of the FNN as the
weights, fi are the activation functions and y are the values input of the direct kinetics model.
of the output nodes. In most soft robotics applications, the One other example of a machine learning formulation for
input and the output layers represent the actuation variables the global inverse kinematics of continuum manipulators was
u and the shape parameters q. The learning process results in presented in [144].
an optimization over the network weights, which is performed The data samples are generated by continuous motor bab-
through back-propagation, a form of chain rule. The idea is bling and a single hidden-layer artificial neural network is
that, after each forward step through the network, a backward employed to learn directly the mapping (xi+1 , qi ) 7→ qi+1 ,
pass is performed that computes the network Jacobian and where x is the pose of the end effector. In one other work
adjusts the model’s weights accordingly. by the same authors [145], a dynamic model for open-loop
With regards to the modeling of soft robots, neural networks control of soft robotics manipulators is presented. The PCS
have been used to approximate the kinematics, the statics model described in III-B2 is employed to obtain the dynamics
and also the dynamics description of the considered systems. of a cable driven robot operating underwater. Considering the
Typically, Feed-forward neural network (FNN) are employed case where the robot and the task space have the same number
for kinematics and static modeling. They represent the first of DoF, the forward dynamics of the system (9) is formulated
and simplest example of neural networks where, as the name using only the task space variables and the direct mapping
suggests, the signals travel in only one direction (forward). On between the states of the task space variables and the control
the other end, Recurrent Neural Networks (RNN) are usually inputs can be obtained. A nonlinear auto-regressive network
preferred for dynamical modeling. They use time series data with exogenous inputs (NARX) is used. The samples for the
and they can use their memory to process sequences of inputs, network learning are obtained from the PCS simulations and,
which makes them particularly suitable to model temporal then, the training is carried out in two steps. The first step
problems. consists in an open loop training of the network, by unfolding
The first examples of the employment of neural networks the recurrent network and training by Bayesian Regularization.
algorithms for the control of continuum robots were presented Then, the network is closed and the training is continued using
in [137] [138]. In particular, the authors present a controller for the same network weights. Once the training is completed,
continuum robots which uses a FNN component to compensate an open loop predictive controller is developed through a
the dynamic uncertainties of the system, in an attempt to trajectory optimization that is carried out with an iterative
reduce the uncertainty bound. In [139] the forward kinematic sequential quadratic programming. In [146], the closed-loop
modeling of a bionic assistant trunk is obtained through Mul- controller is implemented.
tilayer Perceptron (MLP) and Radial Basis Function (RBF) The inverse kinematic modeling of a bionic trunk is ad-
neural networks, which are both class of FNNs. The input dressed in [147]. In particular, a learning phase is carried
 17

out considering a volume of desired cartesian position for the while in [33] it is applied for the reduction of self-collision
robot’s tip, defining a finite set of target vertices. The inverse contact forces.
model is asked to estimate a posture that allows to move the
effector to each vertex, and the training is carried out until the
VII. S OFTWARE I MPLEMENTATION
distance between the target and the actual position for each
vertex is minimized. The inverse kinematics is learned using In this paper we addressed the theoretical frameworks
an online goal babbling scheme. Finally, for the learning of that have been developed so far, but a discussion on the
the system, a regression mechanics is defined in the form of a applications of these models is also required. As we said in
local linear map and the estimated inverse model is obtained the Introduction, the entire design process can be driven by
as a linear combination of local linear functions weighted by the employment of the system’s model and, as a result, the
a Gaussian responsibility function. design optimization can be carried out. In general, a standard
design optimization problems consists in finding a vector of
design variables which minimize a cost function, subject to
B. Data-driven Order Reduction equality and inequality constraints [151].
Here we introduce the modeling tools that employ some Any optimization process implies the research of a point in
forms of data-driven order reduction to efficiently approximate the feasible set of design parameters for which the derivative of
the physical model. In [148], a Koopman Operator Theory the function is zero. Clearly, if the cost function (or, in general,
is employed for a data driven controller of soft robots. In the model providing the cost and the constraints functions)
the data-driven Koopman representation, a dynamical system is differentiable, the optimization process is facilitated, as
is represented in a infinite function space F, which is com- gradient-based research methods can be employed. In this sce-
posed of real-valued functions inside the state of the system nario, some differentiable modeling tools have been presented
domain [149]. The elements q ∈ F are called observables. for optimization and control [152]. In [153] the optimization
The Koopman operators, denoted by Ut , is defined as the of the sensor network design is addressed. In particular, the
linear transformation Ut q = q ◦ Tt , where ◦ indicates the robot’s deformation are simulated using FEM and the Hessan
composition operation, while Tt is the flow (or dynamic) map or tangent stiffness matrix is defined to compute the search
of the system. In other words, the Koopman operator lifts the directions. This matrix is used first to perform a first-order
dynamics of the system from the state space to the space of constrained optimization to obtain the reconstruction of the
the observables, describing the evolution of the observables robot’s state from the sensors readings. The user-provided
q along the trajectories of the system. Its main advantage deformation-space pairs is used to initialize a large set of sen-
is that it provides a linear representation of the flow of a sors that are fabricable, and one other optimization problem is
non-linear system, but in the infinite-dimensional space of performed to obtain the minimum number of sensors required
the observables. The discrete approximate of the Koopman for the desired reconstruction accuracy. In [154] a pseudo-
operator can be obtained from a set of experimental measured rigid body model is used to represent continuum structure,
state, given in the form of snapshot pairs. using as a set of rigid links connected by elastic flexures
Some approaches have been focused on the model reduction modeled using the DER approach that is introduced in Section
of FEM that were presented in Section III-A1. Model reduc- V-A. Given an initial design, different objective functions
tion methods are based on the projection of the FE equations of are defined (motion-tracking, ease of actuation, resilience to
motion to attractive sub-spaces of smaller dimensions. In this failure) and the optimization is performed using a sensitivity
way, the size and the computational time of the simulation are analysis.
drastically lowered, allowing the application of FE methods Development and use of modeling software and toolboxes
for control purposes. In [32], an implicit time integration is for soft robotics are essential to avoid the reinvention of
applied to (10), where the internal forces are linearized apply- the wheel. They could potentially avoid the need to write
ing a Taylor series expansion. A snapshot proper orthogonal a custom code script in a programming language (C/C++,
decomposition (POD) is then employed to generate relevant Python, MATLAB, etc.) for each robot design, optimize the
bases to be used for the order reduction. More into details, robot design for a particular task, or be used for the model-
the solution q of equation (10) is expressed as a truncated based control of a specific robot. The modeling framework
expansion of orthonormal vectors, which depend on the con- should be user-friendly (necessary for the multi-disciplinary
straints λ. The orthonormal basis is then set to minimize the soft robotic community), accurate, fast, and reliable. There
sum of all the errors that are generated by the projection of the are several industrial as well as open-source soft robotic soft-
exact solution onto the basis. The snapshot POD provides a ware/toolboxes available. Here we briefly summarize the most
discrete representation of this cost function, which is expressed popular ones: COMSOL, ANSYS, Abaqus, SOFA (available
in terms of a discrete subset of the constraints λi . The set of at [155]), PyElastica [156], SoRoSim [157], TMTDyn [158],
solutions q(ti , λi )) is finally obtained performing a singular SoMo [159], ChainQueen [160], and IPC-Sim [161].
value decomposition. In this way, the minimization process, The SoftRobots plugin of SOFA, one of the earliest open-
which is a computationally heavy step, is performed only once source platforms for physics-based simulation, uses FEM to
for a given set of λi but it provides a reduced basis that is model, simulate, and control soft robots [28]. SOFA employs
valid once for all. In [150], this approach is further extended a multi-model representation of a deformable body, allowing
for the development of a low-order controller and observer, the presence of several representations (mechanical, thermal
 18

and visual) of the same object. The representations are then expansions with respect to the material dimensions. In this
connected together through a mechanism called ”the Map- way, two main continuous reduced approaches are obtained:
pings”. This multi-modal representation allows the simulation rod (1D) and shell (2D) theory. One other option to reduce
of scenarios where different bodies, such as different rigid the material geometry is to represent the material as a
and/or deformable bodies or fluids, are interacting together. Cosserat media, i.e. a continuous collection of rigid micro-
In particular, deformable bodies are modeled employing FEM structures that are staked along some dominant material
with different constitutive laws. In general, a simulation in direction.
SOFA is described as a scene which is composed by different 2) Rheology, i.e. constitutive law. In the more general case,
nodes organized as a tree or as a Directed Acyli Graph (DAG). the stress field is a functional of the history of strains
PyElastica, which is the python implementation of Elastica, (and possibly also of thermo-dynamical variables). In soft
employs Discrete Geometry and the Cosserat rod theory of robotics, the material is usually treated as elastic, thus the
section V-A to model soft bodies with a finite number of stress is a function (and not a functional) of the strain.
lumped degrees of freedom [123]. SoRoSim and TMTDyn are Moreover, assuming that the strains are small, a linear
MATLAB toolboxes that use geometrically exact piecewise function (Hooke’s law) is also typically employed. Both
variable strain models [57] and discretized lumped systems these assumptions can appear as physically questionable,
with reduced-order models [162], respectively. SoMo (Soft since the materials usually employed in soft robotics are
Motion) toolbox couples the pseudo-rigid model discretization not elastic, let alone linear. There are also some examples
of section V-B with a rigid body physics engine of python to of approaches using hyper-elastic constitutive models, such
model soft robots [163]. as [72] and [169].
ChainQueen, and IPC-Sim (Incremental Potential Contact 3) Geometry of deformations. Even when the strains are
Simulator) are simulators oriented towards computer graphics small, soft robots are subject to finite deformations. Thus,
rather than mechanical systems. ChainQueen is a python the equations describing their motions are fundamentally
toolbox that uses the Moving Least Square Material Point nonlinear. With the assumption of linear elasticity and small
Method, a hybrid Eulerian/Lagrangian FEM which uses both strains, the non-linearities are geometric nonlinearities of
particles and grids to simulate soft bodies [164]. IPC-Sim rigid-body type, that can be expressed in different reduced
solves extreme non-linear volumetric elastodynamic models ways, depending on the considered material geometry and
using FEM (section III-A1) [165]. Particular toolboxes are the parametrization used.
developed for a specific soft robotic problem, such as the One other important characteristic of each approach is whether
DiffAqua, an optimization toolbox for soft underwater swim- they rely on absolute or relative coordinates. In general, a
ming robots [166], and Vine Simulator, a toolbox dedicated to relative parametrization eases the modeling of any internal
growing soft robots [167]. actuator and sensing devise, while an absolute one appears
as more suitable for the modeling of the interaction with the
VIII. D ISCUSSION AND P ERSPECTIVE external environment.
Despite the classification that is used in this paper, is it
Except for the few sporadic examples that were discussed possible to find similarities between the different families of
in Section III-B4, all the above approaches, at some level, models. In the Ritz-Galerking approaches (Section III-B2), the
require a form of numerical discretization. It should be noted configuration of the backbone of the rod is given as the sum
that discretization and reduction are two separate concepts of a finite number of shape and time functions, similarly to
that should not be mistaken. A discretization is employed the modal approach that was described in Section IV-A. One
to obtain a numerical solution to the non-linear problem other observation that should be highlighted is that the GE-
and, for example, the finite difference methods, such the one FEM (Section III-B1), the Piecewise Constant Strain (which
developed in [168], fall in this category. On the other side, is a specific case of the PVS approach discussed in Section
a reduction consists in the depiction of the smallest basis of III-B2) and the Piecewise Constant Curvature (Section IV-B)
functions allowing a proper description of the kinematic fields models are all based on the same backbone geometry (helices
of the body. For example, reduction methods include the mode or arcs that are a special case of helices) and the continuum rod
parametrizations that were discussed in Section III-B2 and IV, is reduced in a finite number of sections where this geometry is
but a reduction of the model was also presented in some works repeated. Finally, in the FEM representation (Section III-A1) ,
discussed in Section VI-B. Within this scenario, Finite Element the continuous domain is discretized through a mesh of nodes
Methods seen in Sections III-A1 and III-B1 can be seen as a and a similar idea is followed by the lumped mass approaches
middle ground between the two, since they involve an explicit presented in Section V-C.
discretization of the material domain in finite elements, over For what concerns the actuation, while for the FEM these
which the kinematic fields are reduced on some polynomials, were included as constraints, in the Directors approaches the
or other shape functions. thread-like actuators actions were simply added to the linear
We can conclude that there are three main levels of reduc- and angular balances as external distributed wrenches [63] and,
tion that can be pursued: sometimes, internal wrenches [60]. This simple adjustment is
1) Material geometry. When one or two dimensions are much one of the main benefit of these approaches, which on the
smaller in comparison with the third one, the three dimen- other end struggle in the modeling of other forms of actuation,
sional position fields can be approximated through Taylor which involve 3D problems (such as ballooning actuators and
 19

electroactive polymers). In this case, FEM appears as a more [10] F. Xu and H. Wang. Soft robotics: Morphology and morphology-
suitable choice. On the other end, some PCC approaches, inspired motion strategy. IEEE/CAA Journal of Automatica Sinica,
8(9):1500–1522, 2021.
specialize to a specific actuators configuration and provide [11] R.J. Webster and B.A. Jones. Design and kinematic modeling of
a kinematic or mechanical transformation from the actuating constant curvature continuum robots: A review. The International
tendons to the parameters describing the configuration of the Journal of Robotics Research, 29(13):1661–1683, 2010.
[12] G. S. Chirikjian. Conformational modeling of continuum structures
rod. However, this is valid only when the constant curvature in robotics and structural biology: A review. Advanced robotics,
assumption is actually preserved in the real world problem. 29(3):817–829, 2015.
Soft robotics was developed in the last two decades and, [13] A. Chawla, C. Frazelle, and I.D. Walker. A comparison of constant
curvature forward kinematics for multisection continuum manipulators.
legitimately, the main efforts so far have been put into the In 2018 Second IEEE International Conference on Robotic Computing,
simulation of the robot’s configuration and in the static and pages 217–223, 2018.
dynamic representation of its behaviour. Now the direction [14] D. Kim, S.H. Kim, T. Kim, B. B. Kang, M. Lee, W. Park, S. Ku,
D. Kim, J. Kwon, H. Lee, J. Bae, Y.L. Park, K.J. Cho, and S. Jo.
of this topic and the efforts of many researchers are mostly Review of machine learning methods in soft robotics. PLOS ONE,
towards the extension of these approaches for control [170] 16(2):1–24, 02 2021.
and for their integration with optimization tools [152]. Once [15] X. Wang, Y. Li, and K. Kwok. A survey for machine learning-based
control of continuum robots. Frontiers in Robotics and AI, 8:280, 2021.
more, the main source of inspiration towards the achievements
[16] J. Burgner-Kahrs, D. C. Rucker, and H. Choset. Continuum robots
of these goals are more likely to come from other research for medical applications: A survey. IEEE Transactions on Robotics,
disciplines. Indeed, while going through the papers that have 31(6):1261–1280, 2015.
been presented so far for the modeling of soft robots, we [17] P. Rao, Q. Peyron, S. Lilge, and J. Burgner-Kahrs. How to model
tendon-driven continuum robots and benchmark modelling perfor-
tackled a great number of disciplines that represent their mance. Frontiers in Robotics and AI, 7:223, 2021.
theoretical foundations: continuum and solid mechanics, com- [18] F. Bullo and A.D. Lewis. Geometric Control of Mechanical Systems,
putational mechanics, machine learning, computer graphics, volume 49 of Texts in Applied Mathematics. Springer Verlag, New
York-Heidelberg-Berlin, 2004.
just to name a few. This interdisciplinary is probably what [19] R. Murray, Z. Li, and S.S. Sastry. A Mathematical Introduction to
makes soft robotics so interesting, attracting scientists from Robotic Manipulation. CRC Press, 1994.
different research fields. On the other side, being the topic [20] P. Villaggio. Mathematical Models for Elastic Structures. Cambridge
University Press, 1997.
so widespread, it is not easy to grasp it and this constitutes [21] H. Altenbach, M. Bîrsan, and V. A. Eremeyev. Cosserat-Type Rods,
the main motivation behind this manuscript. In this way, we pages 179–248. Springer Vienna, Vienna, 2013.
were able to recognize more clearly the uniqueness and the [22] R.W. Ogden. Non-linear Elastic Deformations. Dover Publication,
1999.
commonalities between the different techniques that have been [23] A. A. Shabana. Continuum-based geometry/analysis approach for
presented so far, in the effort to untangle such a vast research flexible and soft robotic systems. Soft Robotics, 5(5):613–621, 2018.
topic. [24] M.S. Xavier, A.J. Fleming, and Y.K. Yong. Finite element modeling
of soft fluidic actuators: Overview and recent developments. Advanced
Intelligent Systems, 3(2):2000187, 2021.
ACKNOWLEDGEMENT [25] P. Moseley, J. M. Florez, H. A. Sonar, G. Agarwal, W. Curtin,
and J. Paik. Modeling, design, and development of soft pneumatic
This work was supported in part by the Khalifa University of actuators with finite element method. Advanced Engineering Materials,
Science and Technology under Grants CIRA-2020-074, RC1- 18(6):978–988, 2016.
2018-KUCARS, and in part by the Office of Naval Research [26] C. Tawk and G. Alici. Finite element modeling in the design process
of 3d printed pneumatic soft actuators and sensors. Robotics, 9(3),
Global under Grant N62909-21-1-2033. 2020.
[27] J. Allard, S. Cotin, F. Faure, PJ Bensoussan, F. Poyer, C. Duriez,
H. Delingette, and L. Grisoni. Sofa–an open source framework for
R EFERENCES medical simulation. In MMVR 15-Medicine Meets Virtual Reality,
[1] B. Tondu and P. Lopez. Modeling and control of mckibben artificial volume 125, pages 13–18, 2007.
muscle robot actuators. IEEE Control Systems Magazine, 20(2):15–38, [28] F. Faure, C. Duriez, H. Delingette, J. Allard, B. Gilles, S. Marchesseau,
2000. H. Talbot, H. Courtecuisse, G. Bousquet, I. Peterlik, and S. Cotin. Sofa:
[2] C. Laschi, B. Mazzolai, and M. Cianchetti. Soft robotics: Technologies A multi-model framework for interactive physical simulation.
and systems pushing the boundaries of robot abilities. Science Robotics, [29] E. Coevoet, T. Morales-Bieze, F. Largilliere, Z. Zhang, M. Thieffry,
1(1), 2016. M. Sanz-Lopez, B. Carrez, D. Marchal, O. Goury, J. Dequidt, and
[3] G. Bao, H. Fang, L. Chen, Y. Wan, F. Xu, Q. Yang, and L. Zhang. C. Duriez. Software toolkit for modeling, simulation, and control of
Soft robotics: Academic insights and perspectives through bibliometric soft robots. Advanced Robotics, 31(22):1208–1224, 2017.
analysis. Soft Robotics, 5(3):229–241, 2018. [30] G. Zheng, O. Goury, M. Thieffry, A. Kruszewski, and C. Duriez.
[4] D. Trivedi, C. D. Rahn, W. M. Kier, and I. D. Walker. Soft robotics: Controllability pre-verification of silicone soft robots based on finite-
Biological inspiration, state of the art, and future research. Applied element method. In 2019 International Conference on Robotics and
Bionics and Biomechanics, 5, 2008. Automation, pages 7395–7400, 2019.
[5] A. A. Transeth, K. Y. Pettersen, and P. Liljebäck. A survey on snake [31] C. Duriez. Control of elastic soft robots based on real-time finite
robot modeling and locomotion. Robotica, 27(7):999–1015, 2009. element method. In IEEE International Conference on Robotics and
[6] D. Rus and M. T. Tolley. Design, fabrication and control of soft robots. Automation, pages 3982–3987, 2013.
Nature, 521(7553):467–475, 2015. [32] O. Goury and C. Duriez. Fast, generic, and reliable control and simu-
[7] C. Lee, M. Kim, Y. J. Kim, N. Hong, S. Ryu, H. J. Kim, and S. Kim. lation of soft robots using model order reduction. IEEE Transactions
Soft robot review. International Journal of Control, Automation and on Robotics, 34(6):1565–1576, 2018.
Systems, 15(1):3–15, 2017. [33] O. Goury, B. Carrez, and C. Duriez. Real-time simulation for control of
[8] T.G. Thuruthel, Y. Ansari, E. Falotico, and C. Laschi. Control strategies soft robots with self-collisions using model order reduction for contact
for soft robotic manipulators: A survey. Soft Robotics, 5(2):149–163, forces. IEEE Robotics and Automation Letters, 6(2):3752–3759, 2021.
2018. [34] Y. Adagolodjo, F. Renda, and C. Duriez. Coupling numerical de-
[9] F. Chen and M. Yu Wang. Design optimization of soft robots: A review formable models in global and reduced coordinates for the simulation
of the state of the art. IEEE Robotics Automation Magazine, 27(4):27– of the direct and the inverse kinematics of soft robots. IEEE Robotics
43, 2020. and Automation Letters, 6(2):3910–3917, 2021.
 20

[35] E. Cosserat and F. Cosserat. Theorie des corps deformables. Cambridge degrees of freedom. In 2018 IEEE International Conference on
University Press, 1909. Robotics and Automation (ICRA), pages 1567–1574, 2018.
[36] F. Boyer, M. Porez, and W. Khalil. Macro-continuous computed torque [60] F. Renda, F. Boyer, J. Dias, and L. Seneviratne. Discrete cosserat ap-
algorithm for a three-dimensional eel-like robot. IEEE Transactions on proach for multisection soft manipulator dynamics. IEEE Transactions
Robotics, 22(4):763–775, 2006. on Robotics, 34(6):1518–1533, Dec 2018.
[37] F. Boyer, V. Lebastard, F. Candelier, and F. Renda. Dynamics of [61] Costanza Armanini, Irfan Hussain, Muhammad Zubair Iqbal, Dong-
continuum and soft robots: A strain parameterization based approach. ming Gan, Domenico Prattichizzo, and Federico Renda. Discrete
IEEE Transactions on Robotics, pages 1–17, 2020. cosserat approach for closed-chain soft robots: Application to the fin-
[38] A. E. H. Love. A treatise on the mathematical theory of elasticity. ray finger. IEEE Transactions on Robotics, 37(6):2083–2098, 2021.
Dover, New York, 1944. [62] D. Trivedi, A. Lotfi, and C. D. Rahn. Geometrically exact models for
[39] E. Reissner. On one-dimensional finite-strain beam theory: The plane soft robotic manipulators. IEEE Transactions on Robotics, 24(4):773–
problem. Journal of Applied Mathematics and Physics (ZAMP), 780, 2008.
23(5):795–804, 1972. [63] D.C. Rucker and R.J. Webster. Statics and dynamics of continuum
[40] E. Reissner. On finite deformations of space-curved beams. Journal robots with general tendon routing and external loading. IEEE
of Applied Mathematics and Physics, 32:734–744, 1981. Transactions on Robotics, 27(6):1033–1044, 2011.
[41] F. Renda, F. Giorgio-Serchi, F. Boyer, and C. Laschi. Modelling [64] J. Till, V. Aloi, and D.C. Rucker. Real-time dynamics of soft and
cephalopod-inspired pulsed-jet locomotion for underwater soft robots. continuum robots based on cosserat rod models. The International
Bioinspiration & Biomimetics, 10(5):055005, sep 2015. Journal of Robotics Research, 38(6):723–746, 2019.
[42] F. Boyer and F. Renda. Poincaré’s equations for cosserat media: [65] F. Janabi-Sharifi, A. Jalali, and I.D. Walker. Cosserat rod-based
Application to shells. Journal of Nonlinear Science, 27:1–44, 2017. dynamic modeling of tendon-driven continuum robots: A tutorial. IEEE
[43] F. Renda, F. Giorgio-Serchi, F. Boyer, C. Laschi, J. Dias, and L. Senevi- Access, pages 1–1, 2021.
ratne. A unified multi-soft-body dynamic model for underwater soft [66] D.C. Rucker, R.J. Webster, G.S. Chirikjian, and N.J. Cowan. Equi-
robots. The International Journal of Robotics Research, 37(6):648–666, librium conformations of concentric-tube continuum robots. The
2018. International Journal of Robotics Research, 29(10):1263–1280, 2010.
[44] J.C. Simo. A finite strain beam formulation. the three-dimensional [67] D. C. Rucker, B.A. Jones, and R.J. Webster. A model for concentric
dynamic problem. part i. Computer Methods in Applied Mechanics tube continuum robots under applied wrenches. In 2010 IEEE Inter-
and Engineering, 49(1):55–70, 1985. national Conference on Robotics and Automation, pages 1047–1052,
[45] J.C. Simo and L. Vu-Quoc. A three-dimensional finite-strain rod 2010.
model. part ii: Computational aspects. Computer Methods in Applied [68] Y. Shapiro, K. Gabor, and A. Wolf. Modeling a hyperflexible planar
Mechanics and Engineering, 58(1):79–116, 1986. bending actuator as an inextensible euler–bernoulli beam for use in
[46] J.C. Simo and L. Vu-Quoc. On the dynamics in space of rods flexible robots. Soft Robotics, 2(2):71–79, 2015.
undergoing large motions — a geometrically exact approach. Computer [69] C. Armanini, F. Dal Corso, D. Misseroni, and D. Bigoni. From the
Methods in Applied Mechanics and Engineering, 66(2):125–161, 1988. elastica compass to the elastica catapult: an essay on the mechanics of
[47] A. Cardona and M. Geradin. A beam finite element non-linear theory soft robot arm. Proceedings Royal Society A, 4732016087020160870,
with finite rotations. International Journal for Numerical Methods in 2017.
Engineering, 26(11):2403–2438, 1988. [70] Xu K. and Simaan N. Analytic formulation for kinematics, statics
[48] Ad. Ibrahimbegović and M. Al Mikdad. Finite rotations in dynamics and shape restoration of multi-backbone continuum robots via elliptic
of beams and implicit time-stepping schemes. International Journal integrals. ASME Journal Mechanics Robotics, 2:11006, 2009.
for Numerical Methods in Engineering, 41(5):781–814, 1998. [71] K. Xu and N. Simaan. Intrinsic wrench estimation and its performance
[49] F. Boyer and D. Primault. Finite element of slender beams in finite index for multisegment continuum robots. IEEE Transactions on
transformations: a geometrically exact approach. International Journal Robotics, 26(3):555–561, 2010.
for Numerical Methods in Engineering, 59(5):669–702, 2004. [72] P. Polygerinos, Z. Wang, J. T. B. Overvelde, K. C. Galloway, R. J.
[50] V. Sonneville, A. Cardona, and O. Brüls. Geometrically exact beam Wood, K. Bertoldi, and C. J. Walsh. Modeling of soft fiber-reinforced
finite element formulated on the special euclidean group se(3). Com- bending actuators. IEEE Transactions on Robotics, 31(3):778–789,
puter Methods in Applied Mechanics and Engineering, 268:451–474, 2015.
2014. [73] V. Cacucciolo, F. Renda, E. Poccia, C. Laschi, and M. Cianchetti.
[51] C. Meier, A. Popp, and W. A. Wall. Geometrically exact finite Modelling the nonlinear response of fibre-reinforced bending fluidic
element formulations for slender beams: Kirchhoff–love theory versus actuators. Smart Materials and Structures, 25(10):105020, sep 2016.
simo–reissner theory. Archives of Computational Methods in Engineer- [74] G. Olson, R.L Hatton, J.A. Adams, and Y. Mengüç. An euler–bernoulli
ing, 26(1):163–243, 2019. beam model for soft robot arms bent through self-stress and external
[52] İ. Tunay. Spatial continuum models of rods undergoing large defor- loads. International Journal of Solids and Structures, 207:113–131,
mation and inflation. IEEE Transactions on Robotics, 29(2):297–307, 2020.
2013. [75] S. Hirose. Biologically inspired robots: Snake-like locomotors and
[53] S. Grazioso, G. Di Gironimo, and B. Siciliano. A geometrically exact manipulators. Robotica, 12(3):282–282, 1994.
model for soft continuum robots: The finite element deformation space [76] Shigeo Hirose and Hiroya Yamada. Snake-like robots. IEEE Robotics
formulation. Soft Robotics, 6(6):790–811, 2019. Automation Magazine, 16(1):88–98, 2009.
[54] S. M. H. Sadati, S. Elnaz Naghibi, I. D. Walker, K. Althoefer, and [77] G. S. Chirikjian and J. W. Burdick. A modal approach to hyper-
T. Nanayakkara. Control space reduction and real-time accurate mod- redundant manipulator kinematics. IEEE Transactions on Robotics and
eling of continuum manipulators using ritz and ritz–galerkin methods. Automation, 10(3):343–354, 1994.
IEEE Robotics and Automation Letters, 3(1):328–335, 2018. [78] G. S. Chirikjian. Hyper-redundant manipulator dynamics: a continuum
[55] A. L. Orekhov and N. Simaan. Solving cosserat rod models via approximation. Advanced Robotics, 9(3):217–243, 1994.
collocation and the magnus expansion. In 2020 IEEE/RSJ International [79] M. Ivanescu, M. C. Florescu, N. Popescu, and D. Popescu. Coil func-
Conference on Intelligent Robots and Systems (IROS), pages 8653– tion control problem for a hyperredundant robot. In 2007 IEEE/ASME
8660, 2020. international conference on advanced intelligent mechatronics, pages
[56] F. Renda, C. Armanini, V. Lebastard, F. Candelier, and F. Boyer. A 1–6, 2007.
geometric variable-strain approach for static modeling of soft manipu- [80] Co. Della Santina and D. Rus. Control oriented modeling of soft robots:
lators with tendon and fluidic actuation. IEEE Robotics and Automation The polynomial curvature case. IEEE Robotics and Automation Letters,
Letters, 5(3):4006–4013, 2020. 5(2):290–298, 2020.
[57] A.T. Mathew, I.B. Hmida, C. Armanini, F. Boyer, and F. Renda. [81] C. Della Santina. The soft inverted pendulum with affine curvature. In
Sorosim: a matlab toolbox for soft robotics based on the geometric 2020 59th IEEE Conference on Decision and Control (CDC), pages
variable-strain approach. arXiv preprint arXiv:2107.05494, 2021, 2021. 4135–4142, 2020.
[58] F. Renda, C. Messer, C. Rucker, and F. Boyer. A sliding-rod variable- [82] I. Singh, Y. Amara, A. Melingui, P. Mani Pathak, and R. Merzouki.
strain model for concentric tube robots. IEEE Robotics and Automation Modeling of continuum manipulators using pythagorean hodograph
Letters, 6(2):3451–3458, 2021. curves. Soft Robotics, 5(4):425–442, 2018.
[59] F. Renda and L. Seneviratne. A geometric and unified approach for [83] M. Wiese, K. Rüstmann, and A. Raatz. Kinematic modeling of a soft
modeling soft-rigid multi-body systems with lumped and distributed pneumatic actuator using cubic hermite splines. In 2019 IEEE/RSJ
 21

International Conference on Intelligent Robots and Systems, pages [106] F. Renda, M. Cianchetti, H. Abidi, J. Dias, and L. Seneviratne. Screw-
7176–7182, 2019. based modeling of soft manipulators with tendon and fluidic actuation.
[84] K. M. Digumarti, B. Trimmer, A. T. Conn, and J. Rossiter. Quantifying Journal of Mechanisms and Robotics, 9:041012, 2017.
dynamic shapes in soft morphologies. Soft Robotics, 6(6):733–744, [107] C. Della Santina, A. Bicchi, and D. Rus. On an improved state
2019. parametrization for soft robots with piecewise constant curvature and
[85] I.A. Gravagne, C.D. Rahn, and I.D. Walker. Large deflection dynamics its use in model based control. IEEE Robotics and Automation Letters,
and control for planar continuum robots. IEEE/ASME Transactions on 5(2):1001–1008, 2020.
Mechatronics, 8(2):299–307, 2003. [108] D. B. Camarillo, C. F. Milne, C. R. Carlson, M. R. Zinn, and
[86] B.A. Jones and I.D. Walker. Kinematics for multisection continuum J. K. Salisbury. Mechanics modeling of tendon-driven continuum
robots. IEEE Transactions on Robotics, 22(1):43–55, 2006. manipulators. IEEE Transactions on Robotics, 24(6):1262–1273, 2008.
[87] I.D. Walker. Continuous backbone “continuum” robot manipulators. [109] D. B. Camarillo, C. R. Carlson, and J. K. Salisbury. Configuration
International Scholarly Research Notices, 2013:19, 2013. tracking for continuum manipulators with coupled tendon drive. IEEE
[88] C. Escande, R. Merzouki, P. M. Pathak, and V. Coelen. Geometric Transactions on Robotics, 25(4):798–808, 2009.
modelling of multisection bionic manipulator: Experimental validation [110] S. M.H. Sadati, A. Shiva, A. Ataka, S. E. Naghibi, I. D. Walker,
on robotinoxt. In 2012 IEEE International Conference on Robotics K. Althoefer, and T. Nanayakkara. A geometry deformation model
and Biomimetics (ROBIO), pages 2006–2011, 2012. for compound continuum manipulators with external loading. In 2016
[89] C. Escande, T. Chettibi, R. Merzouki, V. Coelen, and P. M. Pathak. IEEE International Conference on Robotics and Automation, pages
Kinematic calibration of a multisection bionic manipulator. IEEE 4957–4962, 2016.
ASME Transactions on Mechatronics, 20(2):663–674, 2015. [111] S. M. H. Sadati, S. E. Naghibi, A. Shiva, Y. Noh, A. Gupta, I. D.
[90] O. Lakhal, A. Melingui, A. Chibani, C. Escande, and R. Merzouki. Walker, K. Althoefer, and T. Nanayakkara. A geometry deformation
Inverse kinematic modeling of a class of continuum bionic handling model for braided continuum manipulators. Frontiers in Robotics and
arm. In 2014 IEEE/ASME International Conference on Advanced AI, 4:22, 2017.
Intelligent Mechatronics, pages 1337–1342, 2014. [112] W.S. Rone and P. Ben-Tzvi. Continuum robot dynamics utilizing the
[91] O. Lakhal, A. Melingui, and R. Merzouki. Hybrid approach for principle of virtual power. IEEE Transactions on Robotics, 30(1):275–
modeling and solving of kinematics of a compact bionic handling 287, 2014.
assistant manipulator. IEEE/ASME Transactions on Mechatronics, [113] W. S. Rone and P. Ben-Tzvi. Mechanics Modeling of Multisegment
21(3):1326–1335, 2016. Rod-Driven Continuum Robots. Journal of Mechanisms and Robotics,
[92] T. Mahl, A. E. Mayer, A. Hildebrandt, and O. Sawodny. A variable 6(4), 06 2014.
curvature modeling approach for kinematic control of continuum [114] P. E. Dupont, J. Lock, B. Itkowitz, and E. Butler. Design and control
manipulators. In 2013 American Control Conference, pages 4945– of concentric-tube robots. IEEE Transactions on Robotics, 26(2):209–
4950, 2013. 225, 2010.
[93] T. Mahl, A. Hildebrandt, and O. Sawodny. A variable curvature [115] P. E. Dupont, J. Lock, and E. Butler. Torsional kinematic model for
continuum kinematics for kinematic control of the bionic handling concentric tube robots. In 2009 IEEE International Conference on
assistant. IEEE Transactions on Robotics, 30(4):935–949, 2014. Robotics and Automation, pages 3851–3858, 2009.
[94] V. Falkenhahn, T. Mahl, A. Hildebrandt, R. Neumann, and O. Sawodny. [116] R. Xu, S. F. Atashzar, and R. V. Patel. Kinematic instability in
Dynamic modeling of bellows-actuated continuum robots using the eu- concentric-tube robots: Modeling and analysis. In 5th IEEE RAS/EMBS
ler–lagrange formalism. IEEE Transactions on Robotics, 31(6):1483– International Conference on Biomedical Robotics and Biomechatron-
1496, 2015. ics, pages 163–168, 2014.
[95] Y. Bailly and Y. Amirat. Modeling and control of a hybrid continuum [117] R.J. Hendrick, H. B. Gilbert, and R. J. Webster. Designing snap-free
active catheter for aortic aneurysm treatment. In Proceedings of the concentric tube robots: A local bifurcation approach. In 2015 IEEE
2005 IEEE International Conference on Robotics and Automation, International Conference on Robotics and Automation, pages 2256–
pages 924–929, 2005. 2263, 2015.
[96] S. Neppalli and B. A. Jones. Design, construction, and analysis of [118] M. Bergou, M. Wardetzky, S. Robinson, and E. Audoly, B.and Grin-
a continuum robot. In 2007 IEEE/RSJ International Conference on spun. Discrete elastic rods. ACM Trans. Graph., 27(3):1–12, August
Intelligent Robots and Systems, pages 1503–1507, 2007. 2008.
[97] I.S. Godage, G.A. Medrano-Cerda, D.T. Branson, E. Guglielmino, and [119] M. Bergou, B. Audoly, E. Vouga, M. Wardetzky, and E. Grinspun.
D.G. Caldwell. Modal kinematics for multisection continuum arms. Discrete viscous threads. ACM Trans. Graph., 29(4), July 2010.
Bioinspiration and Biomimetics, 10(2):035002, 2015. [120] B. Audoly, N. Clauvelin, P.-T. Brun, M. Bergou, E. Grinspun, and
[98] I. S. Godage, E. Guglielmino, D. T. Branson, G. A. Medrano-Cerda, Wardetzky M. A discrete geometric approach for simulating the
and D. G. Caldwell. Novel modal approach for kinematics of multi- dynamics of thin viscous threads. Journal of Computational Physics,
section continuum arms. In 2011 IEEE/RSJ International Conference 253:18–49, 2013.
on Intelligent Robots and Systems, pages 1093–1098, 2011. [121] E. Grinspun, M. Desbrun, K. Polthier, P. Schröder, and A. Stern. Dis-
[99] I. S. Godage, D. T. Branson, E. Guglielmino, G. A. Medrano-Cerda, crete differential geometry: an applied introduction. ACM SIGGRAPH
and D. G. Caldwell. Dynamics for biomimetic continuum arms: A Course, 7:1–139, 2006.
modal approach. In 2011 IEEE International Conference on Robotics [122] M. Gazzola, L. H. Dudte, A. G. McCormick, and L. Mahadevan.
and Biomimetics, pages 104–109, 2011. Forward and inverse problems in the mechanics of soft filaments. Royal
[100] I.S. Godage, D.T. Branson, E. Guglielmino, G.A. Medrano-Cerda, and Society Open Science, 5171628171628, 2018.
D.G. Caldwell. Shape function-based kinematics and dynamics for [123] N. Naughton, J. Sun, A. Tekinalp, T. Parthasarathy, G. Chowdhary, and
variable length continuum robotic arms. In 2011 IEEE International M. Gazzola. Elastica: A compliant mechanics environment for soft
Conference on Robotics and Automation, pages 452–457, 2011. robotic control. IEEE Robotics and Automation Letters, 6(2):3389–
[101] Godage I.S., Medrano-Cerda G.A., Branson D.T., Guglielmino E., and 3396, 2021.
Caldwell D.G. Dynamics for variable length multisection continuum [124] N.N. Goldberg, X. Huang, C. Majidi, A. Novelia, O. M. O’Reilly, D. A.
arms. The International Journal of Robotics Research, 35(6):695–722, Paley, and W. L. Scott. On planar discrete elastic rod models for the
2016. locomotion of soft robots. Soft Robotics, 6(5):595–610, 2019.
[102] R.J. Webster, J.S. Kim, N.J. Cowan, G.S. Chirikjian, and A.M. Oka- [125] W. Huang, X. Huang, C. Majidi, and M. K. Jawed. Dynamic simulation
mura. Nonholonomic modeling of needle steering. The International of articulated soft robots. Nature Communications, 11:2233, 2020.
Journal of Robotics Research, 25(5-6):509–525, 2006. [126] M.D.M. Kutzer, S.M. Segreti, C.Y. Brown, M. Armand, R. H. Taylor,
[103] R.J. Webster, J.M. Romano, and N.J. Cowan. Mechanics of precurved- and S.C. Mears. Design of a new cable-driven manipulator with a
tube continuum robots. IEEE Transactions on Robotics, 25(1):67–78, large open lumen: Preliminary applications in the minimally-invasive
2009. removal of osteolysis. In 2011 IEEE International Conference on
[104] P. Sears and P. Dupont. A steerable needle technology using curved Robotics and Automation, pages 2913–2920, 2011.
concentric tubes. In 2006 IEEE/RSJ International Conference on [127] M. Khoshnam and R. V. Patel. A pseudo-rigid-body 3r model for a
Intelligent Robots and Systems, pages 2850–2856, 2006. steerable ablation catheter. In 2013 IEEE International Conference on
[105] T. F. Allen, L. Rupert, T. R. Duggan, G. Hein, and K. Albert. Closed- Robotics and Automation, pages 4427–4432, 2013.
form non-singular constant-curvature continuum manipulator kinemat- [128] V.K. Venkiteswaran, J. Sikorski, and S. Misra. Shape and contact force
ics. In IEEE International Conference on Soft Robotics (RoboSoft), estimation of continuum manipulators using pseudo rigid body models.
pages 410–416, 2020. Mechanism and Machine Theory, 139:34–45, 2019.
 22

[129] Y.Q. Yu, L. L. Howell, C. Lusk, Y. Yue, and M.G. He. Dynamic [151] J. S. Arora. Introduction to optimum design. New York: McGraw-Hill,
Modeling of Compliant Mechanisms Based on the Pseudo-Rigid-Body 1989.
Model. Journal of Mechanical Design, 127(4):760–765, 02 2005. [152] M. Bächer, E. Knoop, and C. Schumacher. Design and control of soft
[130] H. Mochiyama and T. Suzuki. Dynamical modelling of a hyper-flexible robots using differentiable simulation. Current Robotics Reports, 2(2),
manipulator. In Proceedings of the 41st SICE Annual Conference. SICE 2021.
2002., volume 3, pages 1505–1510 vol.3, 2002. [153] J. Tapia, E. Knoop, M. Mutný, M. A. Otaduy, and M. Bächer.
[131] Y. Ganji and F. Janabi-Sharifi. Catheter kinematics for intracardiac Makesense: Automated sensor design for proprioceptive soft robots.
navigation. IEEE Trans Biomed Eng., 56(3):621–32, 2009. Soft Robotics, 7(3):332–345, 2020.
[132] M. Sfakiotakis, A. Kazakidi, A. Chatzidaki, T. Evdaimon, and D. P. [154] V. Megaro, J. Zehnder, M. Bächer, S. Coros, M. Gross, and
Tsakiris. Multi-arm robotic swimming with octopus-inspired compliant B. Thomaszewski. A computational design tool for compliant mecha-
web. In 2014 IEEE/RSJ International Conference on Intelligent Robots nisms. ACM Trans. Graph., 36(4), July 2017.
and Systems, pages 302–308, 2014. [155] Soft robots plugin for SOFA. https://www.sofa-framework.org/.
[133] H. Habibi, C. Yang, R. Kang, I. D. Walker, I. S. Godage, X. Dong, [156] PyElastica. https://github.com/GazzolaLab/PyElastica.
and D. T. Branson. Modelling an actuated large deformation soft [157] SoRoSim: An open-chain robotics simulation toolbox. https://www.
continuum robot surface undergoing external forces using a lumped- mathworks.com/matlabcentral/fileexchange/83038-sorosim.
mass approach. In 2018 IEEE/RSJ International Conference on [158] TMTDyn. https://github.com/smhadisadati/TMTDyn.
Intelligent Robots and Systems, pages 5958–5963, 2018. [159] SoMo (SoftMotion). https://github.com/graulem/somo.
[134] H. Habibi, C. Yang, I. S. Godage, R. Kang, I. D. Walker, and D. T. [160] The (legacy) ChainQueen differentiable MPM solver (v1). https:
Branson. A Lumped-Mass Model for Large Deformation Continuum //github.com/yuanming-hu/ChainQueen.
Surfaces Actuated by Continuum Robotic Arms. Journal of Mecha- [161] IPC. https://github.com/ipc-sim/IPC.
nisms and Robotics, 12(1), 10 2019. [162] S.M.H Sadati, S.E. Naghibi, A. Shiva, B. Michael, L. Renson,
[135] J. Jung, R. S. Penning, N. J. Ferrier, and M. R. Zinn. A modeling M. Howard, C.D. Rucker, K. Althoefer, T. Nanayakkara, S. Zschaler,
approach for continuum robotic manipulators: Effects of nonlinear C. Bergeles, H. Hauser, and I.D. Walker. Tmtdyn: A matlab package
internal device friction. In 2011 IEEE/RSJ International Conference for modeling and control of hybrid rigid–continuum robots based on
on Intelligent Robots and Systems, pages 5139–5146, 2011. discretized lumped systems and reduced-order models. The Interna-
[136] J. Austin, R. Corrales-Fatou, S. Wyetzner, and H. Lipson. Titan: A tional Journal of Robotics Research, 40(1):296–347, 2021.
parallel asynchronous library for multi-agent and soft-body robotics [163] M.A. Graule, C.B. Teeple, T.P. McCarthy, R.C. St. Louis, G.R. Kim,
using nvidia cuda. In 2020 IEEE International Conference on Robotics and R.J. Wood. Somo: Fast and accurate simulations of continuum
and Automation, pages 7754–7760, 2020. robots in complex environments. In 2021 IEEE International Confer-
[137] D. Braganza, D. M. Dawson, I. D. Walker, and N. Nath. Neural network ence on Intelligent Robots and Systems (IROS), page In Review. IEEE,
grasping controller for continuum robots. In Proceedings of the 45th 2021.
IEEE Conference on Decision and Control, pages 6445–6449, 2006. [164] Y. Hu, J. Liu, A. Spielberg, J. B. Tenenbaum, W. T. Freeman, J. Wu,
[138] D. Braganza, D. M. Dawson, I. D. Walker, and N. Nath. A neural D. Rus, and W. Matusik. Chainqueen: A real-time differentiable
network controller for continuum robots. IEEE Transactions on physical simulator for soft robotics. In 2019 International Conference
Robotics, 23(6):1270–1277, 2007. on Robotics and Automation, pages 6265–6271, 2019.
[165] M. Li, Z. Ferguson, T. Schneider, T. Langlois, D. Zorin, D. Panozzo,
[139] A. Melingui, C. Escande, N. Benoudjit, R. Merzouki, and J.B.
C. Jiang, and D. M. Kaufman. Incremental potential contact:
Mbede. Qualitative approach for forward kinematic modeling of a
Intersection-and inversion-free, large-deformation dynamics. ACM
compact bionic handling assistant trunk. IFAC Proceedings Volumes,
Trans. Graph., 39(4), July 2020.
47(3):9353–9358, 2014. 19th IFAC World Congress.
[166] PK Ma, TK Du, J. Z. Zhang, K. Wu, A. Spielberg, R. K. Katzschmann,
[140] A. Melingui, O. Lakhal, D.B., J. B. Mbede, and R. Merzouki. Adaptive
and W. Matusik. Diffaqua: A differentiable computational design
neural network control of a compact bionic handling arm. IEEE/ASME
pipeline for soft underwater swimmers with shape interpolation. ACM
Transactions on Mechatronics, 20(6):2862–2875, 2015.
Trans. Graph., 40(4), July 2021.
[141] M. Giorelli, F. Renda, M. Calisti, A. Arienti, G. Ferri, and C. Laschi. [167] R. Jitosho, N. Agharese, A. Okamura, and Z. Manchester. A dynamics
Neural network and jacobian method for solving the inverse statics of simulator for soft growing robots^*. In 2021 IEEE Inter-
a cable-driven soft arm with nonconstant curvature. IEEE Transactions national Conference on Robotics and Automation, pages 11775–11781,
on Robotics, 31(4):823–834, 2015. 2021.
[142] M. Giorelli, F. Renda, M. Calisti, A. Arienti, G Ferri, and C. Laschi. [168] F. Renda, M. Giorelli, M. Calisti, M. Cianchetti, and C. Laschi.
Learning the inverse kinetics of an octopus-like manipulator in three- Dynamic model of a multibending soft robot arm driven by cables.
dimensional space. Bioinspiration & Biomimetics, 10(3):035006, may IEEE Transactions on Robotics, 30(5):1109–1122, 2014.
2015. [169] A. Shiva, S.M. H. Sadati, Y. Noh, J. Fraś, A. Ataka, H. Würdemann,
[143] F. Renda, M. Cianchetti, M. Giorelli, A. Arienti, and C. Laschi. A 3d H. Hauser, I. D. Walker, T. Nanayakkara, and K. Althoefer. Elasticity
steady-state model of a tendon-driven continuum soft manipulator in- versus hyperelasticity considerations in quasistatic modeling of a soft
spired by the octopus arm. Bioinspiration & Biomimetics, 7(2):025006, finger-like robotic appendage for real-time position and force estima-
may 2012. tion. Soft Robotics, 6(2):228–249, 2019.
[144] T.G. Thuruthel, E. Falotico, M. Manti, A. Pratesi, M. Cianchetti, and [170] C. Della Santina, C. Duriez, and D. Rus. Model based control of soft
C. Laschi. Learning closed loop kinematic controllers for continuum robots: A survey of the state of the art and open challenges. arXiv
manipulators in unstructured environments. Soft Robotics, 4(3):285– preprint arXiv:2110.01358, 2021.
296, 2017.
[145] T.G. Thuruthel, E. Falotico, F. Renda, and C. Laschi. Learning dynamic
models for open loop predictive control of soft robotic manipulators.
Bioinspiration & Biomimetics, 12(6):066003, 2017.
[146] T. G. Thuruthel, E. Falotico, F. Renda, and C. Laschi. Model-based
reinforcement learning for closed-loop dynamic control of soft robotic
manipulators. IEEE Transactions on Robotics, 35(1):124–134, 2019.
[147] M. Rolf and J. J. Steil. Efficient exploratory learning of inverse
kinematics on a bionic elephant trunk. IEEE Transactions on Neural
Networks and Learning Systems, 25(6):1147–1160, 2014.
[148] D. Bruder, X. Fu, R. B. Gillespie, C. D. Remy, and R. Vasudevan.
Data-driven control of soft robots using koopman operator theory. IEEE
Transactions on Robotics, 37(3):948–961, 2021.
[149] H. Arbabi and I. Mezic. Ergodic theory, dynamic mode decomposition,
and computation of spectral properties of the koopman operator. SIAM
Journal on Applied Dynamical Systems, 16(4):2096–2126, 2017.
[150] M. Thieffry, A. Kruszewski, C. Duriez, and T.M. Guerra. Control
design for soft robots based on reduced-order model. IEEE Robotics
and Automation Letters, 4(1):25–32, 2019.